research paper on groundwater potential

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

• View all journals
• Explore content
• About the journal
• Publish with us
• Sign up for alerts
• Open access
• Published: 10 March 2021

Groundwater recharge potential zonation using an ensemble of machine learning and bivariate statistical models

• Maryam Sadat Jaafarzadeh 1 ,
• Naser Tahmasebipour 1 ,
• Ali Haghizadeh 1 ,
• Hamid Reza Pourghasemi 2 &
• Hamed Rouhani 3  

Scientific Reports volume  11 , Article number:  5587 ( 2021 ) Cite this article

7469 Accesses

55 Citations

1 Altmetric

Metrics details

• Environmental sciences

Many regions in Iran are currently experience water crisis, largely driven by frequent droughts and expanding agricultural land combined with over abstraction of groundwater. Therefore, it is extremely important to identify potential groundwater recharge (GWR) zones to help in prevent water scarcity. The key objective of this research is to applying different scenarios for GWR potential mapping by means of a classifier ensemble approach, namely a combination of Maximum Entropy (ME) and Frequency Ratio (FR) models in a semi-arid mountainous, Marboreh Watershed of Iran. To consider the ensemble effect of these models, 15 input layers were generated and used in two models and then the models were combined in seven scenarios. According to marginal response curves (MRCs) and the Jackknife technique, quaternary formations (Qft1 and Qft2) of lithology, sandy-clay-loam (Sa. Cl. L) class of soil, 0–4% class of slope, and agriculture & rangeland classes of land use, offered the highest percolation potential. Results of the FR model showed that the highest weight belonged to Qft1 rocks and Sa. Cl. L textures. Seven scenarios were used for GWR potential maps by different ensembles based on basic mathematical operations. Correctly Classified Instances (CCI), and the AUC indices were applied to validate model predictions. The validation indices showed that scenarios 5 had the best performance. The combination of models by different ensemble scenarios enhances the efficiency of these models. This study serves as a basis for future investigations and provides useful information for prediction of sites with groundwater recharge potential through combination of state-of-the-art statistical and machine learning models. The proposed ensemble model reduced the machine learning and statistical models’ limitations gaps and promoted the accuracy of the model where combining, especially for data-scarce areas. The results of present study can be used for the GWR potential mapping, land use planning, and groundwater development plans.

Similar content being viewed by others

Machine learning predictive insight of water pollution and groundwater quality in the Eastern Province of Saudi Arabia

Mapping reservoir water quality from Sentinel-2 satellite data based on a new approach of weighted averaging: Application of Bayesian maximum entropy

Future groundwater potential mapping using machine learning algorithms and climate change scenarios in Bangladesh

Introduction.

Over the past half century, the least developed countries population has grown rapidly, and continues to grow slowly over the coming decades 1 . This can lead to a further increase in agricultural activities and consequently, putting unprecedented pressure on the surface water and groundwater resources. Over-extraction of groundwater may result in falling groundwater level 2 , land-subsidence 2 , 3 , salinization 4 , reduced well-yields 5 , increased pumping costs 6 , 7 , and enhances saltwater intrusion 7 . The result of all these changes was the reduction the surface of 60% of major aquifers in many parts of the world 8 ; additionally, one of the great challenges in the twenty-first century is the change in the global weather patterns which likely impose additional pressures 9 , 10 . Another problem concerned increase in water demand, which dropped access to adequate water to meet public needs 11 , 12 .

Groundwater is one of the sources of water stock which can be used to tackle the problem of water scarcity 13 , 14 . The social value of groundwater should not be measured by the volume of withdrawal 12 . Due to its availability on a local scale, the possibility of adjusting withdrawal in response to demand fluctuations, high reliability in periods of drought, and desirable quality with minimal need for purification, groundwater has higher economic value compared to surface water. Therefore, identification of groundwater recharge zones is a critical for the sustainable management of groundwater resources 15 , 16 , 17 , 18 , 19 . Groundwater recharge is commonly determined by the amount of precipitation and the percolation process. Recharge occurs when water moves through the unsaturated zone 20 . In semi-arid areas with dry and wet seasons, recharge occurred occasionally after high rainfall. Therefore, problems caused by droughts on the one hand, and devastating floods on the other, highlight the need for proper water resource management. In this regard, collecting surface water and identifying the groundwater recharge zones for storing water are the most important strategies for water resource management, particularly in semi-arid areas of Iran.

There are several approaches to groundwater characteristics and groundwater recharge assessment. With the variety of techniques available for recharge assessment, selecting the appropriate technique is often not easy. Implementing some of these techniques may be restricted by some factors such as the resources available, precision level needed 21 , the space and time scales, reliability/range of recharge evaluation, as well as time and cost 22 , 23 . The methods can generally be categorized as being either direct or indirect. Direct methods include geological and geophysical explorations (seismic, sonic and magnetic) and drilling tests 21 . These methods in-situ investigation is costly and not feasible for estimation at watershed scale. Indirect methods comprise physical and numerical modeling, physically based spatially distributed method 21 and tracer. Most direct approaches provide information over relatively small scale 24 and the methods often consider on a single affecting factor for GWR 25 . Tracer methods for Groundwater recharge assessment are considered the most promising approach but these methods are generally expensive, time-consuming, and often need to be integrated due to the large size of data 23 . Numerical models simplify the physical process of recharge and requires a certain measure parameter to obtain reliable prediction but the model accuracy relies on possible error in initial data, computational costs and some parameters are rarely available in developing countries. Moreover, these models involved many parameters that makes it hard to obtain a unique solution 26 .

Numerous geospatial techniques have been widely used in the past decade and continue to develop new techniques especially due to their capability to increase spatial and temporal information. A series of studies have been conducted by researchers based on knowledge-driven modeling approach or multi influencing factor with integrating various thematic maps on the GWR potential zones using GIS 24 , 27 , 28 , 29 , 30 , 31 , 32 . Rajaskhar et al. 7 and Dar et al. 33 used AHP (Analytic Hierarchy Process) for artificial recharge sites identification, as well as GWR potential zones in Anantapour and Kashmir Valley, respectively in India. The weights of several thematic maps include geomorphology, slope, drainage density, land use, lithology, lineament density, rainfall, and soil texture were calculated by the AHP. This method utilizes expert-based opinion and has relatively high potential for error 34 . For example, Mogaji et al 35 compared AHP and Evidential Belief Function (EBF) for identification of GWRP zones in southwestern Nigeria. They concluded that based on the area under the curve (AUC) measure, EBF, as a data driven technique, outperformed the AHP. Chenini and Msaddek 36 compared the bivariate statistical analysis and logistic regression model (LR) to produce a GWRP map in Qued Guenniche in the north of Tunisia. The authors found that the bivariate and multivariate statistical approaches provide more accurate results than the AHP. On the other hand, data-driven modeling which is based on statistical and machine learning algorithms can be identified and mapped effectively behavior of phenomena. Therefore, recently, there has been a trend towards using numerous statistical methods and machine learning methods to map the GWR potential. Moving towards more advance algorithms in GWR potential mapping, Pourghasemi et al 37 used machine learning algorithms, namely support vector machines (SVM), multivariate adaptive regression splines (MARS), and random forest (RF) to map groundwater recharge potential zones in Firuzkuh County, Iran. In this research, infiltration rate was measured using double ring infiltrometer with random permeability sampling at eleven lithological types. In an attempt to develop sophisticated approaches in phenomena mapping, researchers have developed ensemble models with combining statistical models and machine learnings techniques in spatial prediction of natural phenomena which improve prediction accuracy. In other words, higher precision and predictive ability in comparison with individual machine learning models is the significant property of ensemble models 38 , 39 . To overwhelming literatures on susceptibility and potential mapping of natural phenomena studies such as gully erosion 40 , groundwater 41 , 42 , 43 , flood 44 , 45 and landslide 46 , 47 , 48 are rich of researches. The advantage of the ensemble methods is that achieve higher prediction accuracy by incorporating ideas of multiple learners and reduce bias and variance.

Iran has one percent of the world’s population but the country contains only 0.63 percent of the world’s renewable fresh water. In terms of climatic conditions major parts of Iran have been considered as arid and semi-arid regions. In addition to low and spatial and temporal distribution of precipitation, the occurrence of precipitation with rather high severity, which cause heavy rainfall and destructive floods is another important feature of these regions. Declining groundwater level in arid ad semi-arid area like Iran, where groundwater provides approximately 80% of the water supply for agricultural and households 49 become a serious threat to sustainability. In these areas, extra exploitation of aquifers in comparison with natural recharges has caused depletion in major aquifers. The semi-arid terrains of Iran such as Varamin, Parishan, Qazvin, Darab, plains are among the most constantly increasing agricultural cultivated land areas for maize, cotton, wheat cultivation. The stable growth in agriculture is possible by irrigation with shallow local groundwater resources 50 . However, intense groundwater abstraction with limited recharge, resulting in aquifer depletion and increasing pumping costs 51 . Moreover, climate change is expected to impose additional limits to water supply 9 . Over the last half century, the groundwater withdrawals in Iran has tripled. According to SZCEC the groundwater levels have declined from by nearly 1 m in the south western to 3 m in the northern and northeastern parts of the Marboreh watershed for the last 20 years (SZCEC, 2012) and negative social and environmental effects becoming increasingly evident.

Due to the importance of recharge zones, the identification of appropriate recharge zone is critical for managing aquifer recharge. So far, the ensemble method is applied in different spatial phenomena mapping but to our knowledge no prior studies have been applied in GWR potential zones mapping. Therefore, the present study aimed to evaluate prediction performance ensemble models based on FR and Maximum Entropy (ME) to identify GWR potential mapping in a mountainous watershed, Marboreh in Lorestan Province, Iran, where the groundwater dependency has been largely increased over the last decades. Specifically, the machine learning (ME) and a bivariate statistical method (FR) were used for assess ensemble of spatial prediction of GWR potential scenarios. Whereas, Jackknife test as a simple but effective method utilized for screening important casual factors based on the random sampling in training dataset. The performance of spatial prediction models was compared by area under curve (AUC) and Correctly Classified Instances (CCI).

Methodology

Marboreh Watershed is located between 33 o 12′ to 33 o 51′ N and 49 o 03′ to 49 o 57′ E in the west of Iran, Lorestan Province (Fig.  1 ). The total area of Marboreh Watershed is about 2560 km 2 with a cold semi-arid climate. The annual rainfall decreases from south (900 mm) towards eastern (280 mm) parts of the watershed with the average annual rainfall and the average annual temperature of 412 mm and 13.8 ℃, respectively (LPMO (Lorestan Province Meteorological Organization.)). Rainfall is not uniformly distributed through the year, with the majority occurring in March and April. In the summer time, the rainfall mostly occurs once or twice nearly every month. Topographically, the region is a part of Zagros Chain with altitude range from 1442 to 4056 m, with average altitude of 2749 m. The main river, Marboreh, originates in the Aligudarz Mountains at the east of watershed, making its 110.22 km journey westwards along the entire length study area. Smaller tributaries contributing to the river discharge such as Azna, Aligudarz, and Kamandan. Finally, Marboreh River joins to Tireh River and form the Sezar, which is one of the main branches of Dez basin that flows in Karun River, the longest river in Iran which finally drains into the Persian Gulf. Geologically, the watershed lies within Zagros Fold and 29 types of geological formation cover this watershed, and the dominate land use is agriculture (49.5%), followed by rangeland (42%).

Location of the study area in Lorestan Province, Iran.

The flowchart shown in Fig.  2 , illustrate the method employed in this study and briefly given in the following sections:

Flowchart of carrying out the methodology in this study.

Permeability samples inventory map

The key link between surface water and groundwater is groundwater recharge 4 and quantifying the Groundwater recharge is difficult at large scale. On the other hand, the gross groundwater recharge rate may be depending on the permeability and soil’s absorption, which make it easier to measure 52 . In order to evaluate the performance of the models in GWR potential mapping, double-ring infiltrometer method and soil sampling (Fig.  3 ) were used for understanding the spatial variability of GWR 37 , 53 . The double-ring infiltrometer consist of an inner and outer ring that are driven 5-cm into the subsurface ground. The inner and the outer rings were filled with water. The head is kept constant during the experiments. The procedure continue until the infiltration rate is considered steady and set up by the Sangab Zagros Consulting Engineering Company 22 standard. Due to the extent of the studied area and the cost of the sampling process, information from previous studies 22 in the studied area were also used. The percolation points were randomly split into two datasets: the first group for model training (70%) and to validate the training dataset, a separate independent set of data (30%) to evaluate modelled and observed maps.

Double-ring infiltrometer method in the study area.

The condition factors maps used in FR and ME models (maps layouts were prepared in ArcGIS 10.6 software).

GWR influencing factors

GWR vary significantly in different places depending on topographic factors, hydrogeology variables, and climate conditions 25 , 54 . Consequently, based on various studies 25 , 27 , 37 , 54 , 55 and data availability, 15 effective factors on GWR potential, namely elevation, slope percent, slope aspect, profile curvature, plan curvature, drainage density, distance from rivers, topographic wetness index (TWI), stream power index (SPI), Rainfall, lithology, fault distance, land use (LU), Normalized Difference Vegetation Index (NDVI), and soil texture were used. Out of these selected thematic maps, elevation, slope degree, slope aspect, drainage density, TWI and SPI were generated from ASTER DEM data with a pixel size of 30 × 30-m, while the remaining maps were extracted from Landsat-8 images and through conventional data, such as soil and geology data.

Topographic influencing factors

Elevation play in key role on vegetation and climate that have connection with recharge distribution area. Here, the elevation ranges from 1442 to 4056 m, and the hypsometric map extracted from DEM and classified into five classes (Fig.  4 a). In groundwater recharge investigations, slope angel is a major determinant factor and highlights the importance of topography and the size of runoff-saturated zones 56 , 57 . The regions of steep slope facilities high runoff, whereas at the lower slope gradients, runoff generation reduced and eventually increase infiltration rate and recharge the saturated zone 58 . The slope map extracted from DEM and was classified into 5 classes (Fig.  4 b). Slope aspect influences GWR as indicator for solar radiation which strongly influences infiltration rate 54 . Slope aspect is the orientation of slope, measured clockwise in degrees from 0 to 360. This factor was derived from the DEM in ArcGIS10.6 ( https://www.arcgis.com ) and classified into five main categories (Fig.  4 c). Flow distribution on the surface depends on topography, which is represented by profile curvature. Profile curvature influences on the overland flow velocity, whereas plan curvature mainly affects flow convergence and divergence (Fig.  4 d–e).

Hydrological influencing factors

Drainage density is another effective factor for controlling GWR. The drainage density demonstrates the signature of surface to subsurface formation 59 and it implies indirectly soil characteristics, which is assigned soil infiltration capacity 55 . Drainage density was extracted from DEM with help of ArcHydro Tool of ArcGIS10.6. Then the extracted drainage networks (km -1 ) were overlaid on vector digitized stream map of LPRWA (Lorestan Province Regional Water Authority.) for improved drainage direction map and the area was categorized into five classes (Fig.  4 f). The low drainage density indicates low surface runoff and hence high infiltration rate 57 and consequently a high GWR potential. Rivers are one of the sources of GWR, and thus influence GWR potential in a watershed. The Euclidean distance tool from spatial analyst tools in ArcGIS10.6 was used to generate distance categories. The distance map was reclassified into five classes (Fig.  4 g).

Topographic wetness index (TWI) and stream power index (SPI) are as secondary topographic indices. In the GWR potential zone, one of the important secondary topographic factors was TWI. TWI was proposed by Moore et al. (1991) to indicate local groundwater potential. TWI relies on upslope area which highlights the potential exfiltration groundwater by topography effects; the lower TWI, the greater GWR potential in the specific class. The TWI of the region ranges from 0.891 to 22.553 and the TWI was further classified in five classes (Fig.  4 h). SPI index describes stream power and its value ranges from 6.257 to 26.100. Five categories of SPI values were generated based on the natural break method as (< 9.04), (9.04–10.75), (10.75–12.46), (12.46–15.02), and (> 15.02). (Fig.  4 i). The mean annual rainfall was used to determine the groundwater recharge investigations of the Marboreh watershed. The mean annual rainfall over the study area was interpolated based on the Kriging method from 12 gauging station. The mean annual rainfall ranges from 280 to 900 mm (Fig.  4 o.).

Geological influencing factors

The geological factors generally demonstrate the distribution of various rock units, which have a notable effect on GWR and water availability through hydraulic conductivity 60 . In this study, the lithological map was extracted from the national geological map of Iran at a 1:100,000-scale, and encompassed 26 major lithological groups (Fig.  4 j). Fault map (Fig.  4 k) was extracted from the national geological map of Iran at a 1:100,000-scale. This layer was used to infer groundwater storage and in regional scale, ground water flow direction is controlled by fault systems.

Ecological influencing factors

Land use (LU), NDVI, and soil are ecological parameters, which are commonly used factors for GWR potential zones 27 , 37 , 54 , 55 . The LU types can be significantly effective on runoff, permeability, and evapotranspiration. As stated by Gee et al 61 , recharge in vegetated areas is much lower than non-vegetated areas. In addition, recharge is greater in agricultural lands and grass lands than perennial lands including, shrub and forest areas 62 . The Landsat8 data (downloaded from https://www.earthexplorerusgs.gov ; column 165 and row 37, 17 June 2018) were used to determine the types of LU. Also, LU was classified by a supervised classification technique according to the maximum likelihood approach into 8 classes, namely; agriculture, bare land, forest, orchard, rangeland, rocky areas, urban, and water. LU within the study area is dominated by agriculture (49.6%) and rangeland (43.4%) (Fig.  4 l). The vegetation density and coverage were represented using the Normalized Difference Vegetation Index (NDVI) map. The NDVI layer was prepared in ENVI 5.3. The value range of NDVI is -1 to 1, and higher value of NDVI represents dense vegetation. NDVI values were grouped based on the natural break method into five classes, namely (< 0.11), (0.11–0.17), (0.17–0.23), (0.23–0.33), and (> 0.33) (Fig.  4 m).

Soil texture is another effective factor used to specify sites suitable for recharge. Soil type is a major criterion in groundwater recharge and agricultural production. Soil texture provides essential information on infiltration rate 29 . For example, sandy soils have high permeability, which consequently might cause in reduced GWR. The soil map was obtained from Agricultural Research, Education, and Promotion Organization of Lorestan Province in 1:250,000-scale. The most prominent soil type in region were Loamy-sandy and Silty-loamy which cover 15.2% and 14.3%, respectively of the study area. The soil texture map of the study area was classified into 9 soil texture classes (Fig.  4 n).

Maximum entropy (ME) model

We applied a family of machine learning techniques based on multinomial logistic regression know as maximum entropy (ME) to yield the potential of GWR from ground base observation and the condition factors. Jaynes introduced ME technique in 1957 based on probability distribution of maximum entropy to derive the probability of given number of individuals occurrence across the any given area based on a set of condition factors. The MaxEnt software (version 3.3.3 k), which is based on the ME approach, is conducted for GWR potential mapping. Groundwater potential mapping 63 , 64 , flood 30 , landslide 65 , 66 , rangeland 67 , 68 , earthquake 69 , species distribution 70 , springs and wells 71 , and groundwater quality 72 .

Frequency ratio (FR) model

The FR model, as the probability of incidence for a special attribute, is a simple geospatial assessment tool and is used to compute the eventual relation between GRW potential and GWR´s effective factors 30 . In specifying the frequency ratio, the recharge occurrence ratio in each conditioning factors sub-class is obtained toward the total recharge. Then, each class’s surface ratio is computed compared to the total area of the watershed.

The FR values were identified using Eq. ( 1 ) for each sub-class of GWR potential effective factors based on their correlation with GWR potential inventory:

where α is the number of recharge pixels in each subclass, β is the total number of pixels in the area, γ is the total number of recharge pixels of the entire area. ɘ indicates the number of pixels in every subclass of conditioning factors, θ represents the percentage of recharge occurrences in any sub-class of conditioning factors, and ɷ is the relative percentage of the area of each subclass.

Ensemble modeling

In data analytics and predictive modeling, an individual model based on one data sample could have great variability, a large number of inaccuracies, or extensive biases, which affect the reliability of results 73 (Rokach, 2010). The effects of these restrictions can be reduced by analyzing multiple samples or combining various models which can help provide better information to decision makers and improve modeling algorithms 39 , 73 . In the current study, we integrated the ME and FR models in various scenarios using basic mathematical operations (Table 1 ) in ArcGIS10.6.

Conditioning factors importance and marginal response curves (MRCs)

The Jackknife test can be used to reduce bias of an estimator. This test removed one factor at a time and model created, here MaxEnt, with the remaining factors. In this study, the Jackknife tests were performed to determine individual conditioning factors importance for MaxEnt predictions. Accordingly, the influencing factor contribution in the analysis can be identified 74 . We also ran an MRC to quantify the conditioning factor’ behavior. MRCs were calculated and plotted for all 15 variables by the ME model to depict the role of each variable in the occurrence of GWRP in Marboreh Watershed. Plots represent the correlations of predicted sites with each independent variable relative to the dependent variable, showing how the individual predictor variables are related to the modeled class. On vertical axes of plots, values closer to 1 demonstrate the preferred range of the class.

Validation of the built models

The accuracy of predictive models should be analyzed by comparing the generated data with a training and testing dataset (existing GWR percolation points). To assess the performance of individual and ensembles models, we utilized the area under the Receiving Operator Curve (AUC-ROC) 74 , 75 and the Correctly Classified Instances (CCI) 76 , 77 . The AUC value ranges from 0 to 1, in which the higher the value, the greater perfect discrimination 37 . The CCI derived from the corresponding confusion matrix that calculated true positive (TP), true negative (TN), false positive (FP), false negative (FN), with a range from 1 to 100. The greater CCI value, the more accurate prediction 78 , 79 .

Conditioning factors importance

The Jackknife test was implemented during MaxEnt model building to identify conditioning factors contribution (Fig.  5 ). According to Fig.  5 , soil provided the most critical conditioning factor in GWR potential prediction, and lithology was second important having high training gain contained most unique information when they used independently. Also, DEM and rain conditioning factors provided high gain when used independently. While, distance from fault, drainage density, distance from river, slope and NDVI had moderate gains respectively, when used alone. Furthermore, soil and lithology reduced training gain the most when was excluded from the model, thus had the most information that were not present in other conditioning factors.

The Jackknife of regularized gain for GWR potential zones in Marboreh watershed.

Application of MaxEnt

Firstly, we examined the response of GWR potential to effective factors. Response curves identified the quantitative relationship between the logistic probability of the GWR potential and effective factors. MRCs include curves representing the separate effect of each independent factor on the dependent variable (i.e. GWR potential). According to these graphs (Fig.  6 ), the highest probability of percolation was seen at elevations between 1810 and 1820 m, with a probability of 77% (Fig.  6 a). These conditions are in accordance with the inverse relationship between elevation and percolation since percolation level is higher at the lowest points or in the plains, and decreases with increasing altitude. Due to the altitude range of the area, this elevation class has the most suitable conditions for GWR. Slope and aspect are also among the significant considerations in site selection for groundwater recharge. Steep slopes concentrate water on the lower slopes and contribute to the buildup of hydrostatic pressure 80 . Runoff on lower slopes has a better opportunity to concentrate and percolate. In Marboreh watershed, the slope layer was classified into five class and the highest probability (67%) was observed for the 1st class, which ranges from 0 to 4% (Fig.  6 b). This slope range comprises the suitable slopes for recharging. Aspect does not directly affect runoff. Its role is in determining the rate of runoff generation due to the difference in microclimate on the different slopes. The highest probability of percolation (66%) was recorded for the southwest aspects (Fig.  6 c). In order to determine the geomorphometric characteristics of shapes, the second derivative is used in digital elevation (curvature) model. This characteristic is associated with geomorphologic processes and has two distinct types of curvature with vertical properties that are called plan and profile curvature. Areas with concave morphology have the most potential for percolation. Concave morphology can concentrate water and moisture and create suitable areas for GWR 81 .

Marginal response curves for the quantitative conditioning factors (y-axis: predicted probability of GWR potential related to each conditioning factors). (a) Elevation; (b ) slope degree; (c) slope aspect; (d) profile curvature, (e) plan curvature, (f) drainage density; (g) distance from river; (h) TWI; (i) SPI; (j) lithology; (k) distance from fault; (l ) LU; (m) NDVI; (n) Soil texture and rain (o) (prepared in maxent 3.3.3 k software).

Permeability will be higher in areas with lower drainage densities because drainage by conduction of water on the ground prevents its penetration into the depths and thus reduces underground recharge 82 . The range for the drainage density factor at the studied watershed was between 0 and 1.39 km. The results showed that most places with recharge potential had drainage densities (with 67% probability) between 0.56 and 0.58 km (Fig.  6 f). Results indicate that the probability of percolation in areas close to rivers was greater. The range of distance from rivers spanned 0–5.36 km (Fig.  6 g). For the variable distance from river, the curve represents a steep decline with increasing distance, indicating that the more percolation area located near the river. In this study, the highest probability of percolation was observed at a distance of 0.52 km from rivers, with 0.67 percent. The curve shows a steep decline with increasing precipitation, indicating that this species likely prefers drier areas.

The TWI index is mainly used for quantitative topographic assessment of hydrological processes as well as showing the effect of topography on the position and size of groundwater flow, soil moisture, and saturated sources of runoff production. The higher index resulted in higher recharge potential (Fig.  6 h). The stream power index, assuming the proportion of drainage to the surface area of the watershed, is a topographical combination feature. In this study, the highest percolation probability (69%) was seen in areas with SPI values in the 4–6 range (Fig.  6 i).

The lithology of the studied area includes 26 geomorphic units. Areas with Qft2 quaternary lithology had the highest percolation potential because these layers are more closely associated with foothills and alluvial fans, and have high absorbance and percolation (Fig.  6 j). This type of lithological unit, covering 33.65% of the watershed, is the largest lithological section in the region. Faults, due to the creation of crushed areas and the formation of water conduits, are very important in the development of morphology and the formation of subsidence in the region 54 , 83 . Faults facilitate the penetration of water to lower levels and increase the potential for the formation of underground liquidation cavities. The range of distance from faults was 0–28.9. The highest probability of percolation was 0.7% (Fig.  6 k).

Croplands and rangelands (constituting 93% of the watershed’s land use) are known to be conducive to GWR due to presence of vegetation and permeable soil textures. According to the results, the highest percolation and recharge probability occurred in agricultural land and rangeland (Fig.  6 l) where the soil crust of farmlands being broken by plowing and vegetation cover reducing the velocity of rain or surface water and providing more time for percolation. Cropland and rangelands, especially in areas with other suitable effective factors such as slope, elevation, and geology, show the highest recharge potential. Vegetation is a good sign of underground water and has a direct relationship with percolation potential; the greater the amount of vegetation in a region, the greater the permeability. NDVI was used to measure vegetation cover (ranging from 0.6 to 0.68). As depicted in Fig.  6 m, the highest probability of recharge potential based on NDVI was found to be 70%. In Marboreh watershed, nine soil texture classes were identified. The sandy clay loam (Sa. Cl. L) class has the highest probability of percolation with 57% (Fig.  6 n). Soil texture greatly influences water percolation, permeability, and water-holding capacity. The Sa. Cl. L class, due to its combination of coarse and fine grains, is one of the most permeable textures. According to the response curve of annual rainfall, the probability of GWRP occurrence (78%) increased in area with 600 mm and decreased smoothly after that (Fig.  6 o).

Application of FR model

The outcomes of FR model calculated for each class of 15 effective factors based on their relationship with GWR potential locations are presented in Fig.  7 , and a larger FR value means a higher probability on GWR potential of the corresponding effective factor. The ratio of the area with GWR potential to the entire area was calculated. In terms of elevation, high potential occurs mainly in the elevation range between 1442 and 1964.8 m. In the case of slope aspect, the FR weights were the highest for the north area (2.26) and flat area (1.19), while the West aspects have the lowest value of 0.01. For slope degree, the values of FR were decreasing through to increasing the slope degree. According to the FR, weights of profile curvature, convex, flat, and concave are 1.53, 0.77, and 0.64, respectively. For plan curvature, concave class had the highest GWR potential. In term of the drainage density, the class 0.28–0.56 had the greatest FR value of 1.96, while the class 1.11–1.39 had the lowest value of 0.24. Among various distance in the river network, the area closes to river had the most correlation with GWR potential, while the greatest distance from river had no GWR potential. TWI with higher values have larger weights of FR., while SPI with higher values have smaller weights of FR. The relationship between GWR potentials and the annual rainfall show that the high annual rainfalls have a high frequency of GWR potentials. The highest FR value (1.99) was observed for class > 622.5 mm.

Spatial correlation between effective factors and GWR potential using FR model.

Lithology FR conditioning factor was the highest in Qft1 (9.39) and has the highest FR value than other conditioning factors. Furthermore, Qft2 also had a high FR value of 2.82. In the study area, 0.40% and 33.7% of the area were covered by Qft1 and Qft2, respectively. Result showed that among the fault distance classes, the second class (5.78–11.56) and the third class (11.56–17.34) with FR values of 1.68 and 1.41 had the highest impact on GWR potential in Marboreh watershed. The results obtained in the present study showed that among the land use types, agricultural land (1.32) and range land (0.78) had the highest probability of GWR potential. According to the NDVI, the greater NDVI value had the higher potential for GWR. Another factor affecting GWR potential is the soil texture. The result revealed that the Sa. Cl. L soil texture weighing 4.66 had the highest probability on GWR potential and might indicate the high possibility of GWR potential in this thematic layer.

Groundwater recharge mapping and validation of the built models

The Maxent model performed well for predicting the GWR potential with the given set of training and test data with overall of 0.993 for AUC and 86.5 for CCI in testing data (Table 1 ). The GWR probability index was divided into four potential classes of low, moderate, high, and very high using a natural break method as seen in Fig.  8 a. The model predicted approximately 7.26% of the studied region included very high GWR potential area (Table 2 ). The GWR with high and moderate potential encompass 8.59 and 29.45 percent of the studied area (Table 2 ). Through the AUC and CCI techniques the obtained map using the FR was evaluated. In the present study, the AUC value show 0.975 which revealed an “excellent” predictive accuracy in the model prediction. Also, the validation using CCI in total was 61.50% which has a satisfactory prediction. Ultimately, the map was generated and classified into four potential classes by natural break classification method, i.e., low, moderate, high and very high (Fig.  8 b). Around 2.92% of the studied area falls into the very high GWR potential zone, while 36.63% of the land belongs to the low GWR potential category (Table 2 ). Based on the FR model, most parts of the studied area (45.20%) have a moderate GWR potential.

GWR potential zone map, (a) ME model and (b) FR model (maps layouts were prepared in ArcGIS 10.6 software).

Application of hybrid model

The ensemble method, where a set of simple mathematic scenarios represented in Table 1 combining different base models to get better predictive performance compared to an individual model. Thus, to provide the GWR potential map with the ensemble model, the weights obtained from FR model and ME model were integrated to acquire a better GWR potential classifier. Seven ensemble classifier models were identified for GWR potential zones. AUC and CCI were considered as measures of models’ effectiveness. Table 1 showed AUC and CCI measures for seven classification models assessed. Compared with individual models (scenario 8 and scenario 9), it is relatively obvious that the proposed ensemble models outperformed with respect to AUC. CCI is also applied to assess the performance of the models, for which our results exhibited a satisfactory performance for GWR potential mapping. For the different scenarios, AUC and CCI ranged from 0.984 to 0.990 and 54.6 to 90.9, respectively (Table 1 ).

Effective classification of ensemble models in comparative to individual models indicated that the best ensemble classifier, scenario 5, is characterized by better validation measures in both training datasets and testing datasets. In addition, the results of both individual models clearly indicated satisfactory output for GWR prediction, while the FR model attained slightly lower predictive performance than ME model.

The best ensemble model in terms of performance was scenario 5. Then, GWR potential map of scenario 5 was developed using ArcGIS10.6 software. Each pixel has a value that represents the occurrence of GWR potential. Consequently, the map was divided into four classes of low, moderate, high and very high based on natural break method (Fig.  9 ). The area of each class in scenario 5, was represented in Table 2 . Based on the scenario 5, most parts of the studied area had a low GWR potential (73.79%). A very high GWR potential is only found in 3.01% studied area.

GWR potential zone map of ensemble method by scenario 5 (map layout was prepared in ArcGIS10.6 software).

Discussions

Nowadays, groundwater depletion became a critical global problem with serious consequences for sustainability of water supplies. Knowledge of spatial ground water recharge potential zones has placed considerable emphasis to addressing effectively planning and better managing water resources. Correspondingly, there has been growing numbers of research addressing zoning of groundwater recharge potential in recent years 19 , 27 , 29 , 36 , 37 , 54 , 55 . Although there still remains much work to understanding the effect of multiple drivers on the distribution and pattern of GWR potential. Population growth along with increased development led to heavy withdrawals from the plain aquifer system of the Lorestan province and subsequently decline in water table. Mountainous areas play a critical role in delivering water flow to the lowlands; therefore, it is extremely important to understand the GWR potential zones. Depending on hydrogeology, climate, physiography, and groundwater consumption, the volume of stored groundwater may vary considerably in different places 19 , 29 , 84 , 85 . While physical models are promising to quantitative estimation of natural recharge rates to aquifer system, they often require large hydrogeological data which are not easily measurable in the field 86 and require numerically intensive simulation of the behavior. Furthermore, in-depth physical knowledge of the relevant hydrological process is requiring when developing physical based models (Kim et al., 2015). Moreover, many developing countries suffer from lack of information, accordingly susceptibility and potential mapping with help of statistical methods and machine learning methods which requiring few parameters without an explicit need about the underlying physical process becomes a useful tool to subdivide area into regions with significant GWR potential zone. Nevertheless, significant improvements in model prediction and decreased the prediction uncertainty through an ensemble approaches were recently published by researchers 32 , 41 , 44 , 46 . The current state of scientific researches on disaster susceptibility and natural potential by an ensemble models are in the early stage of development.

In our study, GWR potential zones were investigated in Marboreh watershed in Lorestan province, Iran, using a combination of the machine learning (ME) and a bivariate statistical method (FR), into one predictive model in order to improve predictions. These two individual methods are commonly used by researchers in the field of Groundwater potential mapping 41 , 43 , landslide 47 , 48 , flood 44 , 45 . In order to make clearly accurate evaluations of the ensemble methods, FR and ME were also performed independently. The result showed that the ensemble method in scenario 5 performed “excellent’ with AUC and CCI values of 0.990 and 0.907, respectively. This supports the previous findings from Tehrany et al 74 ; Razavi-Termeh et al 28 and Di Napoli et al 87 that ensemble classifiers allow more accurate prediction than classical models and alternative to the individual models in susceptibility and potential mapping. A research carried out in Qued Guenniche in the north of Tunisia to produce a GWR potential map by Chenini and Msaddek 36 . The authors found that the bivariate and multivariate statistical approaches provide more accurate results than the AHP. Pourghasemi et al 37 compared a number of machine learning methods to investigate groundwater recharge potential zones in Firuzkuh County, Iran. The results showed that SVM and MARS outperformed RF in terms of accuracy based on the data from 2000 double ring infiltrometer.

Here, we also address importance effective factors of GWR potential. Precisely, to quantify effects of topographical, hydrological, geological and ecological factors on GWR potential using data from 850 observations. Jackknife test was applied to evaluate the relative importance of each causal factors. The result indicated that soil had the highest training gain. This is followed by lithology, aspect and DEM. Our findings are in agree with pervious researches that highlight the influence of the soil, lithology, elevation, aspect and drainage density in GWR potential identification 29 , 55 . Similar results were found by Pourghasemi et al 37 in the Firuzkuh County, Iran, who noted that the mean annual rainfall, drainage density, elevation and slope angel are dominate factors that influence groundwater recharge. Rainfall data have been widely used in understanding the GWR potential zones 21 , 25 , 29 , 37 , 55 , 88 . The spatial distribution of the final GWR potential map is consistent with the findings of several studies. Patil and Mohite 88 , Senanayake et al 29 , Yeh et al 25 , de Costa et al 21 and Pourghasemi et al 37 obtained the high effective GWR potential in gentle slope. In contrast, the least GWR potential can be found in higher drainage density zones 25 , 54 . Additionally, high altitude along with steep slope areas could identified poor GWR potential. Furthermore, based on Jackknife test, SPI, plan and profile curvatures have no particular impact on GWR potential in studied area. Moreover, the final GWR potential maps indicated that the areas with high percolation potential are mainly distributed along the south and southwestern parts of the study area.

Marboreh Watershed continue to an experience water scarcity, largely driven by frequent droughts and expanding agricultural land combined with over abstraction of groundwater. Therefore, there is a need to identify potential groundwater recharge zones to help prevent water scarcity. In this study to address water scarcity issue we discuss the combination of the predictions achieved by FR and MA models in the ensemble in order to develop a GWR potential zones for the Marboreh watershed in Lorestan province, Iran. Moreover, the spatial mapping for managed aquifer recharge based on topographic factors, hydrological factors, geological factors and ecological factors and assessing their likely contribution on GWR is explored. The results of the research reveal that evaluating the individual’s models using the AUC and CCI indices reach relatively high performance. The ensemble classifier with scenario 5 may achieves more accurate result than other ensemble models and individual models with AUC and CCI values of 0.990 and 0.907, respectively. Additionally, GWR potential areas were found to vary largely based on soil, lithology, aspect and elevation. The results of the study revealed that about 11.85% of the Marboreh watershed was found to have a very high to high recharge potential and 73.79% falls in the low zone class.

It has been generally recognized that, the accuracy of the outputs of interest is dependent mainly on the input uncertainty. There are large uncertainties in soil spatial information, insufficient knowledge of underlying groundwater system and inadequate of data measurements. Additionally, higher uncertainties are associated with aquifer characteristics and recharge process needs to be address of further studies. Here, since there were no available data for hydraulic conductivity, double-ring infiltrometer method was used to identify varying groundwater recharge potential. Thus, priority needs for validate model predictions in area with monitoring wells and resistivity imaging of water movement in vadose zone.

The results of this study could be used in the planning, management, and control of surface runoff in high-discharge events. It is recommended to construct aquifer recharge facilities in areas with “high” and “very high” percolation potential. Using these findings, it is possible to obtain higher quantities of surface water and recharge aquifers, as well as managing water deficit. The studied area heavily relies on the abstraction of groundwater to sustain irrigated agriculture. Consequently, identifying the groundwater recharge zones is promising, especially for future droughts, but requires more scientific support and supervision to succeed. The studied area is suffering from a combination of water crisis and groundwater pollution. Therefore, identifying zone with a high GWR potential is critical to enhance water storage and taking acting by local managers to define conservation priorities areas to reduce nutrient losses and enhances aquifer recharge.

Gerland, P., Raftery, A. E., Ševčíková, H., Li, N., Gu, D., Spoorenberg, T., & Bay, G. World population stabilization unlikely this century. Science 346 (6206), 234–237 (2014).

Ortiz‐Zamora, D., & Ortega‐Guerrero, A. Evolution of long‐term land subsidence near Mexico City: Review, field investigations, and predictive simulations. Water Resour. Res. 46 (1) (2010).

Chaussard, E., Wdowinski, S., Cabral-Cano, E. & Amelung, F. Land subsidence in central Mexico detected by ALOS InSAR time-series. Remote Sens. Environ. 140 , 94–106 (2014).

Article   ADS   Google Scholar  

Werner, A. D., Bakker, M., Post, V. E., Vandenbohede, A., Lu, C., Ataie-Ashtiani, B., & Barry, D. A. Seawater intrusion processes, investigation and management: Recent advances and future challenges. Adv. Water Resour. 51 , 3–26 (2013).

Kulkarni, H., Shah, M. & Shankar, P. V. Shaping the contours of groundwater governance in India. J. Hydrol. Region. Stud. 4 , 172–192 (2015).

Article   Google Scholar  

Harou, J. J. & Lund, J. R. Ending groundwater overdraft in hydrologic-economic systems. Hydrogeol. J. 16 (6), 1039 (2008).

Qureshi, M. E., Reeson, A., Reinelt, P., Brozović, N. & Whitten, S. Factors determining the economic value of groundwater. Hydrogeol. J. 20 (5), 821–829 (2012).

Richey, A. S. et al. Uncertainty in global groundwater storage estimates in a T otal G roundwater S tress framework. Water Resour. Res. 51 (7), 5198–5216 (2015).

Article   ADS   PubMed   PubMed Central   Google Scholar  

Haghighi, A. T., Darabi, H., Shahedi, K., Solaimani, K. & Kløve, B. A scenario-based approach for assessing the hydrological impacts of land use and climate change in the Marboreh Watershed, Iran. Environ. Model. Assess. 25 (1), 41–57 (2020).

Rouhani, H. & Jaafarzadeh, M. S. Assessing the climate change impact on hydrological response in the Gorganrood river basin, Iran. J. Water Clim. Change 9 (3), 421–433 (2018).

Adnan, M. S., Ali, N. C., Erfen, Y., Rahmat, S. N., Razi, M. A. M., & Musa, S. Analysis the impact of bridges existance for the segamat river using infowork RS. In IOP Conference Series: Materials Science and Engineering Vol. 136(1) 012080. (IOP Publishing, 2016).

Moe, C. L. & Rheingans, R. D. Global challenges in water, sanitation and health. J. Water Health 4 (S1), 41–57 (2006).

Article   CAS   PubMed   Google Scholar  

Ayob, S., & Rahmat, S. N. Rainwater harvesting (RWH) and groundwater potential as alternatives water resources in Malaysia: A review. In MATEC Web of Conferences Vol. 103 04020. (EDP Sciences, 2017).

Carmon, N. & Shamir, U. Water-sensitive planning: Integrating water considerations into urban and regional planning. Water Environ. J. 24 (3), 181–191 (2010).

Braune, E. & Xu, Y. The role of ground water in Sub-Saharan Africa. Groundwater 48 (2), 229–238 (2010).

Article   CAS   Google Scholar  

Howard, K. W. Sustainable cities and the groundwater governance challenge. Environ. Earth Sci. 73 (6), 2543–2554 (2015).

Machiwal, D. & Jha, M. K. Identifying sources of groundwater contamination in a hard-rock aquifer system using multivariate statistical analyses and GIS-based geostatistical modeling techniques. J. Hydrol. Region. Stud. 4 , 80–110 (2015).

Mogaji, K. A., Lim, H. S. & Abdullah, K. Regional prediction of groundwater potential mapping in a multifaceted geology terrain using GIS-based Dempster-Shafer model. Arab. J. Geosci. 8 (5), 3235–3258 (2015).

Souissi, D. et al. Mapping groundwater recharge potential zones in arid region using GIS and Landsat approaches, southeast Tunisia. Hydrol. Sci. J. 63 (2), 251–268 (2018).

Freeze, R. A., & Cherry, J. A. Groundwater (No. 629.1 F7) (1979).

De Costa, A. M., de Salis, H. H. C., Viana, J. H. M., & Leal Pacheco, F. A. Groundwater recharge potential for sustainable water use in urban areas of the Jequitiba River Basin, Brazil. Sustainability 11 (10), 2955 (2019).

Sangab Zagros Consulting Engineering Company (SZCEC). Comprehensive studies of Azna and Aligodarz watershed, Lorestan Province, Project report; Regional Water Company of Lorestan Province: Lorestan, Iran (2012).

Sashikkumar, M. C., Selvam, S., Kalyanasundaram, V. L. & Johnny, J. C. GIS based groundwater modeling study to assess the effect of artificial recharge: A case study from Kodaganar river basin, Dindigul district, Tamil Nadu. J. Geol. Soc. India 89 (1), 57–64 (2017).

Scanlon, B. R., Healy, R. W. & Cook, P. G. Choosing appropriate techniques for quantifying groundwater recharge. Hydrogeol. J. 10 (1), 18–39 (2002).

Article   ADS   CAS   Google Scholar  

Yeh, H. F., Cheng, Y. S., Lin, H. I. & Lee, C. H. Mapping groundwater recharge potential zone using a GIS approach in Hualian River, Taiwan. Sustain. Environ. Res. 26 (1), 33–43 (2016).

Keese, K. E., Scanlon, B. R., & Reedy, R. C. Assessing controls on diffuse groundwater recharge using unsaturated flow modeling. Water Resour. Res. 41 (6) (2005).

Benjmel, K. et al. Mapping of groundwater potential zones in crystalline terrain using remote sensing, GIS techniques, and multicriteria data analysis (case of the Ighrem Region, Western Anti-Atlas, Morocco). Water 12 (2), 471 (2020).

Razavi-Termeh, S. V., Sadeghi-Niaraki, A. & Choi, S. M. Groundwater potential mapping using an integrated ensemble of three bivariate statistical models with random forest and logistic model tree models. Water 11 (8), 1596 (2019).

Senanayake, I. P., Dissanayake, D. M. D. O. K., Mayadunna, B. B. & Weerasekera, W. L. An approach to delineate groundwater recharge potential sites in Ambalantota, Sri Lanka using GIS techniques. Geosci. Front. 7 (1), 115–124 (2016).

Siahkamari, S., Haghizadeh, A., Zeinivand, H., Tahmasebipour, N. & Rahmati, O. Spatial prediction of flood-susceptible areas using frequency ratio and maximum entropy models. Geocarto Int. 33 (9), 927–941 (2018).

Somodi, I., Lepesi, N. & Botta-Dukát, Z. Prevalence dependence in model goodness measures with special emphasis on true skill statistics. Ecol. Evolut. 7 (3), 863–872 (2017).

Tran, Q. Q., Willems, P. & Huysmans, M. Coupling catchment runoff models to groundwater flow models in a multi-model ensemble approach for improved prediction of groundwater recharge, hydraulic heads and river discharge. Hydrogeol. J. 27 (8), 3043–3061 (2019).

Dar, T., Rai, N., & Bhat, A. Delineation of potential groundwater recharge zones using analytical hierarchy process (AHP). Geol. Ecol. Landsc. 1–16 (2020).

Nohani, E., Moharrami, M., Sharafi, S., Khosravi, K., Pradhan, B., Pham, B. T., & Melesse, A. Landslide susceptibility mapping using different GIS-based bivariate models. Water 11 (7), 1402 (2019).

Mogaji, K. A., Omosuyi, G. O., Adelusi, A. O. & Lim, H. S. Application of GIS-based evidential belief function model to regional groundwater recharge potential zones mapping in hardrock geologic terrain. Environ. Process. 3 (1), 93–123 (2016).

Chenini, I. & Msaddek, M. H. Groundwater recharge susceptibility mapping using logistic regression model and bivariate statistical analysis. Q. J. Eng. Geol. Hydrogeol. 53 (2), 167–175 (2020).

Pourghasemi, H. R. et al. Assessing and mapping multi-hazard risk susceptibility using a machine learning technique. Sci. Rep. 10 (1), 1–11 (2020).

Althuwaynee, O. F., Pradhan, B. & Lee, S. A novel integrated model for assessing landslide susceptibility mapping using CHAID and AHP pair-wise comparison. Int. J. Remote Sens. 37 (5), 1190–1209 (2016).

Chen, W. et al. A comparative study of logistic model tree, random forest, and classification and regression tree models for spatial prediction of landslide susceptibility. CATENA 151 , 147–160 (2017).

Tien Bui, D., Shirzadi, A., Shahabi, H., Chapi, K., Omidavr, E., Pham, B. T., & Bin Ahmad, B. A novel ensemble artificial intelligence approach for gully erosion mapping in a semi-arid watershed (Iran). Sensors 19 (11), 2444 (2019).

Avand, M., Janizadeh, S., Tien Bui, D., Pham, V. H., Ngo, P. T. T., & Nhu, V. H. A tree-based intelligence ensemble approach for spatial prediction of potential groundwater. Int. J. Digit. Earth 1–22 (2020).

Chen, W., Li, H., Hou, E., Wang, S., Wang, G., Panahi, M. & Xiao, L. GIS-based groundwater potential analysis using novel ensemble weights-of-evidence with logistic regression and functional tree models. Sci. Total Environ. 634 , 853–867 (2018).

Kordestani, M. D. et al. Groundwater potential mapping using a novel data-mining ensemble model. Hydrogeol. J. 27 (1), 211–224 (2019).

Choubin, B. et al. An ensemble prediction of flood susceptibility using multivariate discriminant analysis, classification and regression trees, and support vector machines. Sci. Total Environ. 651 , 2087–2096 (2019).

Article   ADS   CAS   PubMed   Google Scholar  

Tien Bui, D., Khosravi, K., Shahabi, H., Daggupati, P., Adamowski, J. F., Melesse, A. M., & Pradhan, B. Flood spatial modeling in northern Iran using remote sensing and gis: A comparison between evidential belief functions and its ensemble with a multivariate logistic regression model. Remote Sens. 11 (13), 1589 (2019).

Hong, H., Liu, J. & Zhu, A. X. Modeling landslide susceptibility using LogitBoost alternating decision trees and forest by penalizing attributes with the bagging ensemble. Sci. Total Environ. 718 , 137231 (2020).

Pham, B. T., Jaafari, A., Prakash, I. & Bui, D. T. A novel hybrid intelligent model of support vector machines and the MultiBoost ensemble for landslide susceptibility modeling. Bull. Eng. Geol. Environ. 78 (4), 2865–2886 (2019).

Pradhan, A. M. S., Kang, H. S., Lee, J. S., & Kim, Y. T. An ensemble landslide hazard model incorporating rainfall threshold for Mt. Umyeon, South Korea. Bull. Eng. Geol. Environ. 78 (1), 131–146 (2019).

Mousavi, S. N. & Gharghani, F. Assessing policies of irrigation for groundwater by positive mathematical programming (PMP) case study: Eghlid. Econ. Res. 11 (4), 65–82 (2012).

Google Scholar  

Döll, P., & Flörke, M. Global-scale estimation of diffuse groundwater recharge: Model tuning to local data for semi-arid and arid regions and assessment of climate change impact (2005).

Aghazadeh, N., Chitsazan, M. & Mirzayi, Y. Assessing the potential and actual recharge in urban aquifer and mapping areas with recharge potential using GIS and AHP (Case study: Urmia city aquifer). Adv. Appl. Geol. 9 (2), 168–179 (2019) ( (in Persian) ).

Juandi, M. & Syahril, S. Empirical relationship between soil permeability and resistivity, and its application for determining the groundwater gross recharge in Marpoyan Damai, Pekanbaru, Indonesia. Water Pract. Technol. 12 (3), 660–666 (2017).

Jang, C. S., Chen, S. K. & Kuo, Y. M. Applying indicator-based geostatistical approaches to determine potential zones of groundwater recharge based on borehole data. CATENA 101 , 178–187 (2013).

Singh, S. K., Zeddies, M., Shankar, U. & Griffiths, G. A. Potential groundwater recharge zones within New Zealand. Geosci. Front. 10 (3), 1065–1072 (2019).

Senthilkumar, M., Gnanasundar, D. & Arumugam, R. Identifying groundwater recharge zones using remote sensing & GIS techniques in Amaravathi aquifer system, Tamil Nadu, South India. Sustain. Environ. Res. 29 (1), 15 (2019).

Nampak, H., Pradhan, B. & Manap, M. A. Application of GIS based data driven evidential belief function model to predict groundwater potential zonation. J. Hydrol. 513 , 283–300 (2014).

Prasad, R. K., Mondal, N. C., Banerjee, P., Nandakumar, M. V. & Singh, V. S. Deciphering potential groundwater zone in hard rock through the application of GIS. Environ. Geol. 55 (3), 467–475 (2008).

Das, S. Delineation of groundwater potential zone in hard rock terrain in Gangajalghati block, Bankura district, India using remote sensing and GIS techniques. Model. Earth Syst. Environ. 3 (4), 1589–1599 (2017).

Rajasekhar, M. et al. Delineation of groundwater potential zones of semi-arid region of YSR Kadapa District, Andhra Pradesh, India using RS, GIS and analytic hierarchy process. Remote Sens. Land 2 (2), 76–86 (2018).

Ayazi, M. H. et al. Disasters and risk reduction in groundwater: Zagros Mountain Southwest Iran using geoinformatics techniques. Disaster Adv. 3 (1), 51–57 (2010).

Gee, G. W. et al. Variations in water balance and recharge potential at three western desert sites. Soil Sci. Soc. Am. J. 58 (1), 63–72 (1994).

Prych, E. A. Using chloride and chlorine-36 as soil-water tracers to estimate deep percolation at selected locations on the US Department of Energy Hanford Site, Washington Vol. 2481. (US Geological Survey, 1998).

Al-Abadi, A. M. & Shahid, S. A comparison between index of entropy and catastrophe theory methods for mapping groundwater potential in an arid region. Environ. Monit. Assess. 187 (9), 576 (2015).

Article   PubMed   Google Scholar  

Razandi, Y., Farokhzadeh, B., Yousefzadeh Chabok, M. & Teimurian, T. Applying maximum entropy algorithm (MAXENT) in groundwater potential mapping, case study: Hamedan-Bahar Plain. J. Irrigation Water Eng. 8 (1), 111–124 (2017) ( (in persion) ).

Constantin, M., Bednarik, M., Jurchescu, M. C. & Vlaicu, M. Landslide susceptibility assessment using the bivariate statistical analysis and the index of entropy in the Sibiciu Basin (Romania). Environ. Earth Sci. 63 (2), 397–406 (2011).

Wang, Q., Li, W., Yan, S., Wu, Y. & Pei, Y. GIS based frequency ratio and index of entropy models to landslide susceptibility mapping (Daguan, China). Environ. Earth Sci. 75 (9), 780 (2016).

Hosseini, S. Z. et al. Modelling potential habitats for Artemisia sieberi and Artemisia aucheri in Poshtkouh area, central Iran using the maximum entropy model and geostatistics. Ecol. Inform. 18 , 61–68 (2013).

Sahragard, H. P. & Ajorlo, M. A comparison of logistic regression and maximum entropy for distribution modeling of range plant species (a case study in rangelands of western Taftan, southeastern Iran). Turk. J. Bot. 42 (1), 28–37 (2018).

Main, I. G., & Naylor, M. Maximum entropy production and earthquake dynamics. Geophys. Res. Lett. 35 (19) (2008).

Suárez-Seoane, S., de la Morena, E. L. G., Prieto, M. B. M., Osborne, P. E., & de Juana, E. Maximum entropy niche-based modelling of seasonal changes in little bustard ( Tetrax tetrax ) distribution. Ecol. Model. 219 (1–2), 17–29 (2008).

Hou, E., Wang, J. & Chen, W. A comparative study on groundwater spring potential analysis based on statistical index, index of entropy and certainty factors models. Geocarto Int. 33 (7), 754–769 (2018).

Alizadeh, Z. & Mahjouri, N. A spatiotemporal Bayesian maximum entropy-based methodology for dealing with sparse data in revising groundwater quality monitoring networks: The Tehran region experience. Environ. Earth Sci. 76 (12), 436 (2017).

Rokach, L. Ensemble-based classifiers. Artif. Intell. Rev. 33 (1–2), 1–39 (2010).

Tehrany, M. S., Kumar, L. & Shabani, F. A novel GIS-based ensemble technique for flood susceptibility mapping using evidential belief function and support vector machine: Brisbane, Australia. PeerJ 7 , e7653 (2019).

Mallick, J. et al. Modeling groundwater potential zone in a semi-arid region of Aseer using fuzzy-AHP and geoinformation techniques. Water 11 (12), 2656 (2019).

Fielding, A. H. & Bell, J. F. A review of methods for the assessment of prediction errors in conservation presence/absence models. Environ. Conserv. 24 (1), 38–49 (1997).

Snoke, A. W., Tullis, J., & Todd, V. R. (Eds.). Fault-Related Rocks: A Photographic Atlas . (Princeton University Press, 2014).

Li, X., Zhao, S., Yang, H., Cong, D., & Zhang, Z. A bi-band binary mask based land-use change detection using Landsat 8 OLI imagery. Sustainability 9 (3), 479 (2017).

Rwanga, S.S. & Ndambuki, J.M. Accuracy assessment of land use/land cover classification using remote sensing and GIS. Int. J. Geosci. 8( 04), 611 (2017).

Michaelsen, T., Heinrich, R., & Frisk Smith, T. A. Watershed management field manual; road design and construction in sensitive watersheds (No. FAO CG-13/5) . (FAO, 1989).

Pirasteh, S. & Li, J. Landslides investigations from geoinformatics perspective: Quality, challenges, and recommendations. Geomat. Nat. Haz. Risk 8 (2), 448–465 (2017).

Zavoianu, I. Morphometry of Drainage Basins . (Elsevier, 2011).

Ben-Zion, Y., & Sammis, C. Mechanics, structure and evolution of fault zones. In Mechanics, Structure and Evolution of Fault Zones 1533–1536. (Birkhäuser Basel, 2009).

Riad, P. H., Billib, M., Hassan, A. A., Salam, M. A. & El Din, M. N. Application of the overlay weighted model and Boolean logic to determine the best locations for artificial recharge of groundwater. J. Urban Environ. Eng. 5 (2), 57–66 (2011).

Tweed, S. O., Leblanc, M., Webb, J. A. & Lubczynski, M. W. Remote sensing and GIS for mapping groundwater recharge and discharge areas in salinity prone catchments, southeastern Australia. Hydrogeol. J. 15 (1), 75–96 (2007).

Nayak, P. C., Sudheer, K. P., Rangan, D. M., & Ramasastri, K. S. Short‐term flood forecasting with a neurofuzzy model. Water Resour. Res. 41 (4) (2005).

Di Napoli, M., Carotenuto, F., Cevasco, A., Confuorto, P., Di Martire, D., Firpo, M., Giacomo, P., Emanoele, R. & Calcaterra, D. Machine learning ensemble modelling as a tool to improve landslide susceptibility mapping reliability. Landslides 1–18 (2020).

Patil, S. G. & Mohite, N. M. Identification of groundwater recharge potential zones for a watershed using remote sensing and GIS. Int. J. Geomat. Geosci. 4 (3), 485–498 (2014).

Download references

Author information

Authors and affiliations.

Department of Watershed Management Engineering, Faculty of Agriculture, Lorestan University, Khorramabad, Iran

Maryam Sadat Jaafarzadeh, Naser Tahmasebipour & Ali Haghizadeh

Department of Natural Resources and Environmental Engineering, College of Agriculture, Shiraz University, Shiraz, Iran

Hamid Reza Pourghasemi

Department of Range and Watershed, Management, College of Agriculture Science and Natural Resource, Gonbad-e-Kavous University, Golestan, Iran

Hamed Rouhani

You can also search for this author in PubMed   Google Scholar

Contributions

This manuscript prepared by team. M.S.J., H.R.P. and H.R. run the models and ensemble scenarios. M.S.J. and H.R. wrote the main manuscript text, figures and tables prepared by M.S.J. and H.R.P., H.R., H.R.P. and A.H. reviewed the manuscript. Field survey done by M.S.J., for some weeks (with 3 participants). N.T. participated in the field operation.

Corresponding author

Correspondence to Maryam Sadat Jaafarzadeh .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Jaafarzadeh, M.S., Tahmasebipour, N., Haghizadeh, A. et al. Groundwater recharge potential zonation using an ensemble of machine learning and bivariate statistical models. Sci Rep 11 , 5587 (2021). https://doi.org/10.1038/s41598-021-85205-6

Download citation

Received : 10 January 2020

Accepted : 14 January 2021

Published : 10 March 2021

DOI : https://doi.org/10.1038/s41598-021-85205-6

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

This article is cited by

Millennium-scale changes in the atlantic multidecadal oscillation influenced groundwater recharge rates in italy.

• Nazzareno Diodato
• Gianni Bellocchi

Communications Earth & Environment (2024)

Application of hybrid model-based machine learning for groundwater potential prediction in the north central of Vietnam

• Huu Duy Nguyen
• Van Hong Nguyen
• Quang-Thanh Bui

Earth Science Informatics (2024)

Integrated machine learning and remote sensing for groundwater potential mapping in the Mekong Delta in Vietnam

• Quoc-Huy Nguyen
• Alexandru-Ionut Petrisor

Acta Geophysica (2024)

Delineation of groundwater potential zones using GIS and AHP techniques in Coimbatore district, South India

• B. Gurugnanam

International Journal of Energy and Water Resources (2024)

GIS based multi-criteria decision making to identify regional groundwater potential zones: a critical review

• Shweta Kodihal
• M. P. Akhtar

Sustainable Water Resources Management (2024)

By submitting a comment you agree to abide by our Terms and Community Guidelines . If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

Quick links

• Explore articles by subject
• Guide to authors
• Editorial policies

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Modeling groundwater level using geographically weighted regression

• Original Paper
• Published: 26 August 2024
• Volume 17 , article number  251 , ( 2024 )

Cite this article

• Yuganshu Badetiya 1 &
• Mahesh Barale   ORCID: orcid.org/0000-0001-8254-3604 1  

54 Accesses

Explore all metrics

An economic development, crop production, and socioeconomic development highly dependent on the availability of groundwater resources in nearby areas. In order to manage groundwater sustainably, it is crucial to predict groundwater levels. Analysis of groundwater levels along with various influential factors becomes possible due to the availability of remotely sensed geospatial data. The spatially differing groundwater level is highly influenced by the geographical factors called influential factors as like elevation and slope. In the present study, we use the spatial regression and geographically weighted regression (GWR) models for predicting the groundwater level. The GWR model gives comparatively satisfactory results as compared to the three variants of the spatial regression models with lower Bayesian information criterion value (1101.04) and highest \(R^2\) value (0.84). It can be noted that the factors of vegetation index, drought index, elevation, and topographic position positively affect the groundwater level. While the factors of roughness, surface temperature, precipitation, and runoff are affected negatively. The current study highlights that GWR model is useful for exploring the spatial relationships between the different influencing factors and the groundwater level.

Graphical abstract

Prediction groundwater level using geographically weighted regression

This is a preview of subscription content, log in via an institution to check access.

Access this article

Subscribe and save.

• Get 10 units per month
• Download Article/Chapter or eBook
• 1 Unit = 1 Article or 1 Chapter
• Cancel anytime

Price includes VAT (Russian Federation)

Instant access to the full article PDF.

Rent this article via DeepDyve

Institutional subscriptions

Similar content being viewed by others

Integrated geospatial, geostatistical, and remote-sensing approach to estimate groundwater level in North-western India

To investigate groundwater potentiality, a GIS-based model was integrated with remote sensing data in the Northwest Gulf of Suez (Egypt)

Geospatial distribution modeling and determining suitability of groundwater quality for irrigation purpose using geospatial methods and water quality index (wqi) in northern ethiopia, data availability.

Data will be made available on request.

Adamowski J, Chan HF (2011) A wavelet neural network conjunction model for groundwater level forecasting. J Hydrol 407(1–4):28–40

Article   Google Scholar  

Bidanset PE, Lombard JR (2014) Evaluating spatial model accuracy in mass real estate appraisal: a comparison of geographically weighted regression and the spatial lag model. Cityscape 16(3):169–182

Google Scholar  

Bourgault G (1997) Spatial declustering weights. Mathematical Geology 29:277–290

Brown S, Versace VL, Laurenson L, Ierodiaconou D, Fawcett J, Salzman S (2012) Assessment of spatiotemporal varying relationships between rainfall, land cover and surface water area using geographically weighted regression. Environmental Modeling & Assessment 17:241–254

Chaudhary V, Satheeshkumar S (2018) Assessment of groundwater quality for drinking and irrigation purposes in arid areas of Rajasthan, India. Appl Water Sci 8:1–17

Chinnasamy P, Maheshwari B, Prathapar S (2015) Understanding groundwater storage changes and recharge in Rajasthan, India through remote sensing. Water 7(10):5547–5565

Choubisa SL, Choubisa D, Choubisa A (2023) Fluoride contamination of groundwater and its threat to health of villagers and their domestic animals and agriculture crops in Rural Rajasthan, India. Environ Geochem Health 45(3):607–628

Article   CAS   Google Scholar  

Elhorst JP, Halleck Vega S (2017) The SLX model: extensions and the sensitivity of spatial spillovers to W. Papeles de Economía Española 152:34–50

Fotheringham AS, Brunsdon C, Charlton M (2003) Geographically weighted regression: the analysis of spatially varying relationships. John Wiley & Sons

Gautam VK, Kothari M, Singh P, Bhakar S, Yadav K (2022) Decadal groundwater level changes in Pratapgarh district of Southern Rajasthan, India. Ecology Environment & Conservation 28(1):283–289

Gleeson T, Cuthbert M, Ferguson G, Perrone D (2020) Global groundwater sustainability, resources, and systems in the anthropocene. Annu Rev Earth Planet Sci 48:431–463

Granger R (1989) An examination of the concept of potential evaporation. J Hydrol 111(1–4):9–19

Jasrotia A, Kumar A, Singh R (2016) Integrated remote sensing and GIS approach for delineation of groundwater potential zones using aquifer parameters in Devak and Rui watershed of Jammu and Kashmir, India. Arab J Geosci 9:1–15

Koh E-H, Lee E, Lee K-K (2020) Application of geographically weighted regression models to predict spatial characteristics of nitrate contamination: implications for an effective groundwater management strategy. J Environ Manage 268

Kumar P, Herath S, Avtar R, Takeuchi K (2016) Mapping of groundwater potential zones in Killinochi area, Sri Lanka, using GIS and remote sensing techniques. Sustainable Water Resources Management 2:419–430

Mansour S, Al Kindi A, Al-Said A, Al-Said A, Atkinson P (2021) Sociodemographic determinants of COVID-19 incidence rates in Oman: geospatial modelling using multiscale geographically weighted regression (MGWR). Sustain Cities Soc 65

Mosavi A, Sajedi Hosseini F, Choubin B, Goodarzi M, Dineva AA, Rafiei Sardooi E (2021) Ensemble boosting and bagging based machine learning models for groundwater potential prediction. Water Resour Manage 35:23–37

Mukherjee S, Joshi PK, Mukherjee S, Ghosh A, Garg R, Mukhopadhyay A (2013) Evaluation of vertical accuracy of open source digital elevation model (DEM). Int J Appl Earth Obs Geoinf 21:205–217

Noguera I, Vicente-Serrano SM, Domínguez-Castro F, Reig F (2022) Assessment of parametric approaches to calculate the evaporative demand drought index. Int J Climatol 42(2):834–849

Pande CB, Moharir KN, Panneerselvam B, Singh SK, Elbeltagi A, Pham QB, Varade AM, Rajesh J (2021) Delineation of groundwater potential zones for sustainable development and planning using analytical hierarchy process (AHP), and MIF techniques. Appl Water Sci 11(12):186

Pourghasemi HR, Sadhasivam N, Yousefi S, Tavangar S, Nazarlou HG, Santosh M (2020) Using machine learning algorithms to map the groundwater recharge potential zones. J Environ Manage 265:110525

Pradhan AMS, Kim Y-T, Shrestha S, Huynh T-C, Nguyen B-P (2021) Application of deep neural network to capture groundwater potential zone in mountainous terrain, Nepal Himalaya. Environ Sci Pollut Res 28:18501–18517

Pranjal P, Kumar D, Soni A, Chatterjee R (2024) Assessment of groundwater level using satellite-based hydrological parameters in North-West India: a deep learning approach. Earth Science Informatics, pages 1–14

Pyrcz M, Deutsch C (2003) Declustering and debiasing. Newsletter 19:1–14

Rajput H, Goyal R, Brighu U (2020) Modification and optimization of drastic model for groundwater vulnerability and contamination risk assessment for Bhiwadi region of Rajasthan, India. Environmental Earth Sciences 79:1–15

Rizeei HM, Pradhan B, Saharkhiz MA, Lee S (2019) Groundwater aquifer potential modeling using an ensemble multi-adoptive boosting logistic regression technique. J Hydrol 579:124172

Sachdeva S, Kumar B (2021) Comparison of gradient boosted decision trees and random forest for groundwater potential mapping in Dholpur (Rajasthan), India. Stoch Env Res Risk Assess 35:287–306

Saikia P, Chetry N (2020) Study of fluctuations in the groundwater level in Rajasthan: a spatio-temporal approach. International Journal of Engineering and Technical Research 9:1188–1192

Sattari MT, Mirabbasi R, Sushab RS, Abraham J (2018) Prediction of groundwater level in ardebil plain using support vector regression and m5 tree model. Groundwater 56(4):636–646

Tao H, Hameed MM, Marhoon HA, Zounemat-Kermani M, Heddam S, Kim S, Sulaiman SO, Tan ML, Sa’adi Z, Mehr AD et al (2022) Groundwater level prediction using machine learning models: a comprehensive review. Neurocomputing 489:271–308

Download references

Author information

Authors and affiliations.

Department of Statistics, Central University of Rajasthan, Ajmer, India

Yuganshu Badetiya & Mahesh Barale

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Mahesh Barale .

Ethics declarations

Conflict of interest.

The authors declare no competing interests.

Additional information

Responsible Editor: Amjad Kallel.

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Badetiya, Y., Barale, M. Modeling groundwater level using geographically weighted regression. Arab J Geosci 17 , 251 (2024). https://doi.org/10.1007/s12517-024-12051-x

Download citation

Received : 12 February 2024

Accepted : 02 August 2024

Published : 26 August 2024

DOI : https://doi.org/10.1007/s12517-024-12051-x

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Groundwater level
• Spatial regression models
• Prediction performance
• Find a journal
• Publish with us
• Track your research

We Trust in Human Precision

20,000+ Professional Language Experts Ready to Help. Expertise in a variety of Niches.

API Solutions

• API Pricing
• Cost estimate
• Customer loyalty program
• Educational Discount
• Non-Profit Discount
• Green Initiative Discount1

Value-Driven Pricing

Unmatched expertise at affordable rates tailored for your needs. Our services empower you to boost your productivity.

• Special Discounts
• Enterprise transcription solutions
• Enterprise translation solutions
• Transcription/Caption API
• AI Transcription Proofreading API

Trusted by Global Leaders

GoTranscript is the chosen service for top media organizations, universities, and Fortune 50 companies.

GoTranscript

One of the Largest Online Transcription and Translation Agencies in the World. Founded in 2005.

Speaker 1: I used to absolutely hate trying to find papers to support my research, but now it is easier than ever to get a sense of what's out there in your research field with connected papers. Connected papers is a really easy and simple tool to use that I use very often, even outside of academia these days. This is what it's like. Essentially, it is a way to visually explore academic papers in a visual graph, and that's what I love about it. It gives you a snapshot of a research. In the past, you would have to read loads of peer-reviewed papers and get a sixth sense of your field. Now, that sixth sense can be obtained from this tool. You search initially for keywords, a title, a DOI, which is a document object identifier, or a URL of a paper. I've got this paper here. This is one of my papers, and I'm going to search for it by title. I'm going to copy and paste that into the connected paper search bar, and I'm going to say, build a graph. This is what the core functionality is all about, building this graph. This is the paper that I want to select. You can see that it's got some other ones that maybe I meant, but I meant this one, and then it creates this graph. If it hasn't created it recently, it will create one, and there will be a waiting page, but ultimately, this is what you end up with, no matter what. What you end up with is this three-panel layout. The first panel over here is a list of all of the papers that it's found. In the middle, you get this kind of network graph, and on the side, you get an overview of the paper that you've selected. Here it's in purple, and that's our seed graph, or origin paper, as they call it. Here you can see this is the abstract, and with this abstract, we can get a sense of what the paper is actually about. However, this is probably the most interesting panel, the one in the middle, but not as all as it seems, because this isn't a citation graph. It's not about these papers citing each other. In fact, it's much more powerful. Just takes a little bit of getting used to, like what it really means. I always click down here, and this is how to read the graph. I pull this up nearly every time I use this, just to refresh my memory, and papers are arranged according to their similarity, not citations. That's the first very important point you need to understand, is this is based on whether they think certain papers have similar understandings and similar research fields. They do that, I think, by looking at the citations across a range of papers, and those that cite the most papers together are kind of grouped into a research field. That means we're not just relying on citations directly, which means that we get a really nice clustering of similar articles. Here you can see, this is the seed paper, we've got this paper, one of my other papers, one of these papers, and you can see that they're connected by these lines. The thicker the line, and we'll go here to see that here, the similar papers have strong connecting lines and cluster together. Look at my sad little lonely paper right here, all alone, whereas this little cluster over here could be something that I could explore because I haven't really got much of a strong connection to those at the moment. When I hover over one of these bubbles, you'll see the connection that it's made. It goes from this one to Stapleton to Stapleton, so there's three connections. I could probably have a look at this cluster and see if I could make my research more aligned with this little sort of information bubble right here, but we're not finished there. First of all, they're collected based on how similar the papers are. The second thing you need to know is that the node size is the number of citations, so that means really heavily cited articles, ones that have got a lot of interest in the scientific realm, are bigger. This one is bigger, this one is smaller, this one is small, this is my paper, it doesn't have many citations. All of these ones are actually more heavily cited than my paper because the node is bigger. The third thing that's really important is the color. The color relates to the time in which or the date on which it was published. The lighter the color, the further in the past. What you really want is a nice big node that is dark in color. That tells you that a lot of people are citing it and it's a relatively recent paper. Those are the ones I'd be looking at if I was in a research field because I want to know those papers and find out why they're so exciting. Maybe there's something in there that can inform my research. But you can see here that this one's large, it's lighter, this one's large, it's lighter. We'd expect that correlation. The longer a research paper has been around, the more citations it could attract. But it's these ones that are more interesting to me, so a nice big dark one that's nice and fat. These ones over here look pretty good. This one's really nice and dark. It's recent but it hasn't got many citations, not huge. The bigger, the better. That's what I'd be looking for in my initial viewpoint and my first exploration of this network graph. That's the important things you need to know. It's not a citation tree. In fact, it's much more powerful than that. This is how I would use it. I would start looking here and I really like the list view. If I go up here to list view, you can see then we've got all of these that are similar to my paper and this is what I'm interested in over here. Similarity to origin, which is how they cluster it. But we can also sort by year, by citations, by references and we can download the lot into any one of my reference managers using the bib file. Probably the two most important places on this for me at the moment are these two areas, prior works and derivative works. If I want to see what happened before a certain paper, this is where I would go. I want to go in here and have a look to see if there's anything that is really heavily cited and therefore I must know because it is seminal work, it is very important foundational work for my field of research. Up here it gives you a little bit of a blurb on what this really means. Then we've also got derivative works. These are papers that are cited by many of the papers in the graph and it tells you that it could be a very important add-on from the work that you are initially interested in. I would want to know that because that is more up-to-date. Now, you'll notice that some of these are in blue and you may be like, why are they in blue? This is the least understood thing about connected papers, I think, in my opinion. If we click on one of them, you can see that these end up in blue. If we click off on this, you can see then we've got these in blue and what it means is that it is cited by the selective derivative work. It means that if we click on something, we can see what in our network actually is citing that piece of work. That tells us then that this is heavily connected and that it is probably something we should be spending a bit more time reading because they are directly citing it. Because remember, this is not a citation graph, this is a similarity graph. So if a paper in this graph is highly connected to my initial research paper and is directly cited by something in this graph, that means I want to know about it. That would be headed straight to Zotero or whatever reference manager you use. The last panel you should know about is this one over here. You've got a range of things you can do. You can click on any one of these references and it will pull up an abstract if it's got it or you can go out and find it in other places. So I clicked on this one by Woo and you can see that we can open it in Semantic Scholar, we can look at the publisher page, we can go to Google Scholar, we can go to Publish, PubMed, we can also report a mistake if it's your paper and you realize they've got it wrong and you can save it. Now I've saved a number of different papers but I don't think I would use this to manage my references. I'd use Mendeley, Zotero, EndNote, something like that. Go check out my other videos where I talk about using all of those. But you can go up here and you can go to Saved Papers and they will appear in your Saved Papers list just in case it's something that you don't want to miss out on. But I'm going to go back to this network graph and yeah, that is ultimately all of the stuff you need to know about connected papers. You can also filter by keyword, whether or not PDF is available, open access and filter by year. So there's two important things you can do with each of these papers in this network graph. The first one is by clicking on it and going to Open Graph, you can create an own graph with its own origin paper. You see how there's only one origin paper here? Essentially, it's created a whole new graph with that one as the origin. The second thing you can do is go to Add Origin over here and what that does is it adds it to this list of origins. So it essentially creates a thicker and more connected and bigger network graph because you're adding more origins, i.e. more papers that you want to find similar papers of. Yes, I think that makes sense. So it means now that this network graph is not only looking for my paper but it's also looking for this one, this was my paper down here, and it's also looking for this one because we've added it as an origin paper and you can see that there's very, very few connections to this paper. In fact, Yang has got more and more connections that are similar to it than my paper all the way out over here. And then if we're not happy with that, you can see the origin ones are in this purple color so we can click it and we can say Remove Origin and then it will reconnect and recalculate that graph without that origin in it. So those are two important tools that you need to know about but I would be getting these graphs and I would be putting them into a reference manager for easy access and use later on. If you like this video and you love references, go check out this one where I talk about how to use Mendeley like a pro. It's a great watch.

• Skip to main content
• Skip to "About this site"

Language selection

• Français
• Search and menus

Analytical Studies Branch Research Paper Series Experimental Estimates of Potential Artificial Intelligence Occupational Exposure in Canada

DOI : https://doi.org/10.25318/11f0019m2024005-eng

Skip to text

Acknowledgements

Executive summary, 1 introduction, 4 conclusion.

Text begins

The authors would like to thank Li Xue, Marc Frenette and Vincent Hardy from Statistics Canada, and Jessica Gallant, Matthew Calver, Jacob Loree and Alan Stark from the Department of Finance Canada for their helpful and constructive comments.

Past studies on technological change have suggested that occupations involving routine and manual tasks will face a higher risk of automation-related job transformation. However, recent advances in artificial intelligence (AI) challenge prior conclusions, as AI is increasingly able to perform non-routine and cognitive tasks. These advances have the potential to affect a broader segment of the labour force than previously thought. This study provides experimental estimates of the number and percentage of workers in Canada potentially susceptible to AI -related job transformation based on the complementarity-adjusted AI occupational exposure index of Pizzinelli et al. (2023), inspired by Felten, Raj and Seamans (2021). Results from the 2016 and 2021 censuses of population suggest that, on average, about 60% of employees in Canada could be exposed to AI -related job transformation, and about half of this group are in jobs that may be highly complementary with AI . Unlike previous waves of automation, which mainly transformed the jobs of less educated employees, AI is more likely to transform the jobs of highly educated employees. Despite facing potentially higher exposure to AI -related job transformation, highly educated employees may be in jobs that could benefit from AI technologies. Compared with employees in other industries, exposure to AI -related job transformation is higher for employees in professional, scientific and technical services; finance and insurance; information and cultural industries; educational services; and health care and social assistance. However, education and health care professionals are more likely to be in jobs that are highly complementary with AI . Employees in industries such as construction, and accommodation and food services face relatively lower exposure to AI -related job transformation. Whether occupations that may benefit from AI will experience relatively higher employment and wage growth remains to be seen, as this depends on factors such as firm productivity and the ability of workers in those occupations to leverage the potential benefits of AI .

Recent developments in the field of artificial intelligence (AI) have fuelled excitement, as well as concerns, regarding its implications for society and the economy. While previous waves of technological transformation raised concerns regarding the future of jobs involving routine and manual tasks, a broader segment of the labour force could be affected in an era when sophisticated large language models such as ChatGPT increasingly excel at performing non-routine and cognitive tasks typically done by highly skilled workers. AI encompasses a lot more than just natural language processing. These technologies not only have the capacity to automate routine tasks but can also augment human decision-making processes and create entirely new opportunities for innovation and efficiency. As AI continues to evolve, it has the potential to reshape industries, redefine job roles and transform the nature of work. With the transformative effects of AI already in motion, it raises renewed concerns about job transformation and the need for workforce adaptation.

This study adopts the complementarity-adjusted AI occupational exposure index of Pizzinelli et al. (2023), inspired by the original AI occupational exposure measure of Felten, Raj and Seamans (2021), and applies it to data from the 2016 and 2021 censuses of population. The experimental estimates presented in this study are largely based on the technological feasibility of automating job tasks. Employers may not immediately replace human labour with AI , even if it is technologically feasible to do so, because of financial, legal and institutional constraints. Consequently, exposure to AI does not necessarily imply a risk of job loss. At the very least, it could imply a certain degree of job transformation (Frenette and Frank, 2020). Additionally, some economists argue that the risks and benefits currently being attributed to AI may be exaggerated (Acemoglu and Johnson, 2024; McElheran et al. , 2024), and productivity increases at the macroeconomic level may be modest at best (Acemoglu, 2024).

Following Pizzinelli et al. (2023), this study groups occupations into three categories based on their exposure to and complementarity with AI : (1) high exposure and low complementarity, (2) high exposure and high complementarity, and (3) low exposure. Results suggest that in May 2021, on average, around 4.2 million employees (31%) in Canada were in the first group, about 3.9 million (29%) were in the second group and about 5.4 million (40%) were in the third group. This distribution was very similar in May 2016. Unlike previous waves of automation, which mainly transformed the jobs of less educated employees performing routine and non-cognitive tasks, AI is more likely to transform the jobs of highly educated employees performing non-routine and cognitive tasks. However, highly educated employees are also more likely to hold jobs that are highly complementary with AI technologies than less educated employees. But workers will still need the skills to be able to leverage the potential benefits of AI . Compared with employees in other industries, exposure to AI -related job transformation is higher for employees in professional, scientific and technical services; finance and insurance; information and cultural industries; educational services; and health care and social assistance. However, education and health care professionals are more likely to be in jobs highly complementary with AI . Employees in industries such as construction, and accommodation and food services face relatively lower exposure to AI -related job transformation.

There is a lot of uncertainty when it comes to predicting the transformative effects of technological changes on the labour market. This study provides a static picture of AI occupational exposure based on employment compositions in May 2016 and May 2021, which were fairly similar. How workers respond and adapt to the potentially evolving labour market in the long run remains to be seen. The index used in this study is subjective and based on judgments regarding some current possibilities of AI . Consequently, the applicability of the index may decrease over time as AI capabilities grow and AI can perform an increasing number of tasks currently done by human workers. Alternative measures of AI exposure could provide further insights. Future research could also attempt to answer the question, “What happened to workers whose jobs were exposed to AI -related job transformation?”

A couple of centuries ago, the Industrial Revolution and the forces of globalization coalesced to fundamentally change the global economy. These forces served as catalysts for the technological progress that has been a cornerstone of economic development. Technological advancements and innovation paved the way for machines to take over some labour-intensive tasks and allowed workers to focus on more cognitive tasks requiring creativity and critical thinking. Adoption of new technologies also led to the obsolescence of some jobs, serving as a pathway toward higher productivity. A prominent example of this is the advent of computers, which undoubtedly replaced some jobs but also created new ones in the process (see, e.g. , Autor, Levy and Murnane [2003] or Graetz and Michaels [2018]). However, higher productivity may not always translate to higher wages for workers (Acemoglu and Johnson, 2024).

More generally, automation has become a defining feature of modern economies, including Canada’s. It has revolutionized various industries by streamlining processes, increasing efficiency and reducing operational costs, among other things. It has also raised concerns about the future of workers. The widely cited study by Frey and Osborne (2013), which estimated automation risks in the United States, has spurred a growing body of literature surrounding automation (see, e.g. , Arntz, Gregory and Zierahn [2016]; Oschinski and Wyonch [2017]; Nedelkoska and Quintini [2018]; Frenette and Frank [2020]; and Georgieff and Milanez [2021]). Frenette and Frank (2020) estimated that approximately 1/10 of employees in Canada could be at high risk (probability of 70% or higher) of automation-related job transformation.

The prevailing thought from the automation literature is that highly educated or highly skilled individuals are less susceptible to automation-related job transformation because they are more likely to perform non-routine and cognitive tasks, which are thought to be less automatable. However, another source of disruption, which has the potential to upend prior notions, is emerging: artificial intelligence (AI) . Note While AI has been around for decades ( e.g. , video games, image recognition), it was not until 2022 when it became mainstream and surged in popularity, partly fuelled by the release of ChatGPT by OpenAI.

The unprecedented pace of advancements in the field of AI and its increasing integration into society and the economy have led some researchers to call this a pivotal moment in history, akin to the transformative shifts brought on by the Industrial Revolution (Cazzaniga et al. , 2024). ChatGPT is just one example of a large language model (LLM) that has unlocked the remarkable possibilities of AI . AI can also perform complex tasks like generating music and videos from text input ( e.g. , Sora by OpenAI). AI encompasses a wide range of applications, including natural language processing, machine learning, computer vision and robotics. These technologies not only have the capacity to automate routine tasks but can also augment human decision-making processes and create entirely new opportunities for innovation and efficiency. As the field of AI continues to evolve, it has the potential to reshape industries, redefine job roles and transform the nature of work. In today’s rapidly evolving technological landscape, the integration of AI into various aspects of society, from virtual assistants and recommendation algorithms to autonomous vehicles and predictive analytics, questions naturally arise regarding its impact on society and the economy. The widespread adoption of AI raises renewed concerns about job transformation, skill mismatches and the need for workforce adaptation.

The primary objective of this study is to quantify the level of potential AI occupational exposure (AIOE) in Canada. By employing experimental methods, this study offers preliminary insights into how AI may affect the Canadian labour market and the potential risks and benefits it holds for workers.

This study adopts the complementarity-adjusted AIOE (C-AIOE) index proposed by Pizzinelli et al. (2023). The original AIOE index, which is often cited in the literature, was proposed by Felten, Raj and Seamans (2021) as a way of measuring how AI applications overlap with the human abilities needed to perform a given job. In light of recent advancements in LLMs, Felten, Raj and Seamans (2023) considered an alternate index that weighted language modelling more heavily and found that it was highly correlated with the original AIOE index. Recognizing that AI can complement human labour, the International Monetary Fund (IMF) study by Pizzinelli et al. (2023) proposed the C-AIOE index, which attempts to account for the potential complementarity of AI across occupations, in addition to direct exposure. These measures focus on “narrow” AI , which refers to “computer software that relies on highly sophisticated algorithmic techniques to find patterns in data and make predictions about the future” (Broussard, 2018; Felten, Raj and Seamans, 2021). This definition encompasses generative AI ( e.g. , LLMs, image recognition) but does not capture exposure to “general” AI , which refers to “computer software that can think and act autonomously and is combined with automation and robot technologies” (Pizzinelli et al. , 2023). International comparisons of AIOE based on the original AIOE index have been done (see, e.g. , Georgieff and Hyee [2021] and OECD [2023]). An IMF study by Cazzaniga et al. (2024) compared AI exposure and potential complementarity across countries using the C-AIOE index but did not analyze Canadian data in detail. They found that around 60% of jobs in advanced economies may be highly exposed to AI -related job transformation. As will be shown, this is similar to the share estimated for Canada.

This study offers Canadian evidence on AIOE and asks the following research questions:

• Which occupations are potentially exposed to AI -related job transformation?
• Which occupations may benefit from AI -related job transformation?
• How does the distribution of AIOE vary by industry, education level, employment income and other worker characteristics?

The experimental AI exposure estimates in this study are largely based on the technological feasibility of automating job tasks. Employers may not immediately replace humans with AI , even if it is technologically feasible, for several reasons (see, e.g. , Bryan, Sood and Johnston [2024]), including financial, legal and institutional factors. Consequently, exposure to AI does not necessarily imply a risk of job loss. At the very least, it could imply some degree of job transformation (Frenette and Frank, 2020). AI could lead to the creation of new tasks within existing jobs or create entirely new jobs. Additionally, some economists argue that the risks and benefits of AI may be exaggerated (Acemoglu and Johnson, 2024; McElheran et al. , 2024), and productivity increases at the macroeconomic level may be modest at best (Acemoglu, 2024). Evidence from the United States suggests that the adoption of AI has been more prevalent in larger firms (McElheran et al. , 2024), as some employers may not yet find it economically optimal to adopt such technologies (Svanberg et al. , 2024). Whether this will contribute to a productivity gap between smaller and larger firms is unclear. Predicting the effects of technological changes on the labour market is not an exact science, as some subjectivity is usually involved. For example, more than a decade after Frey and Osborne (2013), it is still difficult to precisely measure the effect of automation on labour markets, as changes are ongoing (Georgieff and Milanez, 2021). Although the diffusion of new technology can take time (Feigenbaum and Gross, 2023), measuring the impact of AI could be challenging given the rapid pace of advancements. The experimental estimates presented in this study should be interpreted with caution. Only time will tell whether predicted changes brought on by new technologies will come to fruition.

The remainder of this article is organized as follows. Section 2 briefly describes the AIOE index of Felten, Raj and Seamans (2021) and the complementarity-adjusted variant of Pizzinelli et al. (2023). Section 3 presents the results, and Section 4 provides concluding remarks and suggestions for future research.

The objective of this study is to estimate the extent to which occupations in Canada are potentially exposed to AI -related job transformation and the extent to which AI can potentially complement human labour in those occupations. This study uses the novel C-AIOE index of Pizzinelli et al. (2023) to achieve this objective. This measure is computed at the occupational level based on data from the Occupational Information Network (O*NET), which was created in the late 1990s by the United States Department of Labor to quantify and track the skills and abilities used across more than 1,000 different occupations ( https://www.onetonline.org ). Thus, the measure used in this study relies on occupational attribute data from the United States, which has a similar skill profile as Canada.

The C- AIOE index is based on the original AIOE index of Felten, Raj and Seamans (2021), which measures the relationship between 52 human abilities and 10 AI applications, weighted by the degree of complexity and importance of those skills for a given occupation i MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWGPbaaaa@3704@ ,

where j MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWGQbaaaa@3705@ indexes 52 occupational abilities; L j i MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWGmbWdamaaBaaaleaapeGaamOAaiaadMgaa8aabeaaaaa@391E@ is the prevalence score from O*NET and I j i MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWGjbWdamaaBaaaleaapeGaamOAaiaadMgaa8aabeaaaaa@391B@ is the importance score from O*NET for ability j MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWGQbaaaa@3705@ in occupation i MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWGPbaaaa@3704@ ; and A j = ∑ k = 1 10 x k j MathType@MTEF@5@5@+= feaagKart1ev2aqatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWGbbWdamaaBaaaleaapeGaamOAaaWdaeqaaOWdbiabg2da9maa wahabeWcpaqaa8qacaWGRbGaeyypa0JaaGymaaWdaeaapeGaaGymai aaicdaa0WdaeaapeGaeyyeIuoaaOGaamiEa8aadaWgaaWcbaWdbiaa dUgacaWGQbaapaqabaaaaa@4353@ is the exposure to AI of ability j MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWGQbaaaa@3705@ computed as the sum of the relatedness scores, x k j MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWG4bWdamaaBaaaleaapeGaam4AaiaadQgaa8aabeaaaaa@394C@ , of ability j MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWGQbaaaa@3705@ with 10 AI applications. Note This index is a relative measure of AI exposure ( e.g. , A I O E m >   A I O E n MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbiqaaWVbqaaaaa aaaaWdbiaadgeacaWGjbGaam4taiaadweapaWaaSbaaSqaa8qacaWG TbaapaqabaGcpeGaeyOpa4JaaeiiaiaadgeacaWGjbGaam4taiaadw eapaWaaSbaaSqaa8qacaWGUbaapaqabaaaaa@41DD@ implies that occupation m MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWGTbaaaa@3708@ faces greater exposure to AI -related job transformation than occupation n MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWGUbaaaa@3709@ ). See Felten, Raj and Seamans (2021) for details.

Because the AIOE index is agnostic regarding the implications of occupations being exposed to AI , Pizzinelli et al. (2023) proposed a variant of the AIOE index that accounts for the potential complementarity of AI . They make the case that certain occupations may be less conducive to the unsupervised use of AI than others. For example, judges and medical professionals are examples of occupations where job aspects such as the criticality of decisions and the gravity of the consequences of errors may require human workers to make the final decision (Cazzaniga et al. , 2024). The C-AIOE of Pizzinelli et al. (2023) is computed as

where 0 ≤ w ≤ 1 MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaaIWaGaeyizImQaam4DaiabgsMiJkaaigdaaaa@3BF1@ is a weight chosen by the researcher that controls the influence of the complementary parameter ( θ MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacqaH4oqCaaa@37CC@ ), θ i MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacqaH4oqCpaWaaSbaaSqaa8qacaWGPbaapaqabaaaaa@3914@ is the complementarity index of occupation and i MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWGPbaaaa@3704@ and θ M I N MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacqaH4oqCpaWaaSbaaSqaa8qacaWGnbGaamysaiaad6eaa8aabeaa aaa@3A99@ is the minimum observed θ MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacqaH4oqCaaa@37CC@ value among all occupations. A weight of w =   0 MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWG3bGaeyypa0Jaaeiiaiaaicdaaaa@3975@ reverts the C-AIOE back to the original AIOE ( e.g. , no role for AI complementarity), while w =   1 MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWG3bGaeyypa0Jaaeiiaiaaigdaaaa@3976@ allows maximum potential AI complementarity for occupation i MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWGPbaaaa@3704@ . Note Like the AIOE index, the complementarity index is also a relative measure, with a higher value indicating higher potential complementarity. The complementarity index for occupation i MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacaWGPbaaaa@3704@ , θ i MathType@MTEF@5@5@+= feaagKart1ev2aaatCvAUfeBSjuyZL2yd9gzLbvyNv2CaerbuLwBLn hiov2DGi1BTfMBaeXatLxBI9gBaerbd9wDYLwzYbItLDharqqtubsr 4rNCHbGeaGqiVu0Je9sqqrpepC0xbbL8F4rqqrFfpeea0xe9Lq=Jc9 vqaqpepm0xbba9pwe9Q8fs0=yqaqpepae9pg0FirpepeKkFr0xfr=x fr=xb9adbaqaaeGaciGaaiaabeqaamaabaabaaGcbaaeaaaaaaaaa8 qacqaH4oqCpaWaaSbaaSqaa8qacaWGPbaapaqabaaaaa@3914@ , is computed using O*NET data on “work contexts” and “job zones” for that particular occupation. To do so, 11 work contexts (each score ranging from 0 to 100) and the job zone (ranging from 1 to 5) are combined into six components as follows:

• Face to face
• Public speaking

Although AI can play a role in enhancing certain aspects of communication, the nuanced complexities of face-to-face interactions and public speaking could remain predominantly within the realm of human expertise.

• For outcomes
• For others’ health

AI has the potential to transform many sectors in the economy, including health care, where tough decisions are routinely made, and such decisions may still require human oversight and judgment.

• Exposure to outdoor environments
• Physical proximity to others

Jobs requiring substantial outdoor exposure and proximity to others require a certain level of adaptability and teamwork ( e.g. , firefighters, construction workers). Integrating AI into highly advanced machinery in diverse work environments could be costly.

• Consequence of errors
• Freedom of decisions
• Frequency of decisions

The importance of human oversight may become increasingly evident as AI continues to automate decision-making processes. In professions such as air traffic control or nursing, where human judgment is paramount, the combination of data analysis and instinct is essential for responding to unexpected scenarios. While AI can offer valuable data and recommendations, thereby potentially reducing human error and accelerating decision making, the indispensability of human oversight remains clear.

• Degree of automation (100 minus the O*NET score so that occupations with a low degree of automation receive higher values)
• Unstructured versus structured work

Occupations involving routine tasks have historically been more susceptible to technological transformation. Despite differences between AI and previous waves of automation, routine-intensive occupations remain particularly vulnerable to transformation. In contrast, less structured jobs may necessitate more advanced technologies for AI to operate autonomously.

Job zone is an indicator of the extent of preparation required for a job. This value must be rescaled to align with the five other components by multiplying it by 20, so that it ranges from 20 to 100 instead of 1 to 5. A higher value indicates more extensive preparation.

Occupations with high educational or training requirements may be more conducive to integrating the skills complementary with AI , as providing instructions to AI and leveraging it require some level of expertise and proficiency.

A score for each of the six components is computed by averaging the work contexts within each component ( e.g. , the score for communication is the average of face-to-face and public speaking work contexts). For the skills component, the score is the rescaled job zone value. Then, θ is calculated as the average of the six component scores divided by 100. See Pizzinelli et al. (2023) for more details regarding the derivation of the C-AIOE index and the sensitivity analyses.

This index does have some limitations, as pointed out by Pizzinelli et al. (2023). The selection of O*NET variables that serve as inputs of the index is subjective and relies on judgment regarding the factors that matter for the interaction between AI and human workers. However, Pizzinelli et al. (2023) show that the work contexts are not all systematically related to each other and offer a multifaceted take on the potential complementarity of AI with human workers. The index considers how human abilities may overlap with 10 AI applications, but as AI capabilities improve, the likelihood of AI supplanting tasks typically performed by human workers may grow. Consequently, the applicability of the index could decrease over time. Note Moreover, while the index captures the potential exposure of occupational abilities and tasks to AI , it does not account for advances in robotics, sensors and other technologies that could potentially integrate with AI (Felten, Raj and Seamans, 2021).

As O*NET is an American database, the occupations are coded according to the Standard Occupational Classification (SOC) system. The complementarity parameter and the AIOE index were computed based on version 28.2 of the O*NET database, which uses the 2018 SOC . The AIOE index was computed at the six-digit level, while the complementarity parameter was computed at the eight-digit level and then aggregated to the six-digit level by averaging the parameter values ( e.g. , the values associated with SOC codes 12-3456.01 and 12-3456.02 would be averaged to obtain the value for SOC code 12-3456). The six-digit SOC codes were then converted to the four-digit codes of version 1.3 of the Canadian National Occupational Classification (NOC) 2016 so the rich set of dimensions from the 2016 and 2021 censuses of population (reference week in May) could be used to examine AIOE in Canada. Note The sample was restricted to employees aged 18 to 64 living off reserve in private dwellings, excluding full-time members of the Canadian Armed Forces. Employment in some industries, such as accommodation and food services, decreased from May 2016 to May 2021 because of the COVID-19 pandemic, so the 2016 Census of Population was also used as a robustness check. However, results suggest that the share of employees exposed to AI -related job transformation changed very little in general.

Figure 1 presents the AIOE and potential complementarity ( θ ) for Canadian occupations. The median AIOE was around 6.0, while the median complementarity was about 0.6. Following Pizzinelli et al. (2023), an occupation is considered “high exposure” if its AIOE exceeds the median AIOE and “low exposure” otherwise. Likewise, an occupation is considered “high complementarity” if its potential complementarity exceeds the median complementarity and “low complementarity” otherwise. Note Based on this, occupations are grouped into four quadrants in Figure 1: high exposure and low complementarity, high exposure and high complementarity, low exposure and low complementarity, and low exposure and high complementarity. For simplicity, the latter two categories are combined into a single category, “low exposure,” in subsequent analyses. High-exposure, low-complementarity occupations are those that may be highly exposed to AI -related job transformation and whose tasks could be replaceable by AI in the future. High-exposure, high-complementarity occupations are those that may be highly exposed to AI -related job transformation but could be highly complementary with AI . However, workers will still need the necessary skills to leverage the complementary benefits of AI . Low-exposure jobs are those that may be less exposed to AI -related job transformation than others. Note

Potential artificial intelligence occupational exposure (AIOE) and complementarity in Canada

This chart shows a scatter plot with the AI occupational exposure index ranging from 5 to 7 on the horizontal axis and the complementarity index ranging from 0.4 to 0.8 on the vertical axis. There are 490 data points. Each data point represents an occupation as per the 4-digit National Occupation Classification version 2016 and are colour-coded with three different colours. The colours are used to distinguish the occupations according to their minimum educational requirement. Occupations requiring a bachelor's degree or higher are represented by blue, occupations requiring some postsecondary education below bachelor's degree are represented by green, and occupations requiring high school or less education are represented by red. The chart shows the relationship between AI occupational exposure and the extent to which AI can play a complementary role in a given occupation. A higher AI occupational exposure index is associated with greater potential occupational exposure to AI . A higher complementarity index is associated with greater potential complementarity with AI . The median AI occupational exposure index score of 6 and the median complementarity index score of 0.6 are used to group the various occupations into four quadrants. The top-left quadrant contain data points representing occupations which might be relatively less exposed to AI and highly complementary with AI . The majority of occupations in that quadrant require some postsecondary education below bachelor's degree but there are also a few which require high school or less education. Some examples include firefighters, plumbers, and carpenters. The bottom-left quadrant contain data points representing occupations which might also be relatively less exposed to AI but also less complementary with AI . The majority of occupations in that quadrant require high school or less education but there are also a few which require some postsecondary education below bachelor's degree. Some examples include food and beverage servers, labourers in processing, manufacturing and utilities, and welders and related machine operators. The top-right quadrant contain data points representing occupations which might be highly exposed to AI and highly complementary with AI . The majority of occupations in that quadrant require a bachelor's degree or higher education but there are a few which require some postsecondary education below bachelor's degree. Some examples include general practitioners and family physicians, secondary school teachers, and electrical engineers. The bottom-right quadrant contain data points representing occupations which might be highly exposed to AI but less complementary with AI . This quadrant has fewer data points than the other quadrants and the occupations represented by the data points have a mixture of educational requirements. Some examples include data entry clerks, economists, computer network technicians, and computer programmers and interactive media developers.

Notes: The AIOE index and potential complementarity are based on Felten, Raj and Seamans (2021) and Pizzinelli et al. (2023). An occupation is considered high-exposure if its AIOE index exceeds the median AIOE across all occupations (6.0) and considered low-exposure otherwise. Similarly, an occupation is considered high-complementarity if its complementarity parameter exceeds the median complementarity across all occupations (0.6) and considered low-complementarity otherwise. Occupations in this chart are based on the 4-digit National Occupational Classification (NOC) 2016 version 1.3 converted from the United States Standard Occupational Classification (SOC) 2018. Of the 500 NOC occupations, 10 occupations which represented less than 1% of Canadian employment, were excluded due to a lack of Occupational Information Network (O*NET) data for computing the AIOE or complementarity parameter.

Source: Occupational Information Network (O*NET) version 28.2.

Figure 1 shows that jobs potentially highly exposed to AI -related job transformation are generally those that require higher education. Although these jobs could face relatively more exposure to AI -related transformation, occupations such as family physicians, teachers and electrical engineers may be complementary with AI technologies given their relatively high complementarity scores. In contrast, occupations such as computer programming, which may also require relatively high education, have low complementarity scores, suggesting less potential complementarity with AI . There is considerable uncertainty, however, regarding the extent to which AI can actually replace human labour.

Low-exposure occupations appear to be those that usually do not require a high level of education. Some examples of occupations facing relatively low exposure to AI -related job transformation are carpenters; welders; plumbers; food and beverage servers; labourers in processing, manufacturing and utilities; and firefighters. However, as illustrated by Figure 1, AI has the potential to transform a broad set of occupations regardless of skill level. The diffusion of AI could also have downstream general equilibrium effects. For example, although less educated employees may be in jobs potentially less exposed to AI -related job transformation, highly educated employees from high-exposure jobs could transition to low-exposure jobs, displacing less educated employees (see, e.g. , Beaudry, Green and Sand [2016]).

Chart 1 aggregates the various NOC occupations into 28 distinct jobs to simplify the analysis and precisely identify the number and distribution of employees falling into the three AI exposure groups: (1) high exposure and low complementarity, (2) high exposure and high complementarity, and (3) low exposure . In May 2021, on average, roughly 4.2 million employees (31%) in Canada were in the first group, about 3.9 million (29%) were in the second group and about 5.4 million (40%) were in the third group.

Chart 1 start

Chart 1 end

At least three-quarters of employees in the following occupations were in the first group ( i.e. , highly exposed to AI -related job transformation and whose tasks could be replaceable with AI in the future): administrative occupations in finance, insurance and business; office support and co-ordination occupations; sales representatives and salespersons in wholesale and retail trade; service representatives and other customer and personal services occupations; professional occupations in business and finance; and computer and information systems professionals. Interestingly, among the 28 occupations, computer and information systems professionals experienced the highest growth (39%) from May 2016 to May 2021. However, this does not necessarily mean that computer and information systems professionals will be in less demand in the future because of AI . While these professionals may be in high-exposure, low-complementarity jobs, they are integral to maintaining and improving the underlying AI infrastructure, and this may lead to the creation of new tasks or jobs. Around 85% of employees or more in management occupations, professional occupations in education services and professional occupations in health (except nursing), as well as engineers, were in the second group ( i.e. , potentially highly exposed to AI -related job transformation, but AI can complement human labour as long as the worker possesses the necessary skills). Some occupations that could be less susceptible to AI -related job transformation (third group) were support occupations in sales and service; trades helpers, construction labourers and related occupations; assisting occupations in support of health services; and natural resources, agriculture and related production occupations.

Chart 2 shows the AI exposure distribution by industry based on the North American Industry Classification System 2017, at the two-digit level. More than half of employees in the following industries were in high-exposure, low-complementarity jobs: professional, scientific and technical services; finance and insurance; and information and cultural industries. In contrast, educational services, and health care and social assistance employed proportionately more employees who may be beneficiaries of AI . Within the health care and social assistance industry, it is mostly the professional occupations ( e.g. , nurses, physicians) that may be complementary with AI technologies (Figure 1). Employees in industries such as accommodation and food services, manufacturing, construction, and transportation and warehousing may face relatively lower exposure to AI -related job transformation.

Chart 2 start

Chart 2 end

Employees in larger enterprises (in the commercial sector) may face relatively higher exposure to AI -related job transformation (Chart 3), compared with their counterparts in smaller enterprises. Roughly over one-third of workers in enterprises with 500 or more employees were in high-exposure, low-complementarity jobs in May 2016. This compares with 25% to 28% of workers in smaller enterprises. However, employees in larger enterprises were somewhat more likely to be in jobs complementary with AI than their counterparts in smaller enterprises.

Chart 3 start

Chart 3 end

Educational attainment has historically been one of the most important indicators of whether a worker will be resilient to technological shocks. The growing consensus from the labour economics literature is that less educated workers face a higher risk of automation-related job transformation than highly educated workers because the former group is more likely to perform routine and manual tasks that are more susceptible to being automated. However, Chart 4 shows that AI could affect a broader segment of the labour force than previously thought because it has the capacity to perform non-routine and cognitive tasks. Highly educated employees may face higher exposure to AI -related job transformation, as was shown in Figure 1. The highest shares of high-exposure, low-complementarity jobs are held by employees with a bachelor’s degree (37%) or a college, CEGEP or other certificate or diploma below a bachelor’s degree (36%), followed by those with a graduate degree (32%), high school or less education (25%), and an apprenticeship or trades certificate or diploma (15%). However, employees with a bachelor’s degree or higher were more likely to hold jobs that may be highly complementary with AI than those with an education below the bachelor’s degree level, as long as the potential beneficiaries of AI possess the necessary skills. Employees with an apprenticeship or trades certificate or diploma may be less exposed to AI -related job transformation, as 73% were in low-exposure occupations. However, as previously mentioned, a more nuanced view is that while less educated workers may face potentially lower exposure to AI -related job transformation, highly educated workers from high-exposure jobs may transition to low-exposure jobs, displacing less educated workers (see, e.g. , Beaudry, Green and Sand [2016]).

Chart 4 start

Chart 4 end

Many of the results presented so far are contrary to the findings on automation documented in the labour economics literature over the past two decades, raising concerns about the nexus of automation and AI . Frenette and Frank (2020) estimated that around 1/10 of employees in Canada were at high risk (probability of 70% or more) of automation-related job transformation in 2016. Chart 5 suggests that exposure to AI -related job transformation decreases as the risk of automation-related job transformation increases. The majority of employees (60%) in jobs at high risk of automation-related transformation were in jobs that may be least exposed to AI -related transformation (Chart 5). In contrast, 18% of employees in jobs at low risk (probability of less than 50%) of automation were in low-exposure jobs. However, although potentially highly exposed to AI -related job transformation, employees at a lower risk of automation-related job transformation hold jobs that could be highly complementary with AI . Jobs facing a moderate risk (probability of 50% to less than 70%) of automation-related transformation were most likely to be high-exposure, low-complementarity jobs. These findings are important, as they suggest that the distinction between manual and cognitive tasks and between repetitive and non-repetitive tasks used in the last two decades in labour economics to understand automation-related technological transformation may not apply to AI .

Chart 5 start

Chart 5 end

Like previous waves of technological transformation, AI has the potential to boost productivity. But this process can also exacerbate earnings inequality. Chart 6 shows the AI exposure distribution across employment income deciles. More than half of the jobs in the bottom half of the distribution were low-exposure jobs, while around 30% were high-exposure, low-complementarity jobs. The middle of the distribution may be the most vulnerable to AI -related job transformation, with around one-third of jobs being high exposure and low complementarity. Exposure to AI -related job transformation increases with employment income, but higher earners hold jobs that may be highly complementary with AI . Although the top decile had the highest share of jobs potentially exposed to AI -related job transformation, they also had the highest share of jobs (55%) that are highly complementary with AI . If higher earners can take advantage of the complementary benefits of AI , their productivity and earnings growth may outpace those of lower earners, and this could exacerbate earnings inequality (Cazzaniga et al. , 2024). However, the diffusion of AI could also potentially reduce earnings inequality if AI happens to adversely affect high-skill occupations (see, e.g. , Webb [2020]).

Chart 6 start

Chart 6 end

Canada’s record population growth, recently driven by international migration, raises questions about the future of jobs done by immigrants and non-permanent residents. In May 2016, recent immigrants (those who landed from 2011 to 2016) (29%) were just as likely as Canadian-born individuals (29%) to be in high-exposure, low-complementarity jobs (Chart 7). However, by May 2021, while the share of Canadian-born individuals in such jobs remained the same, the share of recent immigrants (those who landed from 2016 to 2021) in these jobs increased to 37%. This was partly driven by the fact that nearly 1/10 of permanent residents who landed from 2016 to 2021 were employed in computer and information systems professions in May 2021—occupations more likely to be high exposure and low complementarity. Less than 5% of permanent residents who landed from 2011 to 2016 were employed in these professions in May 2016. This increasing concentration of recent immigrants in computer and information systems professions has been documented by Picot and Mehdi (forthcoming). Another reason could be the (temporarily) falling share of employment in occupations adversely affected by the COVID-19 pandemic. Non-permanent residents were more likely to be in high-exposure, low-complementarity jobs and low-exposure jobs than Canadian-born individuals. One goal of economic immigration programs is to fill labour and skills shortages. However, perceived labour shortages may eventually incentivize some employers to adopt AI technologies, especially if such shortages are in occupations highly exposed to AI -related job transformation.

Chart 7 start

Chart 7 end

Appendix Table A.1 (May 2016) and Appendix Table A.2 (May 2021) provide further results disaggregated by field of study, age group, gender, activity limitation status, selected census metropolitan area (CMA), racialized group, full-time or part-time status, union membership status, and whether the job can be done from home.

Exposure to AI -related job transformation varies substantially not only across fields of study but also on whether the employee has a bachelor’s degree or higher education. For example, employees who studied engineering and engineering technology or health care at a level below a bachelor’s degree were less likely to face AI -related job transformation than employees who studied the same disciplines at the bachelors’ degree or higher level. However, even with increased exposure, the majority of the latter group held jobs that were highly complementary with AI . Close to 60% of employees or more who studied mathematics and computer and information sciences—regardless of where they received their postsecondary education—were in high-exposure, low-complementarity jobs. Employees who studied construction trades and mechanic and repair trades may face relatively lower exposure to AI -related job transformation.

Employees aged 18 to 24 are overrepresented in low-exposure occupations, likely because they do not yet have the necessary experience to be employed in high-skill occupations. Core working-age employees, those aged 25 to 54 years, are generally more likely to hold jobs highly exposed to AI -related job transformation than their younger and older counterparts. But core working-age employees are also more likely to hold jobs that may be highly complementary with AI .

Slightly over one-fifth of men are employed in high-exposure, low-complementarity jobs, compared with 38% of women. This is because men are more likely to be employed in the skilled trades, which may face relatively lower exposure to AI -related job transformation. However, women (33%) are more likely than men (25%) to be employed in occupations that could be highly complementary with AI .

Occupations facing AI -related job transformation are more likely to be in large population centres. The CMAs of Ottawa–Gatineau (39%) and Toronto (37%) had proportionately more high-exposure, low-complementarity employment relative to other CMAs. But urban areas also had proportionately more jobs that could be highly complementary with AI .

Chinese (45%) and South Asian (38%) employees are more likely to hold high-exposure, low-complementarity jobs than other racialized groups. This is partly driven by their relatively higher representation in computer and information systems professions, which potentially highly exposed to AI -related job transformation and whose tasks may be replaceable by AI in the future. However, as noted earlier, these occupations could be integral to maintaining and improving the underlying AI infrastructure.

Unionized employees are almost as likely as their non-unionized counterparts to be highly exposed to AI -related job transformation. However, non-unionized employees (35%) are more likely to be in high-exposure, low-complementarity jobs than unionized employees (23%). This was largely driven by a higher share of unionized employees in health care and education occupations, which are potentially highly exposed to and complementary with AI .

The COVID-19 pandemic has led to significant increases in working from home (see, e.g. , Mehdi and Morissette [2021a] or Mehdi and Morissette [2021b]). These jobs are usually held by highly educated employees who may be more exposed to AI -related job transformation than their less educated counterparts. Just over half (51%) of employees with jobs that can be done from home were in high-exposure, low-complementarity occupations, compared with 14% of employees in jobs that cannot be done from home. Note However, 47% of the former group holds jobs that could be highly complementary with AI , compared with 14% of the latter group. How the advent of AI could affect the labour market in potential future pandemics is unclear (see, e.g. , Frenette and Morissette [2021]).

This study provides experimental estimates of the number and percentage of employees aged 18 to 64 in Canada potentially susceptible to AI -related job transformation using the C-AIOE index of Pizzinelli et al. (2023) and data from O*NET and the 2016 and 2021 censuses of population. Occupations were grouped into three distinct categories: (1) high exposure and low complementarity, (2) high exposure and high complementarity, and (3) low exposure. Being in the second group does not necessarily reduce AIOE , as workers would still need the necessary skills to be able to leverage the potential complementary benefits of AI .

On average, in May 2021, approximately 4.2 million employees (31%) in Canada were in the first group, about 3.9 million (29%) were in the second group and about 5.4 million (40%) were in the third group. This distribution was similar in May 2016. Employees in the following industries were more likely than others to be in the first group: professional, scientific and technical services; finance and insurance; and information and cultural industries. In contrast, employees in educational services, and health care and social assistance were more likely to be in the second group than other employees. Employees in industries such as accommodation and food services, manufacturing, construction, and transportation and warehousing face relatively less exposure to AI -related job transformation.

Unlike previous waves of automation, which affected routine and non-cognitive jobs, AI could affect a broader segment of the labour force than previously thought. Contrary to previous findings from the technological transformation literature, AI could transform the jobs of highly educated employees to a greater extent than those of their less educated counterparts. However, highly educated employees also hold jobs that may be highly complementary with AI . Previous labour market policy recommendations in response to the threat of automation included supporting upskilling and job transition initiatives. The findings in this article, which reflect the possible role of AI exposure and complementarity for occupations and workers in Canada, may inform future policy discussions on the topic.

The index used in this study is subjective and based on judgments regarding some current possibilities of AI . Consequently, the applicability of the index may decrease over time as AI capabilities grow and AI can perform an increasing number of tasks currently done by human workers. The index is also computed at the occupational level, implicitly assuming that tasks within a given occupation are the same across regions and worker characteristics. However, the ability to adapt and respond to changing skill demands will likely vary across worker characteristics. If tasks vary substantially across regions and worker characteristics, and if some tasks are more vulnerable to AI substitution, the index could be over- or underestimated to a certain extent. For example, computer programmers in one region who spend their work day coding may be more susceptible to AI -related job transformation if AI is proficient in writing that code. In contrast, programmers in another region who spend part of their day interacting face to face with team members may be less susceptible, assuming AI is not yet proficient in face-to-face interactions. To address this, future research could develop alternative measures of AI exposure at the worker level, similar to how Arntz, Gregory and Zierahn (2016) or Frenette and Frank (2020) estimated automation risk. Future studies could also attempt to answer the question, “What happened to workers whose jobs were exposed to AI -related job transformation?”

As AI technologies continue to evolve, they have the potential to reshape industries, redefine job roles and transform the nature of work. AI may also create new challenges and divides and push boundaries. But large-scale AI adoption may take some time, as employers may face financial, legal and institutional constraints. This study provides a static picture of AIOE based on employment compositions in Canada in May 2016 and May 2021, which were fairly similar. How AI affects productivity and how workers and firms adapt to the potentially evolving labour market in the long run remain to be seen.

Acemoglu, D. 2024. The Simple Macroeconomics of AI . NBER, Working Paper no. 32487.

Acemoglu, D. and S. Johnson. 2024. Learning from Ricardo and Thompson: Machinery and Labor in the Early Industrial Revolution, and in the Age of AI . NBER, Working Paper no. 32416.

Arntz, M., T. Gregory and U. Zierahn. 2016. The Risk of Automation for Jobs in OECD Countries: A Comparative Analysis . OECD Social, Employment and Migration Working Papers, no. 189. Paris: OECD Publishing.

Autor, D., H. Levy and R. Murnane. 2003. The skill content of recent technological change: An empirical exploration . The Quarterly Journal of Economics 118 (4): 1279–1333.

Beaudry, P., D. A. Green and B. M. Sand. 2016. The Great Reversal in the Demand for Skill and Cognitive Tasks . Journal of Labor Economics 34 (s1): S199–S247.

Broussard, M. 2018. Artificial Unintelligence: How Computers Misunderstand the World . Cambridge: MIT Press.

Bryan, V., Sood, S. and C. Johnston. 2024. Analysis on artificial intelligence use by businesses in Canada, second quarter of 2024 . Analysis in Brief. Statistics Canada Catalogue no. 11-621-M. Ottawa: Statistics Canada.

Cazzaniga, M., Jaumotte, F., Li, L., Melina, G., Panton, A. J., Pizzinelli, C., Rockall, E. J. and M. M. Tavares. 2024. Gen-AI: Artificial intelligence and the future of work . IMF , Working Paper no. 1.

Dingel, J. I. and B. Neiman. 2020. How many jobs can be done at home? Journal of Public Economics 189: 104235.

Eloundou, T., Manning, S., Mishkin, P. and D. Rock. 2023. GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models . Stanford Digital Economy Lab, Working Paper.

Feigenbaum, J. J. and D. P. Gross. 2023. Organizational and Economic Obstacles to Automation: A Cautionary Tale from AT&T in the Twentieth Century . NBER, Working Paper no. 29580.

Felten, E., Raj, M. and R. Seamans. 2021. Occupational, Industry, and Geographic Exposure to Artificial Intelligence: A Novel Dataset and its Potential Uses . Strategic Management Journal 42(12): 2195-2217.

Felten, E., Raj, M. and R. Seamans. 2023. How will Language Modelers like ChatGPT Affect Occupations and Industries? SSRN Working Paper.

Frenette, M. and K. Frank. 2020. Automation and Job Transformation in Canada: Who’s at Risk? Analytical Studies Branch Research Paper Series. Statistics Canada Catalogue no. 11F0019M. Ottawa: Statistics Canada.

Frenette, M. and R. Morissette. 2021. Job security in the age of artificial intelligence and potential pandemics . Economic and Social Reports (June). Ottawa: Statistics Canada.

Frey, C. B. and M. A. Osborne. 2013. The Future of Employment: How Susceptible Are Jobs to Computerisation? Oxford Martin Programme on the Impacts of Future Technology. Oxford: Oxford Martin School, University of Oxford.

Georgieff, A. and R. Hyee. 2021. Artificial intelligence and employment: New cross-country evidence . OECD Social, Employment and Migration Working Papers, no. 265. Paris: OECD Publishing.

Georgieff, A. and A. Milanez. 2021. What happened to jobs at high risk of automation? OECD Social, Employment and Migration Working Papers, no. 255. Paris: OECD Publishing.

Graetz, G. and G. Michaels. 2018. Robots at work . The Review of Economics and Statistics 100 (5): 753–768.

Kochhar, R. 2023. Which U.S. Workers Are More Exposed to AI on Their Jobs? Pew Research Center.

McElheran, K. Li, J. F., Brynjolfsson, E., Kroff, Z., Dinlersoz, E., Foster, L. S. and N. Zolas. 2024. AI Adoption in America: Who, What, and Where . NBER, Working Paper no. 31788.

Mehdi, T. and R. Morissette. 2021a. Working from home: Productivity and preferences . StatCan COVID-19: Data to Insights for a Better Canada. Statistics Canada Catalogue no. 45280001. Ottawa: Statistics Canada.

Mehdi, T. and R. Morissette. 2021b. Working from home in Canada: What have we learned so far? Economic and Social Reports (October). Ottawa: Statistics Canada.

Nedelkoska, L. and G. Quintini. 2018. Automation, Skills Use and Training . OECD Social, Employment and Migration Working Papers, no. 202. Paris: OECD Publishing.

OECD. 2023. OECD Employment Outlook 2023: Artificial Intelligence and the Labour Market . Paris: OECD Publishing.

Oschinski, M. and R. Wyonch. 2017. Future Shock? The Impact of Automation on Canada’s Labour Market . C.D. Howe Institute Commentary, no. 472. Toronto: C.D. Howe Institute.

Picot, G. and T. Mehdi. Forthcoming. The provision of high and low skilled immigrant labour to the Canadian economy .

Pizzinelli, C., Panton, A. J., Tavares, M. M., Cazzaniga, M. and L. Li. 2023. Labour market exposure to AI : Cross-country differences and distributional implications . IMF , Staff Discussion Notes no. 216.

Svanberg, M. S., Li, W., Fleming, M., Goehring, B. C. and N. C. Thompson. 2024. Beyond AI Exposure: Which Tasks are Cost-Effective to Automate with Computer Vision? MIT FutureTech, Working Paper.

Webb, M. 2020. The Impact of Artificial Intelligence on the Labor Market . SSRN, Working Paper.

Wootton, C. W. and B. E. Kemmerer. 2007. The Emergence of Mechanical Accounting in the U.S., 1880-1930 . Accounting Historians Journal 34(1): 91-124.

More information

Note of appreciation.

Canada owes the success of its statistical system to a long-standing partnership between Statistics Canada, the citizens of Canada, its businesses, governments and other institutions. Accurate and timely statistical information could not be produced without their continued co-operation and goodwill.

Standards of service to the public

Statistics Canada is committed to serving its clients in a prompt, reliable and courteous manner. To this end, the Agency has developed standards of service which its employees observe in serving its clients.

Published by authority of the Minister responsible for Statistics Canada.

© His Majesty the King in Right of Canada, as represented by the Minister of Industry, 2024

Use of this publication is governed by the Statistics Canada Open Licence Agreement .

Catalogue no.  11F0019M

Frequency: Occasional