Performance evaluation of convolutional neural network and vision transformer models for groundwater potential mapping

Abstract

Due to excessive consumption and the increasing warming of the earth’s air, the level of groundwater in the world is decreasing, especially in arid and semi-arid countries that need water supply for various purposes from these sources. In this study, the data of 3546 wells and 15 spatial factors influencing the occurrence of groundwater, elevation, slope, plan curvature, profile curvature, terrain wetness index (TWI), valley depth, slope length (LS), river density, distance from river, distance from fault, geology, land cover, aspect, normalized difference vegetation index (NDVI), and rainfall have been used for modeling and groundwater potential mapping (GWPM). In the feature selection process, the wrapper base method, Boruta-XGBoost, and the variance inflation factor (VIF) test were used, and all factors except LS were confirmed and entered into the model. Convolutional neural network (CNN) and vision transformer (VIT) were used as learning models for Chaharmahal Bakhtiari province, one of Iran’s mountainous provinces. The area under receiver operating characteristic curve (AUC), root mean square error (RMSE), and some statistical metrics such as precision, recall and F1-score have been used for model validation. According to the obtained results, the VIT model is the most efficient with an AUC of 0.8530, RMSE (0.3900), precision (0.7740), recall (0.7600), and F1-score (0.7610) which gives the most promising values model, than the CNN model with an AUC of 0.8370, RMSE (0.4100), precision (0.7650), recall (0.7550) and F1-score (0.7560). These results show the appropriate power of both models in modeling and the relative superiority of the VIT method. Finally, the SHapley Additive exPlanations (SHAP) method was used to enhance model explainability. SHAP analysis highlighted land cover, rainfall, and geology as the most important factors in this study. Preparing the groundwater potential map helps managers and decision-makers manage these resources’ consumption and use the potential of groundwater as one of the practical criteria for allocating land use.