Critical role of climate factors for groundwater potential mapping in arid regions: Insights from random forest, XGBoost, and LightGBM algorithms

Abstract

Groundwater potential mapping (GPM) provides the valuable information on groundwater volume that can be withdrawn from the aquifer without affecting the environmental conditions. In this study, GPM was innovatively explored by including three typical climate factors - precipitation (PRE), evaporation (EVA), and ground surface temperature (GST) - and by taking into consideration a total of 23 conditional groundwater factors during the model construction. Three ensemble learning models are used for the modeling process: random forest (RF), XGBoost, and LightGBM. The study was conducted in the southern regions of Yinchuan Plain, which is known for its temperate continental climate with low precipitation and high evaporation. It was found that all six adopted models (i.e., RF-C, XGBoost-C, LightGBM-C, RF, XGBoost, LightGBM) made reasonable predictions of groundwater potential, with areas with high and very high potential concentrated in the southwest region, where the Yellow River enters the study area. The LightGBM-C model performs the best (OA: 0.769, F1 score: 0.667, AUC: 0.921), while the RF model performs the worst (OA: 0.654, F1 score: 0.4, AUC: 0.757). According to the LightGBM-C model, there are more productive wells in areas with high groundwater potential (22 wells, 0.0152 per km2), while fewer non-productive wells are found (3 wells, 0.0021 per km2). The performance of LightGBM-C and RF-C models has been notably enhanced by climate factors (AUC + 0.073, OA + 0.025, F1 score + 0.052 for LightGBM, and AUC + 0.073, OA + 0.051, F1 score + 0.149 for RF). The further cumulative importance results indicate that the three ensemble models which considered climate factors demonstrated a substantial sensitivity to PRE (27.78%), EVA (28.14%), and GST (23.34%). In arid and semi-arid regions, PRE and EVA are highly recommended factors for GPM, while DD (35.17%), EV (35.22%), NDVI (28.06%) and GWD (25.77%) are also confirmed to be important when climate factors are not taken into account.