Integrating machine learning models to map springshed potential zones using key environmental variables in...

This study presents an integrated machine learning framework to predict potential spring occurrence zones in the high-altitude Kishtwar region of the Indian Himalayas, where springs are rapidly depleting due to climate and anthropogenic pressures. Forty-two geotagged spring locations obtained from field visits combined with pseudo-absence background locations were used to train and test the machine learning models using 80:20 split. Predictor variables-including soil moisture, elevation, slope, Enhanced Vegetation Index (EVI), annual precipitation, and saturated soil hydraulic conductivity-were derived from remote sensing and environmental datasets. Random Forest, MaxEnt, and CART models were trained and combined through linear regression-based ensemble integration to reduce the bias. Elevation, precipitation, and soil parameters were key predictors affecting the springs occurrence. Model outputs delineated 815.6 km² as potential and 467.3 km² as high-probability spring zones in the study area. While the ensemble achieved high discriminatory performance (AUC up to 0.99), the results are based on a limited set of field-observed springs and should be interpreted as screening-level prioritization outputs rather than definitive predictive maps. The study, although designed on a regional scale, demonstrates scalable potential for Himalayan springshed management, enabling data-driven restoration through advanced geospatial-machine learning integration.