Effective delineation of Agricultural Drought Hazard (ADH) zones is crucial for mitigating crop losses and ensuring water security in semi-arid regions. Conventional agricultural drought assessment methods, reliant on single-index approaches or static multi-criteria frameworks, struggle to capture the non-linear interactions between geo-environmental drivers that govern drought severity in semi-arid, rainfed basins. This study introduces a Machine Learning (ML)-geospatial framework integrating satellite-derived indices with soil-hydrological parameters to overcome the limitations of conventional drought assessment methods. Four popular ML models, Random Forest (RF), Artificial Neural Network (ANN), Support Vector Machine (SVM), and Adaptive Regression (AR), are utilized for this purpose, considering eight geo-environmental input variables. Model performance was rigorously evaluated in the Upper Dwarakeshwar River Basin (UDRB), a drought-prone, rainfed catchment in eastern India, using a suite of standard statistical approaches. The RF model excelled with a 97.8% area under the curve-receiver operating characteristic (AUC-ROC) curve and root mean square error (RMSE) of 0.26, followed by the SVM model (94.6%, 0.28). The ANN model, too, yielded promising results (93.8%, 0.32), while the AR model exhibited the least performance (90.0%, 0.31). Based on the outputs from all four ML models, ADH mapping for UDRB revealed that 24.85-44.35% of its area was identified as very high and 16.96-22.86% as high ADH regions. From a practical application point of view, the findings of this study and ADH maps are helpful in various aspects, ranging from early drought warning to emergency preparedness, advancing precision agriculture in rainfed basins, where 60-80% of livelihoods depend on climate-vulnerable farming.