Machine learning based precipitation modeling using multi satellite data for climate resilient water resource management in Bundelkhand India

Precipitation modeling can be improved by using spatially continuous climate data from satellite remote sensing; nevertheless, incorporating heterogeneous sensor-derived variables and understanding machine learning results continue to be significant hurdles. In this work, a multi-source climatic parameter-based satellite-driven machine learning system for precipitation prediction is presented. The dependent variable was precipitation, and the model inputs were satellite-derived predictors such as land surface temperature, atmospheric moisture, surface pressure, wind speed, relative humidity, soil wetness, and temporal indicators. Convolutional Neural Networks (CNN) and Extreme Gradient Boosting (XGBoost), two sophisticated machine learning models, were used to capture multiscale and nonlinear interactions between precipitation and climate factors. Standard statistical measures were used to evaluate the model's performance, and explicable machine learning methods were used to determine the relative significance of the input variables. The findings suggest that both models make good use of satellite-derived climate data, with CNN demonstrating a great capacity to learn intricate feature interactions and XGBoost demonstrating strong predictive ability. With R2 = 0.77, RMSE = 88.79, as well as MAE = 42.06, XGBoost scored far superior than CNN (R2 = 0.60, RMSE = 120.82, MAE = 73.46). According to the interpretability analysis, the main factors influencing precipitation variability are soil wetness, land surface temperature, and atmospheric moisture. The suggested system supports enhanced hydrological forecasting as well as sustainable water resource management by providing a transparent and scalable method for precipitation prediction, especially in areas with limited data. Overall, the findings show that a scalable and efficient framework for precipitation prediction in semi-arid, data-poor areas may be created by combining interpretable machine learning with multi-source Earth observation data. Graphical Abstract