Efficacy of Machine Learning in Simulating Precipitation and Its Extremes Over the Capital Cities in North Indian States

Abstract

Climate change-induced precipitation extremes have become a pressing global concern. This study investigate the predictability of precipitation patterns and its extremes using MERRA2 datasets across North Indian states for the period 1984 to 2022 utilizing machine learning (ML) models. A strong positive correlations of precipitation 0.4 was found with dew point temperature and relative humidity significant at 0.05. In simulating precipitation, Random Forest Classifier (RFC) achieved the highest accuracy (~ 83%) for Rajasthan and Uttar Pradesh, while Support Vector Classifier (SVC) performed best (79–83% accuracy) for other states. However, the ML models exhibited about 5% lower skill in higher elevated stations as compared to the lower elevated stations, its due to the different atmospheric mechanisms control differently over the lower and higher topography. For extreme precipitation events (10th and 95th percentiles of intensity), RFC consistently outperformed SVC across all states. It demonstrated superior ability to distinguish extreme from non-extreme events (Area under curve ~ 0.90) and better model calibration (Brier Scores ~ 0.01). The developed ML models successfully simulated precipitation and extreme patterns, with RFC excelling at predicting extreme precipitation events. These findings can contribute to disaster preparedness and water resource management efforts in the region with varied topography and complex terrain.