Causally Aware Feature Selection for Explainable Prediction of Indian Summer Monsoon Rainfall from Global Climatic Indices

Indian Summer Monsoon Rainfall (ISMR) is one of the key factors responsible for shaping India's socioeconomic landscape. ISMR is calculated by taking the average monthly rainfall over June to September (JJAS) for the Indian region. The seasonal variability of ISMR is driven by several climatic phenomena (El-Niño Southern Oscillation, Indian Ocean Dipole, etc.), which are interconnected through teleconnections. Seasonal forecasting of ISMR at considerable lead times is still an extremely challenging problem. The proposed work aims to identify the global drivers that contribute the most to ISMR variability, so that the prediction accuracy of the Machine Learning models can be improved. In this work, 30 such potential drivers are identified from the state of the art, followed by dropping of correlated drivers. We have 3-monthly mean values of each feature at different lag windows. Then, the Granger Causality score of each feature with the ISMR anomaly is calculated, and the features that show 95%confidence with the best lag values are selected as the predictors for ISMR. The predictive power of these features with adjusted lags are evaluated using Multiple Linear Regression (MLR), XGBoost (XGBoost), Random Forest Regressor (RF), Artificial Neural Network (ANN), and Time-Varying Statistical Regressor (TV-Stat). The results show a improved prediction accuracy by 40% over the original set of drivers. In addition, a post-hoc analysis of the effect of the selected drivers on ISMR predictions and extreme ISMR events are performed using Shapley Additive exPlanations(SHAP). This analysis leads to an important conclusion that the Sea Surface Temperature (SST) and wind speed variability of the southwest Indian Ocean and southwest Pacific Ocean are crucial factors for ISMR predictions for the period 1871-2015.