ENHANCED METEOROLOGICAL DROUGHT FORECASTING USING CNN, LSTM, TRANSFORMER, AND BAYESIAN MODEL AVERAGING

Drought is a multifaceted hydro climatic risk that impacts greatly on water resources, agriculture and livelihoods especially in areas that depend on the monsoon like the Tel River Basin in Odisha, India. The paper investigates deep learning and ensemble modelling that are proposed to predict medium-term meteorological drought with the Standardized Precipitation Index at three and six months (SPI-3 and SPI-6). The NASA POWER database provided monthly data of precipitation from 1971 to 2016 in four stations, namely, Kalahandi, Kandhamal, Kesinga, and Nuapada. The input features that included lagged SPI values to a maximum of 12 months were used in order to adjust temporal dependencies. Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), and Transformer models of deep learning were created and their performances compared against an ensemble of a Bayesian Model Averaging (BMA). Correlation coefficient (R), coefficient of determination (R2), Root Mean Square Error (RMSE), Mean Absolute Error (MAE) and NashSutcliffe Efficiency (NSE) were used to measure model performance. Findings have shown that individual models work quite well but the BMA ensemble always outperforms an individual model in terms of accuracy and stability. The results indicate potential of the ensemble-based deep learning methods in making reliable prediction                                                  of                                                  regional                                                                      droughts