Augmenting the real-time rainfall forecast skills over odisha using deep learning technique

Heavy rainfall events are becoming more frequent throughout the Indian area under a climate change scenario. The monsoon low-pressure systems, Depressions (MDs), and Deep Depressions (DDs) are major rain-bearing systems during the Indian Summer Monsoon and these systems often result in localized occurrences of very heavy to extremely heavy rainfall causing huge socio-economical loss. Currently, dynamical models face huge challenges in forecasting these events with adequate lead time. The proposed deep learning (DL) model forecast outperformed the deterministic Weather Research and Forecast (WRF) in categorical and spatial rainfall prediction till the lead time of 96 h (day 4). For two real-time cases (one MD and another low pressure), it is demonstrated that the proposed kernelize-based spatial attention-based DL models substantially reduced the mean absolute error (MAE) by an average of 8 mm over all districts of Odisha, compared to that of WRF (> 20 mm). For the first time, the intensity, amplitude, and distribution of the rainfall in real-time are shown to be accurately represented by the suggested DL model. Furthermore, rainfall prediction skill scores suggest the DL models demonstrate the ability to predict rainfall in appropriate intensity categories. The study's conclusions directly affect the development of early warning systems and increased readiness to reduce the hazards associated with periods of intense rainfall.