A deep learning framework for rainfall forecasting in mumbai metropolitan region

Mumbai experiences heavy rainfall events that significantly disrupt daily life and infrastructure every year, yet operational forecasts still struggle with accuracy and forecasts. This study explores the skill efficacy in temporal and spatiotemporal supervised machine learning (ML) applications for the prognoses of daily rainfall in the Mumbai Metropolitan region. Based on ERA5-reanalyzed and observed datasets, the learning and predictions associate the daily-averaged atmospheric predictors with observed rainfall accumulations for the 1969-2019 period. The temporal learning showed that deep-learning (DL) long-and-short-memory (LSTM) algorithm exhibits consistent skill merits of training and predictions irrespective of the choice of observed and ERA5-reanalyzed predictors. Following this premise, spatiotemporal learning is conducted using a suite of multi-layered unidirectional and bidirectional convolutional LSTM frameworks with additional physics-aware predictors from ERA5, such as column-integrated moisture and layer stability parameters, targeted to observed rainfall during the monsoon season. The 2-layer bidirectional configuration shows optimal performance and reproduces observed moderate-to-heavy rain amounts with 65% skill. Although extreme rain situations (> 100 mm) are temporally captured, a higher skill is generally noted in reproducing these situations embedded within observed incessant rainy spells exhibiting gradual daily variations compared to the ones associated with rapid variations. The study shows the promising potential of DL implementations to support the operational forecast guidance in heavy rainfall alerts for the Mumbai region.