39HYDRO Improving Extreme Rainfall Forecasts for Mumbai

The increasing frequency and severity of extreme weather events, such as heavy rainfall and flooding, emphasize the need for effective early warning systems. This study proposes a machine learning model utilizing spatial convolution to capture complex spatial rainfall patterns of extreme events. This approach combines Convolutional Neural Networks (CNN) and Transfer Learning (TL) to downscale gridded daily accumulated rainfall to station level for Mumbai, a flood-prone coastal city on India's west coast. The model uses gridded Global Forecast System (GFS) rainfall data at a 0.25° x 0.25° resolution as predictors, with station-level observed rainfall data as the predictand. The combined model was trained on rainfall data from 2015-2020 for the JJAS season, validated with 2021-2022 data, and tested on 2023 data. On the unseen 2023 dataset, the proposed model achieved a 9% reduction in Root Mean Square Error (RMSE), a 14.8% increase in Correlation Coefficient (CC), and a 176.5% improvement in the Critical Success Index (CSI) at the 95th percentile for Lead Day 1. For Lead Day 2, compared to the parent model forecasts, the model showed a 1.3% reduction in RMSE, a 5% increase in CC, and increased CSI values for both Lead Day 2 and Lead Day 3 across rainfall percentiles from 80^th to 95^th. These findings highlight the proposed model's accuracy and its superior ability to forecast extreme rainfall events compared to existing GFS models.

Keywords: Global Forecast System (GFS), Extreme events, Machine Learning, Daily Rainfall Forecasts, Mumbai