Hyperlocal urban extreme rainfall forecast system: A convolutional neural network-transfer learning-based downscaling approach

Global warming has intensified extreme precipitation globally, stressing urban infrastructure. While Numerical Weather Prediction (NWP) models have improved, they remain inadequate for forecasting extreme events at local scales, especially in the tropical urban regions. Mumbai, a coastal Indian city with 21 million residents, experiences multiple extreme precipitation events with significant spatial variability during the monsoon season (June to September). The Global Forecast System (GFS) model struggles to forecast extreme rainfall in Mumbai, resulting in low hit rates and high false alarms in predictions. To address this, we performed statistical downscaling of GFS output to 36 stations across Mumbai using a Convolutional Neural Network (CNN). While the CNN improved overall GFS skill, improvements for extreme events were minimal due to their rarity in training data. We enhanced this using CNN-Transfer Learning (CNN-TL), pre-training the model on the full rainfall dataset and fine-tuning it on heavy rainfall events. This approach improved threat scores (TS) by 12.94% to 498.92% and 76.47% to 616.67% for Lead Days 1 and 2, respectively, at the 95th percentile across different stations. On Lead Day 3, the CNN-TL model achieved TS ≥ 0.2 for most stations at the 95th percentile, while GFS forecasts failed to capture extreme rainfall events at this lead time. Correlation coefficients also showed significant improvements. The CNN-TL approach outperformed Mumbai's operational forecast model. This method is replicable in other cities with available long-term station data.