Reducing the real-time heavy rainfall forecast error associated with Monsoon depressions and deep depressions over the Indian region using deep learning

India received a lot of rain from the two main rain-bearing systems (i.e., Monsoon Depressions (MDs) and Deep Depressions (DDs)) during the monsoon season. The MDs and DDs contributed roughly 60% of the seasonal rainfall and primarily originated over the Bay of Bengal (BoB) before moving northwest to the mainland. As they migrate from the BoB to the mainland, they have significantly impacted the eastern coast of India, especially Odisha. Due to its heavy reliance on intricate, non-linear processes at a smaller scale, rainfall forecasting is very challenging to predict with a reasonable degree of accuracy, particularly at the district level.  For the first time, we attempted to employ high-resolution real-time forecast outputs (i.e., WRF) as the input feature to district-scale DL models to reduce prediction error rather than relying on observation or reanalysis datasets.

 In this study, we used input data from the Weather Research and Forecast (WRF) forecast at high resolution (3km), which was initialized using GFS real-time forecast, to enhance the spatial and categorical rainfall prediction (intensity) at the district scale by introducing two DL-based architectures:  U-Net (+A) (attention-based U-Net architecture) and KU-Net (+A) (Attention-based Kernelized U-Net Architecture). The model was trained using 24 cases of monsoon LPS (12 MDs and 12 DDs) spanning between 2007 and 2018 and tested for two cases, July 2023 and August 2023, over Odisha in real-time.

Our suggested model (KU-Net (+A)) significantly enhanced the Odisha district-level rainfall prediction. While WRF shows the MAE of more than 25 mm and 36 mm for both instances, respectively, the DL models decreased the MAE values to less than 8 mm up to Day 4 for case 1 and less than 15 mm for case 2. The rainfall distribution for case 1 demonstrates that the suggested DL model better captures the spatial rainfall, whereas the WRF underestimates it by less than 10 mm. In the second scenario, WRF underestimates the HREs, while the suggested DL model accurately predicts them in terms of size, intensity, and dispersion. The confusion matrix indicates that there were no HREs in Case 1, and the DL model's TPR (True Positive Rate) for the LR (Light Rain) and MR (Moderate Rain) classes is greater (>80%) than WRF's (52.5%). Only the HR (Heavy Rain), VHR (Very Heavy Rain), and HER (Extremely Heavy Rain) categories' TPRs of 85.7%, 87%, and 100%, respectively, are captured by the suggested DL model in the second real-time scenario. The study's conclusions directly impact how early warning systems can improve using the DL model.