A Deep-learning pipeline to estimate instantaneous rain rates from INSAT-3D Outgoing Longwave Radiation data

This study proposes a deep learning framework to estimate instantaneous rain rates in India and surrounding regions (latitudes $0 \mathrm{~N}-40 \mathrm{~N}$ and longitudes $60 \mathrm{E}-100 \mathrm{E}$) using the INSAT-3D Imager Outgoing Longwave Radiation (OLR) product as input. The study utilizes OLR data with 4 km spatial resolution from June 2022 for training and June 2021 for testing. The estimation is done using a customized conditional Generative Adversarial Network (cGAN). The generator architecture incorporates downsampling convolutional layers, multiple residual blocks, and upsampling transposed convolutional layers, while the discriminator uses a PatchGAN approach. Results show that the proposed model results in a lower RMSE relative to the operational INSAT Multi-Spectral Rain Algorithm (IMSRA) and HydroEstimator (HE) algorithms (0.88 vs. 1.96 and 3.25). The model demonstrates superior threat scores for lower rainfall thresholds, while the operational methods demonstrate better results for rain rates greater than $10 \mathrm{~mm} / \mathrm{hr}$.