Improving Rainfall Prediction with Satellite Sensors and Deep Learning

The proposed deep learning framework aims to enhance the operational precipitation retrieval product on the GOES-R series satellite by leveraging observations from the ABI and GLM measurements. Comprising two convolutional neural network modules, the framework integrates cloud-top brightness temperature from multiple ABI channels and lightning flash rate data from the GLM measurement. Trained with ground-based multiradar multisensory (MRMS) quantitative precipitation estimates, the deep learning model demonstrates superior performance compared to the current operational product, particularly in capturing moderate to heavy precipitation events. Quantitatively, the CNN-based retrieval model exhibits significant improvements, achieving a minimum of 32% enhancement in integrated Heidke skill score and critical success index metrics and reducing estimation errors by at least 31% in terms of mean-squared error, mean absolute error, and the normal mean absolute error. These enhancements underscore the potential of the deep learning framework to revolutionize precipitation retrieval accuracy and contribute to improved understanding and forecasting of precipitation patterns, especially in regions prone to extreme weather events.