Satellite precipitation prediction using GOES-16 data and Convolutional Neural Networks

Abstract

Real-time precipitation information is critical for effective basin monitoring and responding to hazardous events like storms, floods, and landslides. In regions with a complex topography and limited ground-based measurements, reliable real-time data is scarce, hindering decision-making. Satellite Precipitation Products (SPPs) have emerged as valuable data to overcome these limitations, offering precipitation estimates in instances where information is either insufficient or unavailable. Despite their utility, SPPs are constrained by latency, typically spanning several hours. Latency causes the omission of possible precipitation events during this interval. This study addresses the latency gap challenge by developing a model that predicts near-instantaneous IMERG-ER SPP using a U-Net-based Convolutional Neural Network (CNN) and infrared data from the GOES-16 geostationary satellite. Five years of hourly data (2019-2023) were used, with the first four for training and the last year for testing. The used band combination was 6.2, 6.9, 7.3, 8.4, and 11.2 μm wavelengths. The model yielded predicted SPPs with an approximate latency of 11 minutes, exhibiting an RMSE of 0.46 mm/h, a Pearson Correlation Coefficient of 0.60, a bias of -0.007 mm/h, and a Critical Success Index of 0.53. While the model performed well in predicting frequent low-intensity precipitation, it struggled with high-intensity values. These shortcomings can be attributed to data imbalance and model regularization procedures. The results highlight the potential of the proposed approach for providing a timely spatial precipitation prediction, particularly in regions where ground-based data is scarce.