Advancing timely satellite precipitation for IMERG-ER using GOES-16 data and a U-net convolutional neural network modelling approach

Abstract

Timely precipitation information is essential for water resources management and hazard monitoring. In regions with limited ground-based measurements, satellite precipitation products (SPPs) provide a valuable alternative, though data latency often creates an information gap for real-time applications. This study addresses the latency gap of IMERG-ER using a U-Net-based Convolutional Neural Network (CNN) model, trained with near-instantaneous GOES-16 satellite data. The optimal combination of GOES-16 infrared bands (6.2, 6.9, 7.3, 8.4, and 11.2 μm) was determined to enhance IMERG-ER predictions. The CNN model's performance, evaluated with both quantitative and qualitative metrics, showed an RMSE of 0.46 mm/h, a Pearson's correlation coefficient of 0.60, and a Critical Success Index of 0.53. The model performed well in predicting low-intensity precipitation (<3 mm/h), which occurs 97 % of the time, but faced challenges with high-intensity events due to data imbalance. These findings advance the use of SPPs and deep learning for operational hydrology.