PrecipGAN: Merging Microwave and Infrared Data for Satellite Precipitation Estimation using Generative Adversarial Network

Global satellite precipitation estimation at high spatiotemporal resolutions is crucial for hydrological and meteorological applications but is still a challenging task. Beside the fact that precipitation cannot be directly observed by satellite, another challenge lies in the fact that the microwave data are not continuous in space and time. We present a novel approach to merge incomplete passive microwave (PMW) precipitation estimates with the conditional information provided by complete infrared (IR) precipitation estimates based on deep learning, generative adversarial network (GAN), and name the algorithm PrecipGAN. PrecipGAN is designed based on the physical knowledge of the precipitation system by decomposing it into content and evolution subspaces, and trained with PMW and IR data from the Integrated Multi-satellitE Retrievals for GPM (IMERG) product at the 0.1 degree and hourly resolution, and verified through the comparison to IMERG Uncalibration (IMERG Uncal). PrecipGAN can skillfully simulate the spatiotemporal changes of precipitation systems and produce precipitation estimates with overall better statistical performance than the benchmark IMERG Uncal. Specifically, the Pearson correlation coefficient (CC), mean error (ME), and root mean squared error (RMSE) of PrecipGAN in the testing period are 0.40, -0.05 mm/h, and 2.27 mm/h, respectively. The CC of PrecipGAN is slightly worse than IMERG Uncal due to the incorporation of IR intensity estimates in PrecipGAN as well as the better source data and bi-directional propagation used in IMERG Uncal. In general, the large-scale application of PrecipGAN over the Continental US (CONUS) in a whole year demonstrates its generalization ability to accurately reproduce and capture the spatiotemporal variations of precipitation. PrecipGAN will facilitate more accurate and computationally efficient algorithm that can be implemented globally to produce satellite-based precipitation estimates.