An Improved Deep-Learning-Based Precipitation Estimation Algorithm Using Multitemporal GOES-16 Images

A near-real-time precipitation estimation product derived from geosynchronous Earth-orbiting (GEO) satellite data is highly desirable due to its ability to provide extensive coverage with high spatial and temporal resolution. This research presents a novel Deep-Learning-based Precipitation Estimation algorithm using a Multi-SpatioTemporal network (DLPE-MST), to investigate the potential of Geostationary Operational Environmental Satellite-16 (GOES-16) multitemporal images in precipitation estimation. First, a series of Advanced Baseline Imager (ABI) bispectral satellite images (6.19 and
$10.35~\mu \text {m}$
) from GOES-16 are used as inputs. Second, a module based on 3-D convolutional neural networks (3-D CNNs) is proposed to be embedded into the DLPE-MST for extracting motion features within rainfall areas. Third, a novel loss function, separated domain error (SDE), is proposed for DLPE-MST to mitigate the issue of underestimation arising from imbalanced precipitation datasets. Finally, to assess the feasibility of the DLPE-MST, GOES-16 satellite images covering the eastern Continental United States (CONUS) of America during the summer of 2020–2021 are utilized to generate raster maps depicting hourly rainfall rates at a resolution of
$ 0.04^{\circ } \times 0.04^{\circ }$
. The experimental results indicate that our algorithm outperforms others in terms of probability of detection (POD) and correlation coefficient (CC), achieving scores of 91.79% and 0.58, respectively. The statistical analysis of multiple rainfall events also demonstrates that the DLPE-MST outputs are closer to the ground truth compared to other products. Furthermore, the SDE shows significant potential in alleviating the underestimation of heavy rain events. After testing, this algorithm takes only 0.09 s to generate one raster map of the rainfall rate.