Deep neural network high SpatioTEmporal resolution Precipitation estimation (Deep-STEP) using Passive Microwave and Infrared Data

Despite recent and fast-growing advancements in satellite remote sensing, the conventional process-oriented precipitation estimation algorithms suffer from methodological deficiencies, and models have not improved with data availability. This study attempts to ingest direct passive microwave (PMW) brightness temperatures (Tbs) at 18 to 183 GHz frequencies from GPM constellation radiometers combined with infrared (IR) Tbs (10m) from geostationary satellites and surface type information and then engages convolutional neural networks to extract the spatial features related to the precipitable clouds. The end-to-end Deep neural network high SpatioTEmporal resolution Precipitation estimation (Deep-STEP) algorithm instantaneously translates the cloud features into surface precipitation intensity maps in 4-km spatial resolution. The GV-MRMS is used as the reference and all datasets (2017-2019) are preprocessed by a very fast data generator pipeline to train the model from scratch. In addition, Deep-STEP is set up with new training techniques, that mitigate the extreme memory resources utilization in the training process. In the test period (July 2017), Deep-STEP outperforms GPROF and IMERG products in terms of critical success index (CSI) by 10% and 9%, and in terms of Volumetric CSI by 9% and 21%, respectively, that highlights the detection skill of our model. GPROF and IMERG show about 0.48 and 0.29 correlation (C) with respect to GV-MRMS, respectively, while Deep-STEP achieves a better agreement (C=0.59) compared to these well-known operational products. For our experimental setup over the eastern US, our model also performs well over the vegetated and coastal areas. 0.63, 0.47, and 0.20 correlation are reported over vegetated lands and 0.53, 0.38, and 0.25 are reported over coastal regions for Deep-STEP, GPROF, and IMERG, respectively. In general, the results show the effectiveness of (1) integrating a set of Tbs from large PMW footprints with high-resolution IR images, and (2) using automatic neighborhood feature extraction approaches in improving the accuracy and in capturing the fine spatial patterns of precipitation events. We anticipate our model to be a starting point for more sophisticated and efficient precipitation retrieval systems in terms of accuracy and computational cost.