NowCasting-Nets: Representation Learning to Mitigate Latency Gap of Satellite Precipitation Products Using Convolutional and Recurrent Neural Networks

Accurate and timely estimation of precipitation is critical for issuing hazard warnings (e.g., flash floods or landslides). Current remotely sensed products have a few hours latency, associated with the acquisition processing satellite data. By applying robust nowcasting system to these products, it (in principle) possible reduce this latency improve their applicability, value, impact. However, development such complicated by chaotic nature atmosphere, consequent rapid changes that can occur in structures systems In work, we develop two approaches (hereafter referred as Nowcasting-Nets) use Recurrent Convolutional deep neural network address challenge nowcasting. A total five models are trained using Global Precipitation Measurement (GPM) Integrated Multi-satellitE Retrievals GPM (IMERG) data over Eastern Contiguous United States (CONUS) then tested against independent Western CONUS. The were designed provide forecasts lead time up 1.5 and, feedback loop approach, ability extend forecast 4.5 was also investigated. Model performance compared Random Forest (RF) Linear Regression (LR) machine learning methods, persistence benchmark (BM) used most recent observation forecast. Independent IMERG observations reference, experiments conducted examine both overall statistics case studies involving specific events. Overall, provided Nowcasting-Net superior, Nowcasting Network Residual Head (CNC-R) achieving 25%, 28%, 46% improvement test ...