Attention-Based Deep Fusion CNN for Geostationary Satellite Rainfall Estimates over Taiwan

Artificial intelligence/machine learning technology has been considered a good way to extract synergistic value in multiple spectral bands from the new generation geostationary satellites. A deep leaning model (named AIQPE) for estimating surface rainfall based on Himawari-8/AHI (H8) data was developed for producing operational products in Taiwan Central Weather Administration. The AIQPE deep leaning model, which has been meticulously tested by various configurations to identify the optimal model design, employs a three-dimensional (3-D) convolutional encoder-decoder architecture with attention mechanisms to fuse cloud features from multiple spectral channels of H8 for estimating hourly precipitation. The use of attention mechanism modules in this model Include mixed pooling and channel attention to fuse multi-modal features and adapt to short-range and long-range dependencies. The 3-D design allows the ingestion of information on cloud changes over the past 30 minutes. Comparing with an earlier 2-D design version, the accuracy of 3-D version increases about 6% according to a correlation index. The evaluation of AIQPE based on its operational products shows promising results both in the overall skill score and the performance of sampling cases affected by various weather systems. An IR-based product established by the PDF-matching technique (named IRQPE) and two GPM rainfall products(IMERG Late V6 and GSMaP MVK V8) are also used for the inter-comparison. The results show that AIQPE outperforms other satellite rainfall products at all time scales from daily rainfall to monthly rainfall estimates for Taiwan’s land area.