Multiscale Attention-UNet-Based Near-Real-Time Precipitation Estimation From FY-4A/AGRI and Doppler Radar Observations

Abstract

Extreme precipitation events greatly threaten people's daily lives and safety, making accurate and timely precipitation estimation especially critical. However, common methods like radar and satellite remote sensing have limitations due to coverage and environmental factors. Existing deep learning models struggle with complex scenarios and multisource data correlations. These make the precipitation estimation tasks challenging. This article proposes a Multiscale Dual Cross-Attention UNet (MS-DCA-UNet) model for near-real-time precipitation estimation. It integrates Doppler weather radar and FY-4A satellite data to overcome single-source data limitations. To narrow the semantic gap among the encoder feature maps, the MS-DCA-UNet model introduces a dual-cross attention (DCA) module at the skip connections of the backbone network U-Net. The DCA module mainly employs a channel cross-attention and a spatial cross-attention to capture remote dependencies and enable multiscale feature fusion. A multiscale convolution module is designed to reduce the risk of the model falling into local optima. It is a multibranch upsampling strategy that runs parallel to the decoder. Experimental results show that the Critical Success Index (CSI), Root Mean Square Error (RMSE), and Pearson's Correlation Coefficient (CC) of MS-DCA-UNet are 0.6033, 0.5949 mm/h, and 0.8460, respectively, with the hourly CMPAS precipitation as the benchmark. These outperform the other comparisons, such as FY-4A QPE, GPM IMERG, U-Net, Attention-UNet, and DCA-UNet on the CSI, RMSE, and CC metrics. MS-DCA-UNet reduces the RMSE of Attention-UNet, UNet, and DCA-UNet by a margin of 34.68% (0.5949 mm/h versus 0.9107 mm/h), 10.24% (0.5949 mm/h versus 0.6628 mm/h), 6.96% (0.5949 mm/h versus 0.6394 mm/h), respectively.