Precipitation Estimation Method Based on Multi-scale Convolutional Attention Decoder

To address the problem of missing precipitation data caused by the uneven distribution of ground-based meteorological observation stations, a precipitation estimation method based on the multi-scale convolutional attention decoder (MCAD) is proposed. This method leverages the all-weather observation capabilities of the Fengyun-4B (FY-4B) meteorological satellite and utilizes deep learning to estimate precipitation with high accuracy. The proposed approach adopts a multi-scale convolutional attention module (MCAM), which integrates a multi-scale convolutional block (MCB), a channel attention block (CAB), and a spatial attention block (SAB). The MCB extracts multi-scale features using parallel depthwise separable convolutions to capture precipitation information across different spatial scales. The CAB applies a channel attention mechanism to enhance the response of key precipitation regions, while the SAB refines the spatial distribution of features using a spatial attention mechanism to improve the recognition of precipitation patterns. The model achieves high computational efficiency while effectively integrating spatial information at multiple scales, thereby improving both the accuracy and detail of precipitation estimation. Experimental results show that the proposed method significantly outperforms traditional approaches in terms of spatiotemporal resolution and precipitation accuracy, achieving an RMSE of 0.4662 and a CC of 0.8324 compared to GPM IMERG reference data.