Abstract:Current deep learning-based QPE methods using radar reflectivity factors mostly adopt global mapping strategies, which to some extent limits the models’ ability to analyze local precipitation features. To this end, this study proposes a Dual-Branch Composite Wavelet Attention UNet (DCWA-UNet) model. The model is mainly improved from the following two aspects. Firstly, a hybrid architecture consisting of a dual-branch encoder (main branch + simplified convolutional downsampling branch) and a feature aggregation subnetwork is designed, which enables the end-to-end mapping from radar volume scan data to station precipitation intensity, thereby constructing a station-centered sample system. Secondly, a composite wavelet attention module (CWAM) is introduced to enhance the model’s representation capability for radar echoes through multi-scale feature decomposition and dynamic weight allocation. Meanwhile, a weighted mean squared error loss function is adopted to emphasize the gradient contribution of moderate-to-high precipitation. Using ground-based precipitation observation and radar observation data in the Sichuan Basin during the summers of 2019-2021, a dataset containing 4020 samples is constructed for model training and testing, and comparative experiments are conducted with deep learning models such as SimVP. The results show that DCWA-UNet achieves obvious comprehensive performance advantages under different precipitation intensities, with particularly significant improvements in critical success index (CSI) and mean absolute error (MAE) within the precipitation intensity below 30 mm·h-1. For precipitation intensities of ［5, 10) mm·h-1, the CSI of DCWA-UNet is significantly higher than that of SimVP and other comparative models, and the MAE is reduced by 4.9% compared to SimVP; for precipitation intensities of ［10, 30］ mm·h-1, the CSI is improved by 4.0% and the MAE is reduced by 4.0% compared to SimVP. Moreover, the false alarm rate is the lowest among all comparative models.