Applications of Attention‐Enhanced CNN Models to Regional Precipitation Downscaling

Abstract Precipitation downscaling is essential for generating high‐resolution data from coarse‐resolution global climate models and assessing the environmental impacts of climate change at regional and local scales. Convolutional Neural Networks (CNNs) are an emerging critical deep‐learning technique that promises significant improvements over other downscaling methods. This study employs three attention‐enhanced CNNs including Attention‐based Laplacian Pyramid Network (AttLap), Attention and Convolutional Mix Network (ACMix) and Multi‐scale Attention Network (MAN), and further evaluates their performance in regional precipitation downscaling. Focusing on the Middle Reaches of the Yellow River in China (MRYR), we utilize ERA5 atmospheric variables and Global Precipitation Measurement (GPM) data for model training and testing. Our results indicate that all the attention‐enhanced CNNs improve spatio‐temporal precipitation simulations across daily, monthly, and annual timescales compared to the conventional CNN model. Notably, the AttLap model shows the greatest improvements compared to the conventional CNN, reducing root‐mean‐square error of daily precipitation by 10.1% and increasing the correlation coefficient by 16.7% for the regional mean. Moreover, the attention mechanism improves the model's ability to simulate extreme precipitation, showing that the 95th and 99th percentiles of predicted precipitation are much closer to that of GPM data. Meanwhile, the probability density function for daily precipitation in the attention‐enhanced CNNs exhibits better agreement with GPM data, particularly for heavy precipitation, further confirming the advantage of the attention mechanism in simulating extreme precipitation. These findings indicate that the attention‐enhanced CNNs significantly improve the ability to capture the spatio‐temporal precipitation features, thereby enhancing the downscaling accuracy. The study highlights the potential of attention‐enhanced models for regional precipitation downscaling, providing valuable tools for climate projection and water resource management in complex terrain regions.