Higher-resolution data improves deep learning-based precipitation nowcasting

Abstract

Accurate and fine-grained precipitation nowcasting holds paramount importance for weather-dependent decision-making and is facing escalating expectations and challenges. While researchers have notably advanced precipitation nowcasting using deep learning (DL) models with larger sizes and more complicated structures, there is scarce research exploring the potential improvement from employing radar data with higher spatial resolutions—a hundred-meter scale rather than a kilometer scale. To evaluate the improvement of higher-resolution data, two U-Net architecture-based models, one larger and another smaller, were designed and trained with radar data at different spatial resolutions—1000 m, 500 m, and 100 m. Their effectiveness was examined by comparison to two baseline models. The models trained with diverse resolutions of data underwent comparative evaluation through two specific precipitation cases. The results unveil a positive correlation between the precipitation nowcasting performance and the spatial resolution of radar data. Models trained with higher-resolution data demonstrate superior forecasting accuracy, reduced bias, and enhanced spatial alignment between predictions and observations. Higher-resolution data empowers DL models to capture boundaries and local-scale patterns of convective systems more accurately, thereby improving the performance in precipitation nowcasting. More importantly, the comparison indicates that to further promote the performance of DL-based precipitation nowcasting, improving data resolution is more efficient than expanding model size. The use of high-resolution data diminishes computational and development costs by concurrently reducing the size of DL models, underscoring pragmatic benefits for related services. Given limited resources, employing higher-resolution data is recommended for priority consideration over larger-size models.