Point-Based and Spatial Deep Learning for Radar Quantitative Precipitation Estimation Across Precipitation Intensity Ranges over Hainan Island

Abstract

Accurate quantitative precipitation estimation (QPE) is critical for responding to severe weather events such as heavy rainfall. Deep learning (DL) methods, which can establish nonlinear mappings between radar observations and rain rate (R), have been widely applied to reduce QPE errors. This study evaluates point-based and spatial DL approaches for radar QPE across precipitation intensity ranges over Hainan Island using dual-polarization radar and rain gauge data. Four DL-based QPE models, including R-DNNNet, R-IncNet, R-Res-IncNet, and R-DenseNet, are trained using point-based or spatial grid-based datasets, and evaluated across retrospectively defined subsets based on the observed rain-gauge rain rate: low-intensity (R < 10 mm h−1), moderate-intensity (10 ≤ R ≤ 20 mm h−1) and high-intensity precipitation (R > 20 mm h−1). Results show that DL models generally outperform traditional empirical relationships. For overall precipitation, the point-based three-parameter (ZH-ZDR-KDP) R-DNNNet achieves the best performance (CC = 0.94, RMSE = 5.88 mm h−1), improving CC and RMSE by 4% and 23%, respectively, over the best traditional empirical method. R-DNNNet also performs best for low-intensity and moderate-intensity precipitation. For high-intensity precipitation, the spatial convolutional neural network (CNN) model R-Res-IncNet showed the best overall performance, with maximum improvements of 8%, 14%, and 19% in CC, RMSE, and MAE, respectively. R-Res-IncNet achieved an RMSE of 12.45 mm h−1 compared with 12.81 mm h−1 for R-DNNNet, and the paired-bootstrap 95% confidence interval for the model-to-model RMSE difference was 0.018–1.006 mm h−1, indicating a modest but statistically supported improvement. This result suggests that retaining neighborhood radar information may be beneficial under high-intensity conditions. These intensity-specific rankings are based on retrospective evaluation within subsets defined by observed rain-gauge rain rate and characterize conditional model performance across precipitation intensity ranges. Overall, the results show that the relative performance of the evaluated point-based and spatial-grid QPE configurations varies with rainfall intensity, while the three-variable polarimetric input provides consistent benefits across the evaluated datasets. In summary, the findings highlight the precipitation-intensity-dependent performance characteristics of point-based and spatial DL models and suggest their potential for improving radar QPE performance over Hainan Island, particularly for severe convective rainfall.