Enhanced Radar Quantitative Precipitation Estimation in South China Through Deep Learning Methods With Global Features

Kexin Zhu, Weixin Xu, Haonan Chen, Nana Liu, Lejian Zhang

IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing · 01 January 2026 · China Meteorological Administration (CMA), Colorado State University (incl. CIRA), University of Oklahoma (AI2ES)

Paper

In one sentence. Graph neural networks outperform CNNs, MLPs and Z-R relations for radar quantitative precipitation estimation in South China, with the lowest MAE for extreme rainfall above 30 mm/h.

Abstract

Accurate radar quantitative precipitation estimation (QPE) remains challenging in regions with complex terrain and diverse precipitation regimes, such as South China, because traditional methods rely on empirical Z–R relations derived from raindrop size distributions. Using dense rain gauge network and ground-based radar measurements, this study fully investigates the potential of deep learning (DL) methods in improving radar QPE over South China. First, multilayer perceptron, convolutional neural networks (CNNs), and graph neural networks (GNNs) with similar architectures are constructed and compared to traditional Z-R relationships considering precipitation types. DL methods outperform traditional Z-R relationships, and GNNs perform the best. More importantly, this study provides a systematic evaluation of precipitation intensity, type, and organization, and the GNN achieves outstanding results. For extreme precipitation ( $>$ 30 mm/h) and regional total rainfall, GNN achieves the smallest MAE. It also maintains low bias and dispersion for stratiform and organized cases. By contrast, CNN better handles isolated rainfall. Last but not least, the Z-R relationship shows systematic spatial biases, generally overestimating precipitation in plains and underestimating it in some inland high-altitude regions. DL methods alleviate these biases by incorporating spatial information. Overall, this study highlights the advantages of DL methods across different precipitation scenarios and demonstrates their ability to mitigate systematic biases from complex terrain.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.