EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Deep Learning–Based Fusion of Multi‐Source Precipitation in Complex Terrain: A Case Study Over Chongqing, China

Rong Mu, Xinyue Li, Jun Yu, Hongping Gu, Chengzhi Deng

Geoscience Data Journal · 2026-10 · 13 (4)

DOI: 10.1002/gdj3.70104 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Geoscience Data Journal, online erschienen am 21.09.2026. Verfasst von Rong Mu und 4 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Accurately capturing the spatial distribution of precipitation over complex terrain remains a major challenge in meteorology. Satellite observations struggle to simultaneously achieve both high spatiotemporal resolution and retrieval accuracy, while radar observations are easily affected by terrain blockage in mountainous areas. As a result, a single data source is insufficient for high‐resolution precipitation monitoring. In this study, we developed a U‐Net–based deep learning fusion framework incorporating digital elevation model (DEM) information as a terrain constraint. Using hourly GPM IMERG satellite precipitation, SWAN QPE radar precipitation, and rain‐gauge observations collected over Chongqing, China, during the rainy seasons (April–October) of 2016–2021, three multi‐source fusion schemes were constructed: GModel (satellite + gauges), QModel (radar + gauges), and GQModel (satellite + radar + gauges). The performance of the two original precipitation products and the three fused products was comprehensively evaluated. The results show that GPM IMERG tends to overestimate precipitation, and the peak time of precipitation is delayed by approximately 1 h. Although its overall performance is inferior to that of SWAN QPE, the radar product generally underestimates precipitation because of terrain blockage, with substantially better performance over the relatively flat western hills than over the mountainous eastern region in Chongqing. All three fusion products exhibit markedly better agreement with rain‐gauge observations than the original satellite and radar products in both temporal evolution and precipitation magnitude, effectively reducing the systematic biases associated with complex terrain. The generalisation capability of the proposed framework was further evaluated using independent observations from April to October 2023 together with two representative heavy rainfall events that occurred over the western hilly region and the northeastern mountainous region, respectively. The fused precipitation products successfully reproduced both the temporal evolution and rainfall magnitude observed by rain gauges. Overall, GQModel achieved the best comprehensive statistical performance, whereas QModel showed superior skill in estimating extreme precipitation events (> 40 mm h −1 and representative heavy rainfall cases).

Zitieren

APA

Mu, R., Li, X., Yu, J., Gu, H., & Deng, C. (2026). Deep Learning–Based Fusion of Multi‐Source Precipitation in Complex Terrain: A Case Study Over Chongqing, China. Geoscience Data Journal, 13(4). https://doi.org/10.1002/gdj3.70104

Vancouver

Mu R, Li X, Yu J, Gu H, Deng C. Deep Learning–Based Fusion of Multi‐Source Precipitation in Complex Terrain: A Case Study Over Chongqing, China. Geoscience Data Journal. 2026;13(4). doi:10.1002/gdj3.70104

BibTeX
@article{Mu2026,
  author = {Mu, Rong and Li, Xinyue and Yu, Jun and Gu, Hongping and Deng, Chengzhi},
  title = {Deep Learning–Based Fusion of Multi‐Source Precipitation in Complex Terrain: A Case Study Over Chongqing, China},
  journal = {Geoscience Data Journal},
  year = {2026},
  volume = {13},
  number = {4},
  doi = {10.1002/gdj3.70104}
}
Kostenlos starten →