Page 326 - 《高原气象》2026年第3期
P. 326
高 原 气 象 45 卷
934
0364-06. Xiao T G, et al, 2025. Multi-source rainfall data assimilation
周媛媛, 杨晓辉, 肖天贵, 2025. 基于宽度学习系统的云南省多源 based on broad learning system over Yunnan Province[J]. Plateau
降水数据融合模型研究[J]. 高原气象, 44(2): 475-491. DOI: Meteorology, 44(2): 475-491. DOI: 10. 7522/j. issn. 1000-
10. 7522/j. issn. 1000-0534. 2023. 00085. Zhou Y Y, Yang X H, 0534. 2023. 00085.
Fusion of FY-4B Land Surface Temperature Based on XGBoost Algorithm:
A Case Study of Shaanxi Province
1, 2
1, 2
1, 2
ZHANG Xuting , QUAN Wenting , ZHOU Hui , PAN Yuying ,
1, 2
WANG Weidong , LI Meirong , WANG Zhao 1, 2
1, 2
1, 2
(1. Shaanxi Meteorological Service Center of Agricultural Remote Sensing and Economic Crops, Xi’an 710016, Shaanxi, China;
2. China Meteorological Administration Eco⁃Environment and Meteorology for the Qinling Mountains and
Loess Plateau Key Laboratory, Xi’an 710016, Shaanxi, China)
Abstract: Thermal infrared remote sensing technology enables the rapid acquisition of land surface temperature
(LST) data at both regional and global scales. However, its effectiveness is substantially diminished under
cloudy conditions, where it fails to reliably characterize the underlying surface thermal environment. To address
the issue of data gaps in the Fengyun-4B (FY-4B) satellite LST remote sensing products in cloud-covered areas,
we proposed a multi-source data fusion method based on the eXtreme Gradient Boosting (XGBoost) machine
learning algorithm. The method integrated FY-4B LST products with auxiliary datasets, including the CMA Land
Data Assimilation System (CLDAS) LST products, meteorological station observations, as well as topographic
data and vegetation index, to reconstruct and fuse cloud-covered LST in Shaanxi Province on different typical
dates. The results showed that, (1) for selected typical dates, the correlation coefficients between the fused
cloud-covered LST and CLDAS LST exceeded 0. 91, with both mean absolute errors (MAE) and root mean
square errors (RMSE) stable between 2 ℃ and 4 ℃, or reduced by more than 0. 5 ℃ compared to clear-sky ar‐
eas.(2) The fusion method effectively addressed data gaps in cloud-covered areas, while preserving the spatial
characteristics of the original FY-4B clear-sky LST. Moreover, the fused results exhibited high spatial consisten‐
cy with the CLDAS LST across diverse terrains, including the Loess Plateau in Northern Shaanxi, the Guan‐
zhong Plain, and regions characterized by complex topography.(3) Shapley Additive exPlanations (SHAP)
analysis revealed that CLDAS LST, latitude and the normalized difference vegetation index emerged as the key
feature variables influencing the fusion outcomes, with higher latitudes and areas with sparse vegetation exhibit‐
ing pronounced positive contributions to the simulated LST values. For areas situated north of 35°N - 36°N, in‐
creased latitude correlated with higher simulated LST values. Conversely, lower normalized difference vegeta‐
tion index (NDVI) values facilitated higher LST outputs, while NDVI values exceeding 0. 2 reverse this contri‐
bution direction.(4) Comparative validation across different regions under both cloud-covered and all-weather
conditions indicated that the fused LST results were closely aligned with CLDAS LST, with performance in
Northern Shaanxi slightly inferior to that in the Guanzhong and Southern Shaanxi regions, the latter two exhibit‐
ing comparable and relatively superior accuracy. These regional discrepancies may arise from the combined ef‐
fects of cloud distribution patterns, topographic complexity, and the model’s sensitivity to local feature vari‐
ables. The method proposed in this study can provide technical support for operational all-weather land surface
temperature monitoring using satellite remote sensing, thereby enhancing the reliability and applicability of satel‐
lite-based thermal observations.
Key words: FY-4B; land surface temperature; XGBoost machnne learning algorithm; SHAP; CLDAS

