Page 231 - 《高原气象》2026年第5期
P. 231
5 期 郭继富等:基于BMA降尺度模型的西北雨养农业区干旱监测研究 1447
Study on drought monitoring in rain-fed agricultural area
of Northwest China based on BMA Model
1
1
Guo Jifu , Zhang Jifang , Zhang Qiang , Zhao Funian , Zhao Hui 3
1, 2
2
(1. Lanzhou Institute of Arid Meteorology, CMA, Gansu Province Key Laboratory of Arid Climate Change and Reducing Disaster,
Lanzhou 730020, Gansu, China;
2. College of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, Gansu, China;
3. School of Management, Gansu Agricultural University, Lanzhou 730070, Gansu, China)
Abstract: Under the background of global warming, drought has emerged as a critical disaster severely impact‐
ing the security of global terrestrial ecosystems, with pronounced regional heterogeneity. Sun-induced chloro‐
phyll fluorescence (SIF), which captures the physiological status of vegetation, has demonstrated the potential
for early drought detection. However, the coarse spatial resolution of existing SIF products limits their applica‐
tion in regional-scale precision drought monitoring. This study integrates multi-source remote sensing data, in‐
cluding GOSIF, MODIS, and meteorological station observations, selects vegetation indices (NDVI and EVI),
reflectance bands (RED and NIR), and topographic factors as independent variables, and utilizes an adaptive ge‐
netic algorithm to optimize the model parameters of Support Vector Regression (SVR), Random Forest (RF),
and Convolutional Neural Network (CNN). A downscaling model was subsequently constructed by applying
Bayesian Model Averaging (BMA) to integrate the results from these individual models through weighted aver‐
aging, thereby generating a high-resolution SIF product for drought monitoring. The results indicate that the
BMA-based downscaling model outperformed any single model (SVR, RF, or CNN) optimized with the adap‐
tive genetic algorithm, achieving higher accuracy on both training and testing datasets. The BMA model yielded
an R² of 0. 785, an RMSE of 0. 0453, and a Bias of 0. 009, compared to the SVR model, which obtained an R²
of 0. 62, an RMSE of 0. 0603, and a Bias of 0. 0095. The downscaled SIF product derived from the BMA model
accurately captured drought severity characteristics and exhibited the highest correlation and consistency with the
Palmer Drought Severity Index (PDSI), achieving Pearson correlation coefficient of 0. 7408 and consistency in‐
dex of 0. 7639. During a drought event in June 2022, the BMA-downscaled SIF data effectively identified local‐
ized fine-scale drought patches and core drought areas at the county level, thereby overcoming the limitations of
coarse-resolution drought monitoring at sub-provincial scales. This approach provides valuable technical support
for precision drought monitoring and facilitates improved disaster prevention and mitigation at regional scales.
This investigation into the advantages and disadvantages of BMA as a downscaling technique offers critical tech‐
nical insights for the evaluation and application of statistical downscaling approaches in producing high-resolu‐
tion, regional-scale remote sensing data.
Key words: drought monitoring; Bayesian model averaging; downscaling; rain-fed agriculture in Northwest
China

