Publications

Zhang, JY; Ding, JL; Wang, JJ; Lin, H; Han, LJ; Li, XH; Liu, J (2023). Remote sensing drought factor integration based on machine learning can improve the estimation of drought in arid and semi-arid regions. THEORETICAL AND APPLIED CLIMATOLOGY.

Abstract
Drought usually occurs as a result of an imbalance between water supply and demand. Drought indices, an important tool for studying drought, are commonly used for drought monitoring and drought risk assessment. The ground-based drought indices, with high accuracy, are limited in the area monitored. In contrast, the remotely sensed drought index covers a large area but with poor accuracy. Data-driven data fusion-based estimation of ground indices helps to fill this gap. The overall objective is to determine whether various remotely sensed drought factors can effectively monitor drought in arid and semi-arid northern China. In this study, the ground-based drought index SPEI was reconstructed by using remotely sensed drought factors, divided from the Global Precipitation Measuring Mission GPM, GLDAS, and MODIS satellite sensors as well as environmental covariates. Based on climate elements, soil elements, vegetation, and environmental covariate elements, a composite drought index CDIS was established in this study. In this study, the single drought index, multivariate linear integrated drought index, and bias-corrected random forest method were used as the prediction model. Based on ground truth historical climate data as reference data, the performance of the model-predicted composite drought index CDIS in drought monitoring was evaluated. Our results indicate that the drought index on 1-month time scale has no significant advantage over the maximum single drought index. The bias-corrected random forest model with drought index predictions outperformed multiple linear regression and single drought index, with higher prediction accuracy in semi-arid and semi-humid regions than in arid regions. Compared with the ground station drought indices, the drought map based on bias-corrected random forest shows visual and statistical consistency. Under the background of a non-homogeneous and complex surface environment, the model prediction results integrating environmental covariates performed best, and the method used in the study was effective and could be extended and applied. For areas where information is scarce, remote sensing data can be easily extended to regional scale monitoring, which can better achieve high-precision drought monitoring at the regional scale.

DOI:
10.1007/s00704-022-04305-z

ISSN:
1434-4483