Deep learning model for drought prediction based on large-scale spatial causal network in the Yangtze River Basin
(2025) In Journal of Hydrology 654.- Abstract
Developing accurate large-scale drought prediction models is challenging due to the complex temporal and spatial correlation patterns that govern drought dynamics, as well as the compounding effects of anthropogenic activities and global climate change. Although recent advances in deep learning have yielded effective drought prediction models, many struggle to fully capture the heterogeneous spatial linkages over large-scale regions. In this study, we proposed a novel large-scale drought prediction framework that considers spatial heterogeneity and leverages a causal network connecting regions delineated by drought centroids of severe agricultural events identified through dynamic drought analysis. Using the predefined causal network,... (More)
Developing accurate large-scale drought prediction models is challenging due to the complex temporal and spatial correlation patterns that govern drought dynamics, as well as the compounding effects of anthropogenic activities and global climate change. Although recent advances in deep learning have yielded effective drought prediction models, many struggle to fully capture the heterogeneous spatial linkages over large-scale regions. In this study, we proposed a novel large-scale drought prediction framework that considers spatial heterogeneity and leverages a causal network connecting regions delineated by drought centroids of severe agricultural events identified through dynamic drought analysis. Using the predefined causal network, we employed the state-of-the-art deep learning algorithm, the Spatio-Temporal Graph Convolutional Networks (STGCN) model, with a recursive multi-step forecasting strategy to predict root-zone soil moisture (RZSM) -based drought indices (DIs) up to four weeks in advance for the Yangtze River Basin (YRB). The results show that compared to the meteorological drought events, the corresponding agricultural drought has a later onset and smaller affected areas, yet greater intensity. The proposed model demonstrated robust predictive performance in drought predictions with an average root mean square error (RMSE) of 0.45 and an R2 value of 0.66 across the YRB for the spatial weekly agricultural DIs on the test dataset. Applying the STGCN with the recursive multi-step forecasting strategy can significantly improve the prediction performance, improving R2 values by 0.15 and reducing RMSE by 0.1 on average, with the most substantial improvements observed during the first three weeks (R2 increases of 0.32, 0.24 and 0.09, respectively). These findings underscore the importance of incorporating spatial correlations and demonstrate the advantages of the STGCN approach for large-scale agricultural drought prediction and inform water resource management at large-scale watersheds.
(Less)
- author
- Dai, Huihui ; Xiong, Lihua ; Ma, Qiumei and Duan, Zheng LU
- organization
- publishing date
- 2025
- type
- Contribution to journal
- publication status
- published
- subject
- keywords
- Deep learning, Drought index, Soil moisture, Spatio-Temporal Graph Convolutional Network
- in
- Journal of Hydrology
- volume
- 654
- article number
- 132808
- publisher
- Elsevier
- external identifiers
-
- scopus:85217895326
- ISSN
- 0022-1694
- DOI
- 10.1016/j.jhydrol.2025.132808
- language
- English
- LU publication?
- yes
- id
- e14d0d6d-6ec2-4fb7-9806-a2cc6e5d8515
- date added to LUP
- 2025-06-10 08:55:59
- date last changed
- 2025-06-10 10:04:26
@article{e14d0d6d-6ec2-4fb7-9806-a2cc6e5d8515,
abstract = {{<p>Developing accurate large-scale drought prediction models is challenging due to the complex temporal and spatial correlation patterns that govern drought dynamics, as well as the compounding effects of anthropogenic activities and global climate change. Although recent advances in deep learning have yielded effective drought prediction models, many struggle to fully capture the heterogeneous spatial linkages over large-scale regions. In this study, we proposed a novel large-scale drought prediction framework that considers spatial heterogeneity and leverages a causal network connecting regions delineated by drought centroids of severe agricultural events identified through dynamic drought analysis. Using the predefined causal network, we employed the state-of-the-art deep learning algorithm, the Spatio-Temporal Graph Convolutional Networks (STGCN) model, with a recursive multi-step forecasting strategy to predict root-zone soil moisture (RZSM) -based drought indices (DIs) up to four weeks in advance for the Yangtze River Basin (YRB). The results show that compared to the meteorological drought events, the corresponding agricultural drought has a later onset and smaller affected areas, yet greater intensity. The proposed model demonstrated robust predictive performance in drought predictions with an average root mean square error (RMSE) of 0.45 and an R2 value of 0.66 across the YRB for the spatial weekly agricultural DIs on the test dataset. Applying the STGCN with the recursive multi-step forecasting strategy can significantly improve the prediction performance, improving R2 values by 0.15 and reducing RMSE by 0.1 on average, with the most substantial improvements observed during the first three weeks (R2 increases of 0.32, 0.24 and 0.09, respectively). These findings underscore the importance of incorporating spatial correlations and demonstrate the advantages of the STGCN approach for large-scale agricultural drought prediction and inform water resource management at large-scale watersheds.</p>}},
author = {{Dai, Huihui and Xiong, Lihua and Ma, Qiumei and Duan, Zheng}},
issn = {{0022-1694}},
keywords = {{Deep learning; Drought index; Soil moisture; Spatio-Temporal Graph Convolutional Network}},
language = {{eng}},
publisher = {{Elsevier}},
series = {{Journal of Hydrology}},
title = {{Deep learning model for drought prediction based on large-scale spatial causal network in the Yangtze River Basin}},
url = {{http://dx.doi.org/10.1016/j.jhydrol.2025.132808}},
doi = {{10.1016/j.jhydrol.2025.132808}},
volume = {{654}},
year = {{2025}},
}