Xing et al. (2025) A deep learning-based composite agricultural drought index for monitoring and impact assessment in Central Asia

Identification

• Journal: Agricultural Water Management
• Year: 2025
• Date: 2025-12-06
• Authors: Xiuwei Xing, Shujie Wei, Xi Chen, Jing Qian, Shuhong Peng, Jiayu Sun, Bo Sun, Chaoliang Chen
• DOI: 10.1016/j.agwat.2025.110043

Research Groups

• Xinjiang Institute of Ecology and Geography, Chinese Academy of Sciences, Ürümqi, China
• Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China
• College of Geoinformatics, Zhejiang University of Technology, Hangzhou, China
• University of Chinese Academy of Sciences, Beijing, China
• School of Computer Science and Engineering, Huizhou University, Huizhou, China
• Moganshan Geospatial Information Laboratory, Huzhou, China
• College of Tourism, Wuyi University, Nanping, China

Short Summary

This study develops a Composite Agricultural Drought Index (CAEDI) using an unsupervised Convolutional Autoencoder (CAE) to integrate multiple drought indicators with soil moisture as a physical prior. CAEDI effectively monitors agricultural drought in Central Asia, outperforming individual indices and accurately assessing yield losses, particularly during critical crop phenological stages.

Objective

• To construct CAEDI through deep feature fusion to capture nonlinear drought responses.
• To verify the applicability of CAEDI for spatiotemporal drought monitoring.
• To analyze the spatial–temporal impacts of agricultural drought using the CAEDI framework.

Study Configuration

• Spatial Scale: Central Asia (approximately 4 million square kilometers), with data resampled to 1 kilometer spatial resolution. Analysis focused on agriculturally sensitive areas, particularly northern Kazakhstan.
• Temporal Scale: Growing seasons from 2015 to 2024, with data processed into 10-day intervals. Crop yield data spans 1994–2023 for Akmola and Kostanay, and 2004–2024 for North Kazakhstan.

Methodology and Data

• Models used:
  • Convolutional Autoencoder (CAE) for Composite Agricultural Drought Index (CAEDI) development.
  • Extreme Gradient Boosting (XGBoost) for Agricultural Drought Impact Index (ADI) construction.
  • SHapley Additive exPlanations (SHAP) for importance analysis of drought impacts across phenological stages.
  • Comparison models: Random Forest (RF), Light Gradient Boosting Machine (LightGBM), Standardized Precipitation Evapotranspiration Index (SPEI), Standardized Precipitation Index (SPI), Crop Water Stress Index (CWSI), Temperature-Vegetation Condition Index (TVCI), Soil Moisture Anomaly Index (SSMI), PCA-based Drought Index (PCADI), Euclidean Distance-based Drought Index (ODDI).
• Data sources:
  • Satellite/Remote Sensing:
    • Moderate Resolution Imaging Spectroradiometer (MODIS) MOD/MYD13Q1 (Normalized Difference Vegetation Index, NDVI) and MOD/MYD11A2 (Land Surface Temperature, LST) for Vegetation Condition Index (VCI) and Temperature Condition Index (TCI).
    • Soil Moisture Active Passive (SMAP) Level-4 Soil Moisture (L4_SM) for Soil Moisture Condition Index (SMCI) and as physical constraints.
    • Global Precipitation Measurement (GPM) IMERG Final Run Version 07B for Precipitation Condition Index (PCI).
    • MODIS MOD16A2GF (evapotranspiration, ET, and potential evapotranspiration, PET) for CWSI.
    • ESA Climate Change Initiative (CCI) land cover product (300 meter resolution).
  • Reanalysis: SPEIbase (CRU TS 4.08 reanalysis data) for SPEI.
  • In-situ/Observation: Monthly Hydrometeorological Bulletins from Kazhydromet (precipitation-based drought severity classifications).
  • Statistical: Annual wheat and oilseed yield data from Akmola, Kostanay, and North Kazakhstan Provinces (official statistics from stat.gov.kz).

Main Results

• CAEDI exhibits strong correlations with SPEI-2 and SPEI-3 (R > 0.80, p < 0.01) and consistently outperforms individual indices in identifying agricultural drought, particularly at medium- to long-term scales.
• CAEDI demonstrates high spatial and temporal consistency with conventional drought indices (SPI, CWSI, TVDI) and strong responsiveness to soil moisture dynamics (R > 0.80 with SSMI).
• CAEDI shows significant correlations with both wheat and oilseed yields across critical growth periods (e.g., R = 0.85 for oilseeds at period 9), outperforming traditional indices in robustness and temporal alignment.
• CAEDI demonstrates strong spatial agreement with Net Primary Production (NPP) (mean r = 0.80, p < 0.05 over 70% of the area), effectively capturing drought-induced productivity declines.
• CAEDI accurately reproduces in-situ drought classifications from Kazhydromet, showing a clear increasing trend from "extreme drought" (mean CAEDI = 0.18) to "extreme wetness" (mean CAEDI = 0.53).
• The Agricultural Drought Impact Index (ADI), developed using XGBoost, achieved the highest accuracy on the validation set (R² = 0.74, RMSE = 0.51, MAE = 0.39) in quantifying yield anomalies.
• ADI spatial patterns from 2015 to 2024 consistently captured drought impacts, showing significant yield losses (e.g., > 0.45 standard deviation average loss in 2019, 2021, 2023) in northern Kazakhstan, aligning with national statistics.
• SHAP analysis revealed that drought impact on wheat yield is most sensitive during stem elongation to heading (Periods 7–11), with SHAP values peaking at 0.23 and 0.13 (accounting for 18.80% and 10.29% of total contribution) during Periods 7 and 9, respectively.

Contributions

• Proposes a novel Composite Agricultural Drought Index (CAEDI) using an unsupervised Convolutional Autoencoder (CAE) that integrates multiple drought-related indicators through nonlinear multivariate fusion.
• Introduces soil moisture anomalies as a weak physical prior and a correlation loss constraint, ensuring physical consistency and interpretability, which addresses a common limitation in existing data-fusion drought indices.
• Develops a comprehensive remote sensing framework for agricultural drought monitoring and impact assessment in Central Asia, particularly valuable for data-scarce regions.
• Demonstrates superior performance of CAEDI over single-variable and linearly integrated drought indices in capturing multi-scale drought dynamics and their impacts on crop yields.
• Develops an Agricultural Drought Impact Index (ADI) using XGBoost and SHAP to effectively quantify yield losses and identify drought-sensitive phenological stages, enhancing early warning systems and climate-resilient agricultural management.

Funding

• National Key R&D Program of China (2023YFE0103600)
• Key Program of National Natural Science Foundation of China (4236114487)
• Overseas Science and Education Cooperation Center Deployment Project of the Bureau of International Cooperation Chinese Academy of Sciences (131965KYSB20210024)
• Guangdong Provincial Science and Technology Plan Project (2022A0505050059)
• Chinese Academy of Sciences – Alliance of International Science Organizations Sustainable Development Research Program (CAS-ANSO-SDRP-2024–03)
• Tianshan Talent Project of Xinjiang Uygur Autonomous Region, China (2022TSYCLJ0056)

Citation

@article{Xing2025deep,
  author = {Xing, Xiuwei and Wei, Shujie and Chen, Xi and Qian, Jing and Peng, Shuhong and Sun, Jiayu and Sun, Bo and Chen, Chaoliang},
  title = {A deep learning-based composite agricultural drought index for monitoring and impact assessment in Central Asia},
  journal = {Agricultural Water Management},
  year = {2025},
  doi = {10.1016/j.agwat.2025.110043},
  url = {https://doi.org/10.1016/j.agwat.2025.110043}
}