Li et al. (2025) Advancing Agricultural Drought Level Prediction in Guangdong Utilizing ERA5-Land and SMAP-L3 Data

Identification

• Journal: Water
• Year: 2025
• Date: 2025-12-16
• Authors: Xiaoning Li, Zhichao Zhong, Jing Wang, Qingliang Li, Xingyu Zhou, Sen Yan, Jinlong Zhu, Xiao Chen
• DOI: 10.3390/w17243564

Research Groups

• College of Computer Science and Technology, Changchun Normal University, China.
• College of Computer Science and Technology, Changchun University of Science and Technology, China.
• College of Civil Engineering and Architecture, Guangxi University, China.

Short Summary

This study introduces a feature recalibration encoder for LSTM-based models to improve agricultural drought forecasting in Guangdong Province. The research demonstrates that direct prediction of the Soil Water Deficit Index (SWDI) using satellite data (SMAP-L3) provides the most stable and accurate results for medium-to-long-term (7–14 days) drought level forecasting.

Objective

• To advance agricultural drought level prediction by developing a feature recalibration mechanism that dynamically prioritizes informative meteorological and soil variables within deep learning architectures.

Study Configuration

• Spatial Scale: Regional (Guangdong Province, China; 20°–26° N, 109°–118° E).
• Temporal Scale: 2015–2020 (Training: 2015–2019; Testing: 2020) with forecast lead times of 1, 3, 7, and 14 days.

Methodology and Data

• Models used: Long Short-Term Memory (LSTM), Attention-based LSTM (AttLSTM), Encoder–Decoder LSTM (EDLSTM), and Attention-based Encoder–Decoder LSTM (AEDLSTM). A novel Feedforward Attention Mechanism (FAM) was integrated for feature recalibration.
• Frameworks:
  • FWA: Predicts Soil Moisture (SM) to calculate SWDI and drought levels.
  • FWB: Directly predicts SWDI to determine drought levels.
  • FWC: Directly classifies drought severity levels.
• Data sources: ERA5-Land reanalysis (0.1° resolution), SMAP-L3 satellite soil moisture (36 km resolution), AVHRR GIMMS-3G+ (NDVI), MERIT DEM (topography), and SoilGrid (soil texture).

Main Results

• Short-term Performance: ERA5-Land-driven models performed best for 1–3 day leads; the FWB-LSTM achieved an overall accuracy ($ACC$) of 0.805 for 1-day predictions.
• Long-term Performance: SMAP-L3-driven models showed superior stability for 7–14 day leads. The FWB-EDLSTM model maintained an $ACC$ of 0.825 at a 14-day lead time.
• Framework Comparison: Framework B (direct SWDI prediction) generally outperformed FWA and FWC, particularly when paired with attention-based models (AttLSTM and AEDLSTM) for capturing extreme drought events.
• Encoder Impact: The proposed feature recalibration encoder consistently improved accuracy across all architectures by focusing on critical signals like precipitation deficits and temperature anomalies.

Contributions

• Developed a novel feature recalibration encoder that acts as an intelligent gateway to amplify relevant input features for drought modeling.
• Conducted a systematic comparison of three distinct forecasting frameworks, identifying that direct index prediction (FWB) minimizes cumulative errors compared to soil moisture-driven approaches (FWA).
• Provided a comprehensive evaluation of the trade-offs between high-resolution reanalysis data (ERA5-Land) and satellite observations (SMAP-L3) across different forecast horizons.

Funding

• Jilin Provincial Science and Technology Development Plan Project (20230101370JC).
• Scientific Research Project of the Education Department of Jilin Province (JJKH20261802KJ).
• Natural Science Foundation of Changchun Normal University (CSJJ2024008ZK).
• Jilin Province Education Science “14th Five-Year Plan” Project (BRJ25073).

Citation

@article{Li2025Advancing,
  author = {Li, Xiaoning and Zhong, Zhichao and Wang, Jing and Li, Qingliang and Zhou, Xingyu and Yan, Sen and Zhu, Jinlong and Chen, Xiao},
  title = {Advancing Agricultural Drought Level Prediction in Guangdong Utilizing ERA5-Land and SMAP-L3 Data},
  journal = {Water},
  year = {2025},
  doi = {10.3390/w17243564},
  url = {https://doi.org/10.3390/w17243564}
}

Generated by BiblioAssistant using gemini-3-flash-preview (Google API)