Liu et al. (2025) High-resolution remote sensing-driven water management in semi-arid basins: A CNN-Attention-SWAT fusion framework for the Fen River

Identification

• Journal: Science of Remote Sensing
• Year: 2025
• Date: 2025-11-13
• Authors: Jiawen Liu, Xianqi Zhang, Yang Yang, Kaiqiang fu, Kaimin Wang
• DOI: 10.1016/j.srs.2025.100333

Research Groups

• Water Conservancy College, North China University of Water Resources and Electric Power, Zhengzhou, China
• Collaborative Innovation Center of Water Resources Efficient Utilization and Protection Engineering, Zhengzhou, China
• Technology Research Center of Water Conservancy and Marine Traffic Engineering, Henan Province, Zhengzhou, China

Short Summary

This study proposes a physics-embedded deep learning (PIDL) framework with bidirectional coupling between the SWAT hydrological model and a CNN-Attention-BiLSTM deep learning model to enhance water management in the semi-arid Fen River Basin. The framework demonstrates superior predictive accuracy for runoff and pollution loads, enabling robust, multi-objective optimized strategies for water allocation, pollution control, and ecological restoration.

Objective

• To develop a Physics-Embedded Deep Learning (PIDL) framework with bidirectional coupling to eliminate unidirectional error accumulation between mechanistic and data-driven models.
• To establish a multi-scale prediction-management system capable of coordinated operations across daily, monthly, and annual scales.
• To validate the robustness of the derived management strategies for policy translation via Shanxi’s digital water platform.

Study Configuration

• Spatial Scale: Fen River Basin (FRB), Shanxi Province, China, covering an area of 39,471 km². The study utilizes data at various resolutions including 10 m (land use, Sentinel-2 NDVI), 30 m (DEM), 1 km (soil data), 0.05° (radar precipitation), and county-level socio-economic data.
• Temporal Scale: Daily for runoff and meteorological inputs, 1-hour for radar precipitation, and multi-scale (daily, monthly, annual) for prediction and management strategies. The SWAT model was run from January 2000 to December 2023 (2000-2002 spin-up, 2003-2015 calibration, 2016-2023 validation). Remote sensing and water quality data span 2018-2023.

Methodology and Data

• Models used:
  • Hydrological Model: Soil and Water Assessment Tool (SWAT)
  • Deep Learning Model: Convolutional Neural Network (CNN)-Attention-Bidirectional Long Short-Term Memory (BiLSTM)
  • Optimization Algorithm: Non-dominated Sorting Genetic Algorithm II (NSGA-II)
  • Uncertainty Quantification: Monte Carlo simulation, Generalized Likelihood Uncertainty Estimation (GLUE)
  • Data Assimilation: Ensemble Kalman Filter (EnKF)
  • Calibration/Sensitivity Analysis: Sequential Uncertainty Fitting Version 2 (SUFI-2), Morris screening method
  • Deep Learning Optimization: AdamW optimizer, Grey Wolf Optimizer (GWO)
  • Decision Making: Entropy-weighted TOPSIS
• Data sources:
  • Topography: ASTER GDEM v3 (30 m resolution)
  • Land Use: GlobeLand30 (2020, 10 m resolution)
  • Soil: HWSD v1.2 (1 km resolution)
  • Meteorology: Daily precipitation, temperature, wind speed, relative humidity from China Meteorological Administration (7 stations); CMFD reanalysis dataset (solar radiation, 0.1° × 0.1°); CMPA radar-retrieved precipitation (1 h / 0.05°)
  • Remote Sensing: Sentinel-2 MSI (NDVI, 10 m resolution); MODIS ET (500 m resolution)
  • Hydrology: Daily runoff from Shanxi Water Resources Department hydrological stations; Water quality monitoring data (total nitrogen, total phosphorus concentrations) from 23 monitoring transects (Bulletin of Ecological Environment Status in Shanxi Province)
  • Socioeconomics: Irrigated area, groundwater extraction, industrial water use from Statistical Yearbook of Shanxi Province (county scale)
  • Supporting Data: FRB Ecological Protection Plan (ecological flow threshold history)

Main Results

• The bidirectionally coupled SWAT-CNN-Attention-BiLSTM model significantly outperformed standalone and unidirectionally coupled models.
  • Runoff prediction: R² = 0.94, Root Mean Square Error (RMSE) = 0.12 mm/d (37% improvement in RMSE vs. unidirectional coupling, 10% improvement in R²).
  • Total nitrogen load estimation error: Reduced to 14.3% (49% lower than conventional SWAT). Hotspot identification accuracy: ±1.5 km.
  • Ecological flow threshold compliance: Reached 92% (23% improvement over baseline BiLSTM).
• NSGA-II optimized strategies achieved synergistic benefits across economic, ecological, and social objectives:
  • Drip irrigation (65% coverage): Reduced groundwater extraction by 12% (150 million m³/year) and increased irrigation efficiency from 0.52 to 0.67.
  • Vegetative buffers (50 m width): Reduced nitrogen loads by 31% (95% CI: 28–34%).
  • Dynamic ecological flows (dry season: 15 m³/s; wet season: 25 m³/s): Decreased annual river disconnection days from 67 to 21.
  • Overall systemic efficiency gain: 18–25%.
• Uncertainty and robustness analysis confirmed the framework's reliability:
  • Coupled optimization reduced groundwater overdraft to 98% (vs. 145% baseline) and total nitrogen to 1.6 mg/L (vs. 2.8 mg/L).
  • Under RCP8.5 climate change scenarios, the coupled model demonstrated enhanced robustness with strategy fluctuations constrained to ±11% (compared to ±45% for standalone SWAT).
  • Monte Carlo simulations confirmed a 95% probability of meeting critical ecological constraints (groundwater over-extraction ≤100%; total nitrogen concentration ≤1.6 mg/L).

Contributions

• Pioneers a physics-embedded deep learning (PIDL) paradigm with bidirectional coupling, establishing the first fully bidirectional feedback between the SWAT model and a CNN-Attention-BiLSTM architecture at a basin scale.
• Introduces a novel "monitor-simulate-optimize-validate" paradigm for semi-arid basin management, integrating high-resolution remote sensing, advanced deep learning, and multi-objective optimization.
• Implements forward physical constraint embedding (SWAT outputs like Soil Water Stress Index and groundwater depth into DL inputs) and backward Bayesian parameter inversion (DL prediction errors dynamically update SWAT parameters).
• Develops a multi-scale closed-loop policy integration system that generates Pareto-optimal strategies validated through physical models and Monte Carlo simulations, ensuring ecological feasibility and robustness.
• Quantifies complex trade-offs and synergies among economic, ecological, and social objectives, providing a scientific basis for cost-benefit optimization of management interventions.
• Demonstrates successful science-policy integration, with the dynamic ecological flow thresholds recommended by the model formally adopted by the Shanxi Provincial Department of Water Resources.

Funding

• Key Scientific Research Project of Colleges and Universities in Henan Province (CN) [grant numbers 17A570004]
• Henan Provincial Universities Scientific and Technological Innovation Team Supporting Programme Grant [24IRTSTHN012]
• National Natural Science Foundation of China [51779093]
• North China University of Water Resources and Hydropower (NCUWH) Graduate Student Innovation Ability Enhancement Project [NCWUYC-202416011]

Citation

@article{Liu2025Highresolution,
  author = {Liu, Jiawen and Zhang, Xianqi and Yang, Yang and fu, Kaiqiang and Wang, Kaimin},
  title = {High-resolution remote sensing-driven water management in semi-arid basins: A CNN-Attention-SWAT fusion framework for the Fen River},
  journal = {Science of Remote Sensing},
  year = {2025},
  doi = {10.1016/j.srs.2025.100333},
  url = {https://doi.org/10.1016/j.srs.2025.100333}
}

Generated by BiblioAssistant using gemini-2.5-flash (Google API)