Lin et al. (2025) Soil salinity estimation based on satellite hyperspectral and synthetic aperture radar remote sensing image fusion

Identification

• Journal: Agricultural Water Management
• Year: 2025
• Date: 2025-12-12
• Authors: Nan Lin, Xunhu Ma, Yuanyuan Sui, Ruifei Zhu, Hanlin Liu, Menghong Wu, Ranzhe Jiang
• DOI: 10.1016/j.agwat.2025.110076

Research Groups

• Modern Industry College, Jilin Jianzhu University, Changchun, China
• College of Biological and Agricultural Engineering, Jilin University, Changchun, China
• Changguang Satellite Technology Co., Ltd., Changchun, China
• College of Resource and Environmental Science, Jilin Agricultural University, Changchun, China

Short Summary

This study proposes a novel multi-scale, multi-depth Wasserstein Generative Adversarial Network with Gradient Penalty (MSD-WGAN-GP) to fuse hyperspectral (HSI) and Synthetic Aperture Radar (SAR) images, significantly improving soil salinity estimation accuracy by mitigating the coupling effects of soil moisture and surface roughness.

Objective

• To develop an HSI and SAR image fusion method based on MSD-WGAN-GP to improve soil salinity estimation accuracy and provide a robust, widely applicable soil salinity prediction model.

Study Configuration

• Spatial Scale: Da’an City region, Jilin Province, Northeast China, covering two experimental areas (Area 1 and Area 2) with diverse land cover and salinity gradients. Soil samples were collected at least 100 m from surrounding ground features.
• Temporal Scale: Soil samples were collected in mid-November 2023. Satellite imagery was acquired on November 14, 2023 (ZY1–02D HSI) and September 9, 2023 (Sentinel-1A SAR), with efforts made to ensure synchronization.

Methodology and Data

• Models used:
  • Multi-scale and Multi-depth Wasserstein Generative Adversarial Network with Gradient Penalty (MSD-WGAN-GP) for HSI and SAR image fusion.
  • Convolutional Neural Network (CNN) for soil salinity prediction.
  • Competitive Adaptive Reweighted Sampling (CARS) for hyperspectral band selection.
  • Random Forest (RF) for supervised classification and removal of non-soil information.
  • Comparative fusion methods: Generative Adversarial Network (GAN), Multi-scale Convolutional Neural Network (MSCNN), Brovey transformation, Dual-Discriminator Conditional Generative Adversarial Network (DDcGAN), and Multi-scale Heterogeneous Fusion Network (MHF-Net).
• Data sources:
  • Satellite Hyperspectral Image (HSI): Ziyuan1–02D (ZY1–02D), 30 m spatial resolution, 153 spectral bands (after filtering).
  • Synthetic Aperture Radar (SAR) Image: Sentinel-1A, 10 m spatial resolution, Vertical Transmit Vertical Receive (VV) and Vertical Transmit Horizontal Receive (VH) polarization modes.
  • Ground-based soil samples: 123 soil samples collected from Da’an City, Northeast China.
  • Measured soil Electrical Conductivity (EC) values (primary indicator of soil salinity).
  • Measured soil moisture (SM) content.
  • Measured soil surface roughness (Root Mean Square Height, RMSH) derived from high-precision 3D laser scanners.

Main Results

• HSI and SAR image fusion significantly improved soil salinity prediction accuracy. Compared with HSI-only prediction, the R² and RPIQ of the model increased by 0.22 and 1.13, respectively, and the RMSE decreased by 2.68 dS⋅m⁻¹.
• The proposed MSD-WGAN-GP model demonstrated superior performance in HSI and SAR image fusion compared to traditional and other deep learning fusion methods, achieving a peak signal-to-noise ratio (PSNR) of 38.39 dB and a structural similarity index (SSIM) of 0.88 for the entire study region.
• MSD-WGAN-GP significantly improved the correlation between soil salinity and spectra, with an average increase of 0.32 in the correlation coefficient per spectral band, while effectively mitigating prediction bias introduced by soil moisture and surface roughness.
• Ablation experiments confirmed the critical importance of the Multi-scale Convolution (MSC) block, Adaptive Fusion (AF) module, attention module, and Adversarial Learning Mechanism (ALM) within the MSD-WGAN-GP model for retaining detail, spectral fidelity, and overall image quality.
• The mean prediction uncertainty for test set samples based on fused images was 0.84 dS⋅m⁻¹, indicating robust model performance and minimal sensitivity to variations in training data composition.

Contributions

• Developed a novel multi-scale, multi-depth Wasserstein Generative Adversarial Network with Gradient Penalty (MSD-WGAN-GP) for HSI and SAR image fusion, specifically designed to address the coupling effects of soil physical properties on spectral responses.
• Demonstrated that integrating HSI and SAR data through this advanced deep learning framework significantly enhances soil salinity estimation accuracy and mapping capability compared to single-source data or traditional fusion methods.
• Quantified the reduction in prediction bias caused by soil moisture and surface roughness through the fusion process, highlighting the mechanistic advantage of combining optical and radar data.
• Provided a robust and transferable computational framework for interpreting complex land surface signals, with potential applications beyond soil salinity to other soil and vegetation properties affected by similar confounding factors.

Funding

• Natural Science Foundation of Jilin Province [20230101373JC]

Citation

@article{Lin2025Soil,
  author = {Lin, Nan and Ma, Xunhu and Sui, Yuanyuan and Zhu, Ruifei and Liu, Hanlin and Wu, Menghong and Jiang, Ranzhe},
  title = {Soil salinity estimation based on satellite hyperspectral and synthetic aperture radar remote sensing image fusion},
  journal = {Agricultural Water Management},
  year = {2025},
  doi = {10.1016/j.agwat.2025.110076},
  url = {https://doi.org/10.1016/j.agwat.2025.110076}
}