Zhang et al. (2026) Cropland soil salinity retrieval using a spectral-spatial cross-attention deep learning framework with environmental interpretability

Identification

• Journal: Computers and Electronics in Agriculture
• Year: 2026
• Date: 2026-04-09
• Authors: Junyan Zhang, Chong Huang, He Li, Qingsheng Liu, Miao Lu
• DOI: 10.1016/j.compag.2026.111748

Research Groups

• State Key Laboratory of Resources and Environmental Information System, Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, Beijing, China
• College of Resource and Environment, University of Chinese Academy of Sciences, Beijing, China
• State Key Laboratory of Efficient Utilization of Arable Land in China, Key Laboratory of Agricultural Remote Sensing (AGRIRS) Ministry of Agriculture and Rural Affairs, Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Sciences, Beijing, China

Short Summary

This study develops SS-SoilNet, a multimodal deep learning framework, to accurately retrieve cropland soil salinity in the Yellow River Delta by integrating multi-source remote sensing observations, topographic features, and crop growth parameters. The model achieves improved accuracy (RMSE ≈ 3.6 g kg−1, R² ≈ 0.68) and interpretability, revealing strong coupling effects among soil salinity, crop growth, and terrain.

Objective

• To develop and validate a multimodal deep learning framework (SS-SoilNet) that integrates remote sensing spectral sequences, indices, and environmental covariates to accurately retrieve soil salinity in coastal croplands, explicitly accounting for spatial heterogeneity and multi-year temporal variability.

Study Configuration

• Spatial Scale:
  • Study Area: Yellow River Delta, northern Shandong Province, China.
  • Remote Sensing Pixel Resolution: 10 meters (Sentinel-2 bands, 20 m bands up-sampled).
  • DEM Resolution: 30 meters (Copernicus GLO-30).
  • Environmental Feature Extraction Unit: 65 × 65 pixel window (650 meters × 650 meters) centered on each target pixel.
  • Soil Sampling Depth: 0–20 centimeters.
• Temporal Scale:
  • Field Sampling Campaigns: July 2019, March 2020, June 2020, October 2020, October 2022, June 2024.
  • Remote Sensing Imagery Acquisition: Synchronous with sampling dates (e.g., 2019-07-21, 2020-06-03).
  • Soil Salinity Retrieval and Mapping: Annual maps from 2019 to 2025.
  • Independent Test Data (NCSS): Records from the most recent five years (prior to 2021).

Methodology and Data

• Models used:
  • Proposed: Spectral and Spatial Soil Net (SS-SoilNet) - a multimodal deep learning framework comprising:
    • Multi-scale One-Dimensional Convolution (MODC) module for spectral sequence extraction.
    • Index-ResNet Extractor with dual branches for vegetation-covered and bare-soil conditions.
    • Convolutional Block Attention Module (CBAMBlock) for spatial feature extraction.
    • Multi-head Cross-Attention (MCA) mechanism for adaptive spectral-spatial feature fusion.
  • Baselines: Random Forest (RF), Support Vector Regression (SVR).
  • Ablation Models: Index model (only index features), Spec-Index model (index + spectral features), SS-Stack model (index + spectral + spatial features without cross-attention).
• Data sources:
  • Remote Sensing Imagery: Sentinel-2 satellite series (10 selected bands: B2, B3, B4, B5, B6, B7, B8, B8A, B11, B12).
  • Digital Elevation Model (DEM): Copernicus Digital Elevation Model (GLO-30).
  • Derived Features:
    • 19 Salinity Indices and 16 Vegetation Indices (from Sentinel-2).
    • Water Environment Indices: Land Surface Water Index (LSWI), Normalized Difference Water Index (NDWI).
    • Crop Phenological Metrics: Start of Season (SOS), Green-up Speed (GUS), Start Date of Peak Stage (SDPS).
    • Topographic Metrics: Slope, Profile Curvature, Plan Curvature (derived from DEM).
  • Field Observations: 806 surface soil samples (0–20 cm layer) from the Yellow River Delta (salinity range: 0.097 to 39.95 g kg−1).
  • Independent Test Data: 43 valid samples from the NCSS Soil Characterization Database (U.S. coastal zone, 0–20 cm layer).

Main Results

• SS-SoilNet demonstrated improved and more stable predictive performance compared to conventional machine learning models (RF, SVR) and ablation models.
• On the validation dataset, SS-SoilNet achieved an RMSE of 3.63 g kg−1, an MAE of 1.73 g kg−1, and an R² of 0.68.
• The integration of multimodal deep learning with multi-source covariates (hydrological conditions, crop vigor, topographic factors) significantly enhanced retrieval robustness.
• Interpretability analysis revealed strong coupling effects among soil salinity, crop growth, and terrain.
• Spectrally, the red-edge (705–750 nanometers) and shortwave-infrared (1580–2350 nanometers) bands provided the highest contributions to salinity retrieval.
• Spatially, DEM, LSWI, and GUS exhibited consistently high attention weights, indicating their key roles in regulating soil salinity patterns. Topography was identified as a crucial structural driver for spatial continuity.
• Multi-year mapping (2019–2025) showed that low-salinity conditions dominated, with elevated salinity concentrated in the northern coastal reclaimed zone and along irrigation-drainage canals.
• Soil salinity dynamics exhibited a "polarization" pattern: stable low-salinity areas, fluctuating mildly/moderately salinized areas, and persistent highly salinized areas that rarely revert to lower-salinity states.

Contributions

• Developed SS-SoilNet, a novel multimodal deep learning framework that synergistically integrates remote sensing spectral sequences, indices, and environmental covariates for accurate cropland soil salinity retrieval.
• Introduced a multi-head cross-attention mechanism that leverages spatial features as queries to adaptively guide spectral and vegetation index representations, enhancing feature fusion and capturing complex nonlinear interactions.
• Achieved superior and more stable retrieval accuracy and generalization performance compared to conventional machine learning and simpler deep learning models, particularly under heterogeneous land-cover conditions.
• Provided enhanced interpretability through SHAP-based attribution and attention weight analysis, elucidating the specific contributions of spectral bands and environmental factors (topography, hydrology, crop growth) and their coupled mechanisms in shaping soil salinization.
• Demonstrated the framework's capability for multi-year (2019–2025) spatiotemporal mapping of soil salinity in coastal croplands, identifying persistent salinization hotspots and dynamic patterns crucial for monitoring and management.

Funding

• National Key R&D Program of China (Grant Nos. 2023YFD1900100 & 2023YFD1900300).

Citation

@article{Zhang2026Cropland,
  author = {Zhang, Junyan and Huang, Chong and Li, He and Liu, Qingsheng and Lu, Miao},
  title = {Cropland soil salinity retrieval using a spectral-spatial cross-attention deep learning framework with environmental interpretability},
  journal = {Computers and Electronics in Agriculture},
  year = {2026},
  doi = {10.1016/j.compag.2026.111748},
  url = {https://doi.org/10.1016/j.compag.2026.111748}
}