Chaaou et al. (2025) Mapping soil salinity using machine learning and remote sensing data in semi-arid croplands

Identification

• Journal: Frontiers in Soil Science
• Year: 2025
• Date: 2025-11-26
• Authors: Abdelwahed Chaaou, Hamza Ait-Ichou, Said El Hachemy, Mohamed Chikhaoui, Mustapha Naïmi, Mohammed Hssaisoune, Mohammed El Hafyani, Yassine Ait Brahim, Lhoussaine Bouchaou
• DOI: 10.3389/fsoil.2025.1653400

Research Groups

• International Water Research Institute, Mohammed VI Polytechnic University (UM6P), Ben Guerir, Morocco
• Applied Geology and Geoenvironment Laboratory, Faculty of Sciences, Ibnou Zohr University, Agadir, Morocco
• Hassan II Institute of Agronomy and Veterinary Medicine, Rabat, Morocco
• Faculty of Applied Sciences, Ibn Zohr University, Ait Melloul, Morocco

Short Summary

This study evaluated machine learning algorithms combined with satellite-derived predictors for soil salinity mapping in semi-arid croplands of Morocco. The K-Nearest Neighbors (KNN) model achieved the highest accuracy (R² = 0.75; RMSE = 0.61 dS/m), demonstrating the reliability of this approach for monitoring soil salinity.

Objective

• To assess soil salinity in the Béni Amir sub-perimeter of the Tadla plain, Morocco, using machine learning algorithms and Landsat-8 OLI data.
• To compare the performance of four machine learning models (Random Forest, Support Vector Regressor, Multi-Layer Perceptron, and K-Nearest Neighbors) for soil salinity mapping and prediction with limited data.
• To generate soil salinity maps to support sustainable management of irrigated areas.

Study Configuration

• Spatial Scale: Béni Amir irrigated sub-perimeter of the Tadla Plain, central Morocco, covering 674 km². Topsoil samples were collected from 0–10 cm depth.
• Temporal Scale: Soil samples were collected between October 28 and 31, 2021. Landsat-8 OLI imagery was acquired on November 12, 2021.

Methodology and Data

• Models used: Random Forest (RF), K-Nearest Neighbors (KNN), Support Vector Regressor (SVR), Multi-Layer Perceptron (MLP).
• Data sources:
  • Ground Truth: 43 georeferenced topsoil samples (0–10 cm) were collected and analyzed for electrical conductivity (ECe), then resampled to 144 samples for model training and testing.
  • Remote Sensing: Landsat-8 OLI data (November 12, 2021, 0.11% cloud cover). Predictor variables included salinity indices (OLI-SI, SI, SI1), intensity indices (Int1, Int2), brightness index (BI), land degradation index (LDI), and reflectance values of spectral bands (B2-B7).
  • Preprocessing: Data standardization, Principal Component Analysis (PCA) to address multicollinearity (first five principal components retained), and data augmentation (bootstrapping with noise) for the limited dataset.

Main Results

• The electrical conductivity (ECe) values in the study area ranged from 0.84 to 10.28 dS/m, with a standard deviation of 2.29 dS/m, indicating substantial salinity variability.
• Individual predictors showed moderate correlation with ECe (R = 0.34-0.72).
• Among the tested machine learning models, K-Nearest Neighbors (KNN) achieved the highest predictive accuracy (mean coefficient of determination (R²) = 0.75 [95% CI: 0.73–0.77]; Root Mean Square Error (RMSE) = 0.61 dS/m).
• Support Vector Regressor (SVR) exhibited competitive performance (mean R² = 0.59; RMSE = 0.62 dS/m), while Random Forest (RF) and Multi-Layer Perceptron (MLP) performed less effectively (RF: mean R² = 0.54, RMSE = 0.75 dS/m; MLP: mean R² = 0.45, RMSE = 0.90 dS/m).
• The generated salinity maps revealed a consistent southwestward increase in salinity, following the regional hydraulic flow.
• KNN classified 49% of the area as moderately saline, 22% as slightly saline, 20% as non-saline, 8.4% as strongly saline, and 0.6% as extremely saline.

Contributions

• Demonstrated the effectiveness of integrating multiple satellite-derived variables (spectral bands and indices) with machine learning algorithms for soil salinity assessment in the Béni Amir sub-perimeter, a region with limited prior ML-based mapping.
• Provided a comparative evaluation of four machine learning models, identifying K-Nearest Neighbors (KNN) as the most accurate and robust for soil salinity prediction in this semi-arid cropland context, offering insights that contrast with some literature favoring tree-based models.
• Generated reliable and spatially coherent soil salinity maps that offer valuable support for targeted sustainable land management and irrigation planning in irrigated agroecosystems.
• Applied data augmentation (bootstrapping with noise) and Principal Component Analysis to enhance model robustness and mitigate issues of multicollinearity and limited field data.

Funding

The Moroccan Ministry of Higher Education, Scientific Research and Innovation, the OCP Foundation, the Mohammed VI Polytechnic University (UM6P), and the CNRST supported this work through the APRD research program (GEANTech).

Citation

@article{Chaaou2025Mapping,
  author = {Chaaou, Abdelwahed and Ait-Ichou, Hamza and Hachemy, Said El and Chikhaoui, Mohamed and Naïmi, Mustapha and Hssaisoune, Mohammed and Hafyani, Mohammed El and Brahim, Yassine Ait and Bouchaou, Lhoussaine},
  title = {Mapping soil salinity using machine learning and remote sensing data in semi-arid croplands},
  journal = {Frontiers in Soil Science},
  year = {2025},
  doi = {10.3389/fsoil.2025.1653400},
  url = {https://doi.org/10.3389/fsoil.2025.1653400}
}