Alahmad et al. (2025) Enhancing Agricultural Sustainability using AI-Driven Soil Moisture Modeling: A Soil-Type and Depth Approach with SHAP Interpretability

Identification

• Journal: Acta Agronomica Óváriensis
• Year: 2025
• Date: 2025-12-22
• Authors: Tarek Alahmad, Miklós Neményi, Anikó Év Nyéki
• DOI: 10.17108/actagrovar.2025.66.2.5

Research Groups

• Albert Kázmér Faculty of Agricultural and Food Sciences of Széchenyi István University, Department of Bioengineering and Precision Technology, Mosonmagyaróvár, Hungary

Short Summary

This study developed and evaluated depth- and soil-specific Random Forest Regression models for predicting soil moisture content (SMC) in loam and silt loam soils. It found that integrating meteorological data with vegetation indices significantly enhances prediction accuracy, with SHAP analysis revealing soil-dependent feature importance crucial for optimizing irrigation strategies and agricultural sustainability.

Objective

• To develop depth- and soil-specific Random Forest Regression (RFR) models for predicting SMC in loam and silt loam soils at five depths (5, 20, 40, 60, 80 cm).
• To evaluate model performance in two input scenarios using different predictor combinations (vegetation indices only vs. vegetation indices + meteorological data).
• To identify the main factors influencing SMC prediction using SHAP-based feature importance analysis.

Study Configuration

• Spatial Scale: A 23 hectare rainfed maize field in Mosonmagyaróvár, Hungary (47°54'11.8"N 17°15'08.9"E). Two soil types (loam and silt loam) were selected, with soil samples collected from six locations (three per soil type) at five depths (5, 20, 40, 60, 80 cm).
• Temporal Scale: Data collected across two growing seasons (2023-2024) on 23 distinct dates, representing most maize growth stages.

Methodology and Data

• Models used:
  • Random Forest Regression (RFR) for soil moisture content prediction.
  • SHapley Additive exPlanations (SHAP) using TreeExplainer for model interpretability and feature importance analysis.
  • GridSearchCV for hyperparameter optimization and 3-fold cross-validation.
  • MinMaxScaler for data normalization.
• Data sources:
  • Soil Moisture Content (SMC): Measured using the gravimetric method from collected soil samples.
  • Meteorological data: Collected by an IoT-based meteorological sensor at 10 to 15 minute intervals, including temperature (°C), precipitation (mm), wind speed (km/h), humidity (%), and solar radiation (W/m²).
  • Vegetation indices: Normalized Difference Vegetation Index (NDVI) and Normalized Difference Moisture Index (NDMI) derived from Sentinel-2A satellite imagery (10 m spatial resolution), utilizing Near-Infrared (NIR), Red, Red Edge 4, and Shortwave Infrared (SWIR1) bands.

Main Results

• Enhanced Model Accuracy with Meteorological Data: Integrating meteorological data (Scenario 2) significantly improved RFR model performance across both soil types and all depths compared to using only vegetation indices (Scenario 1).
  • Loam Soil: R² values increased from 0.65, 0.61, 0.82, 0.81, 0.85 (Scenario 1) to 0.95, 0.83, 0.87, 0.87, 0.92 (Scenario 2) at 5, 20, 40, 60, and 80 cm depths, respectively. Corresponding RMSE values decreased, e.g., from 2.45 % to 0.94 % at 5 cm depth.
  • Silt Loam Soil: R² values increased from 0.88, 0.94, 0.82, 0.91, 0.88 (Scenario 1) to 0.97, 0.96, 0.94, 0.93, 0.91 (Scenario 2) at 5, 20, 40, 60, and 80 cm depths, respectively. Corresponding RMSE values decreased, e.g., from 1.40 % to 0.70 % at 5 cm depth.
• Soil-Specific Feature Importance (SHAP Analysis):
  • Loam Soil (Scenario 2): Precipitation, humidity, and NDVI were identified as the most influential features. Precipitation had SHAP values of 0.12, 0.08, and 0.05 at 5, 20, and 80 cm depths, respectively. Humidity had SHAP values of 0.05, 0.05, and 0.04 at the same depths.
  • Silt Loam Soil (Scenario 2): Solar radiation and NDVI were most influential, particularly at deeper layers. Solar radiation had SHAP values of 0.1 and 0.08 at 60 and 80 cm depths, respectively. NDVI showed high impact at 20, 60, and 80 cm with values of 0.12, 0.07, and 0.06. Humidity had a higher impact at 5 cm depth (0.08).
• Upper soil layers (5-40 cm) showed higher sensitivity to meteorological features, while deeper layers (60-80 cm) exhibited reduced meteorological dependency.

Contributions

• Introduces a novel approach of depth- and soil-specific Random Forest Regression modeling for soil moisture content prediction, coupled with SHAP-based interpretability, addressing limitations of prior studies that focused on single-layer or generalized predictions.
• Quantitatively demonstrates the critical importance of integrating meteorological data with vegetation indices for significantly enhancing the accuracy of SMC prediction across different soil types and depths.
• Provides actionable insights for precision agriculture by identifying the most influential factors affecting SMC for specific soil types and depths, enabling tailored irrigation strategies to reduce water waste and improve agricultural sustainability.

Funding

• EKÖP-24-3-I-Sze-107 University Research Fellowship Program of the Ministry for Culture and Innovation.
• “Precision Bioengineering Research Group” supported by the “Széchenyi István University Foundation”.
• János Bolyai Research Scholarship (Bo/00578/24) of the Hungarian Academy of Sciences.

Citation

@article{Alahmad2025Enhancing,
  author = {Alahmad, Tarek and Neményi, Miklós and Nyéki, Anikó Év},
  title = {Enhancing Agricultural Sustainability using AI-Driven Soil Moisture Modeling: A Soil-Type and Depth Approach with SHAP Interpretability},
  journal = {Acta Agronomica Óváriensis},
  year = {2025},
  doi = {10.17108/actagrovar.2025.66.2.5},
  url = {https://doi.org/10.17108/actagrovar.2025.66.2.5}
}