Shirali et al. (2026) Assessment and mapping of soil moisture at field capacity using machine learning methods in the Central Chernozem Belt of Russia
Identification
- Journal: Scientific Reports
- Year: 2026
- Date: 2026-11-20
- Authors: Masoud Shirali, Javad Hatamiafkoueieh, Anton Poddubsky, Naeim Rezaeian
- DOI: 10.1038/s41598-026-80490-5
Research Groups
- Department of Mechanics and Control Processes, Academy of Engineering, RUDN University, Moscow, Russian Federation
- RUDN University, Moscow, Russian Federation
Short Summary
This study developed a digital soil mapping framework using machine learning to estimate and map soil moisture at field capacity (SMFC) at 0–0.3 meters depth across the Central Chernozem Belt of Russia. The Random Forest model achieved the highest accuracy, revealing a heterogeneous SMFC distribution primarily influenced by vegetation indices and topographic wetness.
Objective
- To develop and evaluate a digital soil mapping framework using machine learning to predict and map soil moisture at field capacity (SMFC) at 0–0.3 meters depth and a matric potential of −33 kilopascals in the Central Chernozem Belt of Russia, while also quantifying prediction uncertainty and identifying key environmental predictors.
Study Configuration
- Spatial Scale: Central Chernozem Belt of Russia (49°30′–55°10′ N and 34°20′–43°10′ E), covering oblasts of Belgorod, Kursk, Lipetsk, Orel, Ryazan, Tambov, Tula, and Voronezh. Soil depth of 0–0.3 meters. Spatial resolution of 30 meters for environmental covariates.
- Temporal Scale: Soil sampling and satellite data acquisition conducted over a three-year period (2023–2025) during peak growing seasons (June–August).
Methodology and Data
- Models used:
- Pedotransfer functions (PTFs) within RETC software (neural network module) to derive van Genuchten water-retention parameters.
- van Genuchten analytical model to calculate SMFC at −33 kilopascals.
- Machine learning algorithms: Artificial Neural Network (ANN), Support Vector Machine (SVM), and Random Forest (RF).
- Data sources:
- Soil Samples: 2157 geo-referenced soil samples (0–0.3 meters depth) collected between 2023 and 2025. Laboratory analyses included bulk density, sand, silt, clay, pH, electrical conductivity, and soil organic carbon.
- Remote Sensing: Landsat 8 Operational Land Imager (OLI) multispectral imagery (30 meters resolution), processed to derive vegetation indices (NDVI, EVI, SAVI), moisture indices (NDWI, BSI), and a clay minerals index (CMI).
- Terrain: 30 meters resolution Digital Elevation Model (DEM) and derived attributes (Elevation, Slope, Aspect, Profile curvature, Plan curvature, Topographic Wetness Index (TWI), Stream Power Index (SPI), Terrain Ruggedness Index (TRI), Topographic Position Index (TPI), Relative Slope Position (RSP)).
- Climatic: Mean annual temperature (MAT) and Mean annual precipitation (MAP) from regional meteorological grids (30 arc-seconds resolution).
- Geological: Geological maps (proxied by spectral bands).
- Spatial: Projected easting (X) and northing (Y) coordinates.
Main Results
- The Random Forest (RF) model consistently outperformed ANN and SVM, achieving the highest accuracy in both calibration (R² = 0.81, RMSE = 1.93, PRD = 3.17) and independent validation (R² = 0.89, RMSE = 1.45, PRD = 3.25).
- The spatial distribution of SMFC across the Central Chernozem Belt was highly heterogeneous, ranging from 13.68 % to 43.34 % (volumetric water content), exhibiting a complex, patchy pattern rather than a simple regional gradient.
- Variable importance analysis identified Normalized Difference Vegetation Index (NDVI), Topographic Wetness Index (TWI), Soil Adjusted Vegetation Index (SAVI), and Normalized Difference Water Index (NDWI) as the most influential predictors, highlighting the strong association of vegetation condition and topographic moisture redistribution with SMFC.
- Uncertainty assessment, expressed through lower and upper prediction limit maps, revealed that predictive reliability varied spatially, with narrower prediction ranges indicating higher confidence in SMFC estimates.
Contributions
- Developed and evaluated a robust digital soil mapping framework, augmented by pedotransfer functions and explainable machine learning, for regional-scale assessment of soil moisture at field capacity (SMFC) in a large, heterogeneous agricultural region.
- Systematically compared the performance of three machine learning algorithms (ANN, SVM, RF) for SMFC prediction, demonstrating the superior accuracy and stability of the Random Forest model.
- Identified key environmental covariates, particularly remote sensing-derived vegetation and moisture indices, and topographic attributes, as primary drivers of SMFC variability in the Central Chernozem Belt.
- Generated comprehensive SMFC maps that include spatial uncertainty information (lower and upper prediction limits), enhancing the practical utility of the maps for agricultural water management and decision-making.
Funding
- Ministry of Digital Development of the Russian Federation, grant number IGK 000000Ц330325P2J0002.
Citation
@article{Shirali2026Assessment,
author = {Shirali, Masoud and Hatamiafkoueieh, Javad and Poddubsky, Anton and Rezaeian, Naeim},
title = {Assessment and mapping of soil moisture at field capacity using machine learning methods in the Central Chernozem Belt of Russia},
journal = {Scientific Reports},
year = {2026},
doi = {10.1038/s41598-026-80490-5},
url = {https://doi.org/10.1038/s41598-026-80490-5}
}
Original Source: https://doi.org/10.1038/s41598-026-80490-5