Moorthi et al. (2025) Comparative assessment of spectral covariates from Sentinel 1 A, Sentinel 2 A, Landsat 8, and PRISMA for digital soil mapping of infiltration rate and textural classes

Identification

• Journal: Environmental Monitoring and Assessment
• Year: 2025
• Date: 2025-12-26
• Authors: Nivas Raj Moorthi, Kumaraperumal Ramalingam, Pazhanivelan Sellaperumal, D Muthumanickam, K Sivasubramanian, Ragunath Kaliaperumal, Prabu Chidambaram
• DOI: 10.1007/s10661-025-14921-7

Research Groups

• Department of Remote Sensing and GIS, Tamil Nadu Agricultural University, Coimbatore, Tamil Nadu, India
• Centre for Water and Geospatial Studies, Tamil Nadu Agricultural University, Coimbatore, Tamil Nadu, India
• Department of Environmental Sciences, Tamil Nadu Agricultural University, Coimbatore, Tamil Nadu, India

Short Summary

This study compared the effectiveness of spectral covariates from Sentinel 1A, Sentinel 2A, Landsat 8, and PRISMA satellites for digital soil mapping of infiltration rate and textural classes in the Thiruparankundram block, India. It found that PRISMA data, particularly when combined with SCORPAN variables and optimized via embedded feature selection methods, and Landsat 8 data demonstrated the highest efficiency in predicting soil infiltration rates and textural classes, respectively.

Objective

• To compare the utility of different spectral modalities (Sentinel 1A, Sentinel 2A, Landsat 8, and PRISMA) for digital soil mapping of soil infiltration rates and textural classes.
• To test if a modelling framework combining SCORPAN covariates with spectral variables yields significantly higher accuracy in predicting infiltration rate and textural classes (H1: Integration efficiency).
• To evaluate if embedded feature selection methods (Boruta, VSURF) achieve similar or superior predictive performance using fewer bands within PRISMA datasets compared to wrapper methods (RFE, GA) or the full band set (H2: Feature selection efficiency).
• To assess if different validation strategies (data partitioning vs. repeated k-fold cross-validation) yield different prediction results, with cross-validation showing superior performance (H3: Validation schema efficiency).

Study Configuration

• Spatial Scale: Thiruparankundram block of Madurai district, India (302.52 km²). Soil observations were for the topsoil layer (0–30 cm depth). Satellite data were resampled to 10 m (Sentinel 1A, Sentinel 2A) and 30 m (Landsat 8, PRISMA) spatial resolutions.
• Temporal Scale: Mean annual climatic information from 1970–2000. Satellite imagery (Sentinel, Landsat) was acquired as a 3-month composite from March to May 2024. The latest soil profile survey for the study area was conducted in November 2024.

Methodology and Data

• Models used:
  • Random Forest (RF) algorithm for digital soil mapping.
  • Supervised band selection techniques for PRISMA hyperspectral data: Recursive Feature Elimination (RFE), Boruta, Variable Selection Using Random Forest (VSURF), and Genetic Algorithm (GA).
  • Validation techniques: Stratified data partitioning (70% training, 30% testing) and repeated k-fold cross-validation (10 folds, 10 repetitions).
  • Variable importance measure: Permutation Feature Importance (PFI).
• Data sources:
  • Satellite Imagery: Sentinel 1A (Synthetic Aperture Radar), Sentinel 2A (Optical), Landsat 8 (Optical), and PRISMA (Hyperspectral). Data obtained from Google Earth Engine (GEE) and the Italian Space Agency (ASI) portal.
  • Legacy Soil Data: 224 topsoil observations (0–30 cm) for infiltration rate and textural classes, extracted from the National Resource Information System (NRIS) soil database (1:50,000 scale) and Tamil Nadu Agricultural University (TNAU) / Soil Survey and Land Use Organization (SS & LUO) soil profile information.
  • Environmental Covariates (SCORPAN factors):
    • Climate: WorldClim 2.1 (mean annual climatic information, 1970–2000, 30 arc-second resolution).
    • Organisms/Vegetation: Land Use and Land Cover (LULC) from National Remote Sensing Centre (NRSC, 1:50,000 scale), spectral derivatives (e.g., NDVI, NDWI, SAVI, EVI).
    • Parent Material: Geology from NRSC (1:50,000 scale).
    • Relief/Topography: SRTM DEM-derived terrain attributes (e.g., Elevation, Slope, Multiresolution Ridge Top Flatness (MRRTF), Multiresolution Valley Bottom Flatness (MRVBF)) using SAGA GIS 9.2.0. Physiography and Geomorphology from NRSC (1:50,000 scale).
    • Spatial: Land Surface Temperature (LST) computed from GEE (for Landsat subset).

Main Results

• Integration Efficiency (H1): Modelling performance significantly increased when spectral variables were integrated with other SCORPAN covariates. PRISMA subsets selected through embedded methods (Boruta and VSURF) and Landsat 8 spectral datasets showed the highest efficiency for predicting soil infiltration rate and textural classes, respectively, especially when combined with SCORPAN variables.
  • For infiltration rate (continuous), R² values ranged from negligible to 0.30 for data partitioning (Boruta PRISMA + DSM framework) and from 0.11 to 0.56 for cross-validation (VSURF PRISMA + DSM framework). RMSE was below 25% (cross-validation) and 35% (data partitioning) of the observed range.
  • For textural classes (categorical), overall accuracy (OA) ranged from 0.26 to 0.72 (data partitioning) and 0.54 to 0.82 (cross-validation). Kappa values ranged from 0.06 to 0.64 (data partitioning) and 0.43 to 0.77 (cross-validation).
• Feature Selection Efficiency (H2): Embedded methods (Boruta and VSURF) demonstrated higher efficiency in selecting optimal PRISMA spectral bands for both continuous and categorical soil predictions, achieving comparable or superior performance with fewer bands than wrapper methods (RFE, GA) or the full band set.
• Validation Schema Efficiency (H3): Cross-validation generally yielded superior performance in terms of validation metrics compared to data partitioning. However, pixel-by-pixel spatial predictions from both strategies were largely similar, with the exception of Landsat 8-derived results.
• Variable Importance:
  • For infiltration rate, Physiography, Geomorphology, Land Use and Land Cover (LULC), Multiresolution Ridge Top Flatness (MRRTF), Multiresolution Valley Bottom Flatness (MRVBF), Elevation, Slope Height, Wind Speed, and Solar Radiation were the most contributing variables.
  • For textural classes, LULC, Geomorphology, Physiography, Elevation, MRRTF, MRVBF, Precipitation, Lithology, and Carbonate Difference Ratio were influential.
  • Specific PRISMA bands (1142 nm for infiltration rate and 956 nm for textural classes) showed high importance within the full PRISMA dataset.
• Infiltration Rate (SI conversion): The observed infiltration rate ranged from 1.2 to 98.6 mm h⁻¹, with a mean of 38.0 mm h⁻¹ and a standard deviation of 36.7 mm h⁻¹.

Contributions

• First comparative assessment of multiple satellite spectral covariates (Sentinel 1A, Sentinel 2A, Landsat 8, PRISMA) for digital soil mapping of infiltration rate and textural classes in the Thiruparankundram block, India.
• Comprehensive benchmarking of different spectral modalities (Optical, SAR, Hyperspectral) both independently and in combination with SCORPAN environmental variables within a digital soil mapping framework.
• Evaluation of the efficacy of embedded (Boruta, VSURF) versus wrapper (RFE, GA) feature selection methods for hyperspectral data in predicting continuous and categorical soil properties.
• Detailed comparison of data partitioning and repeated k-fold cross-validation strategies, including a pixel-by-pixel analysis of spatial predictions, to assess validation schema efficiency.
• Quantification of the contribution of various environmental covariates to the prediction of soil infiltration rate and textural classes using Permutation Feature Importance (PFI).

Funding

The authors did not receive support from any organization for the submitted work.

Citation

@article{Moorthi2025Comparative,
  author = {Moorthi, Nivas Raj and Ramalingam, Kumaraperumal and Sellaperumal, Pazhanivelan and Muthumanickam, D and Sivasubramanian, K and Kaliaperumal, Ragunath and Chidambaram, Prabu},
  title = {Comparative assessment of spectral covariates from Sentinel 1 A, Sentinel 2 A, Landsat 8, and PRISMA for digital soil mapping of infiltration rate and textural classes},
  journal = {Environmental Monitoring and Assessment},
  year = {2025},
  doi = {10.1007/s10661-025-14921-7},
  url = {https://doi.org/10.1007/s10661-025-14921-7}
}