Soil Texture Classification Using Sentinel-2 and Landsat-9

An end-to-end remote sensing and machine learning workflow for mapping clay- and sand-dominated soils at two depth intervals in Semnan Province, Iran.

Project Overview

This project investigates the use of freely available Sentinel-2 and Landsat-9 imagery for soil texture classification across a 143.9 km² arid region in Semnan Province, Iran.

Spectral bands and indices were combined with 3,000 SoilGrids samples to classify clay and sand content at two soil depths:

• 0–5 cm (surface soil)
• 5–15 cm (subsurface soil)

Random Forest (RF) and Support Vector Machine (SVM) classifiers were trained and evaluated using five-fold cross-validation. Random Forest consistently outperformed SVM across both sensors, soil components, and depth intervals.

Study Data

Dataset	Description
Sentinel-2	Cloud-free Level-2A image acquired on 28 May 2024
Landsat-9	Surface reflectance image acquired on 26 May 2024
SoilGrids	3,000 soil samples containing clay and sand information
Study area	143.9 km² in Semnan Province, Iran
Soil depths	0–5 cm and 5–15 cm

Methodology

The project was implemented through the following workflow:

1. Satellite image acquisition and preprocessing
2. Cloud and cloud-shadow masking
3. Study-area clipping
4. Spectral-band extraction
5. Vegetation and soil index calculation
6. SoilGrids sample preparation
7. Satellite-value extraction at sample locations
8. Correlation and regression analysis
9. Class balancing using SMOTE
10. Random Forest and SVM model training
11. Five-fold cross-validation
12. Accuracy and Cohen’s Kappa evaluation
13. Feature-importance analysis
14. Production of classified soil texture maps

Spectral Features

The analysis included satellite spectral bands and several vegetation and soil-related indices:

• NDVI
• SAVI
• EVI
• MCARI
• IRECI
• MTCI
• S2REP

Machine Learning Models

Random Forest

Random Forest was implemented with 100 decision trees and the Gini impurity criterion. It produced the strongest and most stable results across all configurations.

Support Vector Machine

SVM was implemented using the Radial Basis Function kernel. Its performance was generally lower than Random Forest, particularly for clay classification.

Model Performance

Sentinel-2 Results

Soil component	Depth	Model	Accuracy	Kappa
Clay	0–5 cm	RF	83.51%	0.670
Clay	0–5 cm	SVM	54.46%	0.089
Clay	5–15 cm	RF	82.84%	0.657
Clay	5–15 cm	SVM	53.22%	0.065
Sand	0–5 cm	RF	87.44%	0.749
Sand	0–5 cm	SVM	61.48%	0.230
Sand	5–15 cm	RF	87.45%	0.749
Sand	5–15 cm	SVM	59.90%	0.198

Landsat-9 Results

Soil component	Depth	Model	Accuracy	Kappa
Clay	0–5 cm	RF	83.12%	0.662
Clay	0–5 cm	SVM	53.97%	0.079
Clay	5–15 cm	RF	82.68%	0.654
Clay	5–15 cm	SVM	52.28%	0.046
Sand	0–5 cm	RF	89.49%	0.790
Sand	0–5 cm	SVM	72.67%	0.454
Sand	5–15 cm	RF	89.18%	0.784
Sand	5–15 cm	SVM	69.51%	0.390

Performance Comparison

Surface Soil (0–5 cm)

Subsurface Soil (5–15 cm)

Classified Soil Maps

Landsat-9

Sentinel-2

Key Findings

• Random Forest consistently outperformed SVM across all configurations.
• The best result was obtained using Landsat-9 and Random Forest for surface sand classification.
• The best-performing model achieved 89.49% accuracy and a Kappa coefficient of 0.790.
• Sentinel-2 and Random Forest achieved 87.45% accuracy for subsurface sand classification.
• Landsat-9 showed strong performance for sand mapping, supported by its SWIR bands and EVI.
• Sentinel-2 red-edge bands and IRECI contributed strongly to clay discrimination.
• Both sensors demonstrated potential for scalable soil texture mapping in arid and semi-arid environments.

Technologies

• Python
• Google Earth Engine
• Rasterio
• Pandas
• NumPy
• Scikit-learn
• Matplotlib
• Sentinel-2
• Landsat-9
• SoilGrids

Repository Structure

SoilTexture-Classification/
├── Sentinel-2/
│   ├── preprocessing and index-calculation scripts
│   ├── machine-learning scripts
│   ├── trained models
│   └── classified maps
├── Landsat-9/
│   ├── preprocessing and index-calculation scripts
│   ├── machine-learning scripts
│   ├── trained models
│   └── classified maps
├── feature_importance_plots/
├── Other/
│   └── SoilGrids_Points.csv
└── README.md

Project Context

Developed as part of the MSc in Geoinformatics Engineering at Politecnico di Milano.

• Authors: Ali Moeinkhah and Hafizullah Sarwary
• Supervisors: Prof. Mariagrazia Fugini and Prof. Giovanna Venuti
• Academic year: 2024–2025