Abstract

Soil texture is a critical parameter influencing agricultural productivity, water retention, and nutrient availability. Accurate mapping of soil texture enhances precision agriculture and sustainable land management. This study explores the integration of Sentinel-1 (S-1) synthetic aperture radar (SAR), Sentinel-2 (S-2) multispectral imagery, and Light Detection and Ranging (LiDAR) data to predict soil texture across a 3000 km² semi-arid agricultural region in central Tunisia. Using machine learning algorithms, specifically Random Forest (RF) and Support Vector Machine (SVM), we fused multi-source remote sensing data to estimate clay, silt, and sand fractions. Field samples (n=150) with clay content ranging from 13% to 60% were used for training and validation. Results indicate that the fused dataset achieved a classification accuracy of 85% (RF) and 82% (SVM), with a root mean square error (RMSE) of 5.2% for clay content prediction. The integration of LiDAR-derived topographic features with Sentinel data significantly improved prediction accuracy compared to single-sensor approaches. This study highlights the potential of data fusion for high-resolution soil texture mapping, offering valuable insights for precision agriculture.