Machine Learning Based Prediction of Munsell Soil Color Components Using Multi-Sensor Satellite Data

Basic soil maps are essential for describing soil and land characteristics, supporting effective land-use planning, and preventing improper land use. However, traditional soil survey and mapping are time-consuming and labor-intensive. In this context, identifying soil color variation is critical for defining potential soil boundaries. Recent advances in remote sensing and machine learning have enabled the prediction of Munsell soil color components (hue, value, and chroma) and the production of outputs that support detailed soil surveys. This study evaluated the prediction of Munsell soil color components using modern approaches and compared the results with digital soil maps produced at a 1:5,000 scale by traditional methods. Support vector machine (SVM) and random forest (RF) models were trained using color-related features derived from Sentinel-1, Sentinel-2, and Landsat-8 satellite data, together with soil-series level digital soil map information. To best represent bare soil conditions, imagery from periods with the lowest NDVI and highest bare soil index values was selected. SVM achieved the highest performance for hue, value, and chroma prediction, with overall accuracy values of 81.6%, 79.1%, and 85.4% and κ values of 0.57, 0.57, and 0.40, respectively, whereas RF performed better in predicting the Munsell color code (OA = 69.4%, κ = 0.52).