Soil mineralogical attributes estimated by color as accessed by proximal sensors and machine learning

Detailed mapping is essential for land use and management planning. The mappings require a robust database. Costs and time associated with obtaining the database are high and, therefore, it is not always possbile to obtain it. Soil color is a pedoindicator attribute that can be easily characterized. This study aimed to use soil color, based on the RGB (red-green-blue) system and obtained by diffuse reflectance spectroscopy (DRS) and mobile proximal sensor (MPS) to estimate mineralogical attributes using machine learning techniques for the Western Plateau of Sao Paulo. A total of 600 samples were collected throughout the study area. The samples were analyzed by DRS and then photographed. The color data were obtained by the RGB system after analysis in a computer program. The samples were subjected to laboratory analysis to quantify the contents of crystalline and noncrystalline Fe, hematite, goethite, kaolinite, and gibbsite. The database was subjected to the random forest machine learning algorithm and geostatistics. The use of random forest allowed estimating soil mineralogical attributes based on the RGB system by DRS and MPS. Detailed maps of mineralogical attributes could be constructed using the RGB system by the DRS and MPS techniques. The MPS technique can be used to characterize soil color, reducing the costs associated with analysis and the time required for data collection.