Mapping of cropland humus content of the Bryansk region using machine learning methods

Abstract

The FAO methodology within the Global Soil Nutrient and Nutrient Budget Maps (GSNmap) project was tested for the first time for mapping humus content with a spatial resolution of 250 meters per pixel in soils of the Russian Federation at the regional scale, using the Bryansk Region as an example. The map was created in the R soft ware environment using data from Agrochemical Service and remote sensing, global databases and soil maps. The centroids of the sites from which the composite samples were taken by Agrochemical Service were selected as sampling points. The set of predictors available under the FAO project was expanded by additional data, including soil maps and maps of soil-forming rocks. The importance of the predictors was assessed using the Boruta algorithm, which is usually used as an initial stage for a random forest. The model was created using the caret package with the quantile regression forest method. The modeling efficiency coefficient (MEC) was 55%, the coefficient of determination (R2) was 0.57. The map reflects current information that can be used to monitor the dynamics of organic matter content in the soil and assess the state of the arable soils in the Bryansk region.