Pedometric mapping of key topsoil and subsoil attributes using proximal and remote sensing in midwest Brazil

Abstract

The Midwest region in Brazil has the largest and most recent agricultural frontier in the country, where there is no currently detailed soil information to support the land use intensification. Producing large-extent digital soil maps is resource intensive. We aimed to use pedometric techniques coupled with proximal and remote sensing data to produce digital maps with 30 m resolution of key soil attributes at topsoil and subsoil for 851,000 km2 of Midwest Brazil. For mapping key soil attributes we used multi-resolution covariates based on Earth observations: we produced composites of bare topsoil reflectance and potential natural vegetation reflectance using Landsat time series, which were coupled with terrain attributes, geologic and climate variables to capture short and long-range soil spatial patterns. We acquired soil data from observations at 0−20, 20−60 and 60−100 cm (rooting) depth intervals containing soil attributes, which are commonly used (as key criteria) for soil interpretations: clay, silt and sand, organic matter, pH, aluminum and base saturation. We also determined both the soil color in Munsell notation and the relative abundance of minerals in soil (hematite, goethite, kaolinite, gibbsite and 2:1 clay minerals) from laboratorial reflectance spectra (350−2500 nm). We fitted and validated optimal models for the spatial patterns of each soil attribute at topsoil and subsoil using Random Forest regression and 10-fold cross validation in R software. We identified the covariates that were most relevant to describe the soil spatial patterns in the study area. We mapped the spatial distribution of soil attributes at 30 m resolution for the 0−20, 20−60 and 60−100 cm depth intervals using the optimized models and Google Earth Engine. We made publicly available for download (GeoTIFF) at 250 m resolution the predicted soil maps of clay, silt and sand of the study area. We concluded that physical and chemical soil attributes, as well as soil color and mineralogy derived from spectra at multiple depth intervals, can be mapped using Earth observations data and machine learning methods with good performance.