Mapping at 30 m Resolution of Soil Attributes at Multiple Depths in Midwest Brazil

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 agricultural intensification. Producing large-extent digital soil maps demands a huge volume of data and high computing capacity. This paper proposed mapping surface and subsurface key soil attributes with 30 m-resolution in a large area of Midwest Brazil. These soil maps at multiple depth increments will provide adequate information to guide land use throughout the region. The study area comprises about 851,000 km(2) in the Cerrado biome (savannah) in the Brazilian Midwest. We used soil data from 7908 sites of the Brazilian Soil Spectral Library and 231 of the Free Brazilian Repository for Open Soil Data. We selected nine key soil attributes for mapping and aggregated them into three depth intervals: 0-20, 20-60 and 60-100 cm. A total of 33 soil predictors were prepared using Google Earth Engine (GEE), such as climate and geologic features with 1 km-resolution, terrain and two new covariates with 30 m-resolution, based on satellite measurements of the topsoil reflectance and the seasonal variability in vegetation spectra. The scorpan model was adopted for mapping of soil variables using random forest regression (RF). We used the model-based optimization by tuning RF hyperparameters and calculated the scaled permutation importance of covariates in R software. Our results were promising, with a satisfactory model performance for physical and chemical attributes at all depth intervals. Elevation, climate and topsoil reflectance were the most important covariates in predicting sand, clay and silt. In general, for predicting soil chemical attributes, climatic variables, elevation and vegetation reflectance provided to be the most important of predictive components, while for organic matter it was a combination of climatic dynamics and reflectance bands from vegetation and topsoil. The multiple depth maps showed that soil attributes largely varied across the study area, from clayey to sandy, suggesting that less than 44% of the studied soils had good natural fertility. We concluded that key soil attributes from multiple depth increments can be mapped using Earth observations data and machine learning methods with good performance.