Soil property maps with satellite images at multiple scales and its impact on management and classification

Abstract

Soil maps at appropriate scales can aid decision-making in agriculture and the environment. In this sense, remote sensing products have shown their power to investigate soil properties, but the assessment of its spatial information can be hampered by the presence of other objects than soil. To overcome this issue, soil scientists have been studying spatial patterns from multi-temporal satellite images aiming at at improving soil property maps. In this work, we applied the cubist algorithm to predict topsoil properties (clay, sand, organic matter and iron contents, and soil color components) in south-eastern Brazil using multi-temporal (Landsat8-OLI, and Sentinel2-MSI) and single-date images (PlanetScope, Landsat8-OLI, and Sentinel2-MSI). We aimed to evaluate the influence of satellite’s spatial, spectral and temporal resolutions on soil mapping. Predictive models were constructed with 120 soil samples and using four (vis-NIR) and six (vis-NIR-SWIR) spectral bands as predictors in a 10-fold cross-validation procedure. The multi-temporal image obtained from the Sentinel2-MSI satellite (with 10 m pixel size and six spectral bands), showed the best model performances in cross-validation (R2 between 0.48 and 0.78). The PlanetScope image, which has only four bands in the vis-NIR region and spatial resolution of 3 m did not improve model performances, although the R2 values were higher for soil color components (R2 > 0.5). Satellite images in different spatial, spectral and temporal resolution provides slightly different soil property maps which may promote different strategies regarding soil classification and management. However, satellite images should be used with caution, as they provide only surficial information about the soil variability and confirmation with field surveys is required.