Sentinel-2 image capacities to predict common topsoil properties of temperate and Mediterranean agroecosystems

Abstract

To be fully operational for facilitating decisions made at any spatial level, models and indicators of soil ecosystem functions require the use of precise spatially referenced soil information as inputs. This study aimed at exploring the capacity for Sentinel-2A (S2A) multispectral satellite images to predict several topsoil properties in two contrasted pedoclimatic environments: a temperate region marked by intensive annual crop cultivation patterns and soils derived from loess or colluvium and/or marine limestone or chalk (Versailles Plain, 221 km2); and a Mediterranean region marked by vineyard cultivation and soils derived from lacustrine limestone, calcareous sandstones, colluvium, or alluvial deposits (Peyne catchment, 48 km2). Prediction models of soil properties based on partial least squares regressions (PLSR) were built from S2A spectra of 72 and 143 sampling locations across the Versailles Plain and Peyne catchment, respectively. Eight soil surface properties were investigated in both regions: pH, cation exchange capacity (CEC), texture fractions (Clay, Silt, Sand), Iron, Calcium Carbonate (CaCO3) and Soil Organic Carbon (SOC) content. Predictive abilities were studied according to the root mean square error of cross-validation (RMSECV) tests, cross-validated coefficient of determination (R2cv) and ratio of performance to deviation (RPD). Intermediate prediction performance outcomes (R2cv and RPD greater than or equal to 0.5 and 1.4, respectively) were obtained for 4 topsoil properties found across the Versailles Plain (SOC, pH, CaCO3 and CEC), and near-intermediate performance outcomes (0.5 > R2cv > 0.39, 1.4 > RPD > 1.3) were yielded for 3 topsoil properties (Clay, Iron, and CEC) found across the Peyne catchment and for 1 property (Clay) found across the Versailles Plain. The study results show what can be expected from Sentinel-2 images in terms of predictive capacities at the regional scale. The spatial structure of the estimated soil properties for bare soils pixels is highlighted, promising further improvements made to spatial prediction models for these properties based on the use of Digital Soil Mapping (DSM) techniques.