Improving the prediction of soil organic carbon content using field-acquired hyperspectral data by accounting for soil moisture and surface roughness

International audience

英语. Soil surface conditions such as moisture, roughness, and vegetation complicate accurate Soil Organic Carbon (SOC) prediction by altering spectral reflectance. Most studies consider these factors separately and under controlled conditions. Soil roughness has rarely been included [1,2], and typically not alongside soil moisture, which has mostly been studied in laboratory settings [3]. Common methods to reduce moisture effects on spectra, such as external parameter orthogonalization (EPO) and direct standardization (DS), rely heavily on lab-based datasets [3]. To address this, we assessed the influence of soil moisture and surface roughness as co-variables in models predicting SOC content from reflectance spectra of bare Luvisols near Versailles, France. Spectral data were collected under naturallight at 76 points, along with volumetric soil moisture (θ) and 7 roughness indicators from photogrammetry [4]. SOC was predicted using Partial Least Squares Regression (PLSR) and Random Forest (RF), with 4-fold cross-validation repeated 10 times. Six wavelength-selection (WS) strategies were tested: two from satellite simulations (EnMAP, Sentinel-2), two from model variable importance (PLSR, RF), one expert-based, and one using all wavelengths. Moisture and roughness were added individually. In-field spectra enabled reasonably accurate predictions, with RF outperforming PLSR (SOC RMSE: 1.6–1.8 g.kg⁻¹). WS methods improved accuracy only when co-variables were added.Moisture had little effect, while roughness improved prediction quality in most cases, especially shadow percentage for PLSR and the semivariogram sill parameter for RF. These results highlight the benefit of including surface roughness to improve large-scale SOC prediction from remote sensing.