Sentinel-2 indices for enhanced prediction of soil rock fragments cover in a Spanish vineyard

Soil rock fragments (SRF) strongly influence soil properties, nutrient contents and erosion. Their surface cover (SRF cover) affects soil reflectance which can impacts the accuracy of remotely sensed predictions of soil properties, yet it is rarely quantified. In this study, we assess the ability of Sentinel-2 (S2) indices to capture variability in SRF cover and their potential to enhance SRF cover predictions with S2 spectral models. SRF cover (%) was measured across three field campaigns et 60 points in an 82 ha Spanish vineyard trained on a trellis system. The point-count method applied to nadir photographs taken ~1 m above the soil was used. Six S2 indices time series were analysed using a hierarchical agglomerative clustering (HAC) then a principal component analysis (PCA) to identify the index best capturing SRF cover variability. Its relevance was assessed by comparing its values with the average SRF cover measured across the three campaigns. Partial least squares regression (PLSR) and random forest (RF) models were then developed using individuals and combined S2 dates, both with and without and NDVI threshold of 0.4, considering either S2 spectral bands or in combination with best S2 indices. The Non-Photosynthetic Vegetation Soil Separation Index (NSSI) best captured SRF cover variability, showing a negative correlation with SRF cover (R² = 0.41–0.60) and the strongest correlation for NDVI < 0.4 (R² = 0.48–0.91). Most models achieved moderate to good performance, with PLSR outperforming RF. Combining S2 dates improved model stability and performance for both PLSR (RPD = 1.93, RPIQ = 2.65) and RF (RPD = 1.59, RPIQ = 2.19). These results highlight the potential of Sentinel-2 data to predict SRF cover, and future work could explore integrating remote sensing with geophysical methods to further enhance predictions.