Synergetic use of multi-temporal Sentinel-1, Sentinel-2, NDVI, and topographic factors for estimating soil organic carbon

Abstract

Soil organic carbon (SOC) is a reliable indicator of soil productivity and land management. Despite the widespread use of remote sensing data in estimating SOC variation, the predictive power of single-date and multi-temporal as well as the individual and combined effects of radar and optical images have rarely been investigated. This study created six two-month interval composite layers including VH and VV polarizations of Sentinel-1 (S1), spectral bands of Sentinel-2 (S2), normalized difference vegetation index (NDVI), slope and altitude for the target year of 2019–2020 to estimate the SOC variation. Based on 80 soil samples (0–20 cm) taken in a field survey from northern Iran, random forest (RF) and support vector regression (SVR) were calibrated to estimate the SOC content. All models were fine-tuned using grid search and evaluated using spatial cross-validation. The results showed that multi-temporal data were more predictive than single-date data, and SVR outperformed the RF algorithm. SVR model using the multi-temporal S2 achieved the highest R2 and RMSE values of 57.59% and 0.94%, respectively. Feature selection has shown no benefit of adding S1 data and identified band-2 (green), band-3 (red), and band-10 (shortwave infrared) of S2 imagery as the most explanatory variables of SOC variation. The results of this study highlight the potential of freely available high-resolution S2 for mapping SOC variation using the SVR model. SVR is particularly suggested in small sample size studies, an issue which is dominant in soil studies.