Estimation of Farmland Soil Organic Carbon Stocks by Using Single-Year and Multi-Year Landsat Imagery: A Comparison Across Different Time Periods and Indicator Combinations

Accurate soil organic carbon (SOC) estimation is vital for analyzing the global carbon cycle. Currently the bare soil compositing approaches for multi-temporal images are widely used, however the optimized length of compositing period and influence of different indicators on SOC estimaiton for both bare soil and crop cover conditions is unknown. In this study, a time series of Landsat 8 Operational Land Imager multitemporal images was obtained from 2013–2018, with the aim of generating datasets that represent SOC changes across single dates, single years, and multiple years. Soil properties (S), terrain attributes (T), vegetation conditions (V), and farm management practices (F) were employed to predict the spatial distribution of SOC by using the random forest model for both bare soil and crop cover conditions. The results revealed that multi-temporal images from three years and longer produced accurate SOC predictions, with coefficients of determination (R2) and root mean squared errors (RMSEs) of 0.94-0.95 and 1.75-1.77 g kg-1, respectively. The four types of indicator combinations (S+T+V+F) achieved the best model performance, followed by the T+V+F, S+V+F, and V+F combinations for the bare soil condition in 2016-2018 period. This study provides a possible way for obtaining farmland SOC sequestration under crop cover conditions.