Remote sensing of soil organic carbon in varied tillage-crop systems.

The use of remote sensing (RS) to estimate soil organic carbon (SOC) in cropland has become increasingly important to producers, researchers, and policy makers to assess soil and plant health across spatially variable landscapes. Yet, RS estimation of cropland SOC is challenging, particularly when mixed crop residues and soils are present. Our objective was to develop an RS model to estimate SOC under varied tillage-crop systems typical of Corn Belt, US farming practices and evaluate model performance with respect to each system. Four tillage-crop systems were evaluated: conventional till corn (CT-corn) with one tillage event, CT-corn with two tillage events, no-till soybean (NT-soy), and no-till corn (NT-corn). A random forest (RF) model was developed using SOC measurements, Sentinel-2 early spring images (bands and band ratios), and ancillary data (elevation, yield, soils, peak vegetation), and accuracy and most important variables were assessed for each system. The two CT-corn models had similar predictability and accuracy (R² = 0.65-0.66, root mean square error [RMSE] = 0.13), while the NT-soy had comparable predictability but lower accuracy (R² = 0.69, RMSE = 0.22). The NT-corn model, however, underperformed (R² = 0.14, RMSE = 0.29). Sentinel-2 early spring images dominated most important variables for all models except for NT-corn which relied on ancillary inputs. The RF model was also used to map the spatial distribution of SOC, which showed variability related to human disturbance (historical railroad tracks). This research provided insight into estimation and mapping of SOC in varied tillage-crop systems and highlighted the importance of using early spring RS images to improve results in mixed crop residue and soil areas.