Soil organic carbon prediction and mapping in crop areas from a tropical region using Landsat satellite and clay content

Abstract Soil organic carbon (SOC) quantification in a practical manner with reduced costs has become a necessity due to current agricultural demands. Studies have shown that remote sensing is important for SOC prediction, and its use has become crucial in agricultural management. In this study, a Linear Multiple Regression (LMR) model was constructed to predict SOC in a site in the region of Piracicaba, Sao Paulo, Brazil. We used the optical-satellite data of Landsat OLI sensor (bands 5 and 7), clay concentration, and the Normalized Difference Vegetation Index (NDVI) as predictor variables. For the sample points distribution, 218 samples were collected in the field to quantify clay and SOC in the laboratory as calibration procedure. An exposed soil mask (ESM) was created using the method GEOS3 technology in which pixels with greater variability of bare soil were observed. The pixels were evaluated with their respective surface reflectance values obtained by the satellite sensor and their respective NDVI index values. The predictive performance of the model was evaluated based on the adjusted coefficient of determination (R2), the Root Mean-Squared Error (RMSE), and the Ratio of Performance to Interquartile Range (RPIQ) obtained in data validation. The LMR model presented R2 values of 0.79 and 0.81 for calibration and validation, respectively. We obtained important values of RMSE and RPIQ, 0.14 and 2.32, respectively. The high RPIQ indicated significative sampling distribution around the trendline. After construction, the model was applied to the carbon spatialization using the predictive variables as layers. The analysis presented here enables new possibilities for SOC prediction using Geographic Information Systems (GIS) tools.