Abstract:
Soil Organic Carbon (SOC) is a measure of the total carbon content of the soil and is a vital soil health indicator. Over the decades, SOC has been estimated using sampling followed by rigorous laboratory-based testing methods. Spaceborne Microwave/Synthetic Aperture RADAR (SAR) remote sensing has proven to be a versatile tool for various soil study applications. However, there have been very few studies conducted for SOC estimation using SAR remote sensing. This study utilises time-series, C-band remotely sensed SAR data from Sentinel-1 A satellite for SOC estimation and compared the performances of Random Forest (RF) and Ordinary Least Squares (OLS) Regression models over agricultural areas of Rupnagar district of Punjab in India. A set of 96 soil samples were collected from 32 different agricultural field locations in Rupnagar district between November 2019 to January 2020. SAR backscatter of Vertically emitted and Vertically received (VV) and Vertically emitted and Horizontally received (VH) polarisation channels, from Sentinel-1, soil moisture, electrical conductivity, pH, temperature and SOC from the laboratory-based testing methods were used as regression parameters. The RF regression gave a Root Mean Square Error (RMSE) of 0.78 and R2 statistics of 0.887, while the OLS method performed better with an RMSE of 0.53 and an R2 value of 0.907. It was also observed that the backscatter from VV and VH polarisation channels, when used synergistically with field data, have the highest Feature Importance (FI) score in both RF and OLS regression models for SOC estimation.
