Soil organic carbon measurement and modelling at regional scale using Sentinel-1/2 data based on machine learning approaches

Soil is the largest carbon (C) reservoir in terrestrial ecosystems and soil organic carbon (SOC) is the basis for soil’s biodiversity, health and fertility. So, measurement of SOC becomes necessary for sustainable soil ecosystem management. However, lab based conventional method to measure the SOC is time and energy consuming. Also, there is a health risk because of the hazardous chemicals used in the soil analysis. To overcome such problems, remote sensing (RS) imagery data products have been used to estimate the SOC. Among all the optical RS data sources used in soil studies, Sentinel-2 (S2) Multispectral imagery (MSI) data has been proved to be the best by many of the researchers because of its unique spectral features. However, S2 along with a radar data source such as Sentinel-1(S1) gives more accurate results. Therefore, the main objective of our study was to estimate the SOC using S2 MSI and S1 SAR-C data in the Western catchment of Chilika lagoon, which is one of the first Ramsar sites in India. To achieve this, 167 surface soil samples (0-15cm) was collected from the study area for SOC measurement. We used PLSR and three machine learning models such as RF, Cubist and SVR for the prediction of SOC from the RS data source. Model performance, showed that PLSR using the covariate set containing S1, S2 and topographic attributes performed the best in predicting SOC (RMSE = 0.17 and R2= 0.38) among all other models. While, model accuracy reduced slightly (RMSE = 0.16 and  R2= 0.31) with only S2 bands data. This indicates that using the S1 data and topographic attributes along with S2 data results in better SOC predictions. However, model performance was moderate to poor. Therefore, more studies would be needed for accurate estimation of SOC.

Key words: Soil organic carbon; Sentinel-2; Sentinel-1; Machine learning