Includes bibliographical references.

Coal mining inevitably alters land-use and land-cover patterns which in turn influences soil organic carbon stock buildup and nutrient dynamics in areas adjacent to coal mine. SOC and its distribution characteristics are important indicators of soil health and soil quality which directly or indirectly determines soil fertility and plant productivity, so understanding the geographical distribution pattern and prediction of soil organic carbon is necessary in complex topographical settings. Soil organic carbon generally estimated by ways of conventional techniques are rather laborious and costly so researchers are investigating the utilization of alternative approaches through remote sensing & GIS capacity. These advanced approaches are recommended to be rapid, real time and having the advantage of being applicable in global and regional scale. The current study was conducted in Saipung Block, Meghalaya, India covering 995 km2 area and aimed to develop predictive models of SOC content based on field sampling and Sentinel 2-A multispectral data in consideration with different land-use patterns adjacent to coal mining areas. A total of 160 soil samples were collected from 0-20 cm depth by stratified random sampling strategy. Following this, remotely sensed data (12 spectral indices) were derived as predictor variables to improve SOC prediction of the study site. Exposed soil composite indices values were regressed against topsoil measured SOC contents from the corresponding pixels in the calibration dataset (n=120). Three multivariate regression methods viz, Stepwise Multiple Linear Regression (MLRstepwise), Principal component regression (PCR) and Partial least squares regression (PLSR) was then employed to develop the prediction models for SOC. To evaluate stability of the models, accuracy metrics like Root Mean Square Error (RMSE), and Coefficient of Determination (R2), were used with 75% of the data used for prediction and 25% for validation. Further, the developed model was compared with predefined SOC models considering the validation dataset to access other standard model’s stability in the area of research. The results revealed that the SOC content influenced by mining was found to be moderate to low and exhibited great variability within (CV>20%) different land-uses. Moderate positive correlation was found between SOC and indices like NDVI, GNDVI, EVI-2, SAVI, MSAVI-2, RDVI, TVI, BI-2, SBI and SOCI and a negative correlation with CI and SI. The modelling performance of all the three MLR models were found to be acceptable but prediction accuracy of stepwise multiple linear regression model outperformed (R2=0.70; RMSE=0.28) both partial least square regression model (R2=0.63; RMSE= 0.31) and principal component regression model (R2 = 0.56; RMSE=0.36) in validation set. The predefined SOC models when tested using the validation set failed to show statistical significance with observed SOC. The methodology implemented in this study, which elaborated prediction of SOC by regression models, has the potential to map SOC in similar topographical settings.