Spatial distribution and mapping of soil organic carbon in Amedyia region of Iraq

Soil is the largest carbon pool in terrestrial ecosystems, and organic carbon has been commonly adopted as an indicator of land degradation due to the significant role in regulating the physical, chemical and biological properties of soils. This study was carried out to model soil carbon content (SOC) with ordinary kriging (OK), Gradient Boosting Decision Trees (GBDT) and GBDT-OK hybrid approaches in Amediya region (20.06 km2) of northern Iraq. The covariates of GBDT model were the radiometric indices obtained from Landsat 8 (OLI) and Landsat 9 (OLI-2) optical satellite images, and the dependent variable was organic carbon content of 80 soil samples collected from the corners of 500 x 500 m grids in the study area. The k-fold cross validation loss was reduced from 0.385 to 0.257 by optimization of hyperparameters in the high variation dataset and limited observation points using the GBDT model. In addition, root mean square error (RMSE), mean absolute error (MAE) and R2, error metrics, improved from 0.394 to 0.229, from 0.173 to 0.132, and from 0.209 to 0.654, respectively with the optimization of hyperparameters. The improvement in RMSE value with the GBDT-OK hybrid model approach compared to GBDT and ordinary kriging were 32.75 and 68.96. The importance levels of the covariates in the GBDT estimation model were also determined. The covariates used in the SOC estimation model were ranked from the most important to the least important as NDVI > SAVI > EVI > MSAVI > NDSI > SRCI > GSI > BI, respectively. The results showed that the tree-based hybrid model approach is highly powerful for spatial estimation of SOC content in datasets with a small number of sample points. In addition, the study also revealed that selection of radiometric indices used as covariates should be determined considering the climate and vegetation density of the study area to increase the success of the estimation.