Spatial prediction of soil organic carbon using machine learning techniques in western Iran

Abstract

Estimation of soil organic carbon (SOC) is very useful for accurate monitoring of carbon sequestration. However, there are still significant gaps in the knowledge of SOC reserves in many parts of the world, including western Iran. To partially fill the gap, 865 soil samples were used with 101 auxiliary variables and 5 machine learning (ML) algorithms to digitally map SOC for the plough layer (0–30 cm) at a 90-m resolution in Kurdistan province. Results indicated that the most important auxiliary variables were rainfall (27.09%), valley depth (26.66%), terrain surface texture (23.42%), air temperature (20.18%), channel network base level (16.61%) and terrain vector roughness (14.47%). Results also showed that Random Forests (RF) performed best in predicting the spatial distribution of SOC (RMSE = 0.35% and R2 = 0.60), compared to the other ML algorithms (i.e. Cubist: CU, k-Nearest Neighbor: kNN, Extreme Gradient Boosting: XGBoost and Support Vector Machines: SVM). Furthermore, results estimated the total SOC stocks (SOCS) for the whole study area (~15,208 Tg) and amounts under different land uses. These were bareland (~6 Tg), orchard (~356 Tg), irrigated farming (~782 Tg), forest (~1773 Tg), grassland (~5991 Tg) and dry farming (~6297 Tg). As expected, the SOCS were highest in forest soils (652 g m−2) and lowest in bareland (437 g m−2). This result suggests that the conversion of native land (e.g. Forest) to cultivated land (e.g. Irrigated farming) could lead to significant loss of SOCS and appropriate management of land use could increase SOCS.