Digital mapping of soil organic carbon using ensemble learning model in Mollisols of Hyrcanian forests, northern Iran

• Published in: Geoderma Regional Vol. 20; p. e00256
• Main Authors: Tajik, Samaneh, Ayoubi, Shamsollah, Zeraatpisheh, Mojtaba
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.03.2020
• Subjects: computer-aided mapping; cost effectiveness; deciduous forests; Environmental covariates; Feature selection; forest ecosystems; forest soils; Iran; least squares; linear models; Machine learning; Mollisols; prediction; remote sensing; soil organic carbon; Soil properties; soil sampling; soil surveys; Spatial modeling; support vector machines; topography; vegetation index; Iran; Environmental covariates; Feature selection; Mollisols; Soil properties; Machine learning; Spatial modeling
• ISSN: 2352-0094; 2352-0094
• DOI: 10.1016/j.geodrs.2020.e00256

This study was conducted to evaluate the efficacy of the ensemble machine learning model to predict the spatial variation of soil organic carbon (SOC) concentration in a deciduous forest ecosystem in northern Iran. To do this, a total of 153 soil samples by applying regular systematic sampling grid at two depths (0–10 and 10–20 cm) were collected. Two scenarios through digital soil mapping (DSM) were considered to establish the predictive models for estimating SOC including (i) combination of selected topographic attributers (T), remotely sensed data (R), and soil properties (S) (TRS) and (ii) combination of topographic attributes and remotely sensed data (TR). The ensemble model for predicting of SOC was consisted of six machine learning algorithms: partial least squares regression (PLSR), generalized linear model (GLM), recursive partitioning and regression trees (rpart), support vector machines (SVM), random forest (RF) and k-nearest neighbors (kNN). The 10-fold cross-validation with three replications was employed to evaluate the performance of models by root mean square error (RMSE) and the coefficient of determination (R2). The results showed that SOC was varied from 2.05 to 13.16% with a mean of 5.67% in first depth (0--10 cm) and from 1.56 to 9.56% with a mean of 3.99% in second depth (10–20 cm). According to the RMSE and R2 results ensemble machine learning model, first scenario (TRS) showed higher performance (higher R2 and lower RMSE) than second scenario (TR) for prediction of SOC in first depth (R2 = 0.74, RMSE = 0.78%) and second depth (R2 = 0.65, RMSE = 0.90%). Also, in the best scenario (TRS) for individual models, results of validation revealed that GLM, SVM and RF for the first depth and PLSR, GLM, SVM, and RF for the second depth were the most accurate machine learning algorithms in ensemble modeling based on RMSE. Our finding indicated that soil properties had an important contribution to the spatial variability of SOC in the studied forest soil. Moreover, topographic attributes and vegetation indexes were found to be auxiliary attributes in the modeling of SOC and could use for quick and cost-effective assessment of SOC concentration in order to management practices in forest soils. •Digital mapping of soil organic carbon in mixed Caspian Hyrcanian forests was predicted•Ensemble model with different machine learning algorithm was applied•Terrain attributes and soil properties were applied to model SOC•Topographic parameters, especially flow accumulation had most influence on SOC distribution