Evaluation and Prediction of Topsoil organic carbon using Machine learning and hybrid models at a Field-scale

Abstract Digital mapping of soil organic carbon (SOC) is crucial to evaluate its spatial variability and also to assess environmental factors controlling it at field scale. The current study was conducted to compare one statistical method include partial least squares regression (PLSR), four individual machine learning(ML) algorithms including random forest (RF), quantile regression forest (QRF), cubist (CB), fuzzy logic (SoLIM) along with two hybrid methods including the random forest-ordinary kriging (RF-OK) and quantile random regression forest-ordinary kriging (QRF-OK) to map SOC. A number of 146 soil samples (0–30 cm) were collected in Khorramabad plain (680 ha). For the quantitative evaluation of SOC spatial variability, two scenarios were considered to use remote sensing data (R) and a combination of geo-morphometric and remote sensing data (GR). The data randomly were split into 80% (117 points) for training and 20% (29 points) for validation. The model performances were evaluated by the statistical indices as the coefficient of determination (R2), root mean square error (RMSE). According to principal component analysis (PCA), nine covariates including transformed soil adjusted vegetation index (TSAVI), relative vegetation index (RVI), Band 10, Band 11,Digital Elevation Model (DEM), Standard height(Standard_he), Valley_Dep, terrain surface texture (texture), and terrain surface convexity (Convexity) were selected as the environmental predictors. Results showed that the hybrid model RF-OK (RMSE = 0.05and R2 = 0.93) and SoLIM model (RMSE = 0.47and R2 = 0.41) with scenario GR had the highest and lowest accurate respectively. TSAVI, DEM, and Band 10 were the most important predictors and explanation more than 50% of SOC spatial variability in the study area. Generally, using hybrid machine learning models in combination with geo-morphometric and remote sensing covariates make it possible to model and predict SOC with acceptable accurate in the field-scale croplands.