No abstract

No abstract

No abstract

Summary Soil organic carbon (SOC) is an important carbon pool in terrestrial ecosystems. Prediction of SOC based on soil properties and environmental factors helps to describe the spatial and vertical distribution in SOC; however, the effectiveness and accuracy of various prediction methods, including classical and recently developed model approaches, need to be tested for tropical soil and environments. In this study, random forest (RF) and linear mixed effects model (LMM) approaches were tested to predict the spatial and vertical variation of SOC stocks in Eastern Madagascar. Topography, climate, soil types and vegetation‐based variables were used as predictor variables for modelling SOC stocks at different soil depths to 1 m. The LMM was the most accurate method for predicting SOC stocks for different depth ranges; altitude, soil clay content, land use and precipitation were identified as the most relevant factors for prediction. The accuracy of prediction in SOC modelling decreased with increasing soil depth, resulting in a root mean square prediction error (RMSE) that ranged from 1.98 Mg ha−1 (90–100‐cm depth) to 5.54 Mg ha−1 (10–20‐cm depth) for LMM, which resolved 43–68% of the variation in SOC stocks. Explanatory variables, which contributed to the fixed effect of the model, explained from 2.6 to 28.2% of the total variance, whereas the random effect contributed from 21.7 to 35.0%. This study emphasizes the strength of LMM for predicting SOC stocks in tropical soil taking into account the random effect related to sampling. These results could be used to improve SOC mapping in Madagascar. Highlights Prediction accuracy of vertical variation in SOC stocks was tested with RF and LMM approaches. Linear mixed effects model provides the most accurate predictions of SOC stocks. Altitude, clay content, climate and land use were identified as relevant predictor variables of SOC. The LMM can be used to improve SOC mapping of tropical soil.

No abstract