Korean Journal of Soil Science and Fertilizer ISSN:0367-6315(Print) 2288-2162(Online) 한국토양비료학회 학회지

Random forest models (RFM) are useful in predicting the soil carbon (C) contents because RFM predicts soil C with high accuracy under complicated environmental conditions. However, there are very few studies on prediction of soil C using RFM in Korea. Moreover, there is no case study using RFM to predict soil C content of reclaimed tideland (RTL) soils, which have high C sequestration capacity. Therefore, in this study, the applicability of RFM was evaluated using published soil properties data, including soil C and soil variables, for RTL soils located in southwestern coastal areas of Korea. In the present study, RFM was built using the data of 16 variables (e.g., sand, silt, and clay contents, pH, electrical conductivity of saturated soil paste (EC_e), and nutrient concentrations) obtained from five RTLs with similar climate, topography, and vegetation. The 80% of the total data were trained to build the model, and searched optimal hyper parameters were used to improve accuracy. The determination coefficient (R²) of the model was 0.67, and the difference between measured and predicted soil C content was 25.9% on average. However, when the measured values were out of the range of the data trained for building the model or the measured values were close to the minimum or maximum value, the difference between the predicted and measured values became larger (73.9%). The contribution of the independent variables to the prediction of soil C using the model was the greatest (14.9%) for soil NH₄⁺ concentrations. Meanwhile, the contribution of EC_e, which was highly correlated with soil C content, was not detected, suggesting that the importance of the number and range of training data used to build model. Our study shows the possible application of RFM to predict soil C contents of RTL soils in Korea, and further highlights that a large amount of data should be accumulated for high accuracy prediction of soil C using RFM.

Random forest models can predict soil carbon content with high accuracy under complicated environmental conditions by randomly learning multiple decision trees.