Digital mapping of soil organic matter for rubber plantation at regional scale: An application of random forest plus residuals kriging approach

• RFRK method was used to digitally map spatial pattern of SOM for rubber plantation.
• Compared to MLR, RFRK significantly improved prediction accuracy of SOM.
• GeoU, PREC, and ELE mainly controlled spatial pattern of SOM at rubber plantation.
• Relations between SOM and environmental factors were non-linear and hierarchical.

Soil organic matter (SOM) plays an important role in soil fertility and C cycle. Detailed information about the spatial distribution of SOM is vital to effective management of soil fertility and better understanding of the process of C cycle. To date, however, few studies have been carried out to digitally map the spatial variation of SOM for rubber (Hevea brasiliensis) plantation at the regional scale in Hainan Island, China. In this study, a hybrid approach, random forest plus residuals kriging (RFRK), was proposed to predict and map the spatial pattern of SOM for the rubber plantation. A total of 2511 topsoil (0–20 cm) samples were extracted from a soil fertility survey data set of the Danzhou County. These soil samples were randomly divided into calibration dataset (1757 soil samples) and validation dataset (754 soil samples). In this study, stepwise linear regression (SLR), random forest (RF), and random forest plus residuals kriging (RFRK) were used to predict and map the spatial distribution of SOM for the rubber plantation, while generalized additive mixed model (GAMM) and classification and regression tree (CART) were employed to uncover relationships between SOM and environmental variables and further to identify the main factors influencing SOM variation. The RFRK model was developed to predict spatial variability of SOM on the basis of terrain attributes, geological units, climate factors, and vegetation index. Performance of RFRK was compared with SLR. Mean error (ME), mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R2) were selected as comparison criteria. Results showed that RFRK performed much better than SLR in predicting and mapping the spatial distribution of SOM for the rubber plantation. The RFRK model had much lower prediction errors (ME, MAE, and RMSE) and higher R2 than SLR. Values of ME, MAE, RMSE, and R2 were 0.26 g/kg, 1.35 g/kg, 2.19 g/kg, and 0.86 for RFRK model, and were 0.65 g/kg, 2.99 g/kg, 4.37 g/kg, and 0.43 for SLR equation, respectively. Moreover, RFRK model yielded a more realistic spatial distribution of SOM than SLR equation. The good performance of RFRK model could be ascribed to its capabilities of dealing with non-linear and hierarchical relationships between SOM and environmental variables and of accounting for unexplained information in the random forest (RF) model residuals. These results suggested that RFRK was a promising approach in predicting spatial distribution of SOM for rubber plantation at regional scale.