Prediction of spatial distribution of soil available iron in a typical hilly farm-land using terrain attributes and random forest model

Soil available iron is essential for plant growth. Detailed information on the spatial distribution of soil available iron is critical for effective management of soil fertility. To date, published works on soil available iron have mainly focused on the spatial variability and little has been done on predicting the spatial distribution of soil available iron. To understand the spatial distribution of soil available iron in hilly areas of Southwest China, we conducted a study in 2014 at a 2-km² typical hilly region with uniform soil parent materials in Yongxing Town, Jiangjin County, Chongqing City. A total of 309 soil samples were collected from cultivated lands at the depth of 0-20 cm. The samples were randomly divided into calibration (224) and validation (85) samples. Nine terrain attributes (including elevation, slope, aspect, valley depth, horizontal curvature, profile curvature, convergence index, relation position index and topographic wetness index) were extracted from a digital elevation model of spatial resolution of 2.0 m. Ordinary Kriging (OK), Multiple Linear Regression (MLR) and Random Forest (RF) analyses were used to predict the content of soil available iron based on the terrain attributes. Accuracy indicators, including mean absolute error (MAE), root mean square error (RMSE) and coefficient of determination (R²), were used to evaluate model performance based on validation data. Correlation analysis showed that topographic wetness index and valley depth were significantly positively correlated with soil available iron content. Slope, horizontal curvature, profile curvature, convergence index and relative position index were on the other hand significantly negatively correlated with soil available iron content. Compared with OK and MLR, RF model performed best, with MAE=22.33 mg·kg^-1, RMSE=27.98 mg·kg^-1 and R²=0.76. Additionally, RF analysis indicated that topographic wetness and slope were the main factors controlling the spatial distribution of soil available iron. Soil available iron content in the study area was 3.00-276.97 mg·kg^-1, which was higher for paddy field than for dryland. Semivariance model showed strong spatial autocorrelation of soil available iron, indicating that structural factors were the main driving force of spatial variation of soil available iron. Therefore it was concluded that the RF model together with terrain attributes well explained the spatial variability of soil available iron in the area. The result of the study provided valuable information for studies on predicting the spatial distribution of trace elements in soils in hilly areas.