Soil Property Surface Modeling Based on Ensemble Learning for Complex Landforms

Wei Liu, Yongkun Liu, Mengyuan Yang, Meng Xie

Abstract: It is difficult to simulate soil property with a single global interpolation model. For the characteristics of spatial discontinuity, limited precision of global interpolation model and poor adaptability, a high accuracy surface modeling for soil property based on ensemble learning and fusion geographical environment variables was proposed (HASMSP-EL). The simulation accuracy of different interpolation methods was evaluated by using Mean Error (ME), Mean Relative Error (MRE), Root Mean Square Error (RMSE) and Accuracy (AC). The results showed that: (1) In the interpolation method of fusion geographical environment variables, the estimation deviation of HASMSP-EL was lower. Compared with other interpolation methods, ME, MRE, RMSE and AC of HASMSP-EL were better. HASMSP-EL had more advantages in describing spatial variation and local detail information of soil potassium content, and its accuracy was 6.42%, 7.28%, 11.56% and 9.38% higher than that of Regression Kriging (RK), Bayesian Kriging (BK), Inverse Distance Weighting (IDW) and Ordinary Kriging (OK), respectively. (2) The HASMSP-EL can provide more details in the geographical boundary, which made the simulation results consistent with the real auxiliary variables. HASMSP-EL not only considered the nonlinear relationship between geographical environmental variables and soil property, but also combined the adaptive advantages of multiple models. It is a new method to simulate soil property in complex geomorphological regions with higher precision.