A Complex Terrain Simulation Approach Using Ensemble Learning of Random Forest Regression

Highly realistic terrain surfaces are critical to digital terrain analysis and application. In response to limitations of generating high accuracy complex terrain surfaces by conventional methods, a terrain simulation approach based on random forest regression (RFR) is proposed. The objective of this study was to generate terrain surfaces with high precision and high accuracy, by fully realizing the advantages of RFR algorithm for nonlinear fitting. This paper establishes and explains approaches for terrain surface simulation using RFR and an example of this application is provided. In the presented study, the LiDAR point clouds were used as experimental data, and ordinary kriging (OK) and Gaussian geostatistical simulation (GGS) methods were selected as alternative methods. Furthermore, the root mean square error and mean error of the elevation deviation of checkpoints were calculated and the effects of simulated terrain surface were verified for precision and accuracy, using both approaches. Paired t test and Levene tests were performed on the precision and accuracy of the simulated terrain surface, respectively. The experimental results show that the simulated terrain surface based on RFR approach is not only more accurate but also more precise than that derived using the OK and GGS methods.