Land cover classification using moderate resolution satellite imagery and random forests with post-hoc smoothing

Abstract

Various image classification methods have been developed for land cover mapping. Among them, classification trees and their new modifications, such as random forests (RF), have proven effective. However, these tree-based methods typically perform per-pixel classification, which often produces suboptimal results with scattered misclassifications. This paper recommends applying smoothing techniques to address the problem and combines post-hoc smoothing with RF for land cover classification using moderate resolution remote sensing imagery and ancillary data. RF is used to produce probability maps for each type of land cover, a smoothing technique is employed to smooth the probability maps and then a maximum probability rule is applied on the smoothed probability maps to generate a land cover map by assigning each pixel to the class with highest class probability. This method was applied to classify land cover in the Jiuzhaigou Nature Reserve in China using Landsat Thematic Mapper (TM) Images and topographic data, and the classification accuracies with several different smoothing techniques, including anisotropic diffusion, Gaussian, mean and median filtering, were assessed and compared. The results demonstrated that RF combined with post-hoc smoothing improved the overall accuracy by up to 6 percent and the Kappa statistic by up to 9 percent over the land cover classification without a smoothing process, and at the 5 percent significance level, all the smoothed land cover maps had a statistically significant difference in accuracy based on Kappa compared with the unsmoothed map.