A Novel LIDAR Classification Method Based on Ensemble Random Forest and D-S Evidence Synthesis

Abstract

Light detection and ranging system (LIDAR) can quickly, proactively and automatically acquire point cloud data of large-scale area, which contains three-dimensional land-cover object information, meanwhile the multispectral cameras can acquire multi-band spectral information. This paper extracts and selects sixteen features according to point cloud data and spectral images, then these features are divided into five groups, such as height feature subset, intensity subset, spectral subset and texture subset. The weight of each group of features in the decision-making process is characterized by the features' importance. This paper introduces evidence synthesis theory to overcome the evidence conflict in decision-making and to improve decision precision. Final land-cover objects labels are predicted through ensemble algorithm based on random forest and weighted D-S evidence synthesis. The experiment results indicate that classification performance of joint feature set is superior to single feature set, compound classification framework can optimize final classification results. The overall accuracy reaches to 94 % and other parameters can also be improved more or less.