Binary and Multi-Class Classification of Fused LIDAR-Imagery Data Using an Ensemble Method

Airborne Light Detection and Ranging (LIDAR) data is used for multiple applications, such as urban planning, emergency response, flood control, and city 3D reconstruction. The LIDAR data in its raw form needs to be classified for the above applications. There are two types of classifications: binary and multi-class. In the binary classification, the given LIDAR data is classified into two classes: terrain or non-terrain. In the multi-class classification, the given data is classified into multiple classes, such as ground, vegetation (low, medium, and high), and buildings. Although different techniques have been developed to address the challenges in LIDAR data applications in the last two decades, no single algorithm gives the best result. In this paper, we presented two Ensemble methods (Bagging and AdaBoost) that combine multiple algorithms in an intelligent way. These methods were developed and evaluated using 100 decision trees as the weak classifiers and combination of both point and neighborhood features. The authors were able to achieve the accuracy up to 98.9% for the binary classification and 94.6% for the multi-class classification. While AdaBoost performed slightly better that Bagging, the Bagging was more resistant to over-fitting.