Reduction of training costs using active classification in fused hyperspectral and LiDAR data

Abstract

This paper presents a novel approach for the reduction of training costs in classification with co-registered hyperspectral (HS) and Light Detection and Ranging (LiDAR) data using an active classification framework. Fully automatic classification can be achieved by unsupervised learning, which is not suited for adjustment to specific classes. On the other hand, supervised classification with predefined classes needs a lot of training examples, which need to be labeled with the ground truth, usually at a significant cost. The concept of active classification alleviates these problems by the use of a selection strategy: only selected samples are ground truth labeled and used as training data. One common selection strategy is to incorporate in a first step the current state of the classification algorithm and choose only the examples for which the expected information gain is maximized. In the second step a conventional classification algorithm is trained using this data. By alternating between these two steps the algorithm reaches high classification accuracy results with less training samples and therefore lower training costs. The approach presented in this paper involves the user in the active selection strategy and the k-NN algorithm is chosen for classification. The results further benefit from fusing the heterogeneous information of HS and LiDAR data within the classification algorithm. For this purpose, several HS features, such as vegetation indices, and LiDAR features, such as relative height and roughness, are extracted. This increases the separability between different classes and reduces the dimensionality of the HS data. The practicability and performance of this framework is shown for the detection and separation of different kinds of vegetation, e.g. trees and grass in an urban area of Berlin. The HS data was obtained by the SPECIM AISA Eagle 2 sensor, LiDAR data by Riegl LMS Q560.