A New Binary Encoding Algorithm for the Simultaneous Region-based Classification of Hyperspectral Data and Digital Surface Models

In this paper, an approach is proposed to integrate hyperspectral image data, object and height information into a new region-based binary encoding algorithm for automatically deriving land cover information. After georeferencing the different data sets and deriving a normalized digital surface model (nDSM), connected regions are extracted from the hyperspectral data by applying an edge-based segmentation algorithm. The mean spectrum per region is considered representative for the region. Five parameters are defined to describe the size and shape of the region, namely area, asymmetry, rectangular fit, ratio of length to width, and compactness. Together with the spectral information these parameters and the corresponding height values per region from the nDSM are converted into a binary code. This code is then matched to that of a training data set for classification. In order to evaluate the suggested approach we applied it to a test area in Oberpfaffenhofen, Germany. A manually generated classification served as reference. We also compare our result with the well known support vector machine (SVM) classifier. Based on our test data, we could show that the inclusion of size, shape, and height improves the classification accuracy of binary encoding. We could also show that the new method obtained more accurate and more efficient results when compared to the SVM classification.