AN IMPROVED BINARY ENCODING ALGORITHM FOR THE INTEGRATION OF HYPERSPECTRAL DATA AND DSM

Binary spectral encoding is well known as a simple and effective hyperspectral analysis method in classification, searching of similar spectra and identifying mineral components. If the hyperspectral data consists of L spectral channels, this method represents the spectral amplitude and the spectral slope with a 2L-bit binary code vector. Usually the Hamming distance represents the similarity measure to determine the spectral signature matches.
Although this method is frequently used and offers good performance, it still has some weaknesses. Due to the spatial resolution of modern hyperspectral sensors and because this method mainly operates on pixels the efficiency sometimes is low. Nowadays, with the rapid development of advanced remote sensor technologies with increasing spatial resolution, object-based based data processing methods are more frequently applied. This paper therefore attempts to integrate an object based approach with traditional hyperspectral processing methods to support and enhance the comprehension of remote sensing data.
We therefore attempt to integrate the spectral, shape, texture and height information of remote sensing data of the same area by a modification of the binary encoding spectral matching method.
Several processing steps have to be performed before the integration. First segments have to be established. Several existing segmentation methods using a Digital Surface Model (DSM) are investigated. In this paper, an edge-based segmentation algorithm and the Full Lambda-Schedule algorithm are employed to get segments and merge adjacent segments respectively. As one segment then is composed from many pixels, a method to choose a representative spectrum for this specific object has to be found. This is accomplished by removing the spectral noise in one segment by comparing the majority and mean spectra to extract a representative spectrum for the object. The most important task then is, to find a way to represent shape, texture and height information with binary coding, like the spectral information. Several shape and texture description measures, e.g. roundness, solidity, area, texture Kernel size, will be discussed in this part.
The data set used, is an area in Oberpfaffenhofen, Germany. An airborne HyMap image and a DSM with a spacing of 0.5 meters are two main data sources in our research. The HyMap data has 126 bands in visible and near infrared range and has a ground sampling distance of 4 meters. The HyMap data is atmospheric corrected and both data sets are co-registered. Experiments and results using this data will be presented.
Following an introduction a short description of the study area is given. Then data processing and binary encoding methods used in this paper, especially the formation of object signatures (spectral, shape, texture and height information) with binary coding is presented as well as the computation of the signature distances. Practical tests, illustrating the proposed methods are presented and finally some conclusions are drawn at the end.