Compression and Object-oriented Processing of Segmented Hyperspectral Images in Envi

So far, segmented remotely sensed images can not be processed independent from the software they have been generated in. To overcome this limitation and to provide a flexible exchange format, it is suggested to store spectral mean values of image objects in generic band sequential (BSQ) files. In doing so, the pool of available image processing methods is considerably widened, because these non-proprietary formats are read by any standard image processing software. With each pixel corresponding to one object of the segmented image the new file is considerably smaller than the original. This way, a time efficient way of compressed information processing is created that meets the requirements of image data with increasing spatial and spectral resolution. In the present paper the flexibility of this object-compressed BSQ format is tested and the performance of the format is investigated using selected methods in ENVI. It is shown that standard image processing methods can be applied to object-compressed files with spectral mean values. Still, the validity of the methods and their results need to be considered against the background of the new image statistics and the lost spatial context in the data. INTRODUCTION Over the last years image segmentation has become a frequently applied technique for digital image processing of remote sensing data. Subsequent analysis of the objects, which result from image segmentation, has proved advantageous compared to the analysis of original pixels in many applications (1). This can mainly be explained by two aspects: on the one hand, object information includes spectral values plus additional attributes like heterogeneity measures or shape parameters, whereas pixels contain only spectral values. On the other hand, the spatial generalization of the segmentation process leads to more homogenous results that are easier to interpret. Both aspects have been successfully utilized, especially for classification problems (2). In addition, object-oriented image processing implies a compression of the original data. When neighbouring pixels are merged to objects, their spectral information is averaged to one spectral mean value. The memory space that is needed to store the information of a segmented image is thus smaller by a factor equal to the average object size. Unlike other compression methods, e.g. a rescaling to a lower resolution, the spatial generalisation during segmentation follows borders in the natural environment. Nevertheless, a software independent file format for segmented images is still lacking and the processing of image objects is often limited to the methods of individual software packages. The most known segmentation algorithm is probably the multiresolution segmentation (3) as implemented in the software eCognition. Image objects can be classified using a nearest neighbour classifier or hierarchical membership functions in this software. Results from the segmentation or subsequent classifications can be exported as shape or raster files with attribute tables. This way, a useful link to geographic information systems is provided but an interface to other image processing software that offer a greater variety of methods, e.g. transformations, spectral unmixing, or different classifiers, is missing. © EARSeL and Warsaw University, Warsaw 2005. Proceedings of 4th EARSeL Workshop on Imaging Spectroscopy. New quality in environmental studies. Zagajewski B., Sobczak M., Wrzesień M., (eds) In order to allow maximum flexibility concerning software packages and processing methods, a generic and non-proprietary format seems appropriate. A systematic arrangement of the object information is needed to allow an easy reconstruction of the original spatial information of the objects while making best use of the compression effect. Against the background of steadily increasing spectral and spatial resolution useful compression methods for imaging spectrometer data are important in order to reduce storage space and processing times. METHODS Data Two HyMap test data sets are used for the present work. The data were acquired over Berlin, Germany, on 30 July 2003 within the HyEurope 2003 campaign by the German Aerospace Center. A radiometric pre-processing was carried out based on MODTRAN including water vapour estimates per pixel. After the removal of noisy bands, 116 bands remained in the data sets. Test data set 1 is a 882 x 895 pixel subset of a flight line that was parametrically geo-rectified using the software PARGE (4). The spatial resolution of the subset after resampling is 3.5 m. It covers an area in the centre of Berlin (figure 1). Figure 1: A subset from HyMap data of central Berlin that is used as test data set 1. [R band 27; B band 81; B band 16] Test data set 2 consists of a complete HyMap flight line before geo-rectification with 3047 x 512 pixels at a spatial resolution of 4.6 m. Various types of urban structures are represented in the image, which extents from the central governmental district eastwards to suburban areas. Figure 2: Test data set 2, a complete HyMap flight line [R band 27; B band 81; B Band 16]