CLASSIFICATION OF REMOTELY SENSED IMAGES BY TEXTURAL INFORMATION

摘要

The standard techniques for the classification of remotely sensed multispectral imagerndata make use of statistical information only. Standard classification algorithms likernvector quantization algorithms or the maximum-likelihood classifier are based on standardrnstatistical techniques which in general do not make use of the context information inherentlyrncontained in the image data. When defining textural information as a major part ofrncontext information it becomes clear that normally a valuable part of the gathered informationrnis not incorporated into the data analysis. A major reason for this gap in the datarnanalysis is that it is still a challenging task to extract textural information from image datarnby a reliable and robust feature extraction scheme. By using physiological findings fromrnthe mammal visual cortex and by applying suitable mathematical models we developed arnscheme for the extraction of textural information based on Gabor wavelets which act asrntuned bandpass filters. Invariance of the algorithm with respect to scaling and rotation isrnachieved by the incorporation of the flight altitude and flight track information acquiredrnduring the airborne mission. Furthermore an energy normalization of the Fourier componentsrnextracted by the Gabor bandpass filters allows to correct illumination variationsrnbetween different scenarios. Finally we used a supervised feed-forward neural network tornapproximate the functional relationship between the high-dimensional feature data spacernand the predefined texture classes. We will present results achieved with this algorithm byrnusing data sampled with the airborne imaging spectrometer casi over selected test areas.