Textural classification for ecological research using ATM images

Abstract

Abstract Investigation of spectral and textural classification of high resolution ATM image of a semi-natural scene is presented. Pure spectral classification using bands 5, 7, 9 and the maximum likelihood classifier yielded 56, 63 and 64 per cent overall classification accuracies with 1-25m, 2-5m, and 50m spatial resolution data respectively. Application of combined spectral and textural classification using bands 5, 7, 9 and various texture features from seven texture algorithms ( spatial grey level dependence matrices-SGLDM, grey level run length matrices-GLRLM, busyness, neighbouring grey level dependence matrices-NGLDM, sum and difference histograms-SADH, and fractal analysis), yielded overall classification accuracies from 58-65 per cent at 1-25 m resolution. It is concluded that texturally-based classifications improve overall classification although improvements are not dramatic. The first-order texture measures from algorithms like GLDH and SADH have shown more promise than second-order algorithms, like SGLDM and NGLDM. The energy feature from most of the texture algorithms shows considerable classification potential. A selection of distance metric corresponding to the size of the spatial unit for a given cover type improves the classification of that class. With degradation of spatial resolution the overall accuracy of textural classification improves up to 69 per cent for 5-0 m resolution data.