| This dissertation explores the use of texture for modelling geospatial objects in remote sensed images. Texture based analysis of remote sensed images has not enjoyed the same success as spectral based analysis. This is surprising since texture can be considered as context at the pixel level, and context is acknowledged to be important for analyzing geographic data. This work extends texture based analysis beyond land cover classification, which can be viewed as the current state-of-the-art, to modelling objects composed of multiple characteristic textures or texture motifs. A statistical pattern recognition framework is adopted in which the texture motifs are modelled as mixtures of Gaussians in the high dimensional texture feature space, and the model parameters are learned in an unsupervised manner using the expectation maximization algorithm. The objects and their motifs are assumed to occur at arbitrary orientations in the images which presents a formidable challenge to using orientation selective texture features. The proposed approach overcomes this challenge by exploiting the structure of texture features extracted using banks of Gabor filters tuned at different scales and orientations. |
