Location-adaptive Texture: an Experiment Using Quickbird, Aster and Landsat Etm+ Imagery

One of the enduring problems in remote sensing analysis is how to exploit image texture [2]. A central question in texture analysis is the scale of the texture features analyzed, as determined by the size of the texture kernel (moving window). In a classic study, 90% of texture variability was found to be accounted for by the size of the kernel, compared to only 7% from the texture algorithm [3]. The major challenge in identifying an optimal kernel size is that large kernel sizes are required to produce relatively stable texture measures for characterizing spectral variability, but small kernel sizes are preferable to minimize the tendency of intra-class texture to overwhelm the inter-class texture that is usually of interest [2]. Thus the optimal texture scale is not just class dependent [1], but potentially also varies as function of location within an image class [2]. This paper is an investigation of how texture values vary with kernel size for different locations in an image, such as edges and interiors of classes, and whether such information can be used to develop a locally-adaptive texture measure. The approach is tested empirically with a case study utilizing images of three different spatial scales.