A fully adaptive object extraction technique used for spectral-spatial classification of remotely sensed data

The effectiveness of object-based image classification approaches has been frequently addressed and discussed in literature, especially for remote sensing applications. Unlike the traditional pixel-wise methods, object-based classifiers benefit from a segmentation step before the classification process in order to generate objects. In this paper, we propose to use the Pixon concept for segmentation of the data. Meanwhile, in order to form objects which are spectrally homogenous, spatial smoothing is applied as a preprocessing step through using regularized nonlinear partial differential equations (RegAPDE). The parameters of RegAPDE as well as important thresholds used in the Pixon extraction technique are adaptively tuned using three different adaptation algorithms. We also propose to localize the smoothing process via separately applying the RegAPDE algorithm to individual partitions extracted from each layer of the hyperspectral datasets. To this end, a simple partitioning step based on Watershed transformation is used before the smoothing procedure.