Cascaded classification of high resolution remote sensing images using multiple contexts

We present a novel cascaded classification approach by exploiting various contexts on different levels for high resolution remote sensing (HRRS) images. The contexts mentioned in our article are defined according to objects from a set of regions resulting from segmentation. The cascaded procedure comprises three stages: (1) initializing the classification using the object’s inner context (i.e., the gray constraints of different pixels in an object), (2) correcting the classification using the object’s neighbor context (i.e., the characteristic constraints of different objects adjacent to the concerned object), and (3) refining classification using the object’s scene context (i.e., the distribution constraint of different objects’ labels and their feature vectors in the whole scene). The proposed algorithm has the following distinctions. First, it uses an object’s neighbor context to bridge the gap between its inner context and its scene context because the latter two types of contexts have inevitable drawbacks when being used for classification alone. Second, it carries on a cascaded classification procedure in which the previous stage provides a better initial classification for the following stage, and the result is gradually refined by integrating different contexts. The effectiveness and practicability of the proposed algorithm is demonstrated through a set of completely experimental results and substantiated using quantitative criteria.