Accuracy Enhancement of Object Based Image Classification Using Relaxation Labeling Process for High Resolution Satellite Images

Image classification is an important task for many aspects of global change studies and environmental applications. In this study we compare two different classification approaches, which are Object based and Pixel based. Object Based Classification (OBC) methods are increasingly used for classification of land cover/land use from high resolution images, and often the final result is close to the way a human analyst would interpret the image. A number of properties of the regions were computed in OBC like spectral mean vector, average texture, departure from circularity, length-to-breadth ratio, area, perimeter and compactness. Image is then classified on the basis of the regions instead of the pixels. The fine spatial resolution implies that each object is an aggregation of a number of pixels in close spatial proximity, and accurate classification requires that this aspect be considered. In this paper Relaxation Labeling Processes (RLP) is explored as a post-classification refinement tool. RLP requires initial label probability values to start the refinement process which are generated using Cloud Basis Function based Neural Network (CBFNN) classifier. The nodes in the CBFNN output layer are normalized and treated as initial label probability values along with a thematic image. These label probabilities are updated by RLP on an iterative basis. Each time a small neighborhood around each pixel is employed for probability updating, and the iterative process effectively allows propagation of global information through expanding neighborhood.