Remote sensting image classification approach based on sub-block features

Object-based classification approach offers higher accuracy and smoother classification map over pixel-based classification, especially on high spatial resolution imagery. However, a potential limitation in using object-based classification is the possible negative impact of under-segmentation. Under-segmentation error cannot be adjusted in the unit of object, and can affect the potential accuracy of classification On the contrary, pixel-based contextual classification utilizes spatial information to reduce salt-and-pepper noise, and does not deal with segmentation issues. Though, a challenge of pixel-based contextual classification is to define appropriate contextual neighbors. In this paper, we propose a new approach, which can avoid the disadvantages of these two kinds of methods, combine the advantage of object-based methods and pixel-based methods, and make a tradeoff between these two methods. We validate the combined approach using two sets of airborne different spatial resolution imagery: TM data and Ziyuan3 image. The results indicate that the proposed method generated highest classification accuracy and best visual effect for the classification map. This study demonstrated the potential of the sub-block approach.