Superpixel Context Description based on Visual Words Co-Occurrence Matrix

Abstract

In this paper, we introduce a novel representation to encode contextual information in object-based remote sensing image classification problems. The solution relies on the creation of a visual codebook and its use to compute the co-occurrence of visual words within a superpixel and within its neighboring regions. Performed experiments on the well-known collections (grss_dfc_2014 and ISPRS Potsdam) demonstrate that the proposed approach is effective, yielding comparable or better results than several baselines.