A New Method for Semi-Supervised Segmentation of Satellite Images

Satellite image segmentation is an important topic in many domains. This paper introduces a novel semi-supervised image segmentation method for satellite image segmentation. Unlike the semantic segmentation strategies, this method requires only limited labelled data from small local patches of satellite images. Due to the complexity and large number of land cover objects in satellite images, a fixed-size square window is used for feature extraction from 7 different local areas. The local features are extracted by spectral domain analysis. Then, classification is performed based on similarity of the local features to those of the 7 labelled patches. This also allows efficient selection of the suitable window scale. Furthermore, the labeled features remove the need for iterative clustering for decision making about features. The labelled data also allows learning a subspace of transformed features for segmentation of water and green area based on simple thresholding. Comparison of the segmentation results using the proposed strategy compared to unsupervised techniques such as k-means clustering and Superpixel-based Fast Fuzzy C-Means Clustering (SFFCM) shows the superiority of the proposed strategy in terms of content-based segmentation.