General Semi-supervised Possibilistic Fuzzy c-Means clustering for Land-cover Classification

Satellite images with the advantage of wide coverage, short update times can help to establish land-cover maps quickly and efficiently. However, due to the influence of natural conditions, satellite images often contain noise, outliers, the boundary of the objects on the image is unclear and this makes it difficult for many clustering algorithms. The possibilistic fuzzy c-means clustering (PFCM) algorithm has advantages of both fuzzy c-means clustering (FCM) and possibilistic c-means clustering (PCM) algorithms due to the simultaneous use of both fuzzy and function functions, but it also has limitations such as sensitivity with noise and outliers. The paper proposes a general semi-supervised possibilistic fuzzy c-means clustering (GSPFCM) algorithm to improve the clustering quality of PFCM. Our proposed method can solve problems that labeled data has very little compared to unlabeled data. Results of land-cover classification using satellite images (Landsat-7 ETM+, Sentinel-2A) show that the proposed method can significantly improve the accuracy of classification results when compared to some previous methods.