A FCM algorithm for remote-sensing image classification considering spatial relationship and its parallel implementation

Fuzzy C-Means clustering is one of the most perfective and widely used algorithms based on objective function for unsupervised classification. Considering the spatial relationship of pixels when it is used in remote sensing imagery, Neighbor-based FCM algorithm is put forward with the method of modifying the value of fuzzy membership degrees with the neighbor information during the clustering iterations. We use dominant class, if it can be determined in a fixed neighbor region, or the weighted parameters based on the distance of neighbors to perfect the membership degrees of central pixel. Then parallel implement for the algorithm is also proposed by taking account into the communication complexity and the spatial relationship for image partition. In the end, the experimental data indicate the efficiency of the algorithm in decreasing the amount of clustering iterations and increasing the classified precision; the parallel algorithm also achieves the satisfied linear speedup.