An automatic cloud region classification of satellite image by using clustering in local area

This paper proposes a technique for automatically classifying cloud regions in satellite images. The proposed technique uses visible images and brightness temperature images to automatically classify cloud and noncloud regions. The authors assume that dividing the image into small regions (local regions) enables the feature vector distribution of each category within a local region to be approximated by using a Gaussian distribution, and they divide the distribution into two clusters. In addition, they introduce the "degree of coincidence" for judging whether or not it is appropriate to use each cluster for subsequent category classification and use the mean vectors of clusters that satisfy the degree of coincidence and the EM algorithm to perform classification. By using the degree of coincidence, the authors show that the vector distribution within the feature space becomes distinct and that clustering can be suitably performed. A feature of the proposed technique is that it aims to distribute feature vectors continuously so that they link each category. By considering the straight lines linking each category and using only those straight lines that are close to the mean vectors of each category, threshold values are determined dynamically for each image. Finally, the authors apply the proposed algorithm to NOAA and Himawari (GMS) images. Good classification results are obtained compared with results that were obtained manually by specialists.