Real-time cloud detection in high resolution images using Maximum Response Filter and Principle Component Analysis

In order to maximally make use of the limited capacities of storage and transmission, it's necessary for onboard computer to adaptively compress the cloud regions with lower quality compared with the regions without cloud cover. Therefore, the information processing unit needs to recognize the cloud regions before compress the image. To meet this requirement of satellite imaging payload, a novel approach for real time cloud detection is proposed. First, the visual dictionary is learnt from the training features extracted using Maximum Response (MR) Filter. Second, Principle Component Analysis (PCA) is utilized to reduce the dimensions of the visual words for the quick word search. Third, the MR feature of an image patch is converted into the histogram of visual word, in which the MR feature of a pixel is replaced by the index of the most similar visual word. Finally, the histogram is fed into the trained SVM classifier to detect cloud patch. The experimental results verify that the proposed approach can highly precisely detect the cloud region in patch unit.