Cloud Detection for Pleiades and SPOT 6/7 Imageries Using Modified K-means and Deep Learning

Abstract

Cloud detection is one of the important stages in optical remote sensing activities as the cloud's existence interferes with the works. Many methods have been developed to detect the cloud, but it is still a few methods for high-resolution images, which mostly have limited multispectral bands. In this paper, a novel method of cloud detection for the images is proposed by integrating an unsupervised algorithm and deep learning. This method has three main steps: (1) pre-processing; (2) segmentation using modified K-means; and (3) cloud detection using CNN. In the segmentation step, an unsupervised algorithm, K-means is modified and used to divide pixels values into k clusters. Our modified K-means method can separate thin clouds from relative bright objects in gray clusters that will be grouped into potential cloud pixels. Afterward, a design of convolutional neural network (CNN) is used to extract the multiscale features from each cluster and classify them into two classes: (1) cloud, which consists of thin cloud and thick cloud, and (2) noncloud. The potential cloud area from the first step is used for guiding the result of CNN to provide accurate cloud areas. Several Pleiades and SPOT 6/7 images were used to test the reliability of the proposed method. As a result, our modified K-means has an improvement to increase the accuracy of the results. The results showed that the proposed method could detect cloud and non-cloud accurately and has the highest accuracy of the results compared to the other methods.