Reconstruction of cloud-contaminated satellite remote sensing images using kernel PCA-based image modelling

The presence of clouds can restrict the potential uses of remote sensing satellite imagery in extracting information and interpretation. Automatic detection and removal of clouds which hide significant information in the image is an important task in remote sensing. Hence, our aim is to detect clouds and restore the missing information in order to make the image ready for further analysis and applications. Due to the difference in nature and appearance, thick and thin clouds are dealt separately. Thick cloud is detected using an efficient Fuzzy C-Means (FCM) clustering algorithm, while thin cloud is detected using a simple region growing technique. In order to reconstruct the missing pixels, we utilize the prior knowledge about the statistics of the specific image class. Kernel principal component analysis (KPCA)-based image model is obtained using a set of training images. Missing area in the image is restored after an iterative projection operation and gradient descent algorithm. In short, an image lying out of the modelled image space is iteratively modified to obtain the restored image and that would be in the image space according to the obtained nonlinear low-dimensional and sparse KPCA image model. To illustrate the performance of the proposed method, a thorough experimental analysis on FORMOSAT multi-spectral images is done using MATLAB platform. When compared to the two recent existing techniques, our proposed method is superior and makes a promising tool for thick and thin cloud removal in multi-spectral satellite images.