Multi-temporal cloud detection based on robust PCA for optical remote sensing imagery

Abstract Cloud detection is an essential pre-processing step for optical remote sensing imagery in various applications due to the huge negative effect of cloud occlusion. Multi-temporal cloud detection methods are usually more effective than single-image based methods by providing extra temporal information, which is a worthwhile supplement to spatial-spectral information for distinguishing clouds from clear-sky observations. Nevertheless, most of the existing multi-temporal cloud detection algorithms cannot estimate the cloud-free reference background accurately, especially when pixels have no or very few clear-sky observations in time series data, which limits the performance of cloud detection to a large degree. To deal with this problem, a novel multi-temporal cloud detection method based on robust principal component analysis (MCD-RPCA) is proposed for optical remote sensing imagery. Firstly, several spectral tests are performed to extract spectral features based on the physical attributes of clouds and obtain an initial rough cloud mask. Secondly, a low-rank matrix decomposition model, known as robust principal component analysis (RPCA), is constructed based on multi-temporal images to estimate the cloud-free background. By detecting the change between the cloudy image and the estimated clear-sky background image, a change cloud mask can be obtained by the extraction of temporal information. A refined cloud mask is then acquired by taking the intersection of the initial cloud mask and the change cloud mask. Lastly, multiple spatial morphological processing steps are implemented to incorporate the spatial information to further refine the cloud detection map. By fully exploiting the spectral-temporal-spatial features of clouds in optical remote sensing images, the proposed MCD-PRCA method seeks to facilitate the estimation of the clear background component and support the improved detection of cloudy pixels. The performance of MCD-RPCA was evaluated on four Landsat-8 OLI images from the Biome dataset and three popular cloud detection methods were used as benchmarks for comparisons. In general, a mean overall accuracy of 93.79% was achieved by the proposed MCD-RPCA method, outperforming the other state-of-the-art cloud detection algorithms.