A Fast Method for Cloud Removal and Image Restoration on Time Series of Multispectral Images

Missing information in optical remote sensing data due to sensor malfunction or cloud cover reduces their usability. Although several methods in the literature allow an accurate image restoration, they are usually time-consuming and thus impractical for operational applications. In this paper, we propose a Fast Cloud Removal Approach (FCRA) that exploits the multitemporal information in an efficient way to sharply reduce the computational burden. In greater details, the proposed unsupervised method: (i) generates a short time series (TS) of cloud free images temporally close to the cloudy one (i.e., the target image), (ii) detects ground-clear pixels in the target image similar to the missing pixels using the multitemporal information (i.e., similar temporal patterns), and (iii) restores the occluded pixels using the most similar ground-clear spectral values from the same image. The method reduces the computational effort by focusing on short image TS and by detecting similar temporal patterns in a fast and effective way by using the KD-Trees. Moreover, since the method restores the missing pixels with ground clear pixels belonging to the same image, no computational demanding approaches are needed to adjust the replacement information (i.e., regression). Experimental results obtained on Sentinel-2 images on simulated clouds acquired over two test areas located in Italy demonstrate the effectiveness of the proposed method.