Artificial neural networks for cloud masking of Sentinel-2 ocean images with noise and sunglint

Cloudy regions in optical satellite images prevent the extraction of valuable information by image processing techniques. Several threshold, multi-temporal and machine learning approaches have been developed for the separation of clouds in land and ocean applications, but this task still remains a challenge. Concerning deep water marine applications, the main difficulties are imposed in regions with high noise levels and sunglint. In this study, artificial neural networks (ANNs) with different configurations are evaluated for the detection of clouds in Sentinel-2 images depicting deep water regions with several noise levels. The ANNs are trained on a manual public dataset and on a manual dataset created for the needs of this study, which authors intend to make publicly available. Results are compared with the cloud masks produced by three state-of-the-art algorithms: Fmask, MAJA, and Sen2Cor. It was shown that the ANNs trained on the second dataset perform very favourably, in contrast to the ANNs trained on the first dataset that fails to adequately represent the spectra of the noisy Sentinel-2 images. This study further reinforces the value of the 'cirrus' band and indicates the bands that mitigate the influence of noisy spectra, by defining and examining an index that characterizes the importance of the bands according to the weights produced by the ANNs. Finally, the possibility of improving results by making predictions using the feature scaling parameters of the test set instead of those of the training set is also investigated in cases where the test set cannot be adequately represented by the training set.