Automated segmentation of textured dust storms on mars remote sensing images using an encoder-decoder type convolutional neural network

Abstract

We propose a method for detecting Martian dust storms and recognizing their size and shape on remote sensing images. The method is based on a convolutional neural network, one of algorithms that use deep learning for image categorization and recognition. We trained models with three different structures using images of two regions of Mars in visible wavelengths observed over several seasons, together with ground truth images manually prepared by the authors that give the true shapes of the dust storms. The two regions were the western Arcadia Planitia in the northern hemisphere and the Hellas Basin in the southern hemisphere, both of which are areas where high dust storm activity has been observed. The case study showed that models trained on images of the Arcadia Planitia tended to perform better than comparable models trained by images of the Hellas Basin. While third models trained by images of both regions showed little degradation relative to the dedicated models when tested on image of the Arcadia Planitia, their performances clearly decreased in the case of the Hellas Basin. Furthermore, the performance degradation was more pronounced for a model with moderate depth than for a deepest model. This is partially because the Hellas Basin is brighter than the adjacent areas throughout the year and high optical thickness of dust in its interior makes the textures of dust storms relatively unclear. In contrast, any models showed comparable performances in dust storm segmentation in the Arcadia Planitia and mixing data from the two regions with completely different surface patterns produced only a slight degradation of performance. It suggests that training the model with images from various regions may yield a region-independent model that can be effectively applied to the segmentation of dust storms over a wide area.