DUSTNet: An Unsupervised and Noise-Resistant Network for Martian Dust Storm Change Detection

Abstract

Mars exploration highlights the demand for identifying Martian surface changes, which has sparked research interests in planetary surface changes detection (PSCD). However, the prevailing PSCD algorithms face significant challenges due to the sparse features, low resolution, and high noise levels of captured images data. In this letter, we propose an unsupervised model, the dust unsupervised surface tracking network (DUSTNet), designed to track the surface changes caused by Martian dust storms. Our DUSTNet employs a network architecture with dual input branches to learn the cross-temporal complementary information from pretime and posttime image pairs. A multilevel feature complementary fusion (MFCF) module is utilized to enhance the ability to detect subtle changes. Considering the difficulties in image registration caused by illumination variations, noise, and other factors, we design a noise-resistant module (NRM) that mitigates pseudo-changes and improves the robustness of PSCD. In addition, we construct a dataset of Martian dust storms change detection (CD) based on the images captured by moderate resolution imaging camera (MoRIC) of China’s First Mars Mission TianWen-1 (the dataset is available at https://github.com/Limiyu1123/SDS). The detection performance of DUSTNet performs well on multiple Mars surface datasets, including our Martian dust storm test set. Our model achieves improvements of 2.5% in precision, 7.55% in $F1$ -score, 6.54% in overall accuracy (OA), and 4.57% in Kappa over the state-of-the-art model.