Segmentation and Spatiotemporal Analysis of Martian Dust Storms Using a U‐Net Based Model With Attention Modules

ABSTRACT Martian dust storms have been shown to cause profound changes in various aspects of the atmosphere and may impact the safety of Martian exploration missions. Identifying dust storms from remote sensing images and analyzing their spatiotemporal distribution is crucial for understanding the origins of dust storms and assessing their impacts. Applying deep learning methods can greatly enhance the efficiency of the identification. This study presents a deep learning‐based approach for segmenting Martian dust storms in the Mars Daily Global Maps (MDGMs) using the U‐Net architecture with InceptionV3 backbone, which was trained on dust storm labels generated from the Mars Dust Storm Sequence Dataset (MDSSD). First, we compared the U‐Net model with five different backbones (InceptionV3, DenseNet169, EfficientNetB3, VGG16, and ResNet50) and found that the InceptionV3 backbone delivered optimal results. Second, to enhance the model's attention on the most relevant channel features, emphasize the key regions where dust storms occur, and reduce attention to irrelevant areas such as those with artifacts and missing data, the Convolutional Block Attention Module (CBAM) was integrated into each skip connection of the U‐Net architecture with InceptionV3 backbone. The proposed model achieved an overall accuracy of 0.9411, a macro F 1 score of 0.8100, and a dice similarity coefficient of 0.6542 on the testing dataset. The temporal and spatial distribution of dust storms, based on the segmentation results from the proposed model, revealed that these storms exhibit consistent seasonal patterns across Martian years (MY32, MY33, and MY35). In Ls = 0°–135°, dust storms are generally distributed around the polar caps, with occurrence across various longitudes. In Ls = 245°–295°, they are mainly concentrated near the south polar cap. In both periods, the daily total area of dust storms is typically below 5 × 10 6 km 2 . During Ls = 135°–245° and Ls = 295°–360°, dust storms increase in size and frequency, with total areas exceeding 1 × 10 7 km 2 on some sols, covering more than 7% of the Martian surface. These storms often originate in Acidalia or Utopia and are transported southward. Major peaks in column dust optical depth during the dusty season are associated with large dust storms occurring between Ls = 185°–245° and Ls = 295°–360°, frequently observed in the Argyre‐Solis‐Valles Marineris (ASV) and Hellas regions. These temporal and spatial distribution characteristics align with manual identification results from remote sensing images, demonstrating the capability of the proposed model for segmenting Martian dust storms on a global scale, promoting the advancement of fully automated identification of Martian dust storms from remote sensing images.