Code for deep-learning-based segmentation of scalloped terrain on Mars

Scalloped depressions are important periglacial landforms associated with near-surface ground ice and thermokarst degradation in the Martian mid-latitudes. The code implements a SegFormer-based semantic segmentation workflow (Xie et al., 2021) for identifying scalloped terrain from Context Camera (CTX) images.The source imagery is mainly based on the Murray Lab global CTX mosaic (Malin et al., 2007), with an approximate spatial resolution of 5 m/pixel.

Manually labeled regions were cropped into 1024 × 1024-pixel patches, and the corresponding binary masks label scalloped terrain as foreground and other areas as background. The training dataset was expanded using standard data augmentation, including rotations and horizontal/vertical flips.The model uses the SegFormer-B3 architecture with ImageNet-1K pretrained encoder weights. The package includes scripts for model training, prediction, and evaluation, as well as trained weights and CTX image patches with corresponding binary masks.