Abstract: This dataset provides high-resolution semantic segmentation maps ofaeoliansand dunes across five major Mars-analog regions on Earth: the Namib Desert (Africa), Atacama Desert (South America), Mojave Desert (North America), Qaidam Basin (Asia), and Yilgarn Craton (Australia). These regions are critical terrestrial analogs for studying Martianaeolianprocesses.
The classification results were generated using a novel Vision-Language Model (SegFormer with Text-Aware Fusion) trained on Google Satellite Embeddings (64-dimensional feature vectors) and expert geomorphological knowledge. The model achieves an Overall Accuracy (OA) of 96.11% and a Kappa coefficient of 0.9418. This dataset offers a detailed inventory of dune types, supporting comparative planetary studies and Earth system science research.
Data Content: The dataset consists of 5 GeoTIFF (.tif) raster files and 1 Excel file:
Pixel Value Definitions (Classification Scheme): The raster values correspond to the following dune categories:
Technical Details: