Optical-SAR Fine-Grained Zero-Shot Learning

The time range of the integrated data is from 2017 to 2024, with a geographic coverage mainly covering regions such as Asia, Europe, and North America, and a spatial resolution of better than 20m. The data generation process mainly includes target instance pruning, quality screening, category semantic normalization, and data organization. Firstly, based on the annotation file of the original dataset, crop a single target image from the remote sensing image; Subsequently, samples with poor quality or incomplete targets are removed, and semantically consistent but named categories from different data sources are unified under standard category names, ultimately forming a category level aligned SAR optical cross modal dataset.

The dataset contains a total of 147302 target images, including 66384 SAR images and 80918 optical images, covering 17 fine-grained target categories. The data formats include. png,. jpg, *. tif, and *. tif. The data directory adopts a three-level structure of "category modality image" organization, that is, SAR and Optical subdirectories are set separately under each target category to store the target image samples of the corresponding modality.