GSWSeg4MM: a global solid waste segmentation dataset for multi-modal learning from VHR imagery across 264 cities

This study provides a Global Solid Waste Segmentation dataset for Multi-Modal learning from VHR satellite imagery, i.e., GSWSeg4MM, which contains about 103k multi-modal vison-language samples (i.e., about 20k positive samples and 83k negative samples) collected in 264 cities worldwide with a total of 84k solid waste instances and text descriptions comprising 5.49 million words. Specifically, the satellite image patch has a size of 512×512 pixels with a spatial resolution ranging from 0.3 to 2 meters.

The solid waste mask was generated using X-AnyLabeling assisted by SAM. Meanwhile, textual descriptions were annotated to provide details of the solid waste’s color, morphology and contextual environment within the satellite imagery, aiming to guide the accurate delineation of solid waste boundaries. To the best of our knowledge, GSWSeg4MM is the first open-access multi-modal dataset used to train vision-language foundation models (VLMs) in solid waste remote sensing field.