A nine-class UAV semantic segmentation dataset of Solidago canadensis in the rural-urban fringe of Daishan Island, Zhoushan, China

Description

This dataset contains 7,209 UAV image tiles (256 x 256 pixels, RGB JPEG) of Daishan Island, Zhoushan, Zhejiang Province, China, with pixel-level annotations of nine land-cover classes, including the invasive plant Solidago canadensis. The orthomosaic was acquired on 25 October 2025 at 120 m flight altitude (ground sampling distance about 3.4 cm), during the late flowering stage of the species. Annotations were drawn manually in Labelme following one set of boundary conventions, with the classes: background, S. canadensis, discarded construction materials, other vegetation, concrete pavement, bare ground, buildings, vehicle roofs and asphalt pavement. The deposit includes the tile-level training/validation/test assignment (spatially contiguous blocks, so that no training tile is adjacent to a test tile; 5,046 / 1,442 / 721 tiles), the class definitions and per-class pixel statistics computed from the masks, the acquisition metadata, the PyTorch training pipeline used for all experiments, and trained checkpoints of eight segmentation architectures together with the proposed cross-attention configuration.

Institutions

• Hangzhou Normal University

  Zhejiang, Hangzhou

Categories

Licence