UTC-sparse

We introduce UTC-Sparse, a sparsely annotated urban tree canopy dataset derived from the fully labeled UTC benchmark. The original UTC dataset contains 178 high-resolution aerial images with pixel-level annotations. To support research on sparse supervision, we construct a sparse version in which only the training and validation sets are sparsified, while the test set retains full annotations for unbiased evaluation. Sparse labels are generated by applying randomly sampled masks\u2014composed of point-, line-, and polygon-like patterns\u2014to selectively remove or retain portions of the original annotations. In addition, the background class is manually sparsified to simulate realistic annotation practices in large-scale canopy mapping. The resulting dataset preserves the structural diversity and spatial complexity of urban tree canopies while providing a controlled and highly flexible benchmark for evaluating sparse supervision, pseudo-label generation, and boundary-aware segmentation methods. UTC-Sparse serves as a practical and realistic dataset for developing algorithms under limited supervision and supports broader research on remote sensing segmentation where dense labels are costly to obtain.