CryoFlora: Automated Instance Segmentation of Cushion Plants for Species-Level Monitoring in Moraine–Talus Ecosystem

Cushion plants, renowned for their resilience to alpine and arctic extremes, remain challenging to delineate at the species level due to dense interwoven canopies and high spectral similarity among species. This study introduces CryoFlora, an instance segmentation framework tailored to cushion species, integrating a large-kernel attention backbone, UniRepLKNetBlock, with a bidirectional attention-based fusion neck, BIMAFPN, to enhance feature representation and boundary precision. CryoFlora was trained on 528 annotated UAV image tiles containing 17,488 instances of Thylacospermum caespitosum and Rhodiola rosea, representing the two principal morphological forms of cushion plants, flat and hemispherical. The model achieved a mean average precision of 0.975 at IoU 0.5 and 0.785 at IoU 0.5–0.95, with only 7.65 M parameters. Field validation across four 100 × 100 m plots in the Bayi Glacier forefield confirmed its ability to map total cover (575.5–999.3 m2), patch density (0.14–1.48 ind/m2), and individual canopy metrics, revealing clear elevational gradients in species dominance and morphology. By automating extraction of ecological indicators directly from UAV imagery, CryoFlora provides a scalable tool for dynamic monitoring of glacier-forefield ecosystems, supporting conservation and adaptive management under accelerating climate change.