Identifying Brachyura in tidal flats using UAV imaging and a deep-learning approach: a republic of Korea west coast case study

Brachyura taxa serve as key bioindicators of tidal flat ecosystems, yet traditional monitoring methods are labor-intensive, ecologically invasive, and subject to access constraints. Here, we present a non-invasive monitoring approach combining unmanned aerial vehicle (UAV) imagery with a U-Net semantic segmentation model to detect and classify Brachyura species directly on the tidal flat surface. High-resolution images (~ 0.2 cm ground sampling distance) were acquired over the Baramarae tidal flat (Taeanhaean National Park, Republic of Korea), and a dataset of 103,792 images was used to train and evaluate the model for four species: Austruca lactea (nationally endangered), Tubuca arcuata , Macrophthalmus japonicus , and Scopimera globosa , including sex-level discrimination for the first two species. The model achieved overall precision, recall, and F1-scores of 0.84, 0.83, and 0.83, respectively, with a mean intersection over union (mIoU) of 0.67. Classification performance was high for most taxa but was markedly lower for female T. arcuata (IoU = 0.41), primarily due to limited training samples and morphological similarity to M. japonicus . Population estimates were within 18% of manual counts for all taxa. These results demonstrate that UAV-based deep learning can provide rapid, non-contact population assessments of tidal flat Brachyura, with accuracy sufficient for ecological monitoring while minimising habitat disturbance. Future work should focus on expanding training data for underrepresented taxa and integrating multi-temporal imagery to track population dynamics.