Mapping Lichen Cladonia from UAV Imagery Using Deep Learning and Active Domain Adaptation in Canada’s Boreal and Taiga Ecozones

Lichen Cladonia is the primary source of forage for caribou (Rangifer tarandus) during winter. Accurate, intelligent mapping of Cladonia lichens is essential for effective habitat management across northern Canada’s diverse ecozones. This study develops a robust deep learning framework for the semantic segmentation of lichen Cladonia from ultra-high-resolution UAV imagery (1 cm), addressing key challenges related to label scarcity and domain adaptation. A comprehensive annotated dataset was constructed from UAV orthomosaics in the Boreal Shield ecozone using Gaussian Mixture Model clustering refined by expert editing. Three U-Net-based architectures, Attention U-Net, Recurrent Residual U-Net (R2U-Net), and Recurrent Residual Attention U-Net (R2AU-Net), were trained and evaluated. The R2AU-Net model achieved the highest performance on independent Boreal Shield test sites (F1-score: 83.28%, IoU: 71.34%). To enhance model transferability to ecologically distinct regions, an active domain adaptation framework was implemented, combining entropy-based uncertainty sampling with fine-tuning on selected Taiga Shield imagery. Fine-tuning increased the model’s F1-score from 59.70% to 81.61% and reduced misclassification of grey lichen and spectrally similar substrates such as bare ground and rock, while saving over 90% of annotation costs. The results demonstrate the feasibility of scalable, adaptable lichen mapping in complex northern landscapes, providing a transferable methodology for supporting large-scale ecological monitoring and habitat management.