Early detection of woody plant encroachment in Canadian prairies using UAV imagery and transformer-based deep learning

Accurate and early detection of rapid Woody Plant Encroachment (rWPE) in grasslands is critical for management and conservation. However, this task remains challenging due to the spectral and spatial complexities of multi-species grassland ecosystems. This study evaluates the potential of UAV-RGB imagery and deep learning algorithms for early detection and classification of three dominant woody species (Wolf Willow – Elaeagnus commutata , Western Snowberry – Symphoricarpos occidentalis , and Trembling Aspen - Populus tremuloides ) in the Canadian Prairies. Five semantic segmentation models, including three CNNs (PSPNet, DeepLabV3+, UNet) and two Transformers (SegFormer and Mask2Former), were assessed in Foam Lake Community Pasture. The results indicate that Transformers outperformed CNNs, with the largest SegFormer model (MIT-B5) achieving the highest overall accuracy (92.5 %), mean IoU (68.2 %), and F1-score (79.8 %). Transfer learning improved the model performance in SegFormer by more than 5 % in the mF1-score and 7 % in mIoU. A lightweight variant (MIT-B1) balanced high accuracy (79.2 % F1-score) with high speed (17.4 fps). Spatial resolution degradation (from 0.73 cm to 3 cm) reduced detection accuracy for small shrub patches (diameters ∼10–20 cm), while showing minimal impact on larger patches (diameters >1 m). SegFormer exhibited superior capability in distinguishing woody species using high resolution imagery, even at early growth stages. Our findings highlight the effectiveness of Transformers and high-resolution UAV imagery for precise woody species mapping, offering scalable solutions for grassland conservation and monitoring. • Vision Transformers outperform CNNs significantly in early woody species detection. • SegFormer achieves 92.5 % accuracy for grassland woody encroachment detection. • Transfer learning boosts SegFormer accuracy by 5–7 % with limited training data. • Spatial resolution <3 cm/pixel critical for small shrub detection (diameter < 20 cm). • Framework aids scalable grassland conservation via UAV monitoring.