Wildlife detection for biodiversity monitoring and forest management in forest ecosystems under limited data using lightweight tri modal fusion
Huaqiang Xu, Hua Li, JJ Zhao · Frontiers in Forests and Global Change · 2026
Forest ecosystems play a fundamental role in sustaining terrestrial biodiversity and regulating global ecological processes, yet effective biodiversity monitoring remains challenging under global change and anthropogenic disturbances. In forest environments, large-scale annotation is often impractical due to data scarcity, dense vegetation, low illumination, and complex backgrounds, which hinder reliable wildlife species identification. Few-shot object detection (FSOD) provides a promising solution for wildlife detection with limited labeled samples. In this study, we propose a lightweight tri-modal semantic-guided FSOD framework to support biodiversity monitoring and forest management under limited data conditions. Specifically, OWLv2 and SAM are coupled to generate semantically aligned structural priors, where OWLv2-guided SAM masks provide biologically meaningful cues for identifying target animal regions. These priors are incorporated into a YOLOv11-based detector through a decoupled guidance perceiver (DGP), in which the semantic attention unit (SAU) introduces semantic-aligned attention and the Lightweight Modulated Residual Adapter (LMRA) performs adaptive residual refinement. A feature-driven adaptive fusion weighting mechanism (FAFW) further balances semantic and visual information according to saliency reliability. Experiments on Raccoon.v2, Kangaroo.v2, Tiger.v1, and Horse.v1 demonstrate consistent improvements over five competitive baselines, with gains of +5.6% [email protected], +6.2% Precision, +7.1% Precision, and +5.5% [email protected], respectively. These results indicate that the proposed method improves localization robustness under ecological challenges such as camouflage, occlusion, cluttered backgrounds, and pose variation. Overall, this work provides a scalable and lightweight tool for automated wildlife detection, supporting species occurrence documentation, long-term biodiversity monitoring, management-relevant ecological assessment.