ThermalBird-Mamba: Thermal adaptive state space learning for real time UAV bird detection under low contrast and cluttered backgrounds.

UAV thermal imaging offers a non-invasive tool for wildlife monitoring, but detecting small birds remains difficult because targets often appear as faint heat signatures under low contrast, vegetation occlusion, motion blur, and thermally cluttered backgrounds. This study proposes ThermalBird-Mamba, a real-time thermal bird detection framework that combines a MambaVision state-space backbone, a Thermal Adaptive State Space module, and a lightweight multi-scale detection head with a dedicated P2 layer for small-object localization. The model was evaluated on an existing UAV thermal bird dataset containing 5847 annotated thermal images and 8927 bird instances, including challenging seasonal and background conditions. ThermalBird-Mamba achieved a test mAP@50-95 of 0.712, precision of 0.978, recall of 0.983, and 41.3 FPS on an NVIDIA Jetson Orin Nano. Compared with YOLOv11n, the gain in mAP@50-95 was 0.081 absolute, equivalent to a 12.8% relative improvement. Additional ablation and visual analyses indicate that the thermal-adaptive module and P2 layer improve small-object localization and reduce responses to thermal clutter. These results suggest that ThermalBird-Mamba is a promising candidate for UAV-assisted ecological monitoring, although broader operational validation across independent field deployments remains necessary.