Wildfire reignition after suppression remains a critical challenge in disaster response, as residual embers can persist beneath soil, foliage, or debris and later re-ignite into secondary fires. Detecting these embers is difficult because they produce weak, small-scale thermal signatures that conventional thermal surveillance systems cannot reliably identify. To address this gap, this study focuses on the post-suppression stage of wildfire management and proposes two lightweight thermal object detection models optimized for residual ember monitoring. Both models adopt a Bidirectional Feature Pyramid Network structure to enhance multi-scale feature fusion. The first model, YOLOv8-Nano + BiFPN, achieves a balanced trade-off between detection accuracy and inference speed, while the second model, MobileNetV3 + BiFPN, emphasizes real-time performance in low-power edge environments. Experiments conducted on the RWTH Thermal Drone Dataset demonstrate performance improvements of +3.4% in mAP, +3.0% in precision, and +4.4% in recall compared to the baseline YOLOv8-Nano. Furthermore, real-world validation using drone-mounted Jetson Nano and Xavier NX platforms confirmed real-time operation at 11–13 FPS with power consumption below 5 W. These results demonstrate that the proposed framework effectively achieves high accuracy, energy efficiency, and real-time responsiveness, providing a practical and deployable solution for preventing wildfire reignition through post-suppression ember surveillance.