Mamba-YOLO: A Real-Time Edge Detector for UAV-Based Early Wildfire Detection

Wildfire detection in unmanned aerial vehicle (UAV) imagery requires high recall at a controlled false-alarm rate. This study compares YOLO26-Nano and Mamba-YOLO using a source-aware grouped split of 21,527 Tiny Fire images; the main test set contains 4368 images and 5000 effective ground-truth annotations. Recoverable temporal metadata and visual similarity inform the grouping, but missing original video and event identifiers prevent verification of video-level independence. Mamba-YOLO integrates Mamba-style state-space feature processing, spatial token realignment, a four-scale YOLO26-derived detection head, and small-target loss weighting. An adaptive threshold provides an inference-time decision policy, evaluated separately from confidence-sweep metrics. Both models use random initialization and a matched 100-epoch training protocol. The reported three-seed comparison gives Mamba-YOLO 66.78±0.12% mAP@0.5 and 36.60±0.15% mAP@0.5:0.95, compared with 64.87±0.12% and 35.75±0.15% for YOLO26-Nano; uncertainties denote sample standard deviations. Mean descriptive test FROC recall increases by 0.80, 0.89, and 0.75 percentage points at FPPI budgets of 0.1, 0.2, and 0.3. These differences compare complete model configurations and do not isolate an individual component. The efficiency comparison gives 50.0 versus 48.8 FPS on Jetson Orin Nano for Mamba-YOLO and YOLO26-Nano, respectively, indicating comparable inference efficiency under the evaluated conditions.