REAL-TIME UAV EARLY SMOKE DETECTION WITH A LIGHTWEIGHT, FEATURE-REFINED DETECTOR

Abstract

Wildfires are highly destructive, making early smoke detection crucial for rapid suppression. Existing UAV-based methods face challenges such as slow inference, limited accuracy, and weak sensitivity to small or diffuse smoke. This study introduces a lightweight YOLOv8-based detector enhanced with (1) Wise-IoU v3 for robust localization, (2) Ghost Shuffle Convolution for reduced computation and real-time efficiency, and (3) BiFormer attention to emphasize smoke cues while suppressing clutter. On a UAV smoke dataset, the model achieves 79.4% AP (+3.3 over baseline), with notable performance on small (71.3%) and large (92.6%) targets. The results highlight improved accuracy and efficiency, supporting practical real-time UAV-based wildfire monitoring.

More by Norkobil Saydirasulov

• Valid but Uninformative: AI Process Alarms Beyond the Commissioned Operating RangeProcesses · 2026
• ACE-LTT reproducibility package: Fail-Closed Certification of Multi-Stage AI PoliciesZenodo · 2026
• rtu-conformal-fdd: analysis code for Conformal Fault Detection for Rooftop Units with Incomplete SensingZenodo · 2026
• A Reproducible Benchmark-Validity Audit and Calibration Study for Cross-Home Fault Diagnosis in Smart-Home Sensor SystemsSensors · 2026 · Citations: 1