A Deep Learning Framework for Emergency Drone Landing Zone Detection

This research presents a novel, research-grade pipeline for drone landing safety assessment that integrates a custom YOLOv8s object detector with advanced risk-aware modules. The system addresses real-world challenges such as sensor noise, terrain variability, and environmental drift through several innovations: class-specific risk weights for detected objects (Vehicle, UAP, UAI, Person), Sensor Noise Augmentation embedded during YOLO training, and a Risk-Aware Head directly hooked to YOLO backbone (C2f) features. Three specialized neural networks—TerraSense-Net (terrain roughness estimation), DriftGuard-Net (wind-drift risk prediction), and FusionVerdict-Net (final SAFE/RISKY/UNSAFE classification)—are trained on synthetic labels derived from detection outputs and image statistics, achieving high evaluation metrics (FusionVerdict-Net: ~99.86% accuracy, strong F1 scores).A production-ready Flask web application operationalizes the pipeline, enabling authenticated users to upload aerial images. The system performs relevance filtering, YOLO detection, risk mapping, novelty model inference, and generates annotated visualizations with comprehensive safety verdicts. This end-to-end solution bridges cutting-edge computer vision research with practical deployment for safer autonomous drone operations.

Keywords: Drone Landing Safety, YOLOv8, Risk-Aware Detection, Sensor Noise Augmentation, TerraSense-Net, DriftGuard-Net, FusionVerdict-Net, Risk-Weighted Classification, Computer Vision, Autonomous Systems.