Self-Supervised Memory-Bank Anomaly Detection for UAV Landing Pads: Whole-Pad versus Grid-Hybrid Representations under Synthetic Occlusions

This preprint presents a self-supervised visual anomaly detection framework for identifying unsafe conditions on unmanned aerial vehicle landing pads. The study focuses on two nominal-only approaches: Wholepad SSL, which represents the complete landing-pad region, and Grid-Hybrid SSL, which divides the region into a fixed 2 × 2 grid and evaluates the most anomalous tile. Both approaches use a ResNet-101 encoder trained with the SimCLR contrastive learning objective. Nominal landing-pad representations are stored in a memory bank, and each query is assigned an anomaly score using one minus its maximum cosine similarity to the nominal embeddings. The controlled evaluation contains 630 clean landing-pad regions and 630 paired regions with procedurally generated synthetic obstructions. The synthetic anomaly generator varies obstruction geometry, coverage, position, aspect ratio, rotation, contrast-aware color, texture, contour, and shadow. Wholepad SSL achieves an AUROC of 0.9236, an F1 score of 0.8583, and an accuracy of 0.8563. Grid-Hybrid SSL achieves an AUROC of 0.7555 and demonstrates substantial sensitivity to decision-threshold calibration. The accompanying GitHub repository contains the IEEE and ACM paper sources, experiment code, workflow diagrams, metric figures, integrity checksums, and machine-readable result summaries. The original image dataset, generated samples, trained model weights, and memory-bank arrays are not included. Author: Tolga Ferhan KüçükAffiliation: Department of Computer Engineering, Kocaeli University, Kocaeli, TürkiyeEmail: [email protected]: https://github.com/tolgafkProject repository: https://github.com/tolgafk/uav-landing-anomaly-detection