Abstract
In the environments of the North and South Poles, ice floe drifts lead to excessive errors in the GPS and real-time kinematic (RTK) signals of the unmanned aerial vehicle (UAV)-based cross-domain collaborative monitoring systems deployed on these floes, which makes it impossible to complete related tasks such as UAV return and landing. Moreover, deep learning-based target detection methods suffer from insufficient accuracy and real-time performance when addressing low-contrast, fog- and snow-obscured, and high-elevation polar images. Therefore, this study integrates transformer-based convolution (TransConv) into You Only Look Once (YOLO)v11 to strengthen the ability of the network to model the global information contained in fog- and snow-obscured images and to achieve improved target detection performance in complex environments. A mist global feature pyramid network (MistGFPN) is designed to focus on the ability to extract features from small targets to improve the target detection performance achieved at high altitudes. An efficient asymmetric detection head (EADH) is proposed to improve the frames per second (FPS) and real-time detection performance of the target detection model. Finally, several comparative experiments are designed to verify the effectiveness and real-time performance of the proposed model. The experimental results obtained on homemade datasets show that the proposed model (YOLO-TME) improves the accuracy, recall, average accuracy, and F1 score metrics by 4.2%, 5.8%, 5.5%, and 5.1%, respectively, over those of the original YOLOv11 model, which indicates that YOLO-TME provides substantially improved detection accuracy while maintaining high detection speed, thus satisfying the real-time landing sign detection requirements imposed in polar environments.