Efficiency Cost-Sensitive Loss of Transformer based on Mamba Mechanism for Aircraft Detection in Satellite Imagery

Detecting aircraft in satellite images poses considerable challenges due to complex backgrounds and variable conditions influenced by sensor geometry and atmospheric factors. Despite rapid advancements in deep learning algorithms, their main focus has been on ground-based imagery. This study offers a thorough evaluation and comparison of advanced object detection algorithms specifically designed for aircraft detection in satellite imagery. By leveraging the extensive HRPlanesV2 dataset and a rigorous validation process on the GDIT dataset, we trained a cutting-edge object detection model, YOLO-Mamba, published in June 2024. Additionally, we introduce YOLO-Mamba-TransGhost, which integrates a novel Transformer module SC3T and Ghost Convolution into the YOLO model’s backbone architecture. Furthermore, substituting the WIoU loss function with CIoU in YOLO-Mamba results in significant improvements in accuracy and small object detection. Experimental results on the GDIT dataset indicate that YOLO-Mamba-TransGhost improves mAP@.5 by approximately 2% compared to the original YOLO-Mamba. Similarly, tests on the HRPlanev2 data set reveal a notable reduction in model complexity and an impressive accuracy of 98.7% which is achieved by leveraging a cost-sensitive loss function that dynamically focuses training on higher quality samples, improving convergence and accuracy. Therefore, the proposed YOLO-Mamba-TransGhost model demonstrates superior accuracy and reduced complexity in aircraft detection from satellite imagery, highlighting its potential for practical applications in aerospace monitoring, disaster management, and surveillance systems domain.