AMCA-YOLOv11: an efficient oriented object detection framework with adaptive multi-scale fusion and context-aware attention for remote sensing engineering applications

Abstract

Real-time oriented object detection in remote sensing imagery is a fundamental engineering challenge for applications such as autonomous surveillance, urban infrastructure monitoring, and disaster response systems. Detecting arbitrarily rotated targets under complex aerial conditions requires jointly addressing multi-scale feature representation, background clutter suppression, and angular regression stability. This paper presents AMCA-YOLOv11, an efficient single-stage detection framework engineered upon the YOLOv11 architecture, integrating three purpose-designed modules to overcome these engineering bottlenecks. The adaptive multi-scale feature fusion module introduces trainable per-level scalar coefficients with fast normalization-based weighting to dynamically balance hierarchical semantic and geometric feature contributions, effectively mitigating the spatial information dilution that degrades small-target detection performance. A context-aware attention mechanism is incorporated as a cascaded channel-spatial refinement filter positioned after multi-scale aggregation, systematically attenuating background noise from cloud cover, shadow regions, and terrain textures while enhancing discriminative foreground representations. The regression objective is further optimized through a Kullback–Leibler divergence-based rotation-sensitive loss function that models oriented bounding boxes as 2D Gaussian distributions, resolving angular boundary discontinuities and periodic ambiguity inherent in conventional angle-based parameterizations. Comprehensive evaluation on two standard benchmarks (DOTA-v1.0 and HRSC2016) demonstrates that AMCA-YOLOv11 achieves mAP scores of 79.84% and 90.47% respectively at 68 FPS on a single NVIDIA RTX 3090 GPU, with only 9.82 M parameters and 23.1 G floating-point operations. Component-wise ablation, hyperparameter sensitivity analysis, and computational profiling confirm that the proposed modules deliver significant accuracy gains at marginal computational overhead, establishing a practical and deployable solution for real-time aerial image analysis in engineering systems.