A lightweight remote sensing small-object detection approach with scale-based dynamic loss and efficient multi-scale attention

Abstract

Abstract Accurate small-object detection is crucial for automated processes in remote sensing imagery, such as environmental monitoring and aerial surveillance. Yet it remains challenging due to limited feature representation, complex backgrounds, and localization errors. To address these challenges, this work proposes a lightweight detection approach that integrates a scale-based dynamic loss and efficient multi-scale attention. A lightweight backbone, dubbed A2C2fLite, is designed based on the A2C2f module from the most recent YOLO variant to reduce redundant parameters and computational overhead. An EMA module is embedded before the final detection heads to enhance multi-scale feature representation and improve target-background discrimination. The SD loss is used to effectively address scale variation and occlusion, improving localization accuracy beyond existing Intersection-over-Union-based losses. Evaluated on a representative remote sensing dataset, the proposed approach achieves 93.6% precision, 86.6% recall, and 92.6% $$\hbox {mAP}_{50}$$ , with only 2.0M parameters and 5.7 GFLOPs. Compared to the baseline YOLOv13, it achieves a superior accuracy-efficiency trade-off by reducing parameter counts by 20% and computational overhead by 8.1%, while improving $$\hbox {mAP}_{50}$$ by 1.5%. Visualization results demonstrate that it suppresses missed alarms and refines bounding-box localization. By minimizing memory footprint and computational bottlenecks, this lightweight design provides clear practical advantages for real-time small-object detection on resource-constrained edge devices, such as onboard Unmanned Aerial Vehicle systems and embedded aerial sensors. These findings highlight that the proposed approach offers an efficient, robust, and highly deployable solution for complex remote sensing applications.