Detection of Micro-UAVs in Cluttered Multispectral Scenes Through Attention-Driven Cross-Modal Fusion

To address the critical challenge of insufficient detection accuracy for micro-unmanned aerial vehicles (UAVs) which relying solely on visible-light images cannot achieve effective detection under large-scale complex scenarios with diverse backgrounds, this study pioneers a paradigm shift. We present Kilo-UAV, the first multi-spectral micro-UAV detection dataset specifically designed for kilometer-range imaging, comprising 21,940 synchronously acquired multi-spectral remote sensing observations (visible-light, near-infrared, polarized imagery, and atmospheric attenuation-corrected data). The dataset focuses on UAV targets characterized by sub-meter dimensions and minimal pixel occupancy (average pixel area ratio<0.7%), systematically encompassing 10 types, including urban built-up areas, coastal zones, and montane forests. To resolve the critical bottleneck of multi-source remote sensing data fusion, we propose the Weight-Adaptive Decision Fusion Network for Remote Sensing (ADF-RS). This architecture integrates an anomaly-modality sensitivity assessment module to dynamically optimize sensor-specific contribution weights, thereby enabling spatially coordinated representation of multi-spectral features. Experimental validation demonstrates that ADF-RS achieves state-of-the-art performance with an 86.57% detection accuracy (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mAP</i><sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">50</sub>) on the test set, outperforming conventional visible-light-based approaches by a 23-percentage-point margin, while maintaining real-time processing efficiency (57 FPS on NVIDIA RTX3090 hardware). We further propose a lightweight computational framework optimized for edge computing devices, offering an embeddable solution for UAV airspace regulation within existing remote sensing infrastructures. The dataset’s inclusion of multi-scale geographical scenarios supports extensive derivative studies in surveillance technology development. All associated datasets and implementation codes will be open-sourced via GitHub to accelerate advancements in anti-UAV remote sensing technologies.