Automatic Detection of Scattered Garbage Using Multi-Source Unsupervised Domain Adaptation on Low-Altitude SUAV Remote Sensing Images

In recent years, tourist-related garbage pollution has led to the closure of several prominent high-altitude nature reserves. To support more scientific and modern waste management in these areas, this paper proposes a novel unsupervised domain adaptation framework based on a Multi-source domain Knowledge Transfer (MKT) approach. The framework enables automatic detection of scattered garbage regions using lowaltitude remote sensing imagery captured by small unmanned aerial vehicles (SUAVs). To address the scarcity of labeled remote sensing data, a targeted data augmentation strategy is introduced to reduce domain discrepancies. The framework includes a generation network that produces transitional regions to bridge source and target domains. By leveraging a shallow feature extraction network, it aligns low-level features across domains and further achieves high-level feature alignment through a multi-branch network structure. After feature fusion, image-level and instance-level attention mechanisms are applied to enhance positive cross-domain transfer while mitigating negative transfer effects. Experimental results demonstrate that the proposed method outperforms existing state-of-the- art approaches in both accuracy and overall performance on realworld datasets.