We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Improving cross-flight plastic litter segmentation with ConvNeXt V2 U-Net for UAV SWIR hyperspectral imagery
Summary
Scientists have improved a drone-based imaging technique that spots plastic litter in the environment (like on beaches or in waterways) even when lighting and weather conditions change between flights—a common problem that used to trip up earlier detection systems. Better plastic-tracking tools like this matter because they help researchers and cleanup crews find and remove pollution before it breaks down into microplastics, which can enter water supplies, food chains, and ultimately our bodies. While this study focuses on environmental monitoring rather than direct health testing, catching plastic pollution earlier and more reliably is a key step toward reducing our overall exposure to
Plastic litter represents a persistent form of environmental pollution, and accurate spatial monitoring is essential for ecosystem assessment, pollution mitigation, and evidence-based management. UAV short-wave infrared hyperspectral imaging can detect plastic materials with high spatial and spectral detail, but its operational use remains limited by strong variability between acquisitions collected on different flight days, where changes in illumination, exposure, weather, and surface background introduce cube-specific radiometric differences that reduce cross-flight generalisation. This study addresses plastic litter segmentation as both a remote sensing and data harmonisation problem by standardising each hyperspectral cube using its own spectral statistics before segmentation, reducing acquisition-specific bias without relying on training-set statistics. On this harmonised input, we introduce a compact ConvNeXt V2-based U-Net adapted to the 60 retained SWIR bands, with pyramid pooling to incorporate multi-scale contextual information and CBAM decoder refinement to improve background rejection. The framework is evaluated on a public multi-flight UAV SWIR benchmark and compared with PSPNet, SegFormer, DeepLabV3, U-Net, UNet++, and our previous attention-gated residual U-Net. Compared with our previous work, the proposed model improves the mean Dice score from 64.6% to 78.2% and the mean IoU from 49.0% to 64.5%; under a matched retraining protocol that equalises the training-data budgets of the two models, the mean leave-one-cube-out IoU rises from 0.37 to 0.69. The improvement is especially important under difficult acquisition conditions, where the previous model was more affected by overexposure, reduced spectral contrast, and background confusion. The incremental architectural study further shows that pyramid pooling and CBAM refinement improve the balance between plastic detection and background rejection, increasing precision, specificity, AUC-ROC, AUC-PR, and IoU compared with the ConvNeXt V2 baseline on a challenging out-of-distribution cube. Multi-seed analysis on the development cubes shows low sensitivity to random initialisation, indicating that variation across these runs is driven more by acquisition conditions than by initialisation. Overall, the proposed framework provides a compact and transferable approach for UAV-based hyperspectral plastic litter monitoring under realistic cross-flight conditions.