LS-GS: A Lightweight Hierarchical Transformer for Binary Plastic Detection in UAV SWIR Hyperspectral Imagery

Detecting plastic litter in natural environments is a key step for directing clean-up actions and monitoring pollution over time. One effective solution is short-wave infrared (SWIR) hyperspectral imaging (900–1700 nm), which captures polymer specific absorption features that are difficult to observe in the visible spectrum, enabling reliable discrimination from natural materials even in cluttered outdoor scenes. SWIR sensors are commonly mounted on Unmanned Aerial Vehicles (UAVs), which offer a broader spatial perspective and can survey areas that are difficult to access on foot. However, many deep models for hyperspectral analysis remain heavy for UAV deployment and often rely on dimensionality reduction and dataset-level normalization, which can weaken the link between individual wavelengths and material signatures. This paper introduces LSGS (Local Spectral and Global Spatial), a lightweight hierarchical transformer designed for binary plastic detection from UAV based SWIR hyperspectral patches. LS-GS uses two attention stages: (i) local spectral attention within small spatial sub-patches to learn inter-band absorption patterns, and (ii) global spatial attention over the whole patch to integrate context. The model operates on the original 60 SWIR bands and uses only per sample standardization, avoiding PCA and global normalization. We also adopt a type-aware patch sampling strategy to cope with severe class imbalance and large uninformative regions typical of push-broom mosaics. On drone-acquired SWIR data from the public dataset of Balsi et al., LS-GS reaches 93.97% accuracy and 94.01% F1 on the test split, outperforming 2D CNN, 3D CNN, and a flat spectral transformer baseline (with up to 89,506 parameters) while using only 51,938 parameters, which supports edge deployment.