HyperPlastic datasets

The HyperPlastic (HP) database is a collection of hyperspectral images designed to capture the spectral signatures of plastic and non-plastic items. The images were acquired using two linescan hyperspectral cameras: a VIS-NIR camera covering the 400-1000 nm range and a NIR-SWIR camera capturing the 940-1720 nm range, both operating at 25 fps. The spatial resolutions of these cameras are 1024 pixels and 640 pixels, respectively. The database includes five major plastic types (PET, HDPE, LDPE, PS, PP), as well as organic and other material classes, providing a diverse dataset for spectral analysis. The imaging setup simulated an aquatic environment using a black polypropylene container filled with tap water. A white Spectralon sheet served as the white reference, while the dark reference was obtained by closing the camera aperture. To enhance signal-to-noise ratio, raw reflectance data were processed into relative reflectance values using the mean intensities of the white and dark references. Spectral bands with poor responses and lighting (the lowest and highest) were omitted, resulting in a usable spectral range of 489-850 nm (134 bands) for the VIS-NIR camera and 980-1670 nm (98 bands) for the NIR-SWIR camera. Intensity normalization using min-max scaling was applied to remove illumination differences by equalizing the intensity values of darker and lighter pixels. The database comprises 26 images in total (13 from each camera), with each image containing various plastic types.

To enhance the utility of the data, image registration was applied to fuse images from both cameras. This was achieved using a geometric transformation optimized via the OnePlusOne evolutionary algorithm, which maximized mutual information, a statistical measure of dependency between images. The resulting unified hyperspectral images cover the full spectral range of 480-1670 nm, with 232 spectral bands. From these unified images, 1220 non-overlapping patches of 64x64 pixels were manually extracted. Each patch was prepared in three versions: the VIS-NIR spectral range, the NIR-SWIR spectral range, and the combined fused spectral range. To ensure unbiased model training and evaluation, the datasets were partitioned into training (731 patches), validation (246 patches), and testing sets (243 patches).