5:15 PM - 7:15 PM
[U04-P02] Plastic and Wood Discrimination in Marine Debris Detection Using
Hyperspectral Band Selection
Keywords:Marine plastics pollution
Millions of marine animals die each year due to plastic pollution. Therefore, large-scale
detection from space is essential for monitoring and investigating plastic pollution in marine
environments. In this study, we focus on the discrimination of plastics and wood using optical
satellite imagery. However, different types of plastics—such as low-density polyethylene
(LDPE), high-density polyethylene (HDPE), polypropylene (PP), polystyrene (PS), styrofoam,
and polyvinyl chloride (PVC)—exhibit unique reflectance spectra, making their detection
challenging.
When using Sentinel-2 for the discrimination of plastics and wood, we observed that the spectral
signature of wood often overlaps with certain plastic materials, leading to higher false positive
rates in classification. To address this issue, we conducted controlled laboratory measurements
of plastic samples covering the spectral range of 630–1600 nm using an InGaAs camera
spectrometer. We collected samples from coastal areas, including Otaru Beach, Ehime, and
Amami Islands, and supplemented them with pure virgin plastic samples for reference.
Using the extracted spectra of various plastic types and wood, we propose a novel band
selection approach to enhance their discrimination. This approach includes methods such as
the Normalized Difference Spectral Index (NDSI), second derivative analysis, Recursive Feature
Elimination (RFE), entropy-based selection, and genetic algorithms to identify the most effective
hyperspectral bands. The selected bands are compared against existing literature, including
Sentinel-2 and Landsat-8 bands, to validate their effectiveness in marine plastic detection.
detection from space is essential for monitoring and investigating plastic pollution in marine
environments. In this study, we focus on the discrimination of plastics and wood using optical
satellite imagery. However, different types of plastics—such as low-density polyethylene
(LDPE), high-density polyethylene (HDPE), polypropylene (PP), polystyrene (PS), styrofoam,
and polyvinyl chloride (PVC)—exhibit unique reflectance spectra, making their detection
challenging.
When using Sentinel-2 for the discrimination of plastics and wood, we observed that the spectral
signature of wood often overlaps with certain plastic materials, leading to higher false positive
rates in classification. To address this issue, we conducted controlled laboratory measurements
of plastic samples covering the spectral range of 630–1600 nm using an InGaAs camera
spectrometer. We collected samples from coastal areas, including Otaru Beach, Ehime, and
Amami Islands, and supplemented them with pure virgin plastic samples for reference.
Using the extracted spectra of various plastic types and wood, we propose a novel band
selection approach to enhance their discrimination. This approach includes methods such as
the Normalized Difference Spectral Index (NDSI), second derivative analysis, Recursive Feature
Elimination (RFE), entropy-based selection, and genetic algorithms to identify the most effective
hyperspectral bands. The selected bands are compared against existing literature, including
Sentinel-2 and Landsat-8 bands, to validate their effectiveness in marine plastic detection.