11:30 AM - 11:45 AM
[MIS02-10] Enhancing Marine Plastic Debris Detection: A Methodological Approach Using NDSI
and Few-Shot Learning
Keywords:marine debris
Plastic marine debris poses a significant threat to marine habitats due to inadequate waste
management practices. Detecting marine debris, particularly wood and plastics, using optical
satellite imagery has been the focus of recent studies. However, a persistent data gap in plastic
detection presents a major challenge. Supervised machine learning models rely on training
samples, which are often generated using artificial targets designed to be detectable from
satellites. These artificial targets, however, are limited in scope and fail to represent the full
diversity of plastic types present in the marine environment. As a result, machine learning
models frequently suffer from overfitting, leading to suboptimal performance in real-world
scenarios.
To address this issue, we propose a novel methodological approach that enhances detection
accuracy and ensures broader plastic type coverage. Specifically, we introduce a modified
Normalized Difference Spectral Index (NDSI) and compare its effectiveness with existing methods
such as the Floating Debris Index (FDI). Additionally, to mitigate the challenges posed by limited
datasets, we implement a few-shot learning model, which enables effective plastic detection
even with minimal training data. The accuracy of this approach is evaluated against previous
studies to assess its robustness. Sentinel-2 satellite imagery serves as the primary dataset for
monitoring plastic debris distribution.
management practices. Detecting marine debris, particularly wood and plastics, using optical
satellite imagery has been the focus of recent studies. However, a persistent data gap in plastic
detection presents a major challenge. Supervised machine learning models rely on training
samples, which are often generated using artificial targets designed to be detectable from
satellites. These artificial targets, however, are limited in scope and fail to represent the full
diversity of plastic types present in the marine environment. As a result, machine learning
models frequently suffer from overfitting, leading to suboptimal performance in real-world
scenarios.
To address this issue, we propose a novel methodological approach that enhances detection
accuracy and ensures broader plastic type coverage. Specifically, we introduce a modified
Normalized Difference Spectral Index (NDSI) and compare its effectiveness with existing methods
such as the Floating Debris Index (FDI). Additionally, to mitigate the challenges posed by limited
datasets, we implement a few-shot learning model, which enables effective plastic detection
even with minimal training data. The accuracy of this approach is evaluated against previous
studies to assess its robustness. Sentinel-2 satellite imagery serves as the primary dataset for
monitoring plastic debris distribution.