Machine learning-based detection and mapping of riverine litter utilizing Sentinel-2 imagery

Area of Science:

• Environmental Science
• Remote Sensing
• Machine Learning

Background:

• Rivers significantly contribute to marine litter, yet riverine litter receives insufficient attention.
• Automated detection of riverine litter is crucial for effective management.

Purpose of the Study:

• To detect riverine litter using multispectral satellite imagery and machine learning algorithms.
• To investigate the spatio-temporal distribution of riverine litter in the Tisza River.

Main Methods:

• Utilized Very High Resolution (VHR) images for training data and Sentinel-2 images for detection.
• Trained and validated five supervised machine learning algorithms: Artificial Neural Network (ANN), Support Vector Classifier (SVC), Random Forest (RF), Naïve Bayes (NB), and Decision Tree (DT).
• Evaluated model performance on unseen data under varying hydrological conditions and litter sizes.

Main Results:

• Most models performed well on validation data (F1-scores > 0.83), but showed medium to poor performance on test data (F1-scores ranging from 0.45 to 0.69).
• Model performance was limited by the pixel size of Sentinel-2 images.
• Hydraulic structures, such as the Kisköre Dam, were identified as primary litter accumulation zones.
• Largest litter accumulation occurred upstream of the Kisköre Dam during low summer stages, despite higher transport rates during floods.

Conclusions:

• Machine learning models show potential for riverine litter detection but require finer spatial resolution imagery and larger datasets.
• Hydraulic structures play a critical role in concentrating riverine litter.
• Further research is needed to improve automatic detection accuracy and scale.