Detecting stranded macro-litter categories on drone orthophoto by a multi-class Neural Network

Area of Science:

• Environmental Science
• Computer Science
• Remote Sensing

Background:

• Mapping environmental macro-litter using Unmanned Aerial Systems (UAS) imagery is a growing field.
• Automated detection and categorization of plastic debris are crucial for effective environmental management.

Purpose of the Study:

• To develop and evaluate a multi-class Neural Network (NN) for automatically identifying and categorizing stranded plastic litter from UAS-derived orthophotos.
• To assess the impact of intra-class color variability on the performance of automated litter detection.

Main Methods:

• A multi-class Neural Network (NN) was developed to identify plastic litter categories in UAS orthophotos.
• Performance was evaluated using F-scores for different litter types and a binary litter/non-litter detection approach.

Main Results:

• The NN achieved moderate success (F-score = 61%) for litter types with low color variability, such as octopus pots and fishing ropes.
• Performance was lower (F-score = 37%) for plastic bottles and fragments due to significant intra-class color variations.
• A binary detection approach (litter/non-litter) improved average performance (F-score = 73%) but did not allow for category discrimination.

Conclusions:

• Automated litter detection on drone imagery shows promise, with potential for improved categorization using color-based approaches.
• Color variability within litter categories presents a significant challenge for accurate automated identification and classification.
• Further research into color-invariant features or advanced classification techniques is needed to enhance the performance for diverse litter types.