Unsupervised learning for lake underwater vegetation classification: Constructing high-precision, large-scale aquatic...

Area of Science:

• Environmental monitoring
• Artificial intelligence in ecology
• Aquatic ecosystem assessment

Background:

• Underwater vegetation monitoring is crucial for lake health assessment.
• Automated data collection using unmanned vessels improves efficiency but faces analysis challenges.
• Supervised AI for vegetation identification requires extensive manual annotation, limiting scalability and generalization.

Purpose of the Study:

• To develop an unsupervised method for automatic underwater vegetation classification.
• To reduce manual annotation effort and costs for dataset construction.
• To create unbiased datasets for diverse lake environments efficiently.

Main Methods:

• A two-step dimensionality reduction combining pre-trained models and manifold learning for feature extraction.
• A multi-algorithm voting mechanism to enhance classification confidence.
• Unsupervised classification approach negating the need for prior data annotation.

Main Results:

• Achieved 97.32% accuracy on a public dataset and over 92% on private datasets from Erhai and Wuhan East Lakes.
• Outperformed traditional supervised methods and matched manual classification accuracy.
• Reduced annotation effort to approximately 20 labeled images for thousands of data points.

Conclusions:

• The proposed unsupervised method offers a highly accurate and efficient solution for underwater vegetation monitoring.
• This approach significantly lowers costs and effort associated with dataset creation and model training.
• Integration with unmanned vessels enables scalable, high-frequency monitoring across various lake ecosystems.