Alleviating sample imbalance in water quality assessment using the VAE-WGAN-GP model

Area of Science:

• Environmental Science
• Data Science
• Machine Learning

Background:

• Water resources are vital for life and development, but pollution from urbanization threatens freshwater availability.
• Accurate water quality assessment is critical for ecological balance and human health.
• Current machine learning methods struggle with imbalanced data, leading to inaccurate water quality classifications.

Purpose of the Study:

• To develop a novel deep generative model to overcome data imbalance issues in water quality assessment.
• To improve the accuracy and reliability of machine learning models for water quality evaluation.
• To enhance the management and protection of essential water resources.

Main Methods:

• Proposed a novel Variational Autoencoder-Wasserstein Generative Adversarial Network with Gradient Penalty (VAE-WGAN-GP) model.
• Utilized the VAE-WGAN-GP to generate synthetic data samples, effectively compensating for scarcity in minority classes.
• Introduced and applied the concept of 'compensation degree' for comprehensive experimental validation.

Main Results:

• The VAE-WGAN-GP model demonstrated faster convergence and superior distribution learning capabilities.
• Generated synthetic samples closely mimicked real data, successfully addressing data scarcity.
• Achieved a 9.7% increase in water quality assessment accuracy for imbalanced multi-classification datasets.

Conclusions:

• The VAE-WGAN-GP model offers a powerful solution for imbalanced data in water quality assessment.
• This approach significantly enhances the accuracy of water quality monitoring and management.
• The findings contribute to more effective strategies for safeguarding water resources.