Turbidivision: a machine vision application for estimating turbidity from underwater images

Area of Science:

• Environmental Science
• Computer Vision
• Water Quality Monitoring

Background:

• Turbidity is a critical water quality indicator for industrial, ecological, and public health.
• Current turbidity measurement methods can be expensive and labor-intensive.
• Underwater images offer a potential alternative data source for turbidity estimation.

Purpose of the Study:

• To develop a machine vision model for estimating turbidity from underwater images.
• To provide a cost-effective and accessible method for water quality assessment.
• To enable analysis of historical image datasets and support citizen science initiatives.

Main Methods:

• A two-step machine vision approach was employed, combining image classification and multiple linear regression.
• Data was collected from field sites and lab mesocosms, covering various sediment and colorimetric profiles.
• Images were classified into 11 turbidity classes (0-55 Formazin Nephelometric Units - FNU).

Main Results:

• The image classification model achieved 100% accuracy within one class and 84% exact match accuracy.
• Regression analysis provided continuous turbidity values with accuracy of ±0.7 FNU below 2.5 FNU and ±33% between 2.5 and 55 FNU.
• The model's accuracy is comparable to field-based test kits used in educational and citizen science settings.

Conclusions:

• Underwater image analysis provides a viable method for estimating water turbidity.
• The developed model is accurate enough for applications like citizen science and education.
• A free, open-source application makes this technology widely accessible for various users.