Machine learning for manually-measured water quality prediction in fish farming

Area of Science:

• Aquaculture
• Machine Learning
• Environmental Monitoring

Background:

• High-quality fish farming requires monitoring water variables like dissolved oxygen, pH, and temperature.
• Traditional machine learning (ML) approaches often rely on real-time data acquisition, which is not accessible to all fish farmers.
• Many farmers use manual measurements, resulting in limited data for analysis.

Purpose of the Study:

• To investigate the application of ML techniques in fish farming scenarios with limited data.
• To develop models for estimating unobserved variables and forecasting water quality with sparse measurements.
• To assess the feasibility of implementing these models on accessible mobile technology.

Main Methods:

• Utilized random forests, multivariate linear regression, and artificial neural networks.
• Developed a methodology for model building in two scenarios: variable estimation and low-data forecasting.
• Evaluated model performance using water quality data typically measured twice daily.

Main Results:

• Random forests demonstrated effectiveness in forecasting dissolved oxygen, pond temperature, pH, ammonia, and ammonium with limited data.
• The developed prediction models can be implemented on mobile information systems.
• Models are capable of running on average smartphones, making them accessible to fish farmers.

Conclusions:

• Machine learning, particularly random forests, offers a viable solution for water quality monitoring in data-limited aquaculture settings.
• Accessible mobile technology can host these ML models, bridging the gap for farmers lacking advanced infrastructure.
• This approach enhances decision-making for fish farmers by providing reliable water quality predictions.