Estimation of high frequency nutrient concentrations from water quality surrogates using machine learning methods.

Area of Science:

• Environmental Science
• Water Resource Management
• Data Science

Background:

• Continuous water quality monitoring is essential for effective water management.
• Real-time in-situ monitoring of all water quality variables, especially nutrients, is often challenging and costly.
• Surrogate measures and data-driven models offer a viable solution for estimating unmeasured variables.

Purpose of the Study:

• To assess the effectiveness of using in-situ measured variables as surrogates for estimating nutrient concentrations.
• To compare machine learning models (Random Forests) against linear models for nutrient estimation.
• To determine the optimal number of surrogate sensors for cost-effective water quality monitoring.

Main Methods:

• Employed Random Forest models to estimate nutrient concentrations using surrogate variables from both rural and urban catchments.
• Utilized commonly measured in-situ water quality parameters as surrogate predictors.
• Compared the performance of Random Forest models with linear models using the same surrogate inputs.

Main Results:

• Random Forest models achieved up to a 60.1% reduction in Root Mean Squared Error (RMSE) compared to linear models.
• The study identified diminishing returns in error reduction when exceeding 4-5 surrogate sensors per catchment.
• Cost-benefit analysis indicated that adding more than the optimal number of sensors is not economically viable.

Conclusions:

• Machine learning, specifically Random Forests, provides a powerful and accurate method for estimating nutrient concentrations using surrogate data.
• The strategic selection of a limited number of surrogate sensors can significantly enhance water quality monitoring efficiency and cost-effectiveness.
• This approach supports improved water management decisions by enabling reliable estimation of critical water quality parameters.