Water Quality Forecasting

2019 Feasibility Study

Our objectives for forecasting water quality are the same for our forecasts of water quantity: to provide the decision-support tools necessary to proactively manage watersheds. Through this research, we have identified compelling capabilities of HydroForecast to be extended into a water quality forecast service.

Probability distribution output of model An example prediction demonstrating the probability distributions (displayed as confidence intervals) predicted by the model for a single variable. The model is able to indicate the varying certainty of its predictions as conditions change throughout the year.

In a feasibility study conducted in 2019, we sought to:

  1. Evaluate the feasibility of detecting and subsequently predicting water quality metrics as they relate to algal bloom events using machine learning, satellite imagery, and systems-level data analysis; and
  2. Determine the degree of accuracy with which each item can be assessed using machine learning, satellite imagery, and systems-level data analysis.

This study, conducted in the western basin of Lake Erie and the Sacramento River, confirmed that it is feasible to forecast water quality metrics using satellite data and machine learning.

High predictive accuracy was achieved for water temperature and dissolved oxygen. The average R² across all tested sites is 0.96 for water temperature and 0.90 for dissolved oxygen. Other metrics, such as blue green algae presence, pH, conductivity, turbidity, and nitrogen showed promising results and potential for operational usage with further research and development.

Predictions of various water quality variables


The power behind HydroForecast's architecture is its ability to flexibly support multiple outputs, variable horizons, and new inputs with ease. All of our water quality investigations leverage the existing breakthroughs achieved by HydroForecast.

Land surface temperature near Vermilion, OH Satellite sensed daytime land surface temperature data for the last five years. Land surface temperature provides useful information to our model about soil moisture, snow cover and evaporation conditions.

  1. We synthesize data from a range of sources, including landscape conditions from satellites, meteorological models, and on-the-ground sensors.
  2. We train our proprietary neural networks to forecast multiple parameters (water quantity and quality) in any basin based on these data inputs.
  3. Our team evaluates the models to ensure maximum accuracy and uses additional ground data for localized calibration if available.
  4. The model produces accurate predictions for each parameter along with a probability distribution to inform operational decision-making and minimize risk.

Algal bloom in September of 2017 September 2017. Images show an algal bloom in Lake Erie. The NDCI image on the right shows blue and green values. This indicates a high amount of chlorophyll on the water's surface.