This Artificial Intelligence Model can Predict the Characteristics of the River Water

The Susquehanna River spans over 27,510 square miles across New York, Pennsylvania, and Maryland, providing important natural and recreational resources for local communities. The Susquehanna River Basin Commission stewards the river basin’s resources, which support millions of people and a variety of economies, including agriculture and tourism. However, despite collecting data at multiple sites along the river basin and using a data-driven approach, SRBC required a model-driven approach to predict the water characteristics, such as specific conductance of the river water. 

Therefore, a team of researchers from HU (Siamak Aram,  Maria H. Rivero,  Nikesh K. Pahuja, Roozbeh Sadeghian, Joshua L. R. Paulino, Michael Meyer) and SRBC (James Shallenberger) developed an Artificial Intelligence (AI) model to predict 7-day specific conductance forecast for efficient water quality management. For this purpose, the research team applied Machine Learning to historical data collected by SRBC from 62 sites along the Susquehanna River Basin from 2010 to 2019. The study used supervised learning techniques, such as Linear Regression, Decision Tree, Random Forest, Lasso Regression, XGBoost, and Neural Network model. The model predicted a 7-day forecast of ‘specific conductance’ using a daily time series of four predictors – temperature, pH, dissolved oxygen, and turbidity – and a site indicator.

Machine Learning models were trained and evaluated using the 10-fold Cross-Validation (CV) and 90:10 Train: Test split. With the highest value of R square and least standard deviation, the researchers found that the Random Forest (RF) model recorded the best performance. It outperformed the other models with a mean score of 0.95 ± 0.001. The second-best performance was recorded by XGBoost (XG) model, followed by the Neural Network and the Decision Tree (DT) model. Linear Regression (LR) and Lasso Regression (LSR) tied in last place. Fig. 1 shows the actual vs. predicted daily specific conductance by the Random Forest (RF) model on the test dataset sample.

Further, to provide a new insightful and intuitive tool for decision-making, the researchers integrated the trained machine learning model into a Geographic Information System (GIS) operational dashboard. Fig 2. Displays the resulting Machine Learning model on a seven-day prediction embedded on ARCGISPro.

Thus, this unprecedented study developed a model-driven approach using machine learning techniques for SRBC to predict the specific conductance of water on 62 sites along the Susquehanna river Basin. Further, the researchers integrated the machine learning model into a GIS operational dashboard to display the predictions. This data-driven tool can help SRBC monitor and protect water resources, enhance water quality management, and positively impact communities and millions of people that depend on those resources. Finally, this data-driven tool could potentially be scaled across the nation.  

Check out the Paper. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our 15k+ ML SubReddit, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.