Determination of Optimal Predictors and Sampling Frequency to Develop Nutrient Soft Sensors Using Random Forest

Area of Science:

• Environmental Science
• Water Quality Monitoring
• Sensor Technology

Background:

• Real-time, in situ nutrient monitoring remains challenging and costly.
• Soft sensors, driven by data models, offer a promising alternative to direct measurements.
• High data requirements for soft sensor development present logistical hurdles.

Purpose of the Study:

• To optimize predictor subsets and sampling frequencies for nutrient soft sensors.
• To develop effective nutrient monitoring using random forest models.
• To reduce data handling complexities in soft sensor development.

Main Methods:

• Utilized 15-min interval water quality data from two automatic stations on the Main River, Germany.
• Employed random forest models with dissolved oxygen, temperature, conductivity, pH, streamflow, and time features as predictors.
• Applied forward subset selection and knee-point determination for optimization.

Main Results:

• Models achieved R² > 0.95 for nitrate, orthophosphate, and ammonium with optimal predictors.
• Increased sampling frequency improved model performance (RMSE).
• Identified optimal sampling frequencies: Nitrate (3.6/2.8h), Orthophosphate (2.4/1.8h), Ammonium (2.2h).

Conclusions:

• Nutrient soft sensors are effective for water quality monitoring.
• Optimized models function well with fewer predictors and lower sampling frequencies.
• This approach reduces data handling burdens and costs in environmental monitoring.