A cluster-based virtual sensing framework for estimating total nitrogen and total phosphorus using sensor-measurable...

Area of Science:

• Environmental Science
• Water Resource Management
• Machine Learning Applications

Background:

• Spatial heterogeneity in large river basins complicates water-quality monitoring.
• Existing single-model virtual-sensor approaches struggle with site-specific nutrient responses (TN, TP).

Purpose of the Study:

• To develop a novel cluster-based virtual-sensing framework for improved nutrient estimation.
• To enhance the spatial precision and interpretability of total nitrogen and total phosphorus monitoring.

Main Methods:

• Applied Hierarchical Cluster Analysis (HCA), K-means, and Gaussian Mixture Model (GMM) for spatial clustering.
• Developed independent machine learning virtual sensors (RF, XGB, MLP) for each cluster.
• Utilized Water Quality Monitoring Network (WQMN) data from 2016-2025.

Main Results:

• The cluster-based framework significantly improved R-squared values for TN (up to 0.045) and TP (up to 0.029) compared to no-clustering.
• Improvements were most evident under restricted input conditions (e.g., excluding direct nutrient data).
• The Random Forest (RF) model within the cluster framework demonstrated superior performance.

Conclusions:

• Combining spatial clustering with machine learning virtual sensors enhances nutrient estimation accuracy in large river basins.
• This approach improves spatial precision and interpretability for water quality monitoring.
• The framework is particularly effective for managing water quality with limited input data.