Machine learning water quality inversion via optimal band selection from highly correlated subsets

Area of Science:

• Environmental Science
• Remote Sensing
• Water Quality Management

Background:

• Rapid urbanization leads to significant water contamination in urban rivers.
• Efficient monitoring of water quality parameters like chemical oxygen demand (COD) and ammonia nitrogen is crucial for urban fluvial ecosystems.
• Shanghai's Lianqi River serves as a case study for assessing water quality issues.

Purpose of the Study:

• To explore effective water quality retrieval methods for COD and ammonia nitrogen using UAV multispectral remote sensing data.
• To investigate the impact of feature selection strategies and machine learning algorithms on retrieval accuracy.
• To analyze the relationship between pollutant distribution and land-use patterns in urban rivers.

Main Methods:

• Utilized Unmanned Aerial Vehicle (UAV) multispectral remote sensing for data acquisition.
• Employed exhaustive search for generating feature combinations and six machine learning algorithms (CatBoost, XGBoost, LightGBM, KNN, RF, MBLERM) to build optimized model clusters.
• Developed a multi-dimensional feature evaluation system assessing correlation, distance, and spectral morphology.

Main Results:

• Linear correlation alone is insufficient for determining model precision in water quality retrieval.
• Model accuracy decreases significantly with more than seven features, indicating a dimensional catastrophe.
• Spatial distribution of pollutants exhibits clear heterogeneity, strongly linked to surrounding land-use patterns.

Conclusions:

• UAV multispectral remote sensing combined with advanced machine learning offers a viable approach for urban water quality monitoring.
• Optimal feature selection is critical, as excessive features can degrade model performance.
• Understanding the link between land use and pollutant distribution is essential for targeted water management strategies.