Machine learning algorithm inversion experiment and pollution analysis of water quality parameters in urban small and...

Area of Science:

• Environmental Monitoring
• Remote Sensing
• Water Quality Assessment

Background:

• Urban rivers face pollution challenges impacting ecosystem health.
• Traditional water quality monitoring is labor-intensive and provides limited spatial coverage.
• Unmanned Aerial Vehicles (UAVs) offer a flexible platform for high-resolution environmental data acquisition.

Purpose of the Study:

• To assess the applicability of UAV-based multispectral imagery for urban river water quality monitoring.
• To develop and compare predictive models for key water quality parameters using spectral data.
• To evaluate the seasonal performance of these models.

Main Methods:

• Acquisition of UAV multispectral imagery and concurrent water sample collection for Fuyang River.
• Calculation of 51 spectral indices (DI, RI, NDI) from single bands and band combinations.
• Development of water quality parameter models using Partial Least Squares (PLS), Random Forest (RF), and Lasso algorithms.

Main Results:

• Random Forest (RF) and other machine learning models outperformed PLS for water quality parameter inversion.
• RF demonstrated robust accuracy and generalization capabilities across different seasons.
• Inversion accuracy varied seasonally, with summer yielding the best results and winter the poorest.

Conclusions:

• UAV multispectral imagery combined with machine learning algorithms provides a viable approach for seasonal urban river water quality prediction.
• The Random Forest model is particularly effective for accurate and generalized water quality assessment.
• Model prediction accuracy and stability correlate with sample value standard deviation.