Estimating low concentration heavy metals in water through hyperspectral analysis and genetic algorithm-partial least...

Area of Science:

• Environmental Science
• Remote Sensing
• Analytical Chemistry

Background:

• Assessing heavy metal concentrations in water bodies is crucial for environmental health.
• Existing hyperspectral research often overlooks low heavy metal concentrations.
• In situ hyperspectral data offers potential for detailed water quality monitoring.

Purpose of the Study:

• To develop and validate a machine learning model for accurately determining low concentrations of copper (Cu) and iron (Fe) in water using hyperspectral data.
• To identify optimal spectral features and wavelengths for heavy metal retrieval.
• To compare the performance of different machine learning models for heavy metal inversion.

Main Methods:

• Collected in situ hyperspectral data from Dalian Lake, Shanghai.
• Employed empirical analysis and algorithm-based (Random Forest) feature selection, considering original, first-order, and second-order derivative reflectance.
• Utilized the Genetic Algorithm-Partial Least Squares Regression (GA-PLSR) for developing prediction models for Cu and Fe.
• Selected empirical features based on correlations with TOC, Chl-a, and TP.

Main Results:

• The integration of empirical and algorithm-selected features improved model performance.
• Optimal wavelengths for Cu retrieval were identified around 497, 665, 686, 831, and 935 nm.
• Optimal wavelengths for Fe retrieval were identified around 700, 746, 801, 948, and 993 nm.
• The GA-PLSR model demonstrated superior performance over PLSR and RF models, with R² values of 0.75 for Cu and 0.73 for Fe.

Conclusions:

• Hyperspectral remote sensing, coupled with machine learning (GA-PLSR), is effective for monitoring low concentrations of heavy metals in water.
• Combining empirical and algorithm-based feature selection enhances prediction accuracy.
• This study provides a robust methodology for water quality assessment of heavy metals.