Hyperspectral spectrum enables assessment of heavy metal content, but research on low concentration in water is limited. This study employed in situ hyperspectral data from Dalian Lake, Shanghai to develop a machine learning model for accurately determining heavy metal concentrations. Initially, we employed a combination of empirical analysis and algorithm-based analysis to identify the optimal features for retrieving Cu and Fe ions. Based on the correlation coefficients between heavy metals and water quality, the feature bands for TOC, Chl-a and TP were selected as empirical features. Algorithm-based feature selection was conducted by employing the random forest (RF) approach with the original spectrum (OR), first-order derivative reflectance (FDR), and second-order derivative reflectance (SDR). For the development of a prediction model, we utilized the Genetic Algorithm-Partial Least Squares Regression (GA-PLSR) approach for Cu and Fe ions inversion. Our findings demonstrated that the integration of both empirical features and algorithm-selected features resulted in superior performance compared to using algorithm-selected features alone. Importantly, the crucial wavelength data primarily located at 497, 665, 686, 831 and 935 nm showed superior results for Cu retrieval, while wavelengths of 700, 746, 801, 948, and 993 nm demonstrated better results for Fe retrieval. These results also displayed that the GA-PLSR model outperformed both the PLSR and RF models, exhibiting an R2 of 0.75, RMSE of 0.004, and MRE of 0.382 for Cu inversion. For Fe inversion, the GA-PLSR model outperformed other models with an R2 of 0.73, RMSE of 0.036, and MRE of 0.464. This research provides a scientific basis and data support for monitoring low concentrations of heavy metals in water bodies using hyperspectral remote sensing techniques.
Keywords: GA-PLSR; In situ hyperspectral; Low concentration heavy metal; Water heavy metal retrieval.
Copyright © 2024 The Authors. Published by Elsevier B.V. All rights reserved.