Detection of composite heavy metal content in rape leaf using feature clustering and hyperspectral imaging technology

Exploring composite heavy metal content in rape is significant for crop growth and human health. The focus of this paper was to assess the viability of detection of composite heavy metal content in rape leaf utilizing hyperspectral imaging technology (HSI). Furthermore, a hybrid feature selection based on feature clustering and symmetric uncertainty (HFCSU) was proposed for spectral data to reduce dimensionality. Firstly, hyperspectral images of rape leaf stressed by different composite heavy metal concentrations were collected. Then, the spectral data in the wavelength range of 480-1000 nm was extracted. Subsequently, the spectral data was preprocessed utilizing Savitzky-Golay (SG) smoothing, standard normalized variable (SNV) and its combination (SG-SNV). Competitive adaptive reweighted sampling (CARS), random frog (RF), genetic algorithm-partial least squares (GA-PLS) and HFCSU were utilized for feature selection. Ultimately, the support vector machine regression (SVR) was utilized to build predictive models of Cd and Pb content. The results demonstrated that the SVR model using HFCSU provided the optimal prediction performance, the R_P², RMSEP and RPD for prediction of Cd content were 0.9392, 0.1494 mg·kg^-1 and 3.915, respectively, and the R_P², RMSEP and RPD for prediction of Pb content were 0.9442, 0.1806 mg·kg^-1 and 4.702, respectively. The results indicated that HFCSU can effectively mine features relevant to heavy metals, and HFCSU combined with HSI has a greater potential in the determination of composite heavy metal content in rape leaves.