Monitoring heavy metal contamination of wheat soil using hyperspectral remote sensing technology

An effective monitoring of soil heavy metal content can greatly contribute to the remediation and treatment of heavy metal pollution. This study aims to explore the potential level in the indirect estimation of soil heavy metal content using wheat leaf hyperspectra. 22 sampling plots of wheat farmland were evenly selected in the study area of Xushe Town, Yixing City, Jiangsu Province, China. Four sample squares were selected for each sample plot, each of which the hyperspectral data was collected from the soil samples and 16 wheat leaves. Firstly, Savitzky-Golay smoothing (SG) was applied to the wheat leaf spectral data, where the SG-smoothed spectra were marked as the original spectra R. Secondly, seven mathematical transformation methods were selected as the first derivative (FD), second derivative (SD), absorbance transformation (AT), first derivative of absorbance (AFD), second derivative of absorbance (ASD), multiple scatter correction (MSC), and standard normal variate (SNV) for the spectral pre-processing of wheat leaf spectra R. Thirdly, the different pre-processing spectra were filtered by genetic algorithm (GA) for the feature bands, and then the heavy metal content was analyzed using partial least squares regression (PLSR). Finally, the accuracy of the estimation model was evaluated to compare the coefficient of determination (R2), root mean square error (RMSE), and relative percent difference (RPD) of cross-validation and external validation. The results show that: 1) The spectral pre-processing technique highlighted some hidden information in the spectra. The differential transformation, multiple scatter correction, and standard normal variate on the wheat leaf spectra were more favorable to extract spectrally sensitive information. 2) The genetic algorithm was used to screen 17-25 characteristic bands of soil Cd, and 16-30 characteristic bands of soil As from 230 full bands, which effectively reduced the band redundancy. Meanwhile, GA-PLSR better improved the model accuracy, compared with the general PLSR. It indicated that the genetic algorithm was used to select the spectral wavelength, and then to optimize the model accuracy and stability for the spectral estimation of soil heavy metal content. 3) The best estimation model for the soil Cd content was the combination of standard normal transform pre-processed spectrum and GA-PLSR, with an external validation R2 of 0.87, RMSE of 0.04 mg/kg, and RPD of 2.72. The best estimation model for the soil As content was a multiple scatter correction pre-processed spectrum of GA-PLSR with an externally validated R2 of 0.91, RMSE of 0.32 mg/kg, and RPD of 3.25. The spectral transform and GA-PLSR performed better to estimate the soil Cd and As content. Therefore, it is possible to indirectly estimate the soil Cd and As contents of heavy metals using hyperspectral wheat leaves. This finding can provide a strong reference for the future realization of quantitative, dynamic and nondestructive remote sensing monitoring of soil heavy metal contamination in a large area of farmland.