Development of deep learning method for lead content prediction of lettuce leaf using hyperspectral images

The validity and reliability of visible-near infrared (Vis-NIR) hyperspectral imaging were investigated for the determination of lead concentration in lettuce leaves. Besides, a method involving wavelet transform and stacked auto-encoders (WT-SAE) is proposed to decompose the spectral data in the multi-scale transform and obtain the deep spectral features. The Vis-NIR hyperspectral images of 1120 lettuce leaf samples were obtained and the whole region of lettuce leaf sample spectral data was collected and preprocessed. In addition, WT-SAE the deep spectral features using db5 as wavelet basis function, and support vector machine regression (SVR) was used for regression modelling. Furthermore, the best prediction performances for detecting lead (Pb) concentration in lettuce leaves was obtained from raw data set, with coefficient of determination for calibration (R-c(2)) of 0.9911, root mean square error for calibration (RMSEC) of 0.05187, coefficient of determination for prediction (R-p(2)) of 0.9590, root mean square error for prediction (RMSEP) of 0.05587 and residual predictive deviation (RPD) of 3.251 using db5 as wavelet basis function with wavelet fifth layer decomposition. The results of this study indicated that WT-SAE can effectively select the optimal deep spectral features and Vis-NIR hyperspectral imaging has great potential for detecting lead content in lettuce leaves.