Hyperspectral technique combined with deep learning algorithm for detection of compound heavy metals in lettuce

Abstract

The aim of this research was to develop a deep learning method which involved wavelet transform (WT) and stack convolution auto encoder (SCAE) for extracting compound heavy metals detection deep features of lettuce leaves. WT was used to decompose the visible-near infrared (400.68–1001.61 nm) hyperspectral image of lettuce sample in the multi-scale transform to acquire the optimal wavelet decomposition layers of cadmium (Cd) and lead (Pb) content prediction, and then using SCAE to perform deep feature learning on spectral data under optimal wavelet decomposition layer. Support vector machine regression (SVR) models established by the deep features obtained by WT-SCAE achieved reasonable performance with coefficient of determination for prediction (Rp2) of 0.9319, root mean square error for prediction (RMSEP) of 0.04988 mg/kg and the relative percent different (RPD) of 3.187 for Cd content, and with Rp2 of 0.9418, RMSEP of 0.04123 mg/kg and RPD of 3.214 for Pb content. The results of this study confirmed the great potential for detecting compound heavy metals by the combination of hyperspectral technique and deep learning algorithm.