高光谱和卷积神经网络的大白菜农残检测

针对大白菜农药残留传统化学检测手段存在前期处理过程繁琐、检测周期长等不足，提出了一种快速无损识别大白菜农药残留种类的方法。以1组无农药残留和4组含有均匀喷洒农药（毒死蜱、乐果、灭多威和氯氰菊酯）的大白菜样本为研究对象（药液浓度配比分别为0.10，1.00，0.20和2.00 mg·kg-1），经12小时自然吸收后，利用高光谱成像系统获取400～1 000 nm高光谱图像，并选取ROI感兴趣区域后经多元散射校正（MSC）预处理；分别采用竞争性自适应重加权算法（CARS）、主成分分析算法（PCA）和离散小波变换（DWT）降维（分别基于db1，sym2，coif1，bior2.2和rbio1.5小波基函数）；最后，将降维后的高光谱数据分别输入卷积神经网络（CNN）、多层感知机（MLP）、K最邻近算法（KNN）和支持向量机（SVM）建立模型并比较。结果显示，CNN，MLP，KNN和SVM算法均在降维算法DWT（小波基函数及变换层数分别为coif1-2，coif1-4，bior2.2-2和sym2-2）取得最优总体精度分别为91.20%，83.20%，66.40%和90.40%，Kappa系数分别为0.89，0.79，0.58和0.88，预测集用时分别为86.01，63.23，20.02和14.03 ms，总体精度和Kappa指标均优于基于CARS和PCA降维算法建模结果。可见，高光谱与离散小波变换和卷积神经网络相融合显著提高分类识别精度，改善“休斯”现象，为实现无损和快速检测识别大白菜农残提供一个新的方法。

Identification of Pesticide Residue Types in Chinese Cabbage Based on Hyperspectral and Convolutional Neural Network

Traditional chemical detection methods for analyzing pesticide residues in chinese cabbage are slow and destructive. In this study, a rapid, non-destructive method for identifying the types of pesticide residues in chinese cabbage samples was developed. First, the hyperspectral imaging system was used to analyze chinese cabbage samples exposed to one of four pesticides chlorpyrifos, dimethoate, methomyl and cypermethrin. The pesticide concentration ratios were 0.10, 1.00, 0.20 and 2.00 mg·kg-1, respectively; and the data was compared to a pesticide-free sample. After 12 hours of natural degradation at room temperature, a hyperspectral imaging system corrected by a black and white plate was used to obtain 400~1 000 nm hyperspectral images of chinese cabbage samples, and the target area was selected by ENVI software. The specific regions of interest (ROI) in samples were further investigated, and the pre-processing by multiple scattering correction (MSC). Secondly, three algorithms such as competitive adaptive reweighting algorithm (CARS), principal component analysis (PCA), discrete wavelet transform (DWT) (based on db1, sym2, coif1, bior2.2, and rbio1.5 wave base functions) were then used to screen for dimensionality reduction from optimally pre-processed results. Finally, the screening results and the samples divided by the Kennard-Stone algorithm were adopted to construct three recognition models separately. Such as k-nearest-neighbor (KNN), support vector machine (SVM), multilayer perceptron (MLP) and convolutional neural network (CNN) were used to determine the best screening method for the dimension of pesticide residues and the optimal hyperspectral recognition model. Our results showed that the CNN, MLP, KNN, and SVM algorithms achieve the best overall accuracy (91.20%, 83.20%, 66.40%, and 90.40%, respectively), Kappa coefficient (0.89, 0.79, 0.58, and 0.88), and the prediction set time (86.01, 63.23, 20.02 and 14.03 ms) under the dimensionality reduction algorithm DWT, respectively; the wavelet basis function and the number of transform layers are coif1-2, coif1-4, bior2.2-2 and sym2-2. All three indicators are better than the modeling results based on CARS and PCA dimensionality reduction algorithms. It showed that the combination of discrete wavelet transform and convolutional neural network shortens the time of classification and identification and significantly improves the classification and identification accuracy, and improves the Hughes phenomenon, providing a new method for non-destructive and rapid detection and identification of chinese cabbage pesticide residues.