• Spectroscopy and Spectral Analysis
  • Vol. 42, Issue 5, 1385 (2022)
Rong-chang JIANG1、*, Ming-sheng GU2、2;, Qing-he ZHAO1、1;, Xin-ran LI1、1;, Jing-xin SHEN1、1; 3;, and Zhong-bin SU1、1; *;
Author Affiliations
  • 11. Institute of Electrical and Information, Northeast Agricultural University, Harbin 150030, China
  • 22. Harbin City Data Center, Harbin 150030, China
  • show less
    DOI: 10.3964/j.issn.1000-0593(2022)05-1385-08 Cite this Article
    Rong-chang JIANG, Ming-sheng GU, Qing-he ZHAO, Xin-ran LI, Jing-xin SHEN, Zhong-bin SU. Identification of Pesticide Residue Types in Chinese Cabbage Based on Hyperspectral and Convolutional Neural Network[J]. Spectroscopy and Spectral Analysis, 2022, 42(5): 1385 Copy Citation Text show less
    Schematic diagram of hyperspectral imaging system
    Fig. 1. Schematic diagram of hyperspectral imaging system
    Schematic representation of selection of ROI on Chinese cabbage sample
    Fig. 2. Schematic representation of selection of ROI on Chinese cabbage sample
    Spectral data preprocessing by MSC(a): Without preprocessing; (b): With MSC preprocessing
    Fig. 3. Spectral data preprocessing by MSC
    (a): Without preprocessing; (b): With MSC preprocessing
    CNN structure
    Fig. 4. CNN structure
    CNN hyperparameter selection(a): TLA for different LR; (b): TLA for different BS; (c): OA for different Epochs
    Fig. 5. CNN hyperparameter selection
    (a): TLA for different LR; (b): TLA for different BS; (c): OA for different Epochs
    Average spectra of chinese cabbage samples
    Fig. 6. Average spectra of chinese cabbage samples
    Low frequency portions of wavelet transform based on db1 function(a)—(f) corresponding to 1~6 layers of DWT, respectively
    Fig. 7. Low frequency portions of wavelet transform based on db1 function
    (a)—(f) corresponding to 1~6 layers of DWT, respectively
    Process of selecting characteristic wavelength by CARS
    Fig. 8. Process of selecting characteristic wavelength by CARS
    Flow chart of hyperspectral image classification based on DWT and deep learning
    Fig. 9. Flow chart of hyperspectral image classification based on DWT and deep learning
    Modeling results based on DWT, PCA and CARS(a)—(d) Overall accuracies of CNN, MLP, KNN and SVM models based on DWT;(e) Overall accuracies of four models based on PCA; (f) OAs of four models based on CARS
    Fig. 10. Modeling results based on DWT, PCA and CARS
    (a)—(d) Overall accuracies of CNN, MLP, KNN and SVM models based on DWT;(e) Overall accuracies of four models based on PCA; (f) OAs of four models based on CARS
    神经元个数激活函数
    全连接层FC1512relu
    全连接层FC2256relu
    全连接层FC3256relu
    全连接层FC4128relu
    输出层5softmax
    Table 1. MLP network structure
    离散小波变换层数
    1~45~6
    卷积炽C13×3×163×3×16
    激活层A1relurelu
    舍弃层D10.30.3
    卷积层C25×5×323×3×32
    正则层L210.0010.001
    激活层A2relurelu
    舍弃层D20.50.5
    池化层MaxPooling2×2×22×2×2
    全连接层FC1256256
    全连接层FC2256256
    正则层L220.001〗〗0.001
    输出层55
    激活层A3softmaxsoftmax
    Table 2. Parameter settings of the CNN structure
    模型降维算法
    NoneCARSPCADWT
    OA
    /%
    KappaTime
    /ms
    OA
    /%
    KappaTime
    /ms
    OA
    /%
    KappaTime
    /ms
    OA
    /%
    KappaTime
    /ms
    KNN12.00-0.1016.9811.20-0.1110.0056.000.494.0066.400.5820.02
    SVM89.600.8774.7726.400.084.9968.000.604.9990.400.8814.03
    MLP67.200.5980.7872.800.6659.9966.400.5867.0183.200.7963.23
    CNN82.400.78698.1326.400.084.9964.000.5594.0191.200.8986.01
    Table 3. Overall accuracies and Kappa coefficients of different algorithms
    降维
    算法
    模型药物种类
    毒死蜱氯氰菊酯灭多威乐果无残留
    NoneKNN16.0012.008.0024.000.00
    CARS16.0012.004.0020.004.00
    PCA58.8242.0063.7944.4469.56
    DWT60.0080.0048.0064.0080.00
    NoneSVM96.0092.0084.0084.0092.00
    CARS100.008.0020.004.000.00
    PCA68.6266.0070.6966.6667.39
    DWT96.00100.0080.0084.0092.00
    NoneMLP88.0088.0052.0012.0096.00
    CARS56.0052.0092.0072.0092.00
    PCA60.7868.0075.8646.6778.26
    DWT88.0084.0072.0076.0096.00
    NoneCNN76.0084.0076.0080.0096.00
    CARS76.0080.0076.0080.00100.00
    PCA60.7874.0063.7944.4476.08
    DWT88.0088.0084.0096.00100.00
    Table 4. Overall accuracies of the various prediction algorithms
    Rong-chang JIANG, Ming-sheng GU, Qing-he ZHAO, Xin-ran LI, Jing-xin SHEN, Zhong-bin SU. Identification of Pesticide Residue Types in Chinese Cabbage Based on Hyperspectral and Convolutional Neural Network[J]. Spectroscopy and Spectral Analysis, 2022, 42(5): 1385
    Download Citation