• Infrared and Laser Engineering
  • Vol. 51, Issue 3, 20210252 (2022)
Kewang Deng1, Huijie Zhao1、2, Na Li1、*, and Hui Cai3
Author Affiliations
  • 1School of Instrumentation Science and Opto-Electronic Engineering, Beihang University, Beijing 100191, China
  • 2Beihang University Qingdao Research Institute, Beihang University, Qingdao 266101, China
  • 3Unit 96901 of the People's Liberation Army of China, Beijing 300140, China
  • show less
    DOI: 10.3788/IRLA20210252 Cite this Article
    Kewang Deng, Huijie Zhao, Na Li, Hui Cai. Improved data-driven compressing method for hyperspectral mineral identification models[J]. Infrared and Laser Engineering, 2022, 51(3): 20210252 Copy Citation Text show less
    Schematic of network pruning
    Fig. 1. Schematic of network pruning
    Flow chart of improved sample-driven compression method for hyperspectral mineral identification model
    Fig. 2. Flow chart of improved sample-driven compression method for hyperspectral mineral identification model
    Hyperspectral datasets of the Cuprite mine in Nevada
    Fig. 3. Hyperspectral datasets of the Cuprite mine in Nevada
    Importance diagram of neurons in hidden layer
    Fig. 4. Importance diagram of neurons in hidden layer
    Output results of the pruned neurons
    Fig. 5. Output results of the pruned neurons
    Parameters of the compressed identification models obtained in the iterative pruning process
    Fig. 6. Parameters of the compressed identification models obtained in the iterative pruning process
    Identification results of the Nevada mining area
    Fig. 7. Identification results of the Nevada mining area
    Class nameTraining samples Testing samples Diagnostic bands/nm
    Muscovite1004002 200, 2 350
    Halloysite1002402 170, 2 210
    Calcite1002402 160, 2 340
    Kaolinite1004002 170, 2 210
    Montmorillonite1004002 230
    Alunite1004002 170, 2 320
    Chalcedony1002402 250
    Total7002320
    Table 1. Information of identified minerals
    hOverall accuracy
    1091.98%
    1592.54%
    2093.01%
    2593.36%
    3093.62%
    3593.14%
    4092.76%
    4592.50%
    Table 2. Model identification accuracy corresponding to the different number of hidden units
    Importance criteriaSequence number of the pruned neuronNumber of pruned unitsCompression rateIdentification accuracy after retraining
    Proposed C-APoZ1, 4, 6, 12, 17, 20, 23, 2781.3694.61%
    APoZ1, 4, 6, 12, 17, 20, 2371.3094.57%
    Table 3. Result chart of different pruning methods
    IterationC-APoZ (Proposed method) APoZ
    Compression rateThresholdIdentification accuracyCompression rateThresholdIdentification accuracy
    000.81793.62%00.81493.62%
    11.360.65194.61%1.300.64394.57%
    21.760.50294.66%1.760.59794.40%
    32.310.46495.04%2.140.46994.00%
    42.730.44594.66%2.500.42694.00%
    53.330.44894.35%2.730.44494.22%
    64.290.37993.41%3.000.43593.84%
    76.000.31690.13%3.330.41293.32%
    Table 4. Iterative pruning results based on different importance criteria
    Kewang Deng, Huijie Zhao, Na Li, Hui Cai. Improved data-driven compressing method for hyperspectral mineral identification models[J]. Infrared and Laser Engineering, 2022, 51(3): 20210252
    Download Citation