Optics and Precision Engineering, Volume. 32, Issue 9, 1395(2024)

Active learning-clustering-group convolutions network for hyperspectral images classification

Jing LIU1,*... Yinqiao LI1 and Yi LIU2 |Show fewer author(s)
Author Affiliations
  • 1School of Electronic Engineering, Xi’an University of Posts and Telecommunications, Xi'an702, China
  • 2School of Electronic Engineering, Xidian University, Xi'an710071, China
  • show less
    Figures & Tables(15)
    Overall scheme of AL-CGNet
    Schematic diagram of clustering averaging process
    Composition of training sample set
    Schematic diagram of grouped convolutional network
    Ground truth map of IP dataset and classification maps of each classification method on IP dataset
    Ground truth map of BW dataset and the classification maps of each classification method on BW dataset
    Ground truth map of HT dataset and the classification maps of each classification method on HT dataset
    • Table 1. Classification results on IP dataset

      View table
      View in Article

      Table 1. Classification results on IP dataset

      MethodOA±SD/%KappaTime/s
      ClusterCNN95.84±0.430.952 5186.18
      SSRN93.07±0.920.920 557.50
      HybridSN94.90±0.520.941 7257.70
      FC-3DCNN70.13±1.230.655 465.00
      3D-DenseNet85.04±1.760.828 4336.10
      CGNet98.28±0.170.980 3172.66
    • Table 2. Classification results on BW dataset

      View table
      View in Article

      Table 2. Classification results on BW dataset

      MethodOA±SD/%KappaTime/s
      ClusterCNN97.12±0.970.968 756.32
      SSRN96.44±0.810.961 417.80
      HybridSN94.32±1.670.938 477.30
      FC-3DCNN94.15±1.910.936 621.11
      3D-DenseNet92.25±1.380.916 0111
      CGNet98.20±0.680.980 554.34
    • Table 3. Classification results on HT dataset

      View table
      View in Article

      Table 3. Classification results on HT dataset

      MethodOA±SD/%KappaTime/s
      ClusterCNN96.33±0.600.960 3260.50
      SSRN96.99±0.330.967 591.40
      HybridSN95.99±1.110.956 7380.2
      FC-3DCNN92.51±0.840.919 090.82
      3D-DenseNet93.62±0.970.931 0486.9
      CGNet96.91±0.280.966 5255.89
    • Table 4. Number of trainable parameters and FLOPs on IP, BW, and HT datasets

      View table
      View in Article

      Table 4. Number of trainable parameters and FLOPs on IP, BW, and HT datasets

      MethodParameters(IP/BW/HT)FLOPs(IP/BW/HT)
      ClusterCNN201,488/201,358/201,42326,196,864/26,196,608/26,196,736
      SSRN63,520/63,470/63,4951,754,920/1,754,824/1,754,872
      HybridSN4,845,696/4,845,438/4,845,56751,346,384/51,345,872/51,346,128
      FC-3DCNN995,456/995,198/995,3274,220,000/4,219,488/4,219,744
      3D-DenseNet2,549,584/2,548,878/2,549,2319,690,464/9,689,056/9,689,760
      CGNet58,176/57,982/58,1115,524,480/5,508,736/5,524,352
    • Table 5. Classification results of each method on IP dataset

      View table
      View in Article

      Table 5. Classification results of each method on IP dataset

      MethodOA±SD/%KappaTime/sTraining data
      ClusterCNN95.84±0.430.952 5186.186%
      CGNet98.28±0.170.980 3172.666%
      AL-ClusterCNN98.51±0.350.982 9186.046%
      97.12±0.430.967 1148.315%
      AL-CGNet99.57±0.120.995 0176.846%
      98.90±0.350.987 4145.685%
    • Table 6. Classification results of each method on BW dataset

      View table
      View in Article

      Table 6. Classification results of each method on BW dataset

      MethodOA±SD/%KappaTime/sTraining data
      ClusterCNN97.12±0.970.968 756.326%
      CGNet98.20±0.680.980 554.346%
      AL-ClusterCNN98.61±0.550.984 956.476%
      97.68±0.600.974 948.145%
      AL-CGNet99.23±0.470.991 758.166%
      98.81±0.580.987 147.335%
    • Table 7. Classification results of each method on HT dataset

      View table
      View in Article

      Table 7. Classification results of each method on HT dataset

      MethodOA±SD/%KappaTime/sTraining data
      ClusterCNN96.33±0.600.960 3260.506%
      CGNet96.91±0.280.966 5255.896%
      AL-ClusterCNN98.50±0.530.983 8274.396%
      97.81±0.580.976 3223.695%
      AL-CGNet98.82±0.160.987 2256.776%
      97.88±0.370.977 1214.885%
    • Table 8. Classification results of AL-CGNet method on IP/BW/HT dataset using fewer training samples

      View table
      View in Article

      Table 8. Classification results of AL-CGNet method on IP/BW/HT dataset using fewer training samples

      Index4%3%
      OA±SD/%KappaTime/sOA±SD/%KappaTime/s
      IP98.48±0.350.982 7110.6396.42±0.820.959 181.93
      BW98.24±1.180.981 037.2597.22±1.430.969 928.79
      HT96.92±0.500.966 7170.5895.81±0.630.954 7121.31
    Tools

    Get Citation

    Copy Citation Text

    Jing LIU, Yinqiao LI, Yi LIU. Active learning-clustering-group convolutions network for hyperspectral images classification[J]. Optics and Precision Engineering, 2024, 32(9): 1395

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category:

    Received: Nov. 2, 2023

    Accepted: --

    Published Online: Jun. 2, 2024

    The Author Email: LIU Jing (zyhalj1975@163.com)

    DOI:10.37188/OPE.20243209.1395

    Topics