• Laser & Optoelectronics Progress
  • Vol. 62, Issue 2, 0228004 (2025)
Fu Lü1,2,*, Yuxuan Xie1, and Yongan Feng1
Author Affiliations
  • 1School of Software, Liaoning Technical University, Huludao 125105, Liaoning , China
  • 2Department of Basic Teching, Liaoning Technical University, Huludao 125105, Liaoning , China
  • show less
    DOI: 10.3788/LOP241179 Cite this Article Set citation alerts
    Fu Lü, Yuxuan Xie, Yongan Feng. Incorporation of Multiscale Hierarchical Features for Remote-Sensing Classification[J]. Laser & Optoelectronics Progress, 2025, 62(2): 0228004 Copy Citation Text show less
    MHFNet structure diagram
    Fig. 1. MHFNet structure diagram
    Structure diagrams of two attention mechanisms. (a) MHSA; (b) MSIA
    Fig. 2. Structure diagrams of two attention mechanisms. (a) MHSA; (b) MSIA
    Various FFN structure diagrams. (a) FFN; (b) IRFFN; (c) GDFN; (d) MGFN
    Fig. 3. Various FFN structure diagrams. (a) FFN; (b) IRFFN; (c) GDFN; (d) MGFN
    Structure diagrams of two token extractors. (a) STE; (b) CTE
    Fig. 4. Structure diagrams of two token extractors. (a) STE; (b) CTE
    Scene instances of two datasets. (a) Airplane; (b) beach; (c) ship; (d) cloud; (e) island; (f) bridge; (g) parking; (h) center; (i) storage tank; (j) playground
    Fig. 5. Scene instances of two datasets. (a) Airplane; (b) beach; (c) ship; (d) cloud; (e) island; (f) bridge; (g) parking; (h) center; (i) storage tank; (j) playground
    Confusion matrix of AID dataset under 50% training ratio
    Fig. 6. Confusion matrix of AID dataset under 50% training ratio
    Confusion matrix of NWPU-RESISC45 dataset under 20% training ratio
    Fig. 7. Confusion matrix of NWPU-RESISC45 dataset under 20% training ratio
    Visual CAM comparison. (a) Basketball court; (b) bridge; (c) church; (d) freeway; (e) roundabout; (f) ship; (g) parking
    Fig. 8. Visual CAM comparison. (a) Basketball court; (b) bridge; (c) church; (d) freeway; (e) roundabout; (f) ship; (g) parking
    ParameterValue
    CPUi7-12700
    GPUNVIDIA GeForce RTX 3090Ti 24 GB
    LanguagePython 3.9.12
    Operating systemWindows 11
    Deep learning frameworkPyTorch 1.11.0+CUDA 11.3
    Table 1. Experimental configuration
    ParameterValue
    Epochs300
    Initial learning rate0.0002
    Final learning rate0.00001
    OptimizerAdamW
    Batch size128
    Warmup20
    Weight decay0.003
    Random seed42
    Table 2. Key parameter setting
    MethodParameters /106NWPU(1∶9)NWPU(2∶8)AID(2∶8)AID(5∶5)
    VGGNet1627138.3676.4779.7986.5989.64
    GoogleNet2754.4076.1978.4883.4486.39
    ResNet502923.5886.2388.9392.3994.96
    DenseNet121307.9888.3190.4793.7694.73
    ViT-B1185.6387.5990.8791.1694.44
    PVT-M3143.3290.5192.6692.8495.93
    DeiT-B3285.9791.8693.8393.4196.04
    Swin-B3389.7491.8094.1494.8697.80
    EMTCAL3491.6393.6594.6996.41
    GCSANet358.1193.3994.6595.9697.53
    MGSNet3692.4094.5795.4697.18
    EMSCNet(ViT-B)37173.6493.5895.3796.0297.35
    MHFNet41.8594.0995.7397.1298.63
    Table 3. OA of different models on AID dataset and NWPU-RESISC45 dataset
    ModuleParameters /106NWPU(1∶9)NWPU(2∶8)AID(2∶8)AID(5∶5)
    None31.1892.2794.1095.3896.93
    ATM32.5392.8894.7496.0597.57
    ATM+MSIT41.8594.0995.7397.1298.63
    Table 4. Impact of MSIT and ATM on model performance
    OptionParameters /106NWPU(1∶9)NWPU(2∶8)AID(2∶8)AID(5∶5)
    N=032.5392.8894.7496.0597.57
    N=137.6893.7995.5796.7498.37
    N=241.8594.0995.7397.1298.63
    N=346.0293.8295.6696.8998.46
    N=450.1893.7495.5996.7698.38
    Table 5. Impact of number of MSITs in each branch on model performance
    ModuleParameters /106NWPU(1∶9)NWPU(2∶8)AID(2∶8)AID(5∶5)
    None32.5392.8894.7496.0597.57
    MSIA38.4393.7395.4196.7598.30
    MSIA+MGFN41.7693.9795.6596.9498.52
    MSIA+MGFN+LRC41.8594.0995.7397.1298.63
    Table 6. Effectiveness of various components in MSIT
    MethodParameters /106FLOPs /GNWPU(1∶9)NWPU(2∶8)AID(2∶8)AID(5∶5)
    MHFNet(ResNet50)34.614.7192.6394.5895.3996.94
    MHFNet(ViT-B)96.1756.3093.2894.8395.2196.47
    MHFNet(Swin-B)102.3816.0993.7495.8096.8398.92
    MHFNet(FasterNet-S)41.855.1294.0995.7397.1298.63
    Table 7. Impact of different feature extractors on model performance
    CategoryOAPrecisionRecallF1 scoreCategoryOAPrecisionRecallF1 score
    199.6399.26100.0099.6316100.00100.0099.5599.77
    299.12100.00100.00100.001795.0495.9394.1995.05
    3100.00100.00100.00100.0018100.00100.00100.00100.00
    498.8199.7899.3899.581999.8399.0498.5998.81
    599.4499.0099.5099.252099.4599.3998.2298.80
    699.4799.2290.8194.832199.61100.00100.00100.00
    797.2297.67100.0098.822299.80100.00100.00100.00
    897.0897.6198.3797.992396.3396.3595.6095.97
    997.8797.23100.0098.602498.0998.7898.1198.44
    1097.9498.78100.0099.392596.9197.7297.8397.77
    1198.9198.72100.0099.362697.3299.5998.6599.12
    1299.08100.00100.00100.002795.1595.2599.6097.38
    1398.5299.6599.4199.532899.4499.1696.1997.65
    1497.6698.15100.0099.0729100.00100.0099.0199.50
    1599.1899.26100.0099.633099.28100.00100.00100.00
    Table 8. OA, precision, recall rate, and F1 score for each category in AID dataset
    CategoryOAPrecisionRecallF1 scoreCategoryOAPrecisionRecallF1 score
    198.6799.47100.0099.732491.9691.5788.2789.89
    297.6098.7598.1698.452597.3898.16100.0099.07
    394.1494.5694.9394.742697.0198.9396.3597.62
    498.8899.1591.6295.242796.8797.4294.1295.74
    598.4599.4396.1797.772890.7190.9184.7587.72
    697.5798.1697.3397.742995.4896.88100.0098.42
    7100.00100.0099.8199.903096.9396.1492.1394.09
    886.4886.9288.0587.483193.2693.6893.6993.68
    998.5598.19100.0099.093296.7998.7693.4796.04
    1098.8198.4599.2798.863396.1796.8193.1994.97
    1189.9089.9896.7693.253497.5298.0694.1396.05
    1293.1993.1292.1192.613597.5898.0397.5697.79
    1391.3091.2298.6594.7936100.00100.00100.00100.00
    1497.2798.1598.7098.423796.1598.5697.3197.93
    1596.3197.0790.6293.733899.22100.0099.2399.61
    1699.23100.00100.00100.003995.5796.2797.5696.91
    1785.1885.5097.2991.014093.4094.9098.2296.53
    1899.63100.00100.00100.004195.7096.52100.0098.23
    1993.4193.1993.8693.524286.5186.4487.7487.09
    2092.4992.2692.8192.534393.9793.2098.0895.58
    2193.5793.0497.3295.134498.1999.6597.1698.39
    2291.4091.2995.8593.514597.4598.4789.9194.00
    2396.1397.5095.9696.72
    Table 9. OA, precision, recall rate, and F1 score for each category in NWPU-RESISC45 dataset