Spacecraft Recovery & Remote Sensing, Volume. 46, Issue 2, 157(2025)

A VHR Remote Sensing Change Detection Model with Structure Prior Perception

Jiangwei CHEN and Xiaoliang MENG*
Author Affiliations
  • School of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430000, China
  • show less
    Figures & Tables(9)
    [in Chinese]
    [in Chinese]
    [in Chinese]
    [in Chinese]
    [in Chinese]
    [in Chinese]
    [in Chinese]
    • Table 1. The quantitative comparison of the SPP-CD framework and SPPFormer model proposed in this paper and other state-of-the-art benchmark methods on the LEVIR-CD dataset

      View table
      View in Article

      Table 1. The quantitative comparison of the SPP-CD framework and SPPFormer model proposed in this paper and other state-of-the-art benchmark methods on the LEVIR-CD dataset

      分类方法参数量FLOPs发表年份P/%R/%F1/%IoU/%OA/%
      注:1)该方法使用 RemoteCLIP[20] 等遥感大模型的中间输出作为先验对模型进行辅助,采用 Li 等[27]所报告的指标数值。2) 基于原方法对编码器部分进行扩展,使得模型能够接受 RGB 和 D 两种模态的影像输入。
      视觉模态FC-Siam-conc [3]1.53×10621.7×109201891.9976.7783.6971.9698.49
      FC-Siam-diff [3]1.35×10618.67×109201889.5383.3186.3175.9298.67
      STANet [15]12.21×10626.11×109202083.8191.0087.2677.4098.66
      DTCDSCN [25]41.07×10631.38×109202088.5386.8387.6778.0598.77
      SNUNet [11]3.01×10622.04×109202192.7090.0491.3584.0899.08
      BIT [13]8.99×10645.71×109202291.9788.6290.2682.2599.03
      ChangeFormer [14]11.60×10613.95×109202293.0088.1790.5282.6899.06
      Changer [26]11.40×10611.83×109202392.8690.7891.8184.8699.13
      HANet[27]2.63×10672.40×109202391.7390.0690.8983.2998.96
      CGNet [29]38.99×106174.80×109202393.1590.9092.0185.2199.20
      SPPFormer(RGB)18.6931.6393.1390.8991.9985.1799.15
      视觉模态+先验信息BAN [28]1)202493.6191.0292.3085.69
      SNUNet (RGBD)2)5.9943.3591.4990.5691.0283.5299.08
      ChangeFormer (RGBD)2)20.9633.6193.2187.6290.3382.3699.01
      CGNet (RGBD)2)77.97349.0692.2491.2391.7484.8199.11
      SPPFormer(RGBD)18.6931.6393.3191.4392.3685.8199.21
    • Table 2. Ablation study on key submodules in the SPPFormer model

      View table
      View in Article

      Table 2. Ablation study on key submodules in the SPPFormer model

      实验编号方法P/%R/ %F1/%IoU/%
      输入模态全局路径局部路径双路径融合训练时增强训练设定1)
      批次大小训练轮数
      注:1)批次大小(Batch Size)代表每一次反向传播前参与计算梯度的等效样本数量,单位为:样本数量 / 步(Number of Samples per Step)。训练轮数(Iterations)代表训练时反向传播次数,单位为:步(Step)。2)Identity 模块保持输入和输出不变,模块中不进行任何操作,是一种占位模块。3)本方法得到的最佳模型结构 SPPFormer。4)最佳模型结构的训练时增强版本。所有增强方法仅在训练时应用,不影响推理过程。
      #1RGBSelf-Attention[5]Identity2)Identity随机裁剪、随机翻转164,00090.6791.9089.4983.39
      #2RGB-DSelf-AttentionIdentityIdentity随机裁剪、随机翻转91.1291.9390.3484.62
      #3CMAIdentityIdentity随机裁剪、随机翻转91.5092.0490.9784.92
      #4CMALFEIdentity随机裁剪、随机翻转91.3392.2191.9284.95
      #53)CMALFEPA随机裁剪、随机翻转92.0393.0692.0385.25
      #64)CMALFEPA随机裁剪、随机翻转、随机光度畸变93.3191.4392.3685.81
    Tools

    Get Citation

    Copy Citation Text

    Jiangwei CHEN, Xiaoliang MENG. A VHR Remote Sensing Change Detection Model with Structure Prior Perception[J]. Spacecraft Recovery & Remote Sensing, 2025, 46(2): 157

    Download Citation

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

    Category:

    Received: Dec. 15, 2024

    Accepted: --

    Published Online: May. 23, 2025

    The Author Email: Xiaoliang MENG (xmeng@whu.edu.cn)

    DOI:10.3969/j.issn.1009-8518.2025.02.014

    Topics