A Smart Noise Reduction System and Method for SAR Interferogram Atmospheric Phase Screens Based on Variation Function Spatial Attention and Multi-Scale Fractal Adversarial Loss
Through innovative modules in the FANet-V.2 architecture, the problems of insufficient APS suppression accuracy and cross-scene generalization ability in existing technologies are solved, achieving high-precision APS suppression and improved stability, which is suitable for intelligent noise reduction of SAR interferogram atmospheric phase screens in the field of remote sensing image technology.
Patent Information
- Application Number
- CN202611248558.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2026-08-04
- Filing Date
- 2026-08-18
- Publication Date
- 2026-11-10
AI Technical Summary
Existing deep learning methods struggle to effectively suppress atmospheric phase screen noise (APS) in SAR interferograms. This is due to the irreversible loss of spatial details in the network architecture, the inability of the loss function to be adaptively adjusted, and the failure to fully utilize the semivariogram structure and time-series information of APS, resulting in insufficient denoising stability and cross-scene generalization ability.
The FANet-V.2 architecture, based on variational function-guided spatial attention and multi-scale fractal adversarial loss, is adopted. It includes the HRNet-w32 backbone network, domain adaptation gateway, deep supervision module, and four innovative modules: variational function-guided spatial attention module (VSA), temporal coherence constraint module (TCM), multi-scale fractal adversarial loss module (MFAL), and sparse-fractal joint regularization operator module (SFR), to improve APS suppression accuracy and cross-scene generalization capability.
It significantly improves the APS suppression accuracy and cross-scenario generalization ability, reduces MSE compared to the original FANet, improves SSIM, achieves physical consistency modeling of APS spatial related structures, eliminates latent space regulation conflicts, and is lightweight and highly stable.
Smart Images

Figure CN123071179A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of remote sensing image technology, specifically to an intelligent noise reduction system and method for SAR interferogram atmospheric phase screens based on variogram spatial attention and multi-scale fractal adversarial loss. Background Technology
[0002] nSAR technology is widely used in fields such as surface deformation monitoring, topographic mapping, and disaster early warning, but APS noise is the core bottleneck restricting its accuracy. APS is caused by the spatiotemporal variation of water vapor, temperature, and air pressure in the atmosphere. Its spatial distribution has autocorrelation, low-frequency dominance, and statistical fractal characteristics. It highly overlaps with surface deformation signals in the spectral and spatial domains, making it difficult for traditional filtering and meteorological model methods to achieve accurate correction of single maps.
[0003] Existing deep learning methods (such as ARU-Net, U-Net+KAN, etc.) have the following main shortcomings:
[0004] (1) In terms of network architecture: Existing methods usually use symmetric codecs, which irreversibly lose spatial details during downsampling, while the local texture features of APS are the key information to distinguish atmospheric signals from deformation signals.
[0005] (2) Regarding the loss function: The λ switching threshold (0.01) in the existing fractal dimension alignment loss is a fixed empirical value, which cannot be adaptively adjusted according to different SAR scenarios, resulting in unreasonable loss weight allocation for heavily contaminated samples.
[0006] (3) Spatial statistical prior: Existing methods do not make full use of the semi-variogram structure of the APS signal—that is, the power-law decay of spatial covariance with lag distance—as attention guide, thus ignoring the statistical constraints of APS spatial correlation.
[0007] (4) Utilization of temporal information: There is temporal coherence of atmospheric noise between multi-temporal interference pairs. The existing single-image input framework cannot utilize the temporal consistency constraint of adjacent interference pairs to improve the denoising stability.
[0008] (5) Regarding regularization strategies: sparsity penalty and fractal constraint are applied independently, and there is a potential conflict between the two in the direction of regulating feature activation patterns in the latent space, lacking a unified joint regularization framework.
[0009] In summary, existing deep learning methods suffer from sampling loss, unreasonable weight allocation, difficulty in improving denoising stability, and lack of a unified joint regularization framework. Summary of the Invention
[0010] The purpose of this invention is to address the aforementioned technical problems by providing an intelligent noise reduction system and method for SAR interferogram atmospheric phase screens based on variogram spatial attention and multi-scale fractal adversarial loss. Through four fundamental innovations (VSA, TCM, MFAL, and SFR), while maintaining a lightweight model, it significantly improves APS suppression accuracy and cross-scene generalization ability, further reducing MSE compared to the original FANet.
[0011] To achieve the above objectives, the present invention adopts the following technical solution: a smart noise reduction system for SAR interferogram atmospheric phase screens based on variogram-guided spatial attention and multi-scale fractal adversarial loss, including the FANet-V.2 architecture. The FANet-V.2 architecture includes an HRNet-w32 high-resolution backbone network, a domain adaptation gateway, a deep supervision module, four innovative modules, and a fusion output module FCNHead2. The four innovative modules include: variogram-guided spatial attention module (VSA), temporal coherence constraint module (TCM), multi-scale fractal adversarial loss module (MFAL), and sparse-fractal joint regularization operator module (SFR).
[0012] A smart denoising method for atmospheric phase screens of SAR interferograms based on variogram spatial attention and multi-scale fractal adversarial loss includes the following steps:
[0013] S1. The input SAR differential interferogram is preprocessed and transformed into an input specification acceptable to HRNet through two layers of MLP;
[0014] S2. Extract multi-resolution parallel feature representations through the HRNet-w32 backbone network and output a high-resolution feature map of 128×128×48.
[0015] S3. Feature domain migration is performed via the domain adaptation gateway;
[0016] S4. After inserting the VSA module into the domain adaptation gateway, a spatial attention mask is generated based on the power-law exponent β of the semivariogram, and the features are weighted for physical consistency.
[0017] S5. Geometric, spectral, and contextual features are extracted using a deep supervised three-way branching method.
[0018] S6. The interferogram after the three-way feature fusion is output by FCNHead2 to remove atmospheric noise;
[0019] S7. Calculate the adversarial loss for multi-scale fractal dimension alignment using MFAL, apply sparse-fractal joint regularization using SFR, and constrain temporal coherence consistency using TCM.
[0020] S8. Supervise training using the complete loss function defined in step S5 to output a high-precision surface deformation interferogram.
[0021] In step S1, the preprocessed raw sampling data uses Sentinel-1 SAR SLC imagery with a spatial resolution of 30 meters and VV polarization. After processing using the standard InSAR workflow, it is cropped to 512×512 pixel tiles. The input training set contains 3647 pure atmospheric sample tiles and 2636 pure surface deformation tiles. 5000 input / label pairs are constructed by random overlay, and 2000 pure deformation negative samples are added for model training. During the training phase, TCM uses pairs of adjacent temporal interferometry pairs (12-day time interval) as input to calculate the temporal consistency loss. The two sets of prediction results are constrained by the temporal coherence constraint module and updated synchronously during training. During the inference phase, TCM is automatically masked to maintain a single image input interface.
[0022] The Temporal Coherence Constraint Module (TCM) introduces consistency constraints on paired time-series samples (time t and day t+12) during the training phase. The formula for calculating the temporal consistency loss is as follows:
[0023] L_TCM = ||f(I_t) − f(I_{t+Δt})||_1 · w_coh;
[0024] Where f(·) is the APS estimate output of the model;
[0025] w_coh is the pixel-wise weight calculated from the coherence coefficient map.
[0026] Step S4 specifically includes the following steps:
[0027] S41. Calculate the power-law exponent β of the semivariogram γ(h) at each spatial location of the feature map using three window sizes: 8×8, 16×16, and 32×32, and generate a β mask image of 480×H×W.
[0028] S42. After normalizing the β mask image with Sigmoid, multiply it element-wise with the feature map to form a geometrically perceptual attention response.
[0029] In step S41, the variation function γ(h) is embedded into the attention weight generation path, and its calculation formula is as follows:
[0030] γ(h) = (1 / 2N(h)) · Σ[Z(x+h) − Z(x)]²;
[0031] Where Z(x) is the activation value of the feature map at position x.
[0032] h is the lag vector, and N(h) is the corresponding quantity.
[0033] In step S7, MFAL changes the fixed λ switching threshold in the original FANet to an end-to-end learnable scalar parameter λ̂, and independently calculates the fractal dimension difference at three spatial scales: 512×512, 256×256, and 128×128, and sums them up by weighting α=[0.6,0.3,0.1].
[0034] Step S7 specifically includes:
[0035] S71, Adaptive λ learning: Replace λ with an end-to-end learnable scalar parameter, and automatically converge to the scene's optimal value during training through gradient backpropagation, eliminating hyperparameter sensitivity;
[0036] S72. Multi-scale fractal dimension alignment: The fractal dimension differences are calculated and weighted at three spatial scales (512×512, 256×256, 128×128). The multi-scale loss calculation formula is as follows:
[0037] MFAL = Σ_s α_s · Huber(g_s(Pred), g_s(GT)) + λ̂·E[logD(GT)+log(1−D(Pred))]
[0038] Where α_s is the weight of the s-th scale,
[0039] g_s(·) is the fractal dimension estimate at the s-th scale.
[0040] λ̂ is a learnable parameter.
[0041] In step S7, SFR unifies the sparse penalty KL(ρ||ρ̂_j) and the fractal gradient smoothing penalty μ·||∇g(F_j)||² into a single regularization term, which is jointly applied to the latent space activation with μ=0.15.
[0042] By using the sparse-fractal joint regularization operator (SFR), the two are unified to the same optimization objective, and the calculation formula is as follows:
[0043] ;
[0044] Where F_j is the activation output of the j-th hidden layer neuron.
[0045] g(·) is the fractal dimension estimation operator.
[0046] μ is the balance coefficient (0.15 in the experiment).
[0047] 10. According to claim 8, in step S8, the complete loss function is:
[0048] .
[0049] Compared with existing technologies, the advantages of this invention are as follows: through four fundamental innovations (VSA, TCM, MFAL, SFR), while maintaining the model's lightweight nature, it significantly improves the APS suppression accuracy and cross-scene generalization ability, further reducing MSE compared to the original FANet; and SSIM is effectively improved, achieving physical consistency modeling of APS spatial related structures, improving cross-scene generalization stability, and eliminating latent space regulation conflicts. Attached Figure Description
[0050] Figure 1 This is the overall network architecture diagram of FANet-V.2 of the present invention; it shows the connection relationship of the HRNet-w32 backbone, domain adaptation gateway, three branches of deep supervision, and four innovative modules (VSA, TCM, MFAL, SFR).
[0051] Figure 2 This is a visual comparison of the APS denoising effects of SAR interferograms in this invention; it includes two types of samples: those with mild and those with severe APS contamination, comparing the reconstruction quality of ARU-Net, the original FANet, and the FANet-V.2 of this invention.
[0052] Figure 3 This is a bar chart comparing the quantitative performance of various methods in this invention; it includes a comparison of the three core indicators: MSE, MAE, and SSIM. FANet-V.2 achieves the best performance across all indicators. Detailed Implementation
[0053] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0054] like Figure 1-3 As shown, an intelligent noise reduction system for SAR interferogram atmospheric phase screens based on variogram-guided spatial attention and multi-scale fractal adversarial loss is presented. The system includes the FANet-V.2 architecture, which comprises an HRNet-w32 high-resolution backbone network, a domain adaptation gateway (DA-Gate), a deep supervision module (including fractal texture pooling (FTP), channel attention (CA), and three branches of dilated convolution (FCNHead), four innovative modules, and a fusion output module FCNHead2. The four innovative modules include: variogram-guided spatial attention module (VSA), temporal coherence constraint module (TCM), multi-scale fractal adversarial loss module (MFAL), and sparse-fractal joint regularization operator module (SFR).
[0055] Based on the above system, a smart noise reduction method for SAR interferogram atmospheric phase screens based on variogram spatial attention and multi-scale fractal adversarial loss is provided, including the following steps:
[0056] S1. The input SAR differential interferogram (512×512×1) is preprocessed and transformed into an input specification acceptable to HRNet through two layers of MLP.
[0057] S2. Extract multi-resolution parallel feature representations through the HRNet-w32 backbone network and output a high-resolution feature map of 128×128×48.
[0058] S3, Feature domain migration is performed via the Domain Adaptive Gateway (DA-Gate);
[0059] S4. After inserting the VSA module into the domain adaptation gateway, a spatial attention mask is generated based on the power-law exponent β of the semivariogram, and the features are weighted for physical consistency.
[0060] S5. Geometric, spectral, and contextual features are extracted through a three-way branching process with deep supervision (FTP branch, CA branch, and dilated convolution branch).
[0061] S6. The interferogram after the three-way feature fusion is output by FCNHead2 to remove atmospheric noise;
[0062] S7. Calculate the adversarial loss for multi-scale fractal dimension alignment using MFAL, apply sparse-fractal joint regularization using SFR, and constrain temporal coherence consistency using TCM.
[0063] S8. Supervise training using the complete loss function defined in step S5 to output a high-precision surface deformation interferogram.
[0064] In step S1, the preprocessed raw sampling data uses Sentinel-1 SAR SLC imagery with a spatial resolution of 30 meters and VV polarization. After processing using the standard InSAR workflow (registration, interferogram generation, flat-land phase removal, terrain phase correction, Goldstein filtering, and phase unwrapping), it is cropped to 512×512 pixel tiles. The input training set contains 3647 pure atmospheric sample tiles and 2636 pure surface deformation tiles. 5000 input / label pairs are constructed by random overlay, and 2000 pure deformation negative samples are added, for a total of 7000 training samples. Model training is then performed. During the training phase, TCM uses pairs of adjacent temporal interferograms (12-day time interval) as input to calculate the temporal consistency loss. The two sets of prediction results are constrained by the temporal coherence constraint module and updated synchronously during training. During the inference phase, TCM automatically masks the input interface, maintaining a single image input interface.
[0065] 4. According to claim 3, in multi-temporal InSAR data, although the APS components of adjacent interferometric pairs are not the same, their temporal variations follow low-frequency meteorological processes and have a certain degree of temporal coherence. This invention proposes that the Temporal Coherence Constraint Module (TCM) introduces consistency constraints on paired temporal samples (time t and day t+12) during the training phase. The formula for calculating the temporal consistency loss is:
[0066] L_TCM = ||f(I_t) − f(I_{t+Δt})||_1 · w_coh;
[0067] Where f(·) is the APS estimate output of the model;
[0068] w_coh is the pixel-wise weight calculated from the coherence coefficient map.
[0069] This is used to reduce the contribution of low-coherence regions (water bodies, forests) to the constraints. TCM makes the model more stable when generalizing across time phases, improving NSE by approximately 0.063 on time-series datasets.
[0070] Step S4 specifically includes the following steps:
[0071] S41. Calculate the power-law exponent β of the semivariogram γ(h) at each spatial location of the feature map (480 channels) using three window sizes: 8×8, 16×16, and 32×32 (achieved by linear regression log-lag versus log-semivariance), generating a 480×H×W β mask map.
[0072] S42. After normalizing the β mask image with Sigmoid, multiply it element-wise with the feature map to form a geometrically perceptual attention response.
[0073] In step S41, existing attention mechanisms (such as CBAM and SE) only utilize inter-channel or spatially local statistical features, ignoring the spatial semi-variogram structure of the APS signal. This invention proposes embedding the variogram γ(h) into the attention weight generation path, and its calculation formula is as follows:
[0074] γ(h) = (1 / 2N(h)) · Σ[Z(x+h) − Z(x)]²;
[0075] Where Z(x) is the activation value of the feature map at position x.
[0076] h is the lag vector, and N(h) is the corresponding quantity.
[0077] The VSA module first estimates the power-law exponent β of γ(h) (i.e., the spatial correlation scale of APS) on the feature map using a multi-scale window. Then, it uses β as a spatial mask to weight and modulate the output of the domain adaptation gateway, thereby making the attention response physically consistent with the spatial distribution structure of APS. Compared with ordinary spatial attention, VSA can significantly reduce false suppression of high-frequency terrain signals in heavy APS scenarios (false alarm rate reduced by about 18%).
[0078] In step S7, the λ switching threshold in the original FANet's fractal loss is a fixed value of 0.01 (judged according to the batch mean MSE). This invention proposes a multi-scale fractal adversarial loss (MFAL), with two core improvements: MFAL replaces the fixed λ switching threshold in the original FANet with an end-to-end learnable scalar parameter λ̂, and independently calculates the fractal dimension difference at three spatial scales: 512×512, 256×256, and 128×128, and sums them weighted by α=[0.6,0.3,0.1].
[0079] Based on the original FANet discriminator, two downsampling branches of 256×256 and 128×128 are added to independently calculate the fractal dimension difference Huber loss; the scale loss is weighted and summed by α=[0.6, 0.3, 0.1], and λ̂ is initialized to 0.5. During training, it is automatically learned by the Adam optimizer (lr=0.0001).
[0080] Step S7 specifically includes:
[0081] S71, Adaptive λ learning: Replace λ with an end-to-end learnable scalar parameter, and automatically converge to the scene's optimal value during training through gradient backpropagation, eliminating hyperparameter sensitivity;
[0082] S72. Multi-scale fractal dimension alignment: The fractal dimension differences are calculated and weighted at three spatial scales (512×512, 256×256, 128×128). The multi-scale loss calculation formula is as follows:
[0083] MFAL = Σ_s α_s · Huber(g_s(Pred), g_s(GT)) + λ̂·E[logD(GT)+log(1−D(Pred))]
[0084] Where α_s is the weight of the s-th scale (determined by the normalized resolution ratio).
[0085] g_s(·) is the fractal dimension estimate at the s-th scale.
[0086] λ̂ is a learnable parameter.
[0087] The multi-scale design enables the model to maintain a stable fractal constraint gradient under different APS intensities, avoiding the gradient vanishing problem when the APS is weak under a single scale. In the experiment, the MSE decreased by about 0.0023.
[0088] In step S7, SFR unifies the sparse penalty KL(ρ||ρ̂_j) and the fractal gradient smoothing penalty μ·||∇g(F_j)||² into a single regularization term, which is jointly applied to the latent space activation with μ=0.15.
[0089] In the original FANet, the sparsity penalty (KL divergence) is applied to the latent space S, and the fractal constraint (MFAL) is applied to the output space. These two constraints potentially conflict in their control over the latent activation modes. This invention proposes a sparse-fractal joint regularization operator (SFR) to unify these two constraints to a single optimization objective. The calculation formula is as follows:
[0090] ;
[0091] Where F_j is the activation output of the j-th hidden layer neuron.
[0092] g(·) is the fractal dimension estimation operator.
[0093] μ is the balance coefficient (0.15 in the experiment).
[0094] The second term is the fractal gradient smoothness penalty: an excessively large fractal dimension gradient of the activation value means that the feature changes drastically in space, which does not conform to the low-frequency fractal statistical characteristics of APS. SFR, by jointly constraining "activation sparsity" and "activation fractal smoothness", makes the latent space representation both compress redundant channels and maintain the statistical consistency of APS features, resulting in an SSIM improvement of 0.0253 in the ablation experiment.
[0095] In step S8, the complete loss function is:
[0096] .
[0097] The FANet-V.2 architecture uses an NVIDIA RTX 4090 (24GB VRAM) and the optimizer is Adam (β1=0.90, β2=0.999, ε=1×10⁻). 8 Weight decay 1×10⁻ 4 The initial learning rate was 0.001, with 100 training epochs and a 10-epoch early stopping window. The batch size was 16, and the training / validation / test partition ratio was 6:2:2.
[0098] Online data augmentation: random rotation, random cropping, random lighting perturbation, and random noise injection.
[0099] The model has a total of approximately 10.8M parameters and approximately 18.5G FLOPs (input 1×1×512×512).
[0100] Comparison of experimental performance:
[0101] Table 1 presents the quantitative comparison results of FANet-V.2 with existing mainstream methods on all 1000 test tiles.
[0102] method MSE↓ MAE↓ SSIM↑ NSE↑ ARU-Net 0.2691 0.5078 0.0092 -32.33 U-Net + KAN 0.0193 0.1105 0.5824 -1.757 FANet (Original Version) 0.0074 0.0544 0.6371 0.465 FANet-V.2 (This invention) 0.0051 0.0389 0.7124 0.528
[0103] Table 1. Comparison of quantitative performance of various methods (mean of test set, 1000 images in total)
[0104] Note: ↓ indicates the lower the better, ↑ indicates the higher the better; bold indicates the optimal value.
[0105] As shown in Table 1, FANet-V.2 outperforms all the comparison methods in four metrics: MSE, MAE, SSIM, and NSE. Compared with the original FANet (i.e., V.1.), MSE decreased by about 31.1%, SSIM increased by 11.8 percentage points, and NSE increased by 0.063, verifying the effectiveness of the four innovative modules.
[0106] Ablation experiment:
[0107] Table 2 presents the ablation experiment results with the four innovative modules added one by one, with the original FANet as the base model.
[0108] Configuration MSE↓ SSIM↑ Basic FANet 0.0074 0.6371 + MFAL (Innovation Point 3) 0.0063 0.6589 + MFAL + VSA (Innovation Point 1) 0.0057 0.6842 + MFAL + VSA + SFR (Innovation Point Four) 0.0054 0.6971 + MFAL + VSA + SFR + TCM (Complete FANet-V.2.) 0.0051 0.7124
[0109] Table 2. Ablation experiments of innovative modules (test set MSE / SSIM)
[0110] Note: Each line is built upon the previous line, with each item added to verify the independent contribution of each module.
[0111] The principle behind this embodiment is as follows: based on the existing FANet framework, four fundamental innovations are proposed:
[0112] (1) The variation function guides the spatial attention module (VSA), which embeds the power law exponent of the semivariogram into the attention weight path to achieve physical consistency modeling of the spatially related structures of APS;
[0113] (2) Temporal Coherence Constraint Module (TCM) introduces temporal consistency loss of adjacent temporal interference pairs to improve cross-scenario generalization stability;
[0114] (3) Multiscale fractal adversarial loss (MFAL) replaces the fixed switching threshold with a learnable parameter and jointly aligns the fractal dimensions across the three spatial scales;
[0115] (4) The sparse-fractal joint regularization operator (SFR) unifies the sparse penalty and the fractal gradient smoothing penalty into a single regularization term, eliminating latent space regulation conflicts.
[0116] Experimental results show that FANet-V.2 outperforms the original FANet in three core metrics: MSE (0.0051 vs 0.0074), MAE (0.0389 vs 0.0544), and SSIM (0.7124 vs 0.6371), and the model has only about 10.8M parameters, which meets the requirements for lightweight engineering deployment.
[0117] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.
Claims
1. A smart noise reduction system for SAR interferogram atmospheric phase screens based on variogram spatial attention and multi-scale fractal adversarial loss, characterized in that, The FANet-V.2 architecture includes the HRNet-w32 high-resolution backbone network, a domain adaptation gateway, a deep supervision module, four innovative modules, and a fusion output module FCNHead2. The four innovative modules are: Variation Function Guided Spatial Attention (VSA), Temporal Coherence Constraint (TCM), Multi-Scale Fractal Adversarial Loss (MFAL), and Sparse-Fractal Joint Regularization Operator (SFR).
2. A method applicable to the system of claim 1, characterized in that, Includes the following steps: S1. The input SAR differential interferogram is preprocessed and transformed into an input specification acceptable to HRNet through two layers of MLP; S2. Extract multi-resolution parallel feature representations through the HRNet-w32 backbone network and output a high-resolution feature map of 128×128×48. S3. Feature domain migration is performed via the domain adaptation gateway; S4. After inserting the VSA module into the domain adaptation gateway, a spatial attention mask is generated based on the power-law exponent β of the semivariogram, and the features are weighted for physical consistency. S5. Geometric, spectral, and contextual features are extracted using a deep supervised three-way branching method. S6. The interferogram after the three-way feature fusion is output by FCNHead2 to remove atmospheric noise; S7. Calculate the adversarial loss for multi-scale fractal dimension alignment using MFAL, apply sparse-fractal joint regularization using SFR, and constrain temporal coherence consistency using TCM. S8. Supervise training using the complete loss function defined in step S5 to output a high-precision surface deformation interferogram.
3. The method according to claim 2, characterized in that, In step S1, the preprocessed raw sampling data uses Sentinel-1 SAR SLC imagery with a spatial resolution of 30 meters and VV polarization. After processing using the standard InSAR workflow, it is cropped into 512×512 pixel tiles. The input training set contains 3647 pure atmospheric sample tiles and 2636 pure surface deformation tiles. 5000 input / label pairs are constructed by random overlay, and 2000 pure deformation negative samples are added for model training. During the training phase, the TCM uses pairs of adjacent temporal interferometry pairs (with a time interval of 12 days) as input to calculate the temporal consistency loss. The two sets of prediction results are constrained by the temporal coherence constraint module and updated synchronously during training. During the inference phase, the TCM is automatically masked to maintain a single image input interface.
4. The method according to claim 3, characterized in that, The Temporal Coherence Constraint Module (TCM) introduces consistency constraints on paired time-series samples (time t and day t+12) during the training phase. The formula for calculating the temporal consistency loss is as follows: L_TCM = ||f(I_t) − f(I_{t+Δt})||_1 · w_coh; Where f(·) is the APS estimate output of the model; w_coh is the pixel-wise weight calculated from the coherence coefficient map.
5. The method according to claim 4, characterized in that, Step S4 specifically includes the following steps: S41. Calculate the power-law exponent β of the semivariogram γ(h) at each spatial location of the feature map using three window sizes: 8×8, 16×16, and 32×32, and generate a β mask image of 480×H×W. S42. After normalizing the β mask image with Sigmoid, multiply it element-wise with the feature map to form a geometrically perceptual attention response.
6. The method according to claim 5, characterized in that, In step S41, the variation function γ(h) is embedded into the attention weight generation path, and its calculation formula is as follows: γ(h) = (1 / 2N(h)) · Σ[Z(x+h) − Z(x)]²; Where Z(x) is the activation value of the feature map at position x. h is the lag vector, and N(h) is the corresponding quantity.
7. The method according to claim 6, characterized in that, In step S7, the MFAL changes the fixed λ switching threshold in the original FANet to an end-to-end learnable scalar parameter λ̂, and independently calculates the fractal dimension difference at three spatial scales: 512×512, 256×256, and 128×128, and sums them up by weighting α=[0.6,0.3,0.1].
8. The method according to claim 7, characterized in that, Step S7 specifically includes: S71, Adaptive λ learning: Replace λ with an end-to-end learnable scalar parameter, and automatically converge to the scene's optimal value during training through gradient backpropagation, eliminating hyperparameter sensitivity; S72. Multi-scale fractal dimension alignment: The fractal dimension differences are calculated and weighted at three spatial scales (512×512, 256×256, 128×128). The multi-scale loss calculation formula is as follows: MFAL = Σ_s α_s · Huber(g_s(Pred), g_s(GT)) + λ̂·E[logD(GT)+log(1−D(Pred))] Where α_s is the weight of the s-th scale, g_s(·) is the fractal dimension estimate at the s-th scale. λ̂ is a learnable parameter.
9. The method according to claim 8, characterized in that, In step S7, the SFR unifies the sparse penalty KL(ρ||ρ̂_j) and the fractal gradient smoothing penalty μ·||∇g(F_j)||² into a single regularization term, which is jointly applied to the latent space activation with μ=0.15; By using the sparse-fractal joint regularization operator (SFR), the two are unified to the same optimization objective, and the calculation formula is as follows: ; Where F_j is the activation output of the j-th hidden layer neuron. g(·) is the fractal dimension estimation operator. μ is the balance coefficient (0.15 in the experiment).
10. The method according to claim 8, characterized in that, In step S8, the complete loss function is: 。
