Synthetic aperture radar interferometry elevation inversion method based on convolutional network

Abstract

The invention discloses an elevation inversion method, and particularly relates to a synthetic aperture radar interferometry elevation inversion method based on a convolutional network, which comprises the following steps of: 1, receiving an interferometric phase diagram and an intensity diagram as input data; 2, processing input data by using a convolutional neural network model of an improved U-Net architecture, and training the model; and 3, inputting to-be-tested data into the trained model to obtain a reasoning result. The interference phase diagram and the SAR intensity diagram are utilized at the same time, contribution of different modes in different areas is dynamically weighted through the self-adaptive feature fusion module, the model depends on an intensity structure in a smooth area and focuses on phase change in a stripe dense area, and the adaptability to complex terrains is remarkably enhanced.

Description

Technical Field

[0001] This invention relates to the field of next-generation information technology, and in particular to a method for inverting the elevation of synthetic aperture radar interferometry based on convolutional networks. Background Technology

[0002] Interferometric Synthetic Aperture Radar (InSAR) is an advanced remote sensing technique that uses the phase difference between SAR images acquired from two or more different perspectives to generate three-dimensional elevation information of the Earth's surface. However, due to factors such as atmospheric conditions, noise interference, and complex terrain features, accurately extracting a Digital Elevation Model (DEM) from the original interferometric phase images has always faced numerous challenges.

[0003] Traditional InSAR elevation inversion methods primarily rely on interferometric phase unwrapping and physical model-based elevation mapping algorithms. While these methods can provide accurate elevation information under ideal conditions, they are often limited in practical applications due to the following limitations: First, the 2π ambiguity problem (i.e., the "wrapping" phenomenon) in interferometric phase data requires complex phase unwrapping algorithms to resolve, a process prone to introducing errors, especially in the presence of high noise or discontinuous terrain. Second, traditional methods typically require fine corrections for atmospheric delay and orbital errors, which not only increases the complexity of the processing workflow but also limits its automation and efficiency. Finally, when dealing with large-scale and diverse terrains, traditional algorithms struggle to guarantee consistently high accuracy and high resolution output.

[0004] In recent years, with the development of deep learning technology, especially the successful application of convolutional neural networks (CNNs) in image recognition and processing, new approaches have been provided to address the aforementioned challenges. CNNs possess powerful feature extraction capabilities and can automatically learn the nonlinear relationship between input data and target variables without explicitly defining complex physical models. For InSAR elevation inversion, using CNNs to directly map interferometric phase maps and intensity maps to the DEM can not only effectively avoid potential problems during phase unwrapping but also improve processing speed and robustness, especially demonstrating better adaptability and generalization ability in high-noise environments. Summary of the Invention

[0005] In order to solve the problems existing in the prior art, the present invention provides a synthetic aperture radar interferometry elevation inversion method based on convolutional networks to solve the current technical problems.

[0006] The technical solution adopted by this invention to solve its technical problem is:

[0007] This invention provides a method for inverting the elevation of synthetic aperture radar interferometry based on convolutional networks, comprising the following steps:

[0008] Step 1: Receive the interferometric phase diagram and intensity diagram as input data;

[0009] Step 2: Process the input data using a convolutional neural network model based on the improved U-Net architecture and train the model;

[0010] Step 3: Input the test data into the trained model to obtain the inference results.

[0011] Preferably, in step 1: the interferometric phase diagram represents the phase difference after the interference of two SAR images and includes ground elevation information; the intensity diagram represents the square of the amplitude of the SAR echo signal.

[0012] Preferably, step 1 further includes standardizing the interference phase map and intensity map to obtain a normalized image.

[0013] Preferably, in step 2, the model includes an encoder, a bottleneck layer, and a decoder. The contribution weights of phase features and intensity features are dynamically adjusted through an adaptive feature fusion module to achieve effective fusion of phase and intensity information.

[0014] Preferably, in step 2, the convolutional neural network model of the improved U-Net architecture outputs a single-channel digital elevation model, and the training process is optimized through a multi-objective composite loss function to improve elevation accuracy, terrain structure fidelity, physical rationality, and noise robustness.

[0015] Preferably, the multi-objective composite loss function includes, but is not limited to, at least one of L1 absolute error loss, gradient consistency loss, structural similarity loss, frequency domain consistency loss, physical constraint loss, and uncertainty perception loss.

[0016] Preferably, in step 3, the output digital elevation model is subjected to inverse normalization processing during the inference stage, including but not limited to median filtering, slope constraint smoothing, and hole filling.

[0017] The beneficial effects of this invention are:

[0018] Simultaneously, by utilizing interferometric phase maps (containing elevation information) and SAR intensity maps (containing surface texture and scattering characteristics), the contribution of different modes in different regions is dynamically weighted through an adaptive feature fusion module. This enables the model to depend on the intensity structure in smooth regions and focus on phase changes in densely striped regions, significantly enhancing its adaptability to complex terrains (such as cities, mountains, and forests).

[0019] The design incorporates six joint losses: L1 error, gradient consistency, SSIM structural similarity, frequency domain matching, InSAR physical constraints, and uncertainty weighting. This not only ensures the accuracy of elevation numerical values ​​but also effectively preserves terrain edges, large-scale undulations, and local textures. Furthermore, embedding physical priors enhances extrapolation reliability and automatically reduces label dependence in low-coherence regions, thereby improving model robustness.

[0020] The entire process requires no manual intervention for atmospheric correction, baseline refinement, or multi-step post-processing, enabling fully automated, high-throughput generation from raw interferometric pairs to DEMs. It is suitable for rapid mapping of large-scale remote sensing data and is significantly superior to traditional serial processing links.

[0021] The output DEM has the same spatial resolution as the input image (supporting typical sizes such as 512×512 or 1024×1024) and high detail fidelity. After being trained with multi-source and multi-terrain data, the model has good generalization performance and can maintain stable inversion accuracy in unseen geographical areas (such as migrating from plains to mountains). Attached Figure Description

[0022] The above-described aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0023] Figure 1 This is a schematic diagram of the synthetic aperture radar interferometry elevation inversion method based on convolutional networks according to an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] This application proposes a method for elevation inversion based on synthetic aperture radar interferometry using convolutional networks. It is an end-to-end, dual-modal input, adaptive feature fusion convolutional neural network architecture that directly maps the original interferometric phase map and intensity map to the digital elevation model (DEM), achieving high-precision and high-fidelity elevation inversion in high-noise environments.

[0026] It supports direct input of the original wrapped phase (range [–π, π]) and SAR intensity map; effectively suppresses noise while preserving the boundaries of interferometric fringes and details of terrain structure; automatically completes implicit phase unwrapping and nonlinear elevation mapping; the output DEM spatial resolution is consistent with the input image (typical size: 512×512 or 1024×1024 pixels); it has cross-regional and multi-terrain generalization capabilities and is suitable for complex scenes such as mountains, plains, cities and forests. 1. Data preparation and preprocessing Input data processing

[0027] The system receives two types of single-channel images as input: an interferometric phase map, representing the phase difference between two interferometric SAR images, with a numerical range of [–π, π] and units in radians (rad). This image contains surface elevation information but suffers from 2π ambiguity (i.e., "wrap-around"). and an intensity map, representing the squared amplitude of the SAR echo signal, reflecting surface scattering characteristics and indicating terrain texture and land cover type. The two images must be strictly registered (spatial aligned) and have the same geographic reference frame, spatial resolution, and coverage. It is recommended to use interferometric pairs that have undergone precise baseline correction and atmospheric phase compensation. Tag data processing

[0028] The labels used for training are the actual elevation data of the corresponding area, which can be obtained from high-precision DEM products such as SRTM, ALOS World 3D, TanDEM-X, or airborne LiDAR. The labels must be pixel-aligned with the input image and use a unified coordinate system (such as WGS84+UTM projection). Preprocessing

[0029] To improve model training stability and generalization ability, all inputs and labels need to be standardized: Phase normalization: normalize the original phase... The linear mapping to the interval [–1, 1] is given by the following formula:

[0030] .

[0031] Intensity normalization: First take the natural logarithm of intensity I (add a small constant ε = 1 × 10⁻⁶). -6 To prevent log(0), Z-score normalization is then performed:

[0032] I_norm = (ln(I + ε) – μ_I) / σ_I,

[0033] Where μ_I and σ_I are the global mean and standard deviation of the logarithmic values ​​of the training set intensity.

[0034] DEM label normalization: Z-score standardization was used.

[0035] h_norm = (h – μ_h) / σ_h,

[0036] μ_h and σ_h are the statistical mean and standard deviation of the elevation of the training area.

[0037] Data augmentation (training phase only): includes random rotation (0°, 90°, 180°, 270°), horizontal / vertical flipping, addition of Gaussian phase noise (standard deviation 0.1–0.3 rad), simulated multiplicative speckle noise, and random cropping (for large images). 2. Network architecture setup

[0038] This scheme uses an improved U-Net as its basic architecture. Its core innovation lies in the introduction of a dual-modal input mechanism and an adaptive feature fusion module to achieve dynamic weighted fusion of phase and intensity information. 2.1 Overall Structure:

[0039] The network uses a 512×512 pixel image as the standard input size (1024×1024 is supported, but the corresponding number of layers needs to be expanded). The input is a two-channel tensor: the first channel is a normalized interferometric phase map (range [-1, 1]) and the second channel is a normalized intensity map (Z-score normalized).

[0040] The entire network consists of an encoder (downsampling path), a bottleneck layer, and a decoder (upsampling path), and contains five feature scale levels (Level 0 to Level 4).

[0041] Encoder structure (downsampling path)

[0042] Level 0: Input size is 512×512, number of channels is 2. First, two 3×3 convolutional layers are passed (each convolutional layer is followed by batch normalization and ReLU activation function), the output size remains 512×512, and the number of channels increases to 64. Then, a 2×2 max pooling operation is performed to halve the size to 256×256, while the number of channels remains unchanged at 64.

[0043] Level 1: Input size is 256×256, number of channels is 64. Two more 3×3 convolutional layers (including batch normalization and ReLU) are applied, the output size remains 256×256, and the number of channels increases to 128. Then a 2×2 max pooling operation is performed, halving the size to 128×128, while the number of channels remains unchanged at 128.

[0044] Level 2: Input size is 128×128, number of channels is 128. Continue using two 3×3 convolutional layers (including batch normalization and ReLU), output size remains 128×128, number of channels increases to 256. Then perform a 2×2 max pooling operation to halve the size to 64×64, the number of channels remains unchanged at 256.

[0045] Level 3: Input size is 64×64, number of channels is 256. Repeat the above steps, using two 3×3 convolutional layers (including batch normalization and ReLU), output size remains 64×64, number of channels increases to 512. Then perform a 2×2 max pooling operation to halve the size to 32×32, the number of channels remains unchanged at 512.

[0046] Level 4 (bottleneck layer): The input size is 32×32, and the number of channels is 512. It uses two 3×3 convolutional layers (including batch normalization and ReLU) and adds residual connections at the end (that is, the input is directly added to the output), resulting in a final output size of 32×32 and the number of channels increased to 1024.

[0047] Decoder structure (upsampling path)

[0048] The decoder has four stages, progressively restoring the spatial resolution to 512×512. Each stage includes the following steps:

[0049] Upsampling: The feature map size is increased by a factor of 2 using bilinear interpolation.

[0050] Skip connection fusion: The upsampling result is concatenated with the output feature map of the same level of the encoder.

[0051] Adaptive Feature Fusion Module (AFFM): AFFM is applied immediately after stitching.

[0052] Decoding block: The number of output channels is halved step by step through two 3×3 convolutional layers (including batch normalization and ReLU).

[0053] The specific process is as follows:

[0054] Decode 3: The input size is 32×32 with 1024 channels. First, the size is enlarged to 64×64 using bilinear interpolation, and then concatenated with the output of encoder Level 3 (64×64, 512 channels) to obtain a feature map with 1536 channels. After applying AFFM, two 3×3 convolutional layers (including batch normalization and ReLU) are used to reduce the output size to 64×64 and the number of channels to 512.

[0055] Decode 2: The input size is 64×64 with 512 channels. Similarly, the size is enlarged to 128×128 through bilinear interpolation, and then concatenated with the output of encoder Level 2 (128×128, 256 channels) to obtain a feature map with 768 channels. After applying AFFM, two 3×3 convolutional layers (including batch normalization and ReLU) are used to reduce the output size to 128×128 and the number of channels to 256.

[0056] Decode 1: The input size is 128×128 with 256 channels. The size is enlarged to 256×256 using bilinear interpolation, and then concatenated with the output of encoder Level 1 (256×256, 128 channels) to obtain a feature map with 384 channels. After applying AFFM, two 3×3 convolutional layers (including batch normalization and ReLU) are used to reduce the output size to 256×256 and the number of channels to 128.

[0057] Decode 0: The input size is 256×256 with 128 channels. The size is enlarged to 512×512 using bilinear interpolation, and then concatenated with the output of encoder Level 0 (512×512, 64 channels) to obtain a feature map with 192 channels. After applying AFFM, two 3×3 convolutional layers (including batch normalization and ReLU) are used to reduce the output size to 512×512 and the number of channels to 64.

[0058] Output layer

[0059] The final decoded block output is a 64-channel, 512×512 feature map. A 1×1 convolutional layer (without batch normalization and activation function) is then applied to output a single-channel elevation map. If the label data is Z-score normalized, the output is linear; if the label data range is [-1, 1], a Tanh activation function is added to the output to match the label range.

[0060] Extended mode that supports 1024×1024 input

[0061] To process a 1024×1024 input image, an additional level (Level -1) is added to the front end of the encoder:

[0062] The input size is 1024×1024 with 2 channels. Two 3×3 convolutional layers (including batch normalization and ReLU) are used, maintaining the output size at 1024×1024 while increasing the number of channels to 64. Then, a 2×2 max pooling operation is performed, halving the size to 512×512, while keeping the number of channels unchanged at 64.

[0063] The subsequent structure is exactly the same as above, with the total number of levels becoming 6 (5 downsamplings). 2.2 Adaptive Feature Fusion Module (AFFM)

[0064] Traditional U-Net directly concatenates high- and low-level features at skip connections, without considering the differences in importance between features of different modalities. This approach embeds an Adaptive Feature Fusion Module (AFFM) before each skip connection to dynamically adjust the contribution weights of phase and intensity features.

[0065] The working principle of AFFM is as follows: The phase feature map F_p and the intensity feature map F_i from the encoder are concatenated to form a combined feature F_c = [F_p; F_i]; then, the spatial dimension is compressed by global average pooling, and channel attention weight vectors α ∈ [0,1] are generated through two fully connected layers (with the dimensionality reduced to 1 / 4 of the number of channels in the middle); the final output is F_out = α. F_c, where This indicates multiplication by channel.

[0066] This mechanism enables the network to automatically focus on more discriminative modalities in different regions—for example, relying on intensity texture in smooth regions and emphasizing phase structure in densely striped regions. 2.3 Output Layer

[0067] At the decoder end, the feature map is mapped to a single-channel output through a 1×1 convolution. The activation function is selected based on the label normalization method: if Z-score is used, a linear output is adopted; if Min-Max normalization to [–1,1] is used, Tanh activation is adopted. During the inference stage, the output needs to be denormalized to restore the true elevation value (unit: meters). 2.4 Loss Function and Training Strategy 2.4.1 Multi-objective composite loss function

[0068] To comprehensively improve the model's overall performance in terms of elevation accuracy, terrain structure, physical rationality, and noise robustness, this scheme constructs a weighted composite loss function consisting of six types of loss terms:

[0069] L_total = w1·L_L1 + w2·L_grad + w3·L_SSIM + w4·L_freq + w5·L_phys+ w6·L_uncert

[0070] The weights w1 through w6 can be dynamically adjusted according to task requirements. The design principles and calculation methods for each loss are detailed below:

[0071] (1) L1 absolute error loss (basic fidelity term)

[0072] Measuring predicted elevation The pixel-by-pixel absolute deviation from the true elevation h is insensitive to outliers, ensuring the overall accuracy of the elevation values.

[0073]

[0074] Where N is the total number of valid pixels, and Σ represents the summation over all pixels.

[0075] (2) Gradient consistency loss (edge ​​structure preservation term)

[0076] The forced prediction of the DEM is consistent with the actual DEM on the first-order gradient field (i.e., slope), effectively preserving the terrain boundaries such as ridges, valleys, and cliffs.

[0077]

[0078] in and These represent the first-order differences in the horizontal and vertical directions, respectively (equivalent to the Sobel operator approximation).

[0079] (3) Structural similarity loss (SSIM loss, texture-aware term)

[0080] A complementary form of the image quality evaluation metric SSIM (1 – SSIM) is introduced as a loss to improve the naturalness of terrain texture and the consistency of local contrast as perceived by the human eye:

[0081]

[0082] Where M is the number of sliding windows (e.g., 8×8 pixel non-overlapping blocks), and SSIM calculation includes three parts: brightness, contrast, and structure. The parameters use standard settings (c1 = (0.01L)). 2 c2 = (0.03L) 2 (where L is the dynamic range of elevation). This loss has a strong inhibitory effect on oversmoothing or stripe artifacts in smooth regions.

[0083] (4) Frequency domain consistency loss (spectral fidelity term)

[0084] InSAR interferograms exhibit significant frequency domain characteristics (e.g., fringe frequencies are positively correlated with terrain slope). To ensure that the output DEM matches the actual terrain in the frequency domain distribution, a Fourier amplitude spectrum matching loss is introduced:

[0085]

[0086] Where F(·) denotes the two-dimensional discrete Fourier transform. Let be the amplitude value of the k-th frequency component, and K be the total number of frequency points. To focus on low-to-mid-frequency terrain information, a Gaussian attenuation weight can be applied to the high-frequency components. This loss helps to recover large-scale terrain undulations and avoids the problem of "local accuracy but overall distortion".

[0087] (5) Physical constraint loss (InSAR geometric consistency term)

[0088] Although the network learns the mapping end-to-end, it can still embed an InSAR physical model as a soft constraint. Ideally, the unwrapped phase... The elevation h should satisfy a linear relationship:

[0089]

[0090] Where B⊥ is the vertical baseline, λ is the radar wavelength, R is the slant range, and θ is the angle of incidence. Ignoring atmospheric and noise terms, the following physical residual loss is constructed:

[0091]

[0092] Where α = (4π·B⊥) / (λ·R·sinθ) is the system constant, β is the constant offset (which can be set to 0 or estimated by least squares), and wrap(·) represents the wrapping operation that maps the real phase back to the interval [–π, π]. This loss makes the network output physically consistent with the original interferogram, improving the reliability of extrapolation.

[0093] (6) Uncertainty perception loss (robustness enhancement term)

[0094] In high-noise regions (such as low-coherence areas), the true elevation labels may contain significant errors. To avoid the model overfitting unreliable labels, a weighted loss based on the coherence coefficient γ is introduced:

[0095] ,in

[0096] The coherence coefficient γ ∈ [0, 1] can be calculated from the interference pairs. When γ approaches 0 (complete decoherence), the weight ω i A value close to 1 indicates that the pixel label is unreliable, and its contribution to the loss should be reduced; conversely, the weights in high-coherence regions approach 0, emphasizing accurate fitting. This mechanism makes the model rely more on contextual inference rather than forcibly matching noisy labels in low-quality regions.

[0097] The recommended weights for the initial training phase are as follows: w1 = 1.0 (L1 loss, backbone); w2 = 0.5 (gradient loss); w3 = 0.2 (SSIM loss); w4 = 0.1 (frequency domain loss); w5 = 0.3 (physical constraint loss); w6 = 0.4 (uncertainty-weighted loss). The physical constraint loss and uncertainty loss require additional input system parameters (baseline, wavelength, etc.) and a coherence plot. If the coherence plot is unavailable, L_uncert can be temporarily disabled; if the geometric parameters are unknown, L_phys can be disabled. 2.4.2 Loss Scheduling Strategy

[0098] To balance the learning dynamics of different loss terms, a course-based loss scheduling approach can be adopted: For the first 20 rounds, only L_L1 and L_grad are enabled to quickly establish the basic mapping; From rounds 20 to 50, L_SSIM and L_phys are gradually introduced to improve the consistency between structure and physics; After round 50, full loss is enabled, and w5 and w6 are exponentially decayed (e.g., multiplied by 0.9 every 10 rounds) to avoid over-constraint in the later stages.

[0099] Optimizer: AdamW, initial learning rate 1×10 -4 Weight decay coefficient 1×10 -4 Learning rate scheduling: Linear warmup is used for the first 5 epochs, followed by a cosine annealing restart strategy (30 epochs per cycle). Batch size: 16 (512×512 images, 32 GB GPU memory); gradient accumulation can be enabled if GPU memory is limited. Early stopping mechanism: Monitor the validation set L1 loss; if there is no improvement after 15 consecutive epochs, training is terminated and the optimal model is saved. 3. Reasoning

[0100] The process is as follows: Perform the same normalization on the input phase map and intensity map as in the training phase; feed the dual-channel input into the trained model; obtain the normalized elevation output and apply the inverse normalization formula: h = _norm × σ_h + μ_h; 3×3 median filtering to eliminate isolated outliers; slope constraint smoothing: apply Tikhonov regularization to high gradient regions to suppress non-physical oscillations; hole filling: use inverse distance weighted interpolation to repair invalid values ​​(such as NaN) in the output.

[0101] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for inverting the elevation of synthetic aperture radar interferometry based on convolutional networks, characterized in that, Includes the following steps: Step 1: Receive the interferometric phase diagram and intensity diagram as input data; Step 2: Process the input data using a convolutional neural network model based on the improved U-Net architecture and train the model; Step 3: Input the test data into the trained model to obtain the inference results.

2. The method for synthetic aperture radar interferometry elevation inversion based on convolutional networks according to claim 1, characterized in that: In step 1: the interferometric phase diagram represents the phase difference after the interference of two SAR images and includes ground elevation information; the intensity diagram represents the square of the amplitude of the SAR echo signal.

3. The method for synthetic aperture radar interferometry elevation inversion based on convolutional networks according to claim 1, characterized in that: Step 1 also includes standardizing the interference phase map and intensity map to obtain a normalized image.

4. The method for synthetic aperture radar interferometry elevation inversion based on convolutional networks according to claim 1, characterized in that: In step 2, the model includes an encoder, a bottleneck layer, and a decoder. The contribution weights of phase features and intensity features are dynamically adjusted through an adaptive feature fusion module to achieve effective fusion of phase and intensity information.

5. The method for synthetic aperture radar interferometry elevation inversion based on convolutional networks according to claim 1, characterized in that: In step 2, the convolutional neural network model of the improved U-Net architecture outputs a single-channel digital elevation model, and the training process is optimized through a multi-objective composite loss function to improve elevation accuracy, terrain structure fidelity, physical rationality and noise robustness.

6. The method for synthetic aperture radar interferometry elevation inversion based on convolutional networks according to claim 5, characterized in that: The multi-objective composite loss function includes, but is not limited to, at least one of L1 absolute error loss, gradient consistency loss, structural similarity loss, frequency domain consistency loss, physical constraint loss, and uncertainty-aware loss.

7. The method for synthetic aperture radar interferometry elevation inversion based on convolutional networks according to claim 1, characterized in that: In step 3, the output digital elevation model is subjected to inverse normalization processing during the inference stage, including but not limited to median filtering, slope constraint smoothing, and hole filling.

Images