Adaptive homogeneous pixel offset tracking glacier time sequence monitoring method and system

Abstract

The invention discloses a self-adaptive homogeneous pixel offset tracking glacier time sequence monitoring method and system, and belongs to the technical field of image processing, and the method comprises the steps: obtaining a plurality of groups of dual-time-phase SAR image pairs of a glacier; performing image recognition on the dual-time-phase SAR image pair, and outputting a binary segmentation result of a glacier change boundary; a binary segmentation result is used as mask input of pixel offset tracking, and a self-adaptive homogeneous pixel matching window is constructed by distinguishing a changing area and a non-changing area in a pixel offset tracking window, so that matching windows of master and slave images of a dual-time-phase SAR image pair only contain homogeneous pixels with the same changing characteristics as a central pixel, and the matching windows of the master and slave images of the dual-time-phase SAR image pair only contain homogeneous pixels with the same changing characteristics as the central pixel. Normalization matching is carried out; and when the local SNR of the homogeneous pixel matching window is smaller than a preset signal-to-noise ratio threshold value, performing offset calculation by adaptively adjusting the size of the window in the row and column directions, and outputting LOS-direction and azimuth-direction two-dimensional timing sequence offset results. According to the method, the glacier change detection accuracy and the deformation calculation precision can be comprehensively improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to an adaptive homogeneous pixel offset tracking method and system for glacier time-series monitoring. Background Technology

[0002] Glaciers are an important freshwater resource. With global warming, most glaciers worldwide are retreating and melting, causing sea-level rise and freshwater scarcity. High-precision, spatiotemporally continuous deformation monitoring of glaciers is crucial for understanding glacier mass balance and dynamic processes, assessing the risks of disasters such as floods and debris flows, and providing decision support for watershed management.

[0003] Traditional ground-based observation techniques for glacier monitoring face challenges such as difficult on-site deployment and high costs, while satellite remote sensing offers advantages like wide monitoring range, high efficiency, and long-term monitoring capabilities. Compared to optical remote sensing, which is limited by clouds, fog, and day / night cycles, Synthetic Aperture Radar (SAR), with its all-day, all-weather, and multi-polarization / multi-orbit observation capabilities, has become an important tool for monitoring glacier dynamics and deformation. Glacier monitoring using SAR data mainly involves two key steps: boundary change detection and deformation field inversion.

[0004] In the field of glacier SAR boundary change detection, early studies relied heavily on threshold segmentation or difference map analysis of multi-temporal SAR intensity images, combined with simple clustering algorithms or region growing methods to extract glacier change areas. While these methods are computationally simple, they are extremely sensitive to speckle noise and struggle to adapt to complex terrain variations, resulting in limited accuracy. In recent years, with the development of deep learning technology, semantic segmentation models based on convolutional neural networks have shown great potential in change detection, reducing the impact of noise to some extent. However, they still face challenges such as insufficient model generalization ability and blurred glacier boundaries. Subsequently, researchers gradually shifted their focus to improving model structure and methods. While existing improved models have effectively improved the accuracy of glacier boundary detection by introducing general attention modules and residual networks, they still fail to fully address the limitations of single-band SAR information, insufficient noise suppression, and the extraction of subtle deformations.

[0005] In the inversion of glacier deformation fields, traditional SAR intensity migration tracking can only obtain the displacement between single image pairs. In order to improve the performance of pixel migration tracking, many scholars have conducted in-depth research. Although the solution accuracy of pixel migration tracking has been improved, the offset solution still depends on the regular window or the initial offset value.

[0006] In summary, existing glacier SAR pixel offset tracking methods still use fixed-size regular matching windows or rely on initial offsets for optimization during pixel matching. This can lead to the mismatching of changing and non-changing pixels within the same matching window at glacier boundaries, resulting in decorrelation and mismatches. Consequently, this produces unclear deformation boundaries and reduces the reliability of displacement calculations at the boundaries. Summary of the Invention

[0007] To address the problems existing in the above-mentioned fields, this invention proposes an adaptive homogeneous pixel offset tracking glacier time-series monitoring method and system. The binary segmentation result is used as the mask input for pixel offset tracking. By constructing an adaptive homogeneous pixel matching window, the matching window of the master and slave images of the dual-temporal SAR image pair contains only homogeneous pixels with the same variation characteristics as the center pixel, which can improve the stability and sub-pixel accuracy of pixel offset calculation.

[0008] To address the aforementioned technical problems, this invention discloses an adaptive homogeneous pixel offset tracking method for glacier time-series monitoring, comprising the following steps: Acquire multiple pairs of dual-temporal SAR images of glaciers; Image recognition is performed on dual-temporal SAR image pairs to output the binary segmentation results of glacier change boundaries; The binary segmentation result is used as the mask input for pixel offset tracking. By distinguishing between the changing and non-changing regions within the pixel offset tracking window, an adaptive homogeneous pixel matching window is constructed. This ensures that the matching window of the master and slave images of the dual-temporal SAR image pair contains only homogeneous pixels with the same change characteristics as the center pixel, and normalized matching is performed. When the local SNR of the homogeneous pixel matching window is less than the preset signal-to-noise ratio threshold, the offset is calculated and the orbital error is removed by adaptively adjusting the window size in the row and column directions, and the two-dimensional temporal offset results in the LOS direction and azimuth direction are output.

[0009] Preferably, the step of performing image recognition on each dual-temporal SAR image pair and outputting the binary segmentation result of the glacier change boundary specifically includes: By constructing a multi-source spatiotemporal fusion network model, dual-temporal SAR images are input into the pre-trained multi-source spatiotemporal fusion network model for identification; The multi-source spatiotemporal fusion network model is based on the TransUNet model. Before encoding in the TransUNet model, a feature fusion layer is constructed using a differential spatiotemporal attention mechanism to enhance the features of the dual-temporal SAR image pairs. The output differential spatiotemporal fusion features are used as the input to the encoder. The differential spatiotemporal attention mechanism includes a differential spatiotemporal attention module and a differential map enhancement module, wherein: The differential spatiotemporal attention module, in terms of spatial attention, obtains a dual-channel feature map by stitching together dual-temporal SAR image pairs by channel. After convolution and activation of the dual-channel feature map, a spatial attention map with the same size as the input dual-temporal SAR image pair is generated. This attention map is then used to spatially weight the absolute difference map of the two frames of the dual-temporal SAR image pair to obtain spatially weighted differential features. In terms of channel attention, multi-scale channel features of the dual-temporal SAR images from each branch are extracted and fused through two independent parallel branches. The fused channel features are spatially converged and mapped into channel attention vectors. A scaling strategy is used to weight the multi-scale channel features of the extracted dual-temporal SAR images from each branch to obtain channel-weighted features. Residual projection is then performed on the dual-channel feature map to output residual projection features. Finally, the spatially weighted differential features, channel-weighted features, and residual projection features are element-wise added to obtain the differential spatiotemporal attention features. The difference map enhancement module includes a dual convolutional block. It obtains a difference map by applying anisotropic diffusion filtering with speckle noise suppression and logarithmic differencing to the dual-temporal SAR image pair. The dual convolutional block extracts the subtle variation features of the difference map to obtain the difference map enhancement features. Finally, the differential spatiotemporal attention features and differential map enhancement features are merged to obtain differential spatiotemporal fusion features, which are used as input to the encoder.

[0010] Preferably, the step of using the binary segmentation result as the mask input for pixel offset tracking, constructing an adaptive homogeneous pixel matching window by distinguishing between changing and non-changing regions within the pixel offset tracking window, ensures that the matching window of the master and slave images of the dual-temporal SAR image pair contains only homogeneous pixels with the same change characteristics as the center pixel, and performing normalized matching, specifically includes: The binary segmentation map of the glacier change boundary is used as a mask to distinguish the changed and unchanged areas within the pixel offset tracking window. An adaptive homogeneous pixel matching window is constructed, and the position with the largest similarity measure under the matching window is found by determining the normalized cross-correlation coefficient (NCC) between the master and slave images of homogeneous pixels. The normalized cross-correlation coefficient (NCC) between master and slave images of homogeneous pixels is: ; in, It is the set of pixels in the homogeneous pixel matching window. These are the coordinates of the pixels within the window. For window size, and They are respectively The amplitude values ​​of the main image and the secondary image. and These represent the average amplitude values ​​of the main image and the secondary image within the window, respectively.

[0011] Preferably, when the local SNR of the homogeneous pixel matching window is less than a preset signal-to-noise ratio threshold, the offset is calculated and orbital errors are removed by adaptively adjusting the window size in the row and column directions respectively, and the two-dimensional temporal offset results in the LOS direction and azimuth direction are output, specifically including: The local signal-to-noise ratio of the homogeneous pixel matching window is evaluated by the matching results. When the local signal-to-noise ratio of the homogeneous pixel matching window is less than the preset signal-to-noise ratio threshold, it indicates that the cross-correlation solution result of the current window pixels is unreliable. The pixel reconstruction window is then adaptively adjusted in the row and column directions respectively. When the local SNR in the homogeneous pixel matching window is greater than or equal to the preset signal-to-noise ratio threshold, the constructed adaptive homogeneous pixel window is used for matching, and the orbital error is removed by combining quadratic polynomial fitting to calculate the LOS and azimuth two-dimensional temporal offset results.

[0012] Preferably, the calculation of the two-dimensional temporal offset results in the LOS and azimuth directions includes using singular value decomposition based on the idea of ​​small baseline sets to perform temporal calculation of the two-dimensional offsets of multiple SAR image pairs, thereby obtaining the deformation rate and temporal deformation results of the glacier in the LOS and azimuth directions.

[0013] Preferably, the calculation of the two-dimensional temporal offset results in the LOS and azimuth directions includes using singular value decomposition based on the idea of ​​small baseline sets to perform temporal calculation of the two-dimensional offsets of multiple SAR image pairs, thereby obtaining the deformation rate and temporal deformation results of the glacier in the LOS and azimuth directions.

[0014] Preferably, the differential spatiotemporal attention module, in terms of spatial attention, obtains a dual-channel feature map by stitching the dual-temporal SAR image pairs by channel, and generates a spatial attention map with the same size as the input dual-temporal SAR image pair after convolution and activation of the dual-channel feature map. This attention map is then used to spatially weight the absolute difference map of the two frames of the dual-temporal SAR image pair to obtain spatially weighted differential features. In terms of channel attention, multi-scale channel features of the dual-temporal SAR images of each branch are extracted and fused through two independent parallel branches. The fused channel features are spatially converged and mapped into channel attention vectors. A scaling strategy is used to perform channel weighting on the extracted multi-scale channel features of the dual-temporal SAR images of each branch to obtain channel weighted features. Residual projection is then performed on the dual-channel feature map to output residual projection features. The spatially weighted differential features, channel weighted features, and residual projection features are element-wise added to obtain differential spatiotemporal attention features, specifically including: In terms of spatial attention, dual-temporal SAR images are stitched together channel by channel to obtain dual-channel feature maps. XThis dual-channel feature map X A spatial attention map of the same size as the input bi-temporal SAR image pair is generated by a 7×7 convolution followed by Sigmoid activation. m ; Calculate the absolute difference map of the two frames of the dual-temporal SAR image pair and compare it with m Dot product yields spatially weighted features. ; Regarding channel attention, multi-scale channel features of dual-temporal SAR images are extracted from each branch through two independent parallel branches. The channel features output from the two branches are then added together in the channel dimension to obtain the fused features. F and the fusion features F We obtain the channel attention weight vector by performing global average pooling and two layers of 1×1 convolution mapping, ReLU activation, and Sigmoid mapping. s , with 1+ s The scaling strategy weights the channel features of the two branch outputs separately and then sums them to obtain the channel-weighted features. ; And the dual-channel feature map X Residual projection features are obtained by applying 1×1 convolution. ; Finally, the channel weighted features are... Spatial weighted features and residual projection features Adding these three together yields the differential spatiotemporal fusion feature: ; Differential spatiotemporal fusion features are used as input to the encoder.

[0015] The acquisition of multiple sets of dual-temporal SAR image pairs of glaciers specifically includes: Collect multi-temporal, multi-source SAR images of typical mountain glaciers, and preprocess the images through registration and geometric correction to obtain multiple sets of dual-temporal SAR image pairs; Based on the preprocessed pairs of dual-temporal SAR images, the glacier change areas of some pairs of dual-temporal SAR images are binarized and labeled. A training set is constructed by generating a label map, and the pairs of dual-temporal SAR images are divided into a training set and a test set. The multi-source spatiotemporal fusion network model is trained using the training set.

[0016] The image recognition of the dual-temporal SAR image pairs, and the output of the binary segmentation result of the glacier change boundary, specifically includes: Obtain the weights of the trained multi-source spatiotemporal fusion network model; Normalize the dual-temporal SAR image pairs for a specified area in the test set and generate an absolute difference map; Overlapping sliding window slicing is used to divide the normalized dual-temporal SAR image pair into multiple overlapping blocks. When the block size is insufficient, zero padding is used to meet the input requirements of the multi-source spatiotemporal fusion network model, thus obtaining the sliced ​​dual-temporal SAR image pair. The multi-source spatiotemporal fusion network model is trained by inputting the segmented dual-temporal SAR image pairs and absolute difference maps into the model for inference. The inference results of each segment are weighted, fused and normalized, and finally, the binary segmentation results of the glacier change and non-change regions are obtained by threshold segmentation.

[0017] Preferably, it also includes an adaptive homogeneous pixel offset tracking glacier time-series monitoring system, specifically comprising: The data acquisition module is used to acquire multiple pairs of dual-temporal SAR images of glaciers; The boundary segmentation module is used to perform image recognition on dual-temporal SAR image pairs and output the binary segmentation results of glacier change boundaries; The pixel offset tracking module uses the binary segmentation result as a mask input for pixel offset tracking. By distinguishing between changing and non-changing regions within the pixel offset tracking window, it constructs an adaptive homogeneous pixel matching window. This ensures that the matching window of the master and slave images of the dual-temporal SAR image pair contains only homogeneous pixels with the same change characteristics as the center pixel, and performs normalized matching. When the local SNR of the homogeneous pixel matching window is less than a preset signal-to-noise ratio threshold, the offset is calculated and orbital errors are removed by adaptively adjusting the window size in the row and column directions, outputting two-dimensional temporal offset results in the LOS and azimuth directions.

[0018] Compared with the prior art, the present invention has the following beneficial effects: The adaptive homogeneous pixel offset tracking glacier time-series monitoring method proposed in this invention uses the binary segmentation result of glacier change boundary output from image recognition as the mask input for pixel offset tracking. By constructing an adaptive homogeneous pixel matching window, the window from the image contains only homogeneous pixels with the same change characteristics as the center pixel of the main image. This eliminates the decorrelation phenomenon caused by matching pixels with different motion characteristics within the matching window, improving the stability and sub-pixel accuracy of pixel offset calculation. By constructing a collaborative framework for glacier change detection and homogeneous pixel offset tracking based on the change detection mask, high-precision automated processing from change recognition to two-dimensional time-series deformation inversion can be achieved. Throughout the process, the degree of automation is high, and it can effectively suppress noise interference and weak boundary problems in glacier change detection. It can effectively solve the decorrelation problem of glacier deformation calculation boundary and improve monitoring accuracy. Attached Figure Description

[0019] Figure 1 This is a flowchart of the adaptive homogeneous pixel offset tracking glacier time-series monitoring method proposed in this invention; Figure 2 A schematic diagram of the DSTA structure provided in an embodiment of the present invention; Figure 3 This is a diagram illustrating the architecture of a multi-source spatiotemporal fusion network model provided in an embodiment of the present invention. Figure 4 The figure shows the experimental results of different models provided in the embodiments of the present invention; Figure 5 The figure shows the ablation experiment results of the multi-source spatiotemporal fusion network model provided in the embodiments of the present invention; Figure 6 The figure shows the experimental results comparing the multi-source spatiotemporal fusion network model and the CBAM-TransUNet model provided in the embodiments of the present invention. Figure 7 A research area diagram provided for embodiments of the present invention; Figure 8 This is a binary segmentation map of the changes in each image pair of Glacier A obtained by a multi-source spatiotemporal fusion network model, provided in an embodiment of the present invention. Figure 9 Comparison results of the LOS upward pixel offset tracking method provided in the embodiments of the present invention; Figure 10 Comparison results of the orientation-up pixel offset tracking method provided in the embodiments of the present invention; Figure 11 Statistics on the standard deviations of the offsets of two adjacent stable regions ROIA and ROIB along the LOS and azimuth directions provided in this embodiment of the invention; Figure 12 The deformation rate results of Glacier A from March 6, 2024 to December 11, 2024, provided for embodiments of the present invention; Figure 13 The temporal variation diagram of Glacier A is provided for an embodiment of the present invention. Detailed Implementation

[0020] The following will refer to the appendices in the embodiments of the present invention. Figures 1-13 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0021] Example like Figure 1 As shown, this invention proposes an adaptive homogeneous pixel offset tracking method for glacier time-series monitoring, comprising the following steps: S1: Acquire multiple pairs of dual-temporal SAR images of the glacier; S2: Perform image recognition on the dual-temporal SAR image pairs and output the binary segmentation results of the glacier change boundary; S3: The binary segmentation result is used as the mask input for pixel offset tracking. By distinguishing between the changing and non-changing regions within the pixel offset tracking window, an adaptive homogeneous pixel matching window is constructed. This ensures that the matching window of the master and slave images of the dual-temporal SAR image pair contains only homogeneous pixels with the same change characteristics as the center pixel, and normalized matching is performed. When the local SNR of the homogeneous pixel matching window is less than the preset signal-to-noise ratio threshold, the offset is calculated and the orbital error is removed by adaptively adjusting the window size in the row and column directions, and the two-dimensional temporal offset results in the LOS direction and azimuth direction are output.

[0022] Specifically, in step S1, existing multi-temporal SAR change detection datasets for Himalayan glaciers are relatively scarce. This invention collects multi-source SAR images such as Sentienl-1A and Lutan-1 covering typical Himalayan glaciers. After preprocessing including registration, geometric correction, and multi-look processing, the glacier change areas are binarized and labeled to generate tag maps. All samples are then uniformly cropped to 256×256 pixels. Subsequently, a Himalayan mountain glacier SAR change detection dataset is constructed through data augmentation. This dataset is divided into a training set, a validation set, and a test set, containing 3462 training samples, 436 validation samples, and 440 test samples.

[0023] To overcome the problem of subtle changes in single-band intensity images being masked by speckle noise, in step S2, this invention proposes a Differential Spatio-Temporal Attention Module (DSTA) specifically for SAR images, such as... Figure 2 As shown in Figure (a), DSTA employs an improved combination of spatial attention and channel attention to enhance its representation capability of temporal variation features. The DSTA framework is embedded into the TransUNet model to construct the multi-source spatiotemporal fusion network model DSTA-TransUNet, as shown in Figure (a). Figure 3 As shown, the DSTA-TransUNet model includes a feature fusion layer, an encoder, a skip connection layer, and a decoder, specifically:

[0024] To address the issue that SAR single-band intensity images contain only backscatter information and are susceptible to speckle noise interference, this invention proposes a DSTA framework based on the TransUNet model. In terms of spatial attention, it first stitches together dual-temporal SAR image pairs channel by channel to obtain dual-channel feature maps. X ,like Figure 2 As shown in (b), this dual-channel feature map X A spatial attention map of the same size as the input bi-temporal SAR image pair is generated by a 7×7 convolution followed by Sigmoid activation. mNext, the absolute difference map of the two frames of the dual-temporal SAR image pair is calculated: and m Dot product yields spatially weighted features. : ; Where T1 and T2 are the SAR images acquired at time T1 and T2, respectively, and H and W represent the height and width of the SAR image, respectively. By calculating the absolute difference of the intensity of corresponding pixels in the SAR images acquired at time T1 and T2, an absolute difference map can be obtained. The size of the pixel value in the difference map reflects the degree of change in the backscattering characteristics of the ground objects at the corresponding locations between the two time phases. The larger the value, the more significant the change.

[0025] Weight m The amplification can amplify regions with significant spatiotemporal variations and suppress noise interference.

[0026] Regarding channel attention, for dual-temporal SAR image pairs, multi-scale channel features of each branch of the dual-temporal SAR image are extracted using two parallel branches (each path includes 3×3 convolution, BatchNorm, ReLU, and a dilated convolution to expand the receptive field). The channel features output from the two branches are then added together in the channel dimension to obtain the fused features. F ,like Figure 2 As shown in (c), and regarding the fusion features F We obtain the channel attention weight vector by performing global average pooling and two layers of 1×1 convolution mapping, ReLU activation, and Sigmoid mapping. s , with 1+ s The scaling strategy weights the output channel features of the two branches separately and then sums them to obtain the channel-weighted features. : ; Each channel adaptively amplifies its response in the range [1, 2], thereby adaptively emphasizing the channel responses that contribute most to change detection.

[0027] To preserve the original information, the spliced ​​dual-channel feature maps were also... X Residual projection features are obtained by applying 1×1 convolution. .

[0028] Finally, the channel-weighted features, spatial-weighted features, and residual projection features are added together to obtain the differential spatiotemporal attention features: ; This design highlights significant spatial variations and enhances features through channels, while ensuring no information is lost through residual fusion.

[0029] The feature fusion layer of the DSTA-TransUNet model consists of a DSTA module and a Differential Component Module (Dicomp). The DSTA module performs feature enhancement on the dual-temporal SAR image pairs to obtain differential spatiotemporal attention features. The Dicomp module uses anisotropic diffusion filtering with speckle noise suppression and logarithmic differencing to obtain a differential map from the dual-temporal SAR images. This differential map is then processed by the dual convolutional blocks of the Differential Component Module to extract subtle variation features, resulting in enhanced differential map features. The differential spatiotemporal attention features and the enhanced differential map features are merged in the fusion layer to obtain the differential spatiotemporal fused features.

[0030] The differential spatiotemporal fusion feature is input into the encoder, where a pre-trained convolutional network obtains multi-scale local texture and structural information, and a lightweight self-attention supplements the global association, ultimately outputting a multi-scale deep feature map containing global semantics.

[0031] The decoding stage follows the classic UNet network convolutional decoding approach, progressively upsampling and restoring resolution to multi-scale deep feature maps. During this process, skip connection layers use differential or standard convolutions to finely filter and align features from each encoder layer before fusing them with the upsampled results from the same level of the decoder. This allows the decoder to utilize both high-resolution details from shallow encoder layers and high-level semantic information from deeper layers at each reconstruction step, effectively improving boundary localization accuracy. Finally, after layer-by-layer fusion and upsampling, the decoder outputs a high-resolution feature map of the same size as the input image, which is then used by a convolutional classification head to generate the final binary segmentation map of the glacier change boundary.

[0032] The loss function adopts a strategy that combines Focal Loss and Dice Loss. Focal Loss automatically reduces the weight of easily classified training samples based on cross-entropy to alleviate the problem of class imbalance. Dice Loss directly measures the degree of overlap between the prediction and the true value, optimizing the overall consistency of the segmentation results.

[0033] The loss function is as follows: ; in, and These represent Focal Loss and Dice Loss, respectively.

[0034] The multi-source spatiotemporal fusion network model is trained using the training set to obtain the trained multi-source spatiotemporal fusion network model. In the application phase, the multi-source spatiotemporal fusion network model trained by inputting dual-temporal SAR images from the test set is tested to verify its performance. It is used for glacier change detection and outputs the binary segmentation results of glacier change boundaries (including changed and unchanged areas).

[0035] The training, testing, and application process of the multi-source spatiotemporal fusion network model includes: The dataset was trained using the PyTorch deep learning framework, with the Adamw optimizer (weight decay 1×10⁻⁶) selected. - ²), with an initial learning rate of 1×10. -5 The learning rate scheduler uses cosine annealing and linearly increases from 1×10⁻⁶ in the first 5 rounds. -8 Once raised to the initial value, the learning rate during training will not fall below 1×10. -8 The entire training process used a batch size of 16 and ran for a total of 100 rounds.

[0036] After training, the multi-source spatiotemporal fusion network model was tested using a test set. In this test, five commonly used metrics were used to evaluate the accuracy of the multi-source spatiotemporal fusion network model: pixel accuracy, precision, recall, F1 score, and mean intersection over union (MIoU). The formulas for calculating these metrics are shown below:

[0037] ; ; ; ; ; ; To systematically evaluate the advantages of the DSTA-TransUNet model in SAR glacier change detection tasks, this invention conducted comparative experiments with seven representative segmentation models—UNet, NestedUNet, Deeplabv3+, AttenUNet, DCSAU_Net, FPN, and SegNet—under the same dataset and training configuration. Specific comparison results are shown in Table 1 and... Figure 4 As shown.

[0038] like Figure 4As shown, there are six different glacier regions (a), (b), (c), (d), (e), and (f). The first column is the master image in the dual-temporal SAR imagery, the second column is the slave image in the dual-temporal SAR imagery, the third column represents the UNet model results, the fourth column represents the NestedUNet model results, the fifth column represents the Deeplabv3+ model results, the sixth column represents the AttenUNet model results, the seventh column represents the DCSAU_Net model results, the eighth column represents the FPN model results, the ninth column represents the SegNet model results, the tenth column represents the DSTA-TransUNet model results, and the eleventh column represents the ground truth map.

[0039] Table 1 Performance Comparison of Different Models As can be seen from the data in Table 1, the DSTA-TransUNet model constructed in this invention has higher performance indicators than other models.

[0040] Meanwhile, ablation experiments were conducted on four schemes: the basic TransUNet, adding only the combined loss (Comloss), adding only the DSTA module, and adding both DSTA and the combined loss. The comparison results are shown in Table 2 and [Table data would be inserted here]. Figure 5 As shown.

[0041] like Figure 5 As shown, there are five different glacier regions (a), (b), (c), (d), and (e). The first column represents the master image in the dual-temporal SAR imagery, the second column represents the slave image in the dual-temporal SAR imagery, the third column represents the TransUNet model results, the fourth column represents the model results with TransUNet+Comloss loss, the fifth column represents the DSTA-TransUNet model results, the sixth column represents the model results with DSTA-TransUNet+Comloss loss, and the seventh column represents the ground truth map.

[0042] As can be seen from the data in Table 1, the pixel accuracy of the DSTA-TransUNet model (Ours) constructed in this invention has increased from 96.23% to 97.36%, precision from 95.67% to 96.28%, recall from 93.98% to 96.53%, IoU from 90.15% to 93.06%, and F1 score has reached 96.41%. These improvements indicate that the model has been significantly enhanced in terms of overall recognition accuracy and boundary discrimination.

[0043] Table 2 Ablation Experiments of DSTA-TransUNet Model from Figure 5 It can also be seen that the basic model does not capture edge details well enough, while the introduction of Comloss reduces missed detections and DSTA alleviates false detections. When the two are combined, the coherence and detail accuracy of the glacier boundary contour are both optimal, and the model can effectively filter out noise and robustly capture small real changes.

[0044] To verify the advantages of DSTA over CBAM, in the DSTA-TransUNet model, only the DSTA module was replaced with the CBAM module, while the rest of the model structure (Dicomp module and feature fusion), loss function, dataset, and hyperparameters remained unchanged, and training and testing were completed under the same conditions. The quantitative performance comparisons of the DSTA-TransUNet and CBAM-TransUNet models are shown in Table 3. Figure 6 The visualizations show the differences between the two models and the original TransUNet model (Primary).

[0045] like Figure 6 As shown, there are four different glacier regions (a), (b), (c), and (d). The first column represents the master image in the dual-temporal SAR imagery, the second column represents the slave image in the dual-temporal SAR imagery, the third column represents the TransUNet model results, the fourth column represents the CBAM-TransUNet model results, the fifth column represents the DSTA-TransUNet model results, and the sixth column represents the ground truth map.

[0046] Table 3 Performance Comparison of DSTA-TransUNet and CBAM-TransUNet Models By analyzing the data in Table 3 and Figure 6 The differences reveal that the CBAM-TransUNet model did not show significant improvements in pixel accuracy, F1, or IoU, indicating that its application of uniform spatial and channel attention to the stitched dual-temporal SAR images and lack of residual preservation led to misweighting of noise and limited information. In contrast, the DSTA-TransUNet model outperformed CBAM-TransUNet in all metrics, demonstrating that its dual-branch multi-scale extraction, explicit differential weighting, and residual fusion better enhance subtle changes and suppress noise, thereby more accurately identifying glacier changes and achieving a comprehensive performance improvement.

[0047] The DSTA-TransUNet model is used for inference applications. First, the weights of the trained multi-source spatiotemporal fusion network model are loaded. The dual-temporal SAR image pairs in a specified area of ​​the test set are normalized and an absolute difference map is generated. The normalized dual-temporal SAR image pairs are divided into multiple overlapping blocks using an overlapping sliding window. When the block size is insufficient, zero padding is used to meet the input requirements of the multi-source spatiotemporal fusion network model, resulting in the segmented dual-temporal SAR image pairs. The segmented dual-temporal SAR image pairs and the absolute difference map are input into the trained multi-source spatiotemporal fusion network model for inference. The inference results of each block are weighted, fused, and normalized. Finally, the binary segmentation results of the glacier change and non-change areas are obtained through threshold segmentation, which are used as the mask input for the adaptive homogeneous pixel offset tracking method in the subsequent step S3.

[0048] The basic principle of traditional offset is to use the normalized cross-correlation (NCC) between master and slave images to find the position with the highest similarity metric under the regular window for matching, thereby accurately estimating the sub-pixel offset between master and slave images.

[0049] The NCC calculation in the rules window is as follows: ; in For window size, These are the coordinates of the pixels within the window. and They are respectively The amplitude values ​​of the main image and the secondary image. and These represent the average amplitude values ​​of the main image and the secondary image within the window, respectively.

[0050] In step S3, this invention proposes an adaptive homogeneous pixel offset tracking method based on a SAR change detection mask. The binary segmentation result of step S2 is used as the mask input for pixel offset tracking to distinguish between changed and unchanged regions within the pixel offset tracking window, thereby constructing an adaptive homogeneous pixel matching window. This ensures that the matching window of the master and slave images of the dual-temporal SAR image pair only contains homogeneous pixels with the same change characteristics as the center pixel, excluding pixels with different change characteristics, and performing normalized matching. The normalized cross-correlation coefficient (NCC) between master and slave images of homogeneous pixels is: in, It is the set of pixels in the homogeneous pixel matching window. These are the coordinates of the pixels within the window. For window size, and They are respectively The amplitude values ​​of the main image and the secondary image. and These represent the average amplitude values ​​of the main image and the secondary image within the window, respectively; Constructing an adaptive homogeneous pixel matching window can eliminate the decorrelation phenomenon caused by matching pixels with different motion characteristics within the matching window, thereby improving the accuracy of offset calculation.

[0051] Meanwhile, after constructing a homogeneous pixel matching window, the reduction in the number of pixels leads to a decrease in the signal-to-noise ratio (SNR). The local SNR of the homogeneous pixel matching window is evaluated by the matching results. When the local SNR of the homogeneous pixel matching window is less than a preset SNR threshold of 6dB, it indicates that the cross-correlation solution of the current window pixels is unreliable. The window is reconstructed by adaptively adjusting the pixels in the row and column directions (increasing the pixel count by 16 pixels in each direction), with a maximum size not exceeding 192×192, to avoid excessive smoothing of details in an overly large window.

[0052] When the local SNR in the homogeneous pixel matching window is greater than or equal to the preset signal-to-noise ratio threshold, the constructed adaptive homogeneous pixel window is used for matching, and the orbital error is removed by combining quadratic polynomial fitting to calculate the two-dimensional offsets of the glacier in the LOS direction and azimuth direction.

[0053] This invention presents five pixel offset tracking methods for comparison: ① Traditional pixel offset tracking (POT); ② Mask-based pixel homogeneous offset tracking (M-POT); ③ Mask-based adaptive homogeneous pixel offset tracking (M-APOT) with an adaptive signal-to-noise ratio adjustment window; ④ Change detection mask-based homogeneous pixel offset tracking (CDM-POT); and ⑤ Change detection mask-based adaptive homogeneous pixel offset tracking (CDM-APOT) with an adaptive signal-to-noise ratio adjustment window. The matching windows for POT, M-POT, and CDM-POT were set to 128×128 pixels; the initial windows for M-APOT and CDM-APOT, which can adaptively adjust their windows, were also set to 128×128 pixels. All methods used a uniform oversampling factor of 2. To quantitatively evaluate the accuracy of the offset calculations using different methods, a stable region without deformation near the glacier was selected, and the standard deviations of its offsets in the azimuth and LOS directions were calculated.

[0054] in, This represents the offset value of the i-th pixel in the stable region. The standard deviation is the average of all pixel offsets within the region, where N is the total number of pixels. A smaller standard deviation indicates higher accuracy in pixel offset calculation. Subsequently, an adaptive homogeneous pixel offset tracking method based on a change detection mask is used to process the SAR image pairs, obtaining high-precision two-dimensional offset results for the glacier in the LOS and azimuth directions.

[0055] The Pixel Offset Tracking-Small BaselineSubset (PO-SBAS) technique is employed. Based on the concept of small baseline sets, singular value decomposition is used to perform temporal calculations on the two-dimensional offsets of multiple SAR image pairs to obtain the deformation rate and temporal deformation results of glaciers in the LOS and azimuth directions, thereby enabling temporal monitoring of glaciers.

[0056] This invention also proposes an adaptive homogeneous pixel offset tracking glacier time-series monitoring system, comprising: The data acquisition module is used to acquire multiple pairs of dual-temporal SAR images of the glacier; The boundary segmentation module is used to perform image recognition on dual-temporal SAR image pairs and output the binary segmentation results of glacier change boundaries; The pixel offset tracking module uses the binary segmentation result as a mask input for pixel offset tracking. By distinguishing between changing and non-changing regions within the pixel offset tracking window, it constructs an adaptive homogeneous pixel matching window. This ensures that the matching window of the master and slave images of the dual-temporal SAR image pair contains only homogeneous pixels with the same change characteristics as the center pixel, and performs normalized matching. When the local SNR of the homogeneous pixel matching window is less than a preset signal-to-noise ratio threshold, the offset is calculated and orbital errors are removed by adaptively adjusting the window size in the row and column directions, outputting two-dimensional temporal offset results in the LOS and azimuth directions.

[0057] This invention proposes a differential spatiotemporal attention mechanism, employing bi-branch multi-scale convolution, explicit differential weighting, and residual fusion to enhance the temporal differential information of single-band SAR and suppress speckle noise. Embedding the differential spatiotemporal attention mechanism into the TransUNet model can effectively enhance the representation of subtle changes, suppress speckle noise, and improve the accuracy of boundary extraction.

[0058] This invention proposes an adaptive homogeneous pixel offset tracking technology based on a change detection mask. Combined with an adaptive signal-to-noise ratio adjustment window, it can improve the stability and sub-pixel accuracy of pixel offset calculation. By constructing a collaborative framework for glacier change detection and homogeneous pixel offset tracking based on a change detection mask, it can achieve high-precision automated processing from change recognition to two-dimensional temporal deformation inversion. Throughout the process, the degree of automation is high, and it can effectively suppress noise interference and weak boundary problems in glacier change detection. It can also effectively solve the boundary decorrelation problem in glacier deformation calculation and improve monitoring accuracy.

[0059] Example 1 The study area in this example is Glacier A, such as... Figure 7 The ROIA and ROIB shown are stable regions of the glacier boundary used for accuracy assessment, and are represented by profiles. This invention collected LuTan-1 ascending-orbit SAR images from March 6, 2024 to December 11, 2024. These images were combined and input into the DSTA-TransUNet model trained in step 2 to infer and extract binary segmentation maps of the glacier's changing boundaries. The segmentation results of some image pairs are shown below. Figure 8 As shown, this binary image will serve as the mask input for subsequent pixel offset tracking.

[0060] Subsequently, the initial offsets in the LOS and azimuth directions were calculated using the POT method, and the glacier change boundary was extracted using a threshold segmentation strategy. This boundary, along with the glacier change boundary obtained from change detection, was used as a mask to compare the various pixel offset tracking methods in step 3. The comparison results of the LOS direction calculated by each offset method are shown below. Figure 9 As shown, the azimuth comparison results are as follows: Figure 10 As shown. Traditional POT can reflect overall deformation, but the boundaries are blurred and there is serious noise and mismatch. M-POT, although it uses a mask to suppress some interference, causes some information loss. M-APOT is just an adaptive window that further enhances noise suppression based on M-POT. CDM-POT, which uses a change detection mask, can effectively filter out most boundary noise and make the contour more complete. CDM-APOT further eliminates noise and captures the LOS and weak azimuth deformation that traditional POT cannot resolve. The boundaries are more continuous and smooth in both directions. The standard deviation of the stable region offset of each method in the azimuth and LOS directions is shown in the figure. Figure 11 As shown, CDM-APOT achieved the lowest standard deviation in both directions, 0.3402 in azimuth up and 0.434 in LOS up, which is significantly better than other methods, further verifying the accuracy improvement effect of CDM-APOT in glacier migration calculation.

[0061] The CDM-APOT method was used to calculate the migration values ​​of each SAR image pair of Glacier A, obtaining the two-dimensional deformation of the glacier in the LOS and azimuth directions. Subsequently, based on these two-dimensional migration results, the PO-SBAS method was used to calculate the deformation rate and time series results in the LOS and azimuth directions. The deformation rate results are shown below. Figure 12 As shown, the timing diagram drawn along the cross-section is as follows: Figure 13 As shown in the figure, the results indicate that the maximum LOS deformation rate of Glacier A exceeds 80 mm / d, and the maximum azimuth deformation rate exceeds 40 mm / d. The LOS and azimuth deformation rates are smaller in the central bend and at the terminus, mainly because the accumulation of large amounts of moraine at the bend increases the friction and shear resistance at the ice-bedrock interface, thus slowing glacier movement. The thinning of the terminus tongue reduces the gravitational driving force, resulting in a significant slowdown in the rate. The time-series diagram shows that the cumulative deformation is correlated with elevation. The upper part of the glacier experiences the most dramatic deformation due to its steep slope and high potential energy, with a maximum cumulative LOS deformation of 45.51 m and an azimuth deformation of 43.36 m. The deformation in the middle section decreases due to depression and deposition, with LOS deformation ranging from 17.64 to 32.89 m and azimuth deformation ranging from -17.09 m to 14.60 m. The lower section's slope increases again, causing a second increase in deformation, reaching 34.10 m in the LOS direction and 33.18 m in the azimuth direction. At the end, the deformation weakens and exhibits local fluctuations due to the influence of glacial deposits and freeze-thaw cycles.

[0062] The above examples demonstrate that the adaptive homogeneous pixel offset tracking glacier time-series monitoring proposed in this invention can achieve high-precision inversion of the entire chain from glacier change detection to time-series deformation monitoring, providing technical support for glacier monitoring in the Himalayas and assisting in the scientific prevention and risk management of related disasters.

[0063] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0064] Furthermore, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All references to this specification are incorporated by way of citation to disclose and describe methods relating to those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

Claims

1. An adaptive homogeneous pixel offset tracking method for glacier time-series monitoring, characterized in that, Includes the following steps: Acquire multiple pairs of dual-temporal SAR images of glaciers; Image recognition is performed on dual-temporal SAR image pairs to output the binary segmentation results of glacier change boundaries; The binary segmentation result is used as the mask input for pixel offset tracking. By distinguishing between the changing and non-changing regions within the pixel offset tracking window, an adaptive homogeneous pixel matching window is constructed. This ensures that the matching window of the master and slave images of the dual-temporal SAR image pair contains only homogeneous pixels with the same change characteristics as the center pixel, and normalized matching is performed. When the local SNR of the homogeneous pixel matching window is less than the preset signal-to-noise ratio threshold, the offset is calculated and the orbital error is removed by adaptively adjusting the window size in the row and column directions, and the two-dimensional temporal offset results in the LOS direction and azimuth direction are output.

2. The adaptive homogeneous pixel offset tracking glacier time-series monitoring method according to claim 1, characterized in that, The process of performing image recognition on each dual-temporal SAR image pair and outputting the binary segmentation result of the glacier change boundary specifically includes: By constructing a multi-source spatiotemporal fusion network model, dual-temporal SAR images are input into the pre-trained multi-source spatiotemporal fusion network model for identification; The multi-source spatiotemporal fusion network model is based on the TransUNet model. Before encoding in the TransUNet model, a feature fusion layer is constructed using a differential spatiotemporal attention mechanism to enhance the features of the dual-temporal SAR image pairs. The output differential spatiotemporal fusion features are used as the input to the encoder. The differential spatiotemporal attention mechanism includes a differential spatiotemporal attention module and a differential map enhancement module, wherein: The differential spatiotemporal attention module, in terms of spatial attention, obtains a dual-channel feature map by stitching together dual-temporal SAR image pairs by channel. After convolution and activation of the dual-channel feature map, a spatial attention map with the same size as the input dual-temporal SAR image pair is generated. This attention map is then used to spatially weight the absolute difference map of the two frames of the dual-temporal SAR image pair to obtain spatially weighted differential features. In terms of channel attention, multi-scale channel features of the dual-temporal SAR images from each branch are extracted and fused through two independent parallel branches. The fused channel features are spatially converged and mapped into channel attention vectors. A scaling strategy is used to weight the multi-scale channel features of the extracted dual-temporal SAR images from each branch to obtain channel-weighted features. Residual projection is then performed on the dual-channel feature map to output residual projection features. Finally, the spatially weighted differential features, channel-weighted features, and residual projection features are element-wise added to obtain the differential spatiotemporal attention features. The difference map enhancement module includes a dual convolutional block. It obtains a difference map by applying anisotropic diffusion filtering with speckle noise suppression and logarithmic differencing to the dual-temporal SAR image pair. The dual convolutional block extracts the subtle variation features of the difference map to obtain the difference map enhancement features. Finally, the differential spatiotemporal attention features and differential map enhancement features are merged to obtain differential spatiotemporal fusion features, which are used as input to the encoder.

3. The adaptive homogeneous pixel offset tracking glacier time-series monitoring method according to claim 1, characterized in that, The binary segmentation result is used as a mask input for pixel offset tracking. By distinguishing between changing and non-changing regions within the pixel offset tracking window, an adaptive homogeneous pixel matching window is constructed. This ensures that the matching window of the master and slave images of the dual-temporal SAR image pair contains only homogeneous pixels with the same variation characteristics as the center pixel, and normalized matching is performed. Specifically, this includes: The binary segmentation map of the glacier change boundary is used as a mask to distinguish the changed and unchanged areas within the pixel offset tracking window. An adaptive homogeneous pixel matching window is constructed, and the position with the largest similarity measure under the matching window is found by determining the normalized cross-correlation coefficient (NCC) between the master and slave images of homogeneous pixels. The normalized cross-correlation coefficient (NCC) between master and slave images of homogeneous pixels is: ； in, It is the set of pixels in the homogeneous pixel matching window. These are the coordinates of the pixels within the window. For window size, and They are respectively The amplitude values ​​of the main image and the secondary image. and These represent the average amplitude values ​​of the main image and the secondary image within the window, respectively.

4. The adaptive homogeneous pixel offset tracking glacier time-series monitoring method according to claim 3, characterized in that, When the local SNR of the homogeneous pixel matching window is less than a preset signal-to-noise ratio threshold, the offset is calculated and the orbital error is removed by adaptively adjusting the window size in the row and column directions respectively, and the two-dimensional temporal offset results in the LOS direction and azimuth direction are output, specifically including: The local signal-to-noise ratio of the homogeneous pixel matching window is evaluated by the matching results. When the local signal-to-noise ratio of the homogeneous pixel matching window is less than the preset signal-to-noise ratio threshold, it indicates that the cross-correlation solution result of the current window pixels is unreliable. The pixel reconstruction window is then adaptively adjusted in the row and column directions respectively. When the local SNR in the homogeneous pixel matching window is greater than or equal to the preset signal-to-noise ratio threshold, the constructed adaptive homogeneous pixel window is used for matching, and the orbital error is removed by combining quadratic polynomial fitting to calculate the LOS and azimuth two-dimensional temporal offset results.

5. The adaptive homogeneous pixel offset tracking glacier time-series monitoring method according to claim 4, characterized in that, The calculation of the two-dimensional temporal offsets in the LOS and azimuth directions includes the use of singular value decomposition based on the small baseline set concept to perform temporal calculation of the two-dimensional offsets of multiple SAR image pairs, thereby obtaining the deformation rate and temporal deformation results of the glacier in the LOS and azimuth directions.

6. The adaptive homogeneous pixel offset tracking glacier time-series monitoring method according to claim 2, characterized in that, The differential spatiotemporal attention module, in terms of spatial attention, obtains a dual-channel feature map by stitching together dual-temporal SAR image pairs by channel. After convolution and activation of the dual-channel feature map, a spatial attention map with the same size as the input dual-temporal SAR image pair is generated. This attention map is then used to spatially weight the absolute difference map of the two frames of the dual-temporal SAR image pair to obtain spatially weighted differential features. In terms of channel attention, multi-scale channel features of the dual-temporal SAR images from each branch are extracted and fused through two independent parallel branches. The fused channel features are spatially converged and mapped into channel attention vectors. A scaling strategy is used to weight the multi-scale channel features of the extracted dual-temporal SAR images from each branch to obtain channel-weighted features. Finally, residual projection is performed on the dual-channel feature map to output residual projection features. The spatially weighted difference features, channel-weighted features, and residual projection features are element-wise added to obtain the differenced spatiotemporal attention features, which specifically include: In terms of spatial attention, dual-temporal SAR images are stitched together channel by channel to obtain dual-channel feature maps. X This dual-channel feature map X A spatial attention map of the same size as the input bi-temporal SAR image pair is generated by a 7×7 convolution followed by Sigmoid activation. m ; Calculate the absolute difference map of the two frames of the dual-temporal SAR image pair and compare it with m Dot product yields spatially weighted features. ; Regarding channel attention, multi-scale channel features of dual-temporal SAR images are extracted from each branch through two independent parallel branches. The channel features output from the two branches are then added together in the channel dimension to obtain the fused features. F and the fusion features F We obtain the channel attention weight vector by performing global average pooling and two layers of 1×1 convolution mapping, ReLU activation, and Sigmoid mapping. s , with 1+ s The scaling strategy weights the channel features of the two branch outputs separately and then sums them to obtain the channel-weighted features. ; And the dual-channel feature map X Residual projection features are obtained by applying 1×1 convolution. ; Finally, the channel weighted features are... Spatial weighted features and residual projection features Adding these three together yields the differential spatiotemporal fusion feature: ； Differential spatiotemporal fusion features are used as input to the encoder.

7. The adaptive homogeneous pixel offset tracking glacier time-series monitoring method according to claim 2, characterized in that, The acquisition of multiple sets of dual-temporal SAR image pairs of glaciers specifically includes: Collect multi-temporal, multi-source SAR images of typical mountain glaciers, and preprocess the images through registration and geometric correction to obtain multiple sets of dual-temporal SAR image pairs; Based on the preprocessed pairs of dual-temporal SAR images, the glacier change areas of some pairs of dual-temporal SAR images are binarized and labeled. A training set is constructed by generating a label map, and the pairs of dual-temporal SAR images are divided into a training set and a test set. The multi-source spatiotemporal fusion network model is trained using the training set.

8. The adaptive homogeneous pixel offset tracking glacier time-series monitoring method according to claim 7, characterized in that, The image recognition of the dual-temporal SAR image pairs, and the output of the binary segmentation result of the glacier change boundary, specifically includes: Obtain the weights of the trained multi-source spatiotemporal fusion network model; Normalize the dual-temporal SAR image pairs for a specified area in the test set and generate an absolute difference map; Overlapping sliding window slicing is used to divide the normalized dual-temporal SAR image pair into multiple overlapping blocks. When the block size is insufficient, zero padding is used to meet the input requirements of the multi-source spatiotemporal fusion network model, thus obtaining the sliced ​​dual-temporal SAR image pair. The multi-source spatiotemporal fusion network model is trained by inputting the segmented dual-temporal SAR image pairs and absolute difference maps into the model for inference. The inference results of each segment are weighted, fused and normalized, and finally, the binary segmentation results of the glacier change and non-change regions are obtained by threshold segmentation.

9. An adaptive homogeneous pixel offset tracking glacier time-series monitoring system, characterized in that, include: The data acquisition module is used to acquire multiple pairs of dual-temporal SAR images of the glacier; The boundary segmentation module is used to perform image recognition on dual-temporal SAR image pairs and output the binary segmentation results of glacier change boundaries; The pixel offset tracking module uses the binary segmentation result as a mask input for pixel offset tracking. By distinguishing between changing and non-changing regions within the pixel offset tracking window, it constructs an adaptive homogeneous pixel matching window. This ensures that the matching window of the master and slave images of the dual-temporal SAR image pair contains only homogeneous pixels with the same change characteristics as the center pixel, and performs normalized matching. When the local SNR of the homogeneous pixel matching window is less than a preset signal-to-noise ratio threshold, the offset is calculated and orbital errors are removed by adaptively adjusting the window size in the row and column directions, outputting two-dimensional temporal offset results in the LOS and azimuth directions.

Images