Fixed ice line linear deformation detection method based on space-borne InSAR and deep learning

Abstract

The application discloses a kind of linear deformation detection methods of fixed ice based on space-borne InSAR and deep learning, belong to sea ice remote sensing monitoring technical field.To solve the problem that prior art cannot automatically identify the subtle linear deformation characteristics before fixed ice cracking, the application generates a coherence map by using Sentinel-1 image data for differential interference processing, constructs a pixel-level annotation dataset through manual interpretation, constructs an improved U-Net network framework LDF-UNet model suitable for linear deformation feature detection, and introduces a Dice loss function for training to automatically detect linear deformation features from the coherence map.The application realizes large-scale mapping of fixed ice linear deformation features at 60-meter resolution, and the detected features are highly consistent with the ice arch cracking boundary, which can serve as an early indicator of fixed ice instability.

Description

Technical Field

[0001] This invention belongs to the field of sea ice remote sensing monitoring technology, and particularly relates to a method for detecting linear deformation of fixed ice based on spaceborne InSAR and deep learning. Background Technology

[0002] Landfast ice is a type of sea ice that forms along the Arctic coast during winter and spring and remains largely stationary, playing a vital role in Arctic ecosystems and coastal communities. Since the early 21st century, many Arctic residents have reported that coastal landfast ice breakup has occurred significantly earlier than in the 1990s. With the increasing severity of Arctic climate change, the timing of spring melting and breakup of landfast ice has become increasingly unpredictable, posing growing risks to Arctic shipping and coastal production activities. Therefore, improving the ability to predict landfast ice breakup, even on scales ranging from days to subseasonal, remains of significant practical importance.

[0003] Traditional fixed ice deformation monitoring primarily relies on buoy trajectory observations and differential global navigation satellite system (DGNSS) observations. While these methods provide valuable on-site information, their sparse spatial distribution makes it difficult to characterize deformation structures within large areas of fixed ice. With the development of remote sensing technology, researchers can observe sea ice deformation processes on a larger scale. Existing studies have utilized optical remote sensing imagery and synthetic aperture radar backscattering data to identify linear deformation features based on their elongated geometric morphological characteristics, and analyzed their spatial distribution, density, and directional features. However, these methods typically focus on identifying deformation structures after significant sea ice fracturing or breaking, while automatic identification methods for linear deformation features prior to fracturing remain relatively limited.

[0004] Interferometric Synthetic Aperture Radar (InSAR) provides an effective means of monitoring surface deformation with high spatial resolution and centimeter- to millimeter-level accuracy. In fixed ice environments, deformation processes often lead to a decrease in interferometric coherence, resulting in narrow, low-coherence bands in coherent images corresponding to local deformation zones. Compared to interferometric phase, coherence is less sensitive to atmospheric delay and residual topographic errors, thus potentially serving as a more robust deformation indication in complex sea ice environments. However, automatically identifying these narrow and sparsely distributed deformation structures in coherent InSAR images remains a significant challenge, especially at the regional scale. How to automatically and accurately identify these subtle linear deformation features from massive amounts of data is a pressing technical problem in the field of fixed ice monitoring. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a fixed ice linear deformation detection method based on spaceborne InSAR and deep learning, thereby resolving the issues present in the prior art.

[0006] Firstly, to achieve the above objectives, this invention provides a method for detecting linear deformation of fixed ice based on spaceborne InSAR and deep learning, comprising the following steps:

[0007] Temporal synthetic aperture radar (SAR) image data of the study area is acquired. Differential interferometry is performed on the temporal SAR image data to generate an interferometric coherence map. The linear low coherence structures in the interferometric coherence map are manually interpreted and a pixel-level labeled dataset is constructed. A deep learning semantic segmentation model based on the U-Net network is constructed. The pixel-level labeled dataset is input into the deep learning semantic segmentation model for training to obtain a linear deformation feature detection model. The interferometric coherence map to be detected is input into the linear deformation feature detection model to detect the linear deformation features of fixed ice.

[0008] Secondly, the present invention also provides a fixed ice linear deformation detection system based on spaceborne InSAR and deep learning, for implementing a fixed ice linear deformation detection method based on spaceborne InSAR and deep learning, the system comprising:

[0009] The data acquisition module is used to acquire time-series synthetic aperture radar image data of the study area;

[0010] The differential interferometry processing module is used to perform differential interferometry processing on the time-series synthetic aperture radar image data to generate an interferometric coherence map;

[0011] The dataset construction module is used to manually interpret the linear low-coherence structures in the interferometric coherence map and construct a pixel-level labeled dataset.

[0012] The model training module is used to construct a deep learning semantic segmentation model based on the U-Net network, and input the pixel-level labeled dataset into the deep learning semantic segmentation model for training to obtain a linear deformation feature detection model.

[0013] The feature detection module is used to input the interferometric coherence map to be detected into the linear deformation feature detection model to detect the linear deformation features of fixed ice.

[0014] Thirdly, the present invention also provides a computer terminal device, comprising:

[0015] One or more processors;

[0016] A memory, coupled to the processor, for storing one or more programs;

[0017] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the fixed ice linear deformation detection method based on spaceborne InSAR and deep learning in the first aspect described above.

[0018] Fourthly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the method for detecting linear deformation of fixed ice based on spaceborne InSAR and deep learning in the first aspect described above.

[0019] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method for detecting linear deformation of fixed ice based on spaceborne InSAR and deep learning in the first aspect described above.

[0020] Compared with the prior art, the present invention has the following advantages and technical effects:

[0021] This invention provides a method for detecting linear deformation of fixed ice based on spaceborne InSAR and deep learning. By constructing a deep learning semantic segmentation model based on the U-Net network and introducing the Dice loss function, this invention achieves accurate identification of linear deformation features of fixed ice from interferometric coherence images, effectively extracting narrow, elongated, and sparse deformation structures. Applying the trained linear deformation feature detection model to interferometric coherence images in different Arctic regions, it can stably generate large-scale spatial distribution maps of linear deformation features, demonstrating good detection performance under various sea ice environmental conditions. Spatial overlay and temporal comparison analysis of the detected linear deformation feature distribution with the fixed ice fracture boundary shows that the detected linear deformation features have significant spatial consistency with the final fracture boundary, and appear several days before the final collapse of the ice arch, serving as an early indicator signal of the instability of fixed ice structures. Attached Figure Description

[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0023] Figure 1 This is a schematic diagram of the research area and Sentinel-1 coverage area of ​​this invention according to an embodiment of the invention;

[0024] Figure 2 The flowchart below is an embodiment of the present invention;

[0025] Figure 3 This is an example of a coherence graph and its corresponding LDF annotation results according to an embodiment of the present invention;

[0026] Figure 4 This is a framework diagram of the LDF-UNet model proposed in this embodiment of the invention for detecting linear deformation features from InSAR coherent images;

[0027] Figure 5 This is a schematic diagram of the monthly-scale spatial distribution and temporal variation of D-InSAR coherence in the study area in 2017, according to an embodiment of the present invention.

[0028] Figure 6 Heatmaps of LDF detection results for different models in this embodiment of the invention;

[0029] Figure 7 This is a schematic diagram of the large-scale LDF mapping results obtained by LDF-UNet in multiple Arctic regions according to an embodiment of the present invention. The upper image is a Sentinel-1 InSAR coherence map of multiple Arctic regions, including Nares Strait, the Canadian Arctic Archipelago, the Beaufort Sea, the East Siberian Sea, the Laptev Sea, and the Kara Sea. The lower image is the corresponding LDF distribution results automatically extracted by LDF-UNet. The orange boxes indicate representative sub-regions used for detailed analysis in subsequent images.

[0030] Figure 8 This is a schematic diagram of a magnified local LDF result detected in a typical Arctic region according to an embodiment of the present invention;

[0031] Figure 9 The above are Sentinel-1 SAR amplitude diagrams according to an embodiment of the present invention, wherein (a) is a schematic diagram on May 6, 2017; (b) is a schematic diagram on May 10, 2017; and (c) is a schematic diagram on May 12, 2017.

[0032] Figure 10 This is a schematic diagram of the LDF time-series evolution of the fracture boundary at the leading edge of the ice arch rupture according to an embodiment of the present invention. Detailed Implementation

[0033] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0034] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0035] Interferometric Synthetic Aperture Radar (InSAR) provides an effective means of observing deformation in stationary ice. Linear Deformation Features (LDFs) are narrow structures within stationary ice associated with crack development. However, research on automatic LDF detection based on InSAR coherent images remains limited. To address this issue, this invention proposes a framework that integrates differential InSAR and deep learning semantic segmentation for detecting LDFs from Sentinel-1 coherent images. An improved segmentation model, LDF-UNet, is built on the U-Net network structure to enhance the extraction capability of slender linear structures. Simultaneously, the Dice loss function is introduced to improve the segmentation performance for sparse and slender targets. Experimental results show that compared with various traditional segmentation networks, the proposed LDF-UNet achieves better performance in deformation feature detection, with an F1-score improvement of approximately 14%. This model achieves an overall accuracy of 0.97 and an F1-score of 0.63 in the LDF detection task. Experiments conducted in multiple Arctic regions further validated the robustness of the method and enabled large-scale mapping of LDFs at 60 m resolution. In the Nares Strait region, the detected LDFs showed good spatial consistency with the crack boundaries associated with the 2017 ice arch rupture event and appeared several days before the final collapse of the ice arch, suggesting that LDFs detected based on InSAR coherent imagery may serve as an early indicator of ice instability.

[0036] Example 1

[0037] This embodiment provides a method for detecting linear deformation of fixed ice based on spaceborne InSAR and deep learning, including:

[0038] Temporal synthetic aperture radar (SAR) image data of the study area is acquired. Differential interferometry is performed on the temporal SAR image data to generate an interferometric coherence map. The linear low coherence structures in the interferometric coherence map are manually interpreted and a pixel-level labeled dataset is constructed. A deep learning semantic segmentation model based on the U-Net network is constructed. The pixel-level labeled dataset is input into the deep learning semantic segmentation model for training to obtain a linear deformation feature detection model. The interferometric coherence map to be detected is input into the linear deformation feature detection model to detect the linear deformation features of fixed ice.

[0039] As one implementation method in this embodiment, the process of generating the interferometric coherence map includes: selecting an image pair with a time baseline of 6 days or 12 days as the master image and the slave image, registering the master image and the slave image, multiplying the complex signals of the registered master image and the slave image to generate an interferometric map, estimating the interferometric coherence coefficient using a sliding window in the interferometric map, and then performing geocoding processing on the obtained interferometric coherence map.

[0040] As one implementation method in this embodiment, the process of constructing the pixel-level labeled dataset includes: selecting images with linear low coherence structures in the interferometric coherence map for manual interpretation, labeling pixels belonging to linear deformation structures as deformation categories, labeling other pixels as background categories to generate a labeling mask, dividing the interferometric coherence map into image blocks of a set size, and constructing a training set and a validation set based on the image blocks and their corresponding labeling masks.

[0041] As one implementation method in this embodiment, the structure of the deep learning semantic segmentation model includes: an encoder for progressively extracting high-level semantic features, a decoder for progressively restoring spatial resolution, and a skip connection connecting the encoder and the decoder. The skip connection is used to perform channel concatenation of the feature map output by the encoder and the feature map upsampled by the decoder to achieve multi-scale feature fusion.

[0042] As one implementation method in this embodiment, the training process includes: calculating the loss value between the model prediction result and the labeled mask using a hybrid loss function that combines the cross-entropy loss function and the Dice loss function; iteratively updating the model parameters through the backpropagation algorithm; and evaluating the model performance using the validation dataset after each training cycle to save the optimal model.

[0043] As one implementation method in this embodiment, the method further includes: performing spatial superposition and temporal comparison analysis on the detected linear deformation feature distribution of fixed ice and the fixed ice fracture boundary obtained by synthetic aperture radar amplitude image recognition.

[0044] The study area of ​​this invention covers multiple regions along the Arctic Ocean coast where fixed ice is prone to form. For example... Figure 1 As shown, this invention selected several representative research areas along the Arctic marginal seas to evaluate the effectiveness of the proposed method, including the Laptev Sea, the East Siberian Sea, the Beaufort Sea, and the Canadian Arctic Archipelago. These areas represent typical scenarios where fixed ice forms along complex coastlines and island distribution environments.

[0045] Within the aforementioned region, the Nares Strait, located between Greenland and Ellesmere Island, was selected as the primary study area for time-series analysis. This strait, approximately 500 km long and 30-40 km wide, is a crucial channel for Arctic sea ice export and is known for the frequent formation of ice arches, which play a significant regulatory role in the Arctic sea ice export process. The northern part of the strait typically contains some of the thickest sea ice in the Arctic region.

[0046] Table 1. Sentinel-1 data used in this invention

[0047]

[0048] The satellite data used in this invention comes from the Sentinel-1 synthetic aperture radar mission. Sentinel-1 operates in the C-band and possesses all-weather, all-day imaging capabilities, making it particularly suitable for monitoring sea ice dynamics in polar regions. Figure 1 As shown, this invention selected multiple Sentinel-1 orbital data distributed in the marginal seas of the Arctic. The numbers in each box in the figure represent the relative orbit and frame number of the corresponding Sentinel-1 image. Detailed information on the relevant Sentinel-1 orbital information, acquisition mode, spatial resolution, and corresponding research objectives is summarized in Table 1.

[0049] For the Nares Strait region (Track 26-310), this invention selected Sentinel-1 imagery acquired throughout 2017 to analyze the temporal evolution of linear deformation features (LDFs). The study employed single-look complex (SLC) data in interferometric wide (IW) mode, with a spatial resolution of approximately 5 × 20 m, for differential interferometric processing (D-InSAR) to invert deformation information related to fixed ice dynamics (Torres et al., 2012). Furthermore, this invention also used Sentinel-1 ground range detected (GRD) imagery, with a spatial resolution of approximately 10 × 10 m, to identify fixed ice fracturing processes associated with ice arch collapse.

[0050] The overall detection process for fixed ice linear deformation features (LDFs) is as follows: Figure 2As shown, the proposed framework combines interferometric SAR analysis with a deep learning-based semantic segmentation model to automatically identify deformation features from interferometric coherent images. The framework mainly consists of three parts: interferometric SAR data processing, deep learning-based deformation feature extraction, and accuracy evaluation.

[0051] In the first stage, Sentinel-1 SAR data was processed using Differential Interferometric Synthetic Aperture Radar (D-InSAR) to generate interferometric coherence maps. Interferometric coherence is used to characterize the similarity of radar backscattered signals between two SAR observations. When sea ice deforms, structural disturbances such as cracks, shearing, and changes in surface roughness alter the scattering characteristics within the radar resolution cell, leading to a decrease in interferometric coherence. Therefore, deformed regions typically appear as elongated, low-coherence stripes in the coherence map, serving as important indicators for identifying LDFs (Local Deformations of Sea Ice). To determine the optimal observation period for monitoring fixed ice using InSAR, this invention analyzed the seasonal variation characteristics of coherence in the study area. Monthly-scale coherence maps were generated based on D-InSAR interferograms from January to December 2017, and the spatial distribution and temporal variation of coherence were analyzed to identify periods with relatively high and stable coherence, which are more suitable for detecting fixed ice deformation characteristics.

[0052] In the second stage, a deep learning semantic segmentation model based on the U-Net structure is used to automatically detect LDFs from coherent images. Since LDFs typically exhibit sparse, narrow, and elongated structures, traditional segmentation methods often struggle to extract them accurately. To address this issue, this invention constructs an improved segmentation framework—LDF-UNet—to enhance the detection capability for narrow, linearly deformable structures.

[0053] By manually interpreting coherent images, linear low-coherence structures associated with deformation zones were identified, and a pixel-level labeled dataset was constructed. The dataset was then divided into training and validation sets, and the model was trained using supervised learning. Finally, the trained model was applied to the study area to achieve automatic extraction of LDFs (Local Deformation Fields). Furthermore, the detected deformation features were jointly analyzed with fixed ice fracture regions to investigate the spatial distribution characteristics and temporal evolution of deformation structures before ice sheet fracturing.

[0054] D-InSAR processing:

[0055] In this invention, the D-InSAR processing method is used to process Sentinel-1 C-band SAR data provided by the European Space Agency to generate coherence maps. The D-InSAR processing flow mainly includes steps such as interferometric pair selection, image registration, interferogram generation, and coherence estimation. After multi-look processing, the spatial resolution of the obtained interferogram and coherence map is 60 m, thereby improving the stability of coherence estimation.

[0056] To ensure high interferometric quality, SAR image pairs with time baselines of 6 days or 12 days were selected for interferometric processing. Shorter time baselines effectively reduce temporal decoherence caused by rapid changes in the sea ice surface while maintaining sufficient sensitivity to deformation signals. Furthermore, precise orbital data provided by the Sentinel-1 data product was utilized during interferometric processing to improve geometric positioning accuracy.

[0057] Image registration is a crucial step in the interferometric processing workflow. In this process, geometric corrections are used to precisely align the master and slave images spatially, ensuring that corresponding pixels represent the same ground target. First, coarse registration is performed using orbital information, followed by fine registration based on image cross-correlation methods to achieve sub-pixel level registration accuracy. After image registration is complete, an interferogram is generated by multiplying the complex SAR signal of the master image with the complex conjugate of the slave image. The interferometric signal can be represented as:

[0058]

[0059] in, Represents a complex interferogram. and These represent the complex SAR signals of the main image and the secondary image, respectively. Indicates the complex conjugate of the image; and These represent the phases of the primary image and the secondary image, respectively.

[0060] After generating the interferogram, the coherence coefficient is further calculated to quantify the similarity between the radar signals of the two images. The coherence value ranges from 0 to 1, with a higher value indicating stronger phase correlation. The coherence is calculated using the maximum likelihood estimation method within a sliding window, and the stability of the estimation results is improved by statistically averaging multiple neighboring pixels.

[0061]

[0062] in, Represents the coherence coefficient. To estimate the number of pixels in the window, and These represent the main image and the secondary image at the 1st... The complex values ​​at each pixel are used to obtain the final coherent image, which is then geocoded and used as the main input data for LDF detection.

[0063] Dataset Construction:

[0064] To train a deep learning model for automatic detection of LDFs, this invention constructs a pixel-level labeled dataset based on manual interpretation of coherence maps. First, interferometric coherence maps with obvious linear low-coherence structures are selected for labeling; these structures correspond to deformation bands within fixed ice. Figure 3 This demonstrates an example of a coherent image and its corresponding annotation mask. During the annotation process, pixels belonging to the LDF structure are labeled as deformation classes, while the remaining pixels are labeled as background.

[0065] These deformable structures typically exhibit subtle, fragmented, and low-contrast features, embedded in the complex sea ice background. Furthermore, LDFs usually constitute only a small proportion of the image and present elongated linear geometries. To facilitate model training, the coherence map was divided into 64×64 pixel patches. These patches were used to construct the training and validation sets. After data preprocessing and sample construction, approximately 7,500 labeled samples were generated. The dataset was divided into training and validation sets in a 7:3 ratio.

[0066] LDF detection based on deep learning:

[0067] Automatic detection of LDFs is achieved through a deep learning semantic segmentation model called LDF-UNet, which is built upon the classic U-Net network architecture. Due to its encoder-decoder structure and ability to fuse multi-scale contextual information with high-resolution spatial details, U-Net has been widely used for pixel-level segmentation tasks. These characteristics make it particularly suitable for detecting narrow and elongated deformable structures in coherent sea ice images.

[0068] In the proposed framework, the encoder progressively extracts high-level semantic features through stacked convolutional layers and downsampling operations; the decoder progressively recovers spatial resolution through upsampling layers. Simultaneously, skip connections are introduced between corresponding layers in the encoder and decoder to achieve the fusion of low-level spatial features and high-level semantic features. The feature fusion process can be represented as:

[0069]

[0070] in, Represents the encoder feature map, Represents the decoder feature map, Indicates an upsampling operation. This refers to channel concatenation. This multi-scale feature fusion mechanism can integrate contextual information from different spatial scales, thereby improving the detection capability of narrow deformation structures in complex sea ice environments.

[0071] This network takes a single-channel coherent image as input and outputs a pixel-level probability map to represent the probability that each pixel belongs to an LDF region. To improve segmentation performance for sparse deformable structures, this invention employs a hybrid loss function combining Cross-Entropy Loss and Dice Loss. Figure 4 ):

[0072]

[0073] The cross-entropy loss is defined as follows:

[0074]

[0075] in, Represents pixels The true label, This represents the model's predicted probability for that pixel.

[0076]

[0077] in, Indicates the predicted probability. Indicates the true label, This is a small constant used to avoid numerical instability.

[0078] Dice loss emphasizes the spatial overlap between the predicted region and the ground truth label, and is particularly effective for detecting small or sparse targets. Since LDFs typically exhibit narrow linear structures that occupy only a small proportion of the image, introducing Dice loss can significantly improve the model's ability to recognize these deformation features.

[0079] Model training:

[0080] The LDF-UNet model is implemented using the PyTorch deep learning framework. During training, coherent image patches and their corresponding segmentation annotation masks are input into the network in mini-batches. Model parameters are updated using the Adam optimization algorithm, with an initial learning rate set to... .

[0081] Since LDFs constitute only a small proportion of an image, there is a significant class imbalance between deformed pixels and background pixels. To address this issue, this invention employs a weighted cross-entropy loss function, assigning higher weights to LDF classes, thereby guiding the model to pay more attention to deformation features during training.

[0082] The model parameters are updated iteratively through forward propagation and back propagation. At the end of each training epoch, the model performance is evaluated using a validation dataset.

[0083] Table 2. Training process of the proposed LDF-UNet model

[0084]

[0085] Model accuracy verification:

[0086] To quantitatively evaluate the performance of the proposed LDF-UNet model, this invention employs several commonly used binary semantic segmentation metrics, including Precision, Recall, Intersection over Union (IoU), and F1-score. These metrics are calculated only for LDF categories. Since the primary objective of this invention is to detect linear deformation features in stationary ice, the evaluation focuses solely on the detection performance of LDF pixels, without evaluating background pixels.

[0087] These metrics are calculated based on the confusion matrix between predicted labels and true labels. Let TP, FP, and FN represent the number of true positives, false positives, and false negatives in the LDF category, respectively. Specifically, TP represents the number of pixels correctly identified as LDF, FP represents the number of background pixels incorrectly identified as LDF, and FN represents the number of LDF pixels incorrectly classified as background.

[0088] Precision represents the proportion of correctly predicted LDF pixels out of all pixels predicted as LDF. It reflects the reliability of the detection results and is defined as follows:

[0089]

[0090] Recall represents the proportion of correctly detected pixels out of all true LDF pixels, and is used to measure the model's ability to recognize deformed structures. Since LDFs typically exhibit sparse and narrow structures in coherent images, this metric is particularly important in this invention. Recall is defined as:

[0091]

[0092] IoU (Intersection over Union) is used to quantify the spatial overlap between the predicted LDF region and the true LDF region. It is defined as follows:

[0093]

[0094] Finally, the F1-score, which takes into account both precision and recall, is used to measure the overall detection performance of the model and is defined as follows:

[0095]

[0096] Seasonal variation characteristics of InSAR coherence:

[0097] Figure 5 This paper presents the monthly coherence distribution of the study area obtained based on D-InSAR from January to December 2017. The coherence of fixed ice in the nearshore area exhibits a clear seasonal variation. In January and February, high coherence areas begin to appear along the coastline, but their spatial extent remains relatively limited, with average coherence of 0.47 and 0.50, respectively. In March and April, the high coherence areas expand significantly in the nearshore waters, forming a more continuous distribution band, reaching the largest spatial coverage of the year. The average coherence in March and April reaches 0.72 and 0.65, respectively, the highest levels of the year. During this period, the high coherence areas are more widespread and have stronger spatial continuity, and the spatial distribution of fixed ice is most clearly shown in the imagery.

[0098] Starting in May, the area of ​​high coherence decreased significantly, and overall coherence declined markedly, with the average coherence dropping to 0.35. From June to August, the study area was predominantly characterized by low coherence, with average coherence values ​​of 0.22, 0.23, and 0.22, respectively, and the high coherence area essentially disappeared. Entering September, some nearshore areas again exhibited localized high coherence characteristics, with the average coherence rising to 0.37, but its spatial extent remained small and discontinuous. From October to December, coherence gradually recovered, with average values ​​of 0.39, 0.39, and 0.42, respectively, but the spatial coverage of the high coherence area was still significantly smaller than in March and April. Considering both the spatial distribution characteristics of coherence and the monthly average, March and April not only had the highest average coherence throughout the year, but also the most extensive and continuous spatial distribution of high coherence areas, clearly characterizing the spatial extent of nearshore fixed ice. Therefore, March and April are the optimal observation periods for monitoring and analyzing fixed ice using InSAR.

[0099] LDF-UNet model performance evaluation:

[0100] To evaluate the effectiveness of the proposed LDF-UNet model, this invention selected several commonly used semantic segmentation networks for comparative experiments, including DeepLabV3, PSPNet, and U-Net. For each network architecture, improved versions incorporating the Dice optimization strategy were further constructed, denoted as DeepLabV3-DC, PSPNet-DC, and UNet-DC, respectively. The proposed LDF-UNet actually corresponds to the UNet-DC model optimized for the land margin fixed ice deformation feature detection task.

[0101] The quantitative evaluation results of all models are summarized in Table 3. Although LDF-UNet achieves an overall accuracy (OA) of 0.97, this metric is primarily influenced by the correct classification of a large number of background pixels. Since linear deformation features constitute only a small proportion of an image, OA alone is insufficient to fully reflect the model's ability to recognize small-scale structures. Therefore, this invention focuses on the Precision, Recall, and F1-score of the LDF category, which can more rigorously evaluate the model's ability to detect slender, discrete deformation bands.

[0102] Table 3. Recognition accuracy of different models for linear deformation features

[0103]

[0104] Among the baseline models, PSPNet achieved the highest precision (0.72), while DeepLabV3 and UNet achieved precisions of 0.69 and 0.71, respectively. However, the recall values ​​of these models were relatively low, all falling within the range of 0.45-0.50, indicating that a considerable portion of deformation features remained undetected. Therefore, the IOU values ​​of the baseline models were only between 0.38 and 0.41, reflecting the significant challenge of accurately segmenting narrow and fragmented deformation structures in fixed ice environments along continental margins.

[0105] After introducing the Dice optimization strategy, the performance of each network model was improved to varying degrees. For example, DeepLabV3's precision improved from 0.69 to 0.71, and recall from 0.46 to 0.50; PSPNet-DC's precision improved from 0.72 to 0.73, and recall from 0.50 to 0.53. The UNet architecture showed the most significant improvement, with recall increasing from 0.45 to 0.55 and IOU from 0.40 to 0.46. These results demonstrate that introducing Dice optimization can significantly enhance the model's ability to capture deformation features. Compared to the original UNet, UNet-DC's F1-score improved from 0.55 to 0.63, representing the most significant performance improvement among all tested models.

[0106] Among all models, LDF-UNet achieved the best overall performance, with Precision, Recall, IOU, and F1-score of 0.74, 0.55, 0.46, and 0.63, respectively. Compared with the baseline model, LDF-UNet showed significant advantages in all evaluation metrics. In particular, the improved Recall indicates that the model can detect more deformation features, while the higher IOU and F1-score indicate better spatial consistency between the predictions and the ground truth annotations.

[0107] In addition to quantitative assessment, Figure 6 The presentation also showcases a visual comparison of segmentation results from different models. Two representative examples are selected to illustrate the differences between the models. In the first example, multiple thin deformation features are visible in the SAR image, exhibiting a clear linear structure. While the baseline model can identify the main deformation regions, discontinuities often appear at the boundaries of the deformation lines. For example, although DeepLabV3 and PSPNet can partially capture the deformation structure, several discontinuities and irregular boundaries still exist in the prediction results. In some locations, some deformation lines are not detected, resulting in a noticeable fragmentation in the segmentation results.

[0108] After introducing Dice optimization, the consistency of segmentation results was significantly improved. The spatial consistency between the improved DC models and the ground truth annotations was higher, especially in deformation feature boundary regions. Compared to the original models, DeepLabV3-DC and PSPNet-DC generated more continuous and better connected deformation lines. Simultaneously, the boundaries in the prediction results were clearer, indicating that the models can more effectively distinguish deformation regions from the surrounding background ice.

[0109] Among all models, LDF-UNet achieved the most accurate segmentation results. The predicted deformation features closely matched the ground truth annotations and effectively preserved the boundary morphology of the deformed structures. In contrast, baseline models often produced fragmented or incomplete detection results, while LDF-UNet generated more continuous deformation lines and effectively reduced the occurrence of missed segments. Furthermore, the boundaries of the predicted deformed regions were smoother and more consistent, indicating better stability in spatial structure representation.

[0110] The second example further illustrates the method's advantage in detecting smaller-scale deformation features. In this scenario, the deformation structures in the SAR image are short and inconspicuous. While the DC-improved model enhances the continuity of predictions, gaps still exist at deformation boundaries. In contrast, LDF-UNet successfully identifies most deformation features and better preserves their geometry.

[0111] Overall, both quantitative metrics and visualization results demonstrate the effectiveness of the proposed method. The Dice optimization strategy improves the deformation feature detection capabilities of each segmentation model, particularly in boundary characterization and structural continuity. LDF-UNet further enhances detection performance, achieving the most accurate segmentation results among all tested models. The deformation features detected by LDF-UNet exhibit clearer boundaries, better continuity, and higher consistency with ground truth annotations, indicating that this method has significant application potential in extracting linear deformation features of fixed ice along the land margin in SAR images.

[0112] Large-scale LDF mapping based on LDF-UNet:

[0113] To further evaluate the capabilities of the proposed model in large-scale applications, this invention applies LDF-UNet to the Sentinel-1 InSAR coherence dataset for multiple Arctic regions. Figure 7As shown, large-scale mapping of LDFs was carried out in regions including the Nares Strait, the Canadian Arctic Archipelago, the Beaufort Sea, the East Siberian Sea, the Laptev Sea, and the Kara Sea. These regions encompass diverse Arctic environmental characteristics, including varied coastline morphologies, sea ice conditions, and dynamic processes, thus providing a comprehensive test scenario for evaluating the robustness and generalization ability of the proposed method.

[0114] The results show that LDF-UNet can effectively extract deformation structures across a wide spatial scale. In all studied regions, the detected LDFs correspond well to the elongated, low-coherence structures presented in the InSAR coherent images. Compared to the original coherent maps, the extracted LDF distribution more clearly characterizes the spatial pattern of deformation features within fixed ice on the continental margin, while effectively suppressing background noise and complex sea ice textures. This indicates that the proposed model can efficiently map fixed ice deformation structures over a large area of ​​the Arctic.

[0115] To further demonstrate the spatial characteristics of the test results, Figure 7 Several representative sub-regions that were selected were enlarged and, Figure 8 The results, when magnified, clearly show that the extracted LDFs exhibit a distinct elongated banded morphology, forming linear or quasi-linear structures within fixed ice regions. These deformation features are widely distributed across various Arctic regions (including the Nares Strait, Beaufort Sea, Laptev Sea, and Kara Sea), extending over considerable distances. Some LDFs can reach lengths of tens of kilometers, with the longest deformation structure observed in the Laptev Sea measuring approximately 140 km.

[0116] Furthermore, the detected LDFs exhibited significant directional consistency. In many regions, numerous deformation features appeared as parallel or nearly parallel linear structures, forming clustered deformation bands. This spatial distribution pattern suggests that the formation of LDFs may be influenced by regional-scale dynamic processes, such as dominant wind fields, ocean currents, or large-scale sea ice movements acting continuously in similar directions. However, a thorough analysis of the physical mechanisms underlying the spatial distribution and directional characteristics of these LDFs is beyond the scope of this invention, and related issues will be further explored in future research.

[0117] Overall, large-scale mapping results demonstrate that the proposed LDF-UNet model can stably detect deformation features of fixed ice along the Arctic margin in different regions and at different observation periods. This stable detection performance indicates that the method possesses good robustness and transferability in large-scale monitoring of Arctic fixed ice deformation structures.

[0118] Reliability analysis of LDF test results:

[0119] To further evaluate the reliability of the detected LDFs, this invention analyzes their spatial relationship with the fracture boundary formed during the 2017 Nares Strait ice arch rupture. First, the evolution of this rupture event is identified by analyzing Sentinel-1 SAR amplitude images.

[0120] like Figure 9 As shown, on May 6, 2017, the continental margin fixed ice remained relatively stable. By May 10, obvious fracturing signals began to appear within the ice arch region; by May 12, the cracks had significantly expanded, indicating that the ice arch structure was undergoing continuous fracturing and gradual disintegration. Based on the amplitude images, two main fault boundaries were further interpreted to represent the main fracturing areas of this fracturing event.

[0121] To analyze the relationship between LDFs and these fracture boundaries, the temporal evolution of LDFs and the distribution of fracture boundaries were compared and analyzed (see...). Figure 10 In the earlier observation periods (March 31-April 6 and April 6-April 12), only a few deformation features were detected near the fault boundary, and their distribution was relatively scattered. As the rupture event approached, the number and spatial extent of LDFs gradually increased. Especially during the periods of April 24-April 30 and April 30-May 6, a large number of linear deformation zones began to appear and gradually concentrated along the fault boundary.

[0122] The detected LDFs showed a clear spatial correspondence with the fault boundaries. Multiple LDFs were observed near or intersecting the fault boundaries before final rupture, both of which occurred before fracture. In later observation periods, some deformation zones were oriented substantially parallel to the fault boundaries. Therefore, these deformation structures may represent potential precursors to the formation of large fractures in continental margin fixed ice systems.

[0123] It is noteworthy that the large-scale deformation zone appeared approximately 10-15 days earlier than the final ice arch rupture event, which occurred between May 10th and May 12th, 2017. This temporal relationship suggests that the detected LDFs may provide early indications of instability in continental margin fixed ice structures. Furthermore, the significant spatial consistency between the LDF distribution and the ultimately formed fault boundary further demonstrates that these detected LDFs reflect physically significant deformation processes, rather than random, coherent variations.

[0124] The results demonstrate that LDFs extracted from InSAR coherent images can effectively characterize the evolution of deformation zones prior to the rupture of the Nares Strait ice arch. The ability to identify these deformation zones days to weeks before the formation of large cracks indicates that the proposed method has significant potential for monitoring and early warning of instability in Arctic continental margin fixed ice.

[0125] It should be noted that the appearance of LDFs does not necessarily coincide perfectly with the final crack location. The fracturing of land margin fixed ice is a complex process, influenced not only by external dynamics (such as ocean currents and wind fields) but also by the internal properties of sea ice (such as ice thickness and internal stress state). Future research could integrate deformation characteristics obtained from D-InSAR inversion with more relevant data, such as sea ice thickness, ocean dynamics, and atmospheric conditions, to improve the ability to identify potential crack regions and predict fracturing events. Furthermore, future research could extend the proposed methodological framework to long-term SAR data archives to construct a time-series distribution map of LDF occurrences in the Arctic region. This will help reveal large-scale sea ice deformation patterns and further deepen our understanding of the stability of land margin fixed ice under the background of climate change.

[0126] In this invention, a framework for automatic detection of linear deformation features of fixed ice on the continental margin is constructed by combining Sentinel-1 InSAR coherent images with an improved deep learning model. The main conclusions are as follows:

[0127] 1. LDF-UNet effectively extracts slender and sparse deformation structures. Addressing the geometric characteristics of LDF, such as its narrowness, elongation, and extremely sparse spatial distribution, this invention constructs the LDF-UNet model based on the U-Net architecture. By introducing a Dice-based loss function, the model's ability to identify narrow deformation bands is significantly improved. Experimental results show that this framework achieves an overall accuracy of 0.97 on the test data, with an F1-score improvement of up to approximately 14% compared to the baseline model.

[0128] 2. Regional-scale mapping results validated the model's good transferability. Using temporal coherent data obtained through Sentinel-1 D-InSAR processing, the proposed method was applied to multiple Arctic marginal sea regions, successfully generating 60 m resolution LDF spatial distribution maps. Stable detection performance was demonstrated under different sea ice environmental conditions, indicating that the method has good robustness and transferability in large-scale fixed ice deformation structure monitoring.

[0129] 3. Detected LDFs can serve as potential precursor signals for the fracturing of fixed ice along the continental margin. Time-series analysis in the Nares Strait region shows a strong spatial correlation between the detected LDFs and the fault boundaries in the 2017 ice arch fracturing event. Notably, these LDFs appeared several days before the final collapse of the ice arch, indicating that InSAR-based LDF detection can serve as an early indicator of the instability of fixed ice structures and provide new observational evidence for the prediction of fixed ice fracturing events along the continental margin.

[0130] Example 2

[0131] In this embodiment, a computer terminal device is provided, including:

[0132] One or more processors;

[0133] A memory, coupled to the processor, for storing one or more programs;

[0134] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the above-described method for detecting linear deformation of fixed ice based on spaceborne InSAR and deep learning.

[0135] In this embodiment, a computer-readable storage medium is also provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the above-described method for detecting linear deformation of fixed ice based on spaceborne InSAR and deep learning.

[0136] In this embodiment, an electronic device is also provided, including a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to perform the steps of the above-described method for detecting linear deformation of fixed ice based on spaceborne InSAR and deep learning.

[0137] In this embodiment, a computer program product is also provided, including a computer program that, when executed by a processor, implements the steps of the above-described method for detecting linear deformation of fixed ice based on spaceborne InSAR and deep learning.

[0138] The aforementioned program can run on a processor or be stored in memory (or a computer-readable medium). Computer-readable media includes both permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random-access memory (SRAM), dynamic random-access memory (DRAM), other types of random-access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0139] These computer programs may also be loaded onto a computer or other programmable data processing equipment, causing a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes can be implemented by different modules for different steps.

[0140] This embodiment provides such a device or system. The system, referred to as a spaceborne InSAR and deep learning-based fixed ice linear deformation detection system, includes:

[0141] The data acquisition module is used to acquire time-series synthetic aperture radar image data of the study area;

[0142] The differential interferometry processing module is used to perform differential interferometry processing on the time-series synthetic aperture radar image data to generate an interferometric coherence map;

[0143] The dataset construction module is used to manually interpret the linear low-coherence structures in the interferometric coherence map and construct a pixel-level labeled dataset.

[0144] The model training module is used to construct a deep learning semantic segmentation model based on the U-Net network, and to input the pixel-level labeled dataset into the deep learning semantic segmentation model for training to obtain a linear deformation feature detection model.

[0145] The feature detection module is used to input the interferometric coherence map to be detected into the linear deformation feature detection model to detect the linear deformation features of fixed ice.

[0146] As one implementation method in this embodiment, the differential interferometry processing module includes:

[0147] The image pair selection unit is used to select image pairs with a time baseline of 6 days or 12 days as the master image and slave image;

[0148] A registration unit is used to register the main image and the slave image;

[0149] The interferogram generation unit is used to multiply the complex signals of the registered master image and slave image to generate an interferogram;

[0150] A coherence coefficient estimation unit is used to estimate the interference coherence coefficient in the interferogram using a sliding window to obtain an interference coherence map;

[0151] A geocoding unit is used to perform geocoding processing on the interferometric coherence map.

[0152] As one implementation method in this embodiment, the dataset construction module includes:

[0153] The manual interpretation unit is used to select images with linear low-coherence structures in the interferometric coherence map for manual interpretation;

[0154] The annotation mask generation unit is used to mark pixels belonging to linear deformation structures as deformation categories and other pixels as background categories to generate an annotation mask;

[0155] An image segmentation unit is used to divide the interferometric coherence pattern into image blocks of a set size;

[0156] The sample set construction unit is used to construct training and validation sets based on the image blocks and their corresponding labeled masks.

[0157] As one implementation method in this embodiment, the deep learning semantic segmentation model constructed in the model training module includes:

[0158] The encoder is used to progressively extract high-level semantic features;

[0159] Decoder, used to gradually restore spatial resolution;

[0160] A skip connection unit is used to connect the encoder and the decoder, and to perform channel stitching of the feature map output by the encoder and the feature map upsampled by the decoder to achieve multi-scale feature fusion.

[0161] As one implementation method in this embodiment, the model training module includes:

[0162] The loss calculation unit is used to calculate the loss value between the model prediction result and the labeled mask using a hybrid loss function that combines the cross-entropy loss function and the Dice loss function;

[0163] The parameter update unit is used to iteratively update the model parameters using the backpropagation algorithm.

[0164] The performance evaluation unit is used to evaluate the model performance using the validation dataset after each training cycle in order to save the optimal model.

[0165] As one implementation method in this embodiment, it also includes:

[0166] The results analysis module is used to perform spatial superposition and temporal comparison analysis of the fixed ice linear deformation feature distribution detected by the feature detection module and the fixed ice fracture boundary identified by synthetic aperture radar amplitude image.

[0167] The system or apparatus is used to implement the functions of the methods in the above embodiments. Each module in the system or apparatus corresponds to each step in the method, as has been described in the method and will not be repeated here.

[0168] The above implementation method solves the problem of fixed ice linear deformation detection based on spaceborne InSAR and deep learning in related technologies, thereby ensuring that the problems existing in the prior art are resolved.

[0169] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for detecting linear deformation of fixed ice based on spaceborne InSAR and deep learning, characterized in that, Includes the following steps: Temporal synthetic aperture radar image data of the study area is acquired, and differential interferometry is performed on the temporal synthetic aperture radar image data to generate an interferometric coherence map. The deformation region is manifested as a slender low coherence strip structure in the interferometric coherence map. The slender low coherence strip structure corresponds to the deformation zone inside the fixed ice. The linear deformation feature is a narrow structure inside the fixed ice related to the crack development process. The linear low-coherence structure in the interferometric coherence map is manually interpreted and a pixel-level labeled dataset is constructed. Interferometric coherence maps with obvious linear low-coherence structures are selected for labeling. Pixels belonging to linear deformation features are labeled as deformation categories, and the remaining pixels are labeled as background categories to generate a labeling mask. The linear deformation features are characterized by subtle, fragmented and low-contrast features that are embedded in the complex sea ice background environment, occupying only a small proportion in the image and presenting a slender linear geometric shape. A deep learning semantic segmentation model based on the U-Net network is constructed. The pixel-level labeled dataset is input into the deep learning semantic segmentation model for training to obtain a linear deformation feature detection model. During training, a hybrid loss function combining the weighted cross-entropy loss function and the Dice loss function is used to calculate the loss value between the model prediction result and the labeled mask. The model parameters are iteratively updated through the backpropagation algorithm. After each training cycle, the model performance is evaluated using a validation set to save the optimal model. There is a significant class imbalance between deformed pixels and background pixels, and higher weights are assigned to the deformation class. The interferometry image to be detected is input into the linear deformation feature detection model to detect the linear deformation features of fixed ice. The linear deformation features of fixed ice serve as an early indication signal of the instability of fixed ice.

2. The method according to claim 1, characterized in that, The process of generating the interferometric coherence map includes: selecting an image pair with a time baseline of 6 days or 12 days as the master image and the slave image; registering the master image and the slave image; multiplying the complex signals of the registered master image and the slave image to generate an interferometric map; estimating the interferometric coherence coefficient using a sliding window in the interferometric map; and then geocoding the obtained interferometric coherence map.

3. The method according to claim 1, characterized in that, The process of constructing the pixel-level labeled dataset includes: dividing the interferometric coherence map into image blocks of a set size, and constructing a training set and a validation set based on the image blocks and their corresponding labeled masks.

4. The method according to claim 1, characterized in that, The structure of the deep learning semantic segmentation model includes: an encoder for progressively extracting high-level semantic features, a decoder for progressively restoring spatial resolution, and a skip connection connecting the encoder and the decoder. The skip connection is used to perform channel concatenation of the feature map output by the encoder and the feature map upsampled by the decoder to achieve multi-scale feature fusion.

5. The method according to claim 1, characterized in that, The method further includes: spatially superimposing and temporally comparing the detected linear deformation feature distribution of fixed ice with the fixed ice fracture boundary obtained by synthetic aperture radar amplitude image recognition.

6. A fixed ice linear deformation detection system based on spaceborne InSAR and deep learning, characterized in that, The system for implementing the method of any one of claims 1-5 comprises: The data acquisition module is used to acquire time-series synthetic aperture radar image data of the study area; The differential interferometry processing module is used to perform differential interferometry processing on the time-series synthetic aperture radar image data to generate an interferometric coherence map; The dataset construction module is used to manually interpret the linear low-coherence structures in the interferometric coherence map and construct a pixel-level labeled dataset. The model training module is used to construct a deep learning semantic segmentation model based on the U-Net network, and input the pixel-level labeled dataset into the deep learning semantic segmentation model for training to obtain a linear deformation feature detection model. The feature detection module is used to input the interferometric coherence map to be detected into the linear deformation feature detection model to detect the linear deformation features of the fixed ice.

7. A computer terminal device, characterized in that, include: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the steps of the method as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-5.

Images