Landslide extraction method based on SAR image change detection

Abstract

The invention relates to the technical field of earthquakes, in particular to a landslide extraction method based on SAR image change detection. According to the method, a difference chart is constructed based on changes of polarimetric SAR intensity information of different time phases, an adaptive Gaussian threshold algorithm is adopted to calculate an optimal threshold, a majority voting method is adopted to extract a preliminary landslide candidate area, a gradient value calculated by a digital elevation model (DEM) is fused as a topographic constraint, pseudo-change signals of a plain area are effectively filtered out, and the landslide detection accuracy is improved. And finally extracting the landslide range. Experimental results show that the research method shows significant advantages in landslide detection under complex terrain conditions, the model can successfully recognize most landslide events, the overall precision reaches 90.7%, good robustness is shown, and reliable technical support is provided for post-earthquake disaster assessment and emergency response.

Description

Technical Field

[0001] The present invention relates to the field of earthquake technology, and in particular to a landslide extraction method based on SAR image change detection. Background Art

[0002] Synthetic Aperture Radar (SAR), an advanced remote sensing technology, has been widely used in a variety of fields, including environmental monitoring, resource surveys, and disaster assessment, achieving remarkable results. SAR image change detection leverages the high-resolution echo signal's sensitivity to ground feature changes. By analyzing the multi-temporal intensity characteristics before and after earthquake landslides, it can more quickly obtain comprehensive and accurate assessment responses. However, the high noise content and insufficient extraction of effective information in SAR images impair the accuracy of image change detection. Summary of the Invention

[0003] To solve the above technical problems, the present invention provides a landslide extraction method based on SAR image change detection, which includes:

[0004] Acquire multi-temporal SAR images of the earthquake area before and after the landslide and perform preprocessing;

[0005] For the VV and VH polarization images of multi-temporal SAR images before and after the landslide, the difference maps of the earthquake area were constructed using the ratio method, logarithmic ratio method and mean ratio method respectively.

[0006] For each difference map, an adaptive Gaussian threshold segmentation method is applied to solve the optimal segmentation threshold, and the corresponding difference map is binarized using the optimal segmentation threshold to divide the earthquake area into a change area and a non-change area;

[0007] For a given pixel, the majority voting method is used to fuse the segmentation results of each difference map to make a decision. If the pixel is classified as a change area in at least half of the difference maps, it is marked as a candidate landslide area.

[0008] The slope is calculated based on the pre-acquired DEM data of the earthquake area to obtain the slope distribution map of the earthquake area for terrain constraint, and the candidate landslide areas with slopes less than a preset threshold are eliminated to finally obtain the landslide area.

[0009] Preferably, the preprocessing includes orbit correction, radiation calibration, filtering, terrain correction, decibel processing and image registration.

[0010] Preferably, the difference map is constructed based on a multi-temporal SAR intensity information image obtained by mathematically defining the multi-temporal SAR image. The mathematical definition is expressed as follows:

[0011] I before ={I before(i,j),1< <i<<M,1<<j<<N},

[0012] I after ={I after (i,j),1<<i<<M,1<<j< <N},

[0013] Among them, I before and I after is the mean value of the intensity information of multi-temporal SAR images before and after the landslide, M and N are I before and I after The number of pixels in the length and width directions, i represents the i-th pixel in the length direction, and j represents the j-th pixel in the width direction.

[0014] As an example, the calculation formulas for constructing the difference map using the ratio method, logarithmic ratio method, and mean ratio method are:

[0015]

[0016]

[0017] Where R represents the ratio, LR represents the log ratio, and MR represents the mean ratio.

[0018] Preferably, the formula for binarizing the corresponding difference map using the optimal segmentation threshold is:

[0019]

[0020] Among them, B k (i, j) is the binarization result of pixel (i, j) in the corresponding difference map (1 represents the changed area, 0 represents the non-changed area), T k is the optimal segmentation threshold of the corresponding difference map determined by the adaptive Gaussian threshold method.

[0021] As a preference, the majority voting method is used to fuse the segmentation results of each difference map to make a decision. The formula is:

[0022]

[0023] Among them: B MV (i, j) represents the final voting result of pixel (i, j), N is the total number of difference maps, and if at least half of the difference maps at pixel (i, j) are classified as a change area (i.e., 1), the final voting result is 1, otherwise it is 0; if the final voting result is 1, the pixel (i, j) is marked as a candidate landslide area.

[0024] Preferably, the calculation formula for the slope tangent value tanα of a point on the terrain surface in the slope distribution map of the earthquake area is:

[0025]

[0026] Where Δh is the vertical height change of the point (Δh), and Δd is the horizontal distance change of the point.

[0027] Preferably, the preset threshold for eliminating slopes less than a preset threshold is 15°.

[0028] In a second aspect, the present invention provides a landslide extraction system based on SAR image change detection, which comprises:

[0029] An acquisition module is used to acquire multi-phase SAR images of the earthquake area before and after the landslide and perform preprocessing;

[0030] A construction module is used to construct a difference map of the earthquake area using the ratio method, logarithmic ratio method and mean ratio method for the VV and VH polarization images of the multi-phase SAR images before and after the landslide;

[0031] a segmentation module for applying an adaptive Gaussian threshold segmentation method to each difference map to obtain an optimal segmentation threshold, and using the optimal segmentation threshold to perform binarization processing on the corresponding difference map to divide the earthquake area into a change area and a non-change area;

[0032] The decision module is used to make a decision by fusing the segmentation results of each difference map using the majority voting method for a given pixel. If the pixel is classified as a change area in at least half of the difference maps, it is marked as a candidate landslide area.

[0033] The extraction module is used to calculate the slope based on the pre-acquired DEM data of the earthquake area to obtain the slope distribution map of the earthquake area, and eliminate the candidate landslide areas with slopes less than a preset threshold to finally extract the landslide areas.

[0034] Compared to the existing technology, the present invention has the following beneficial effects: the present invention constructs a difference map based on the intensity information changes of the VV and VH polarization images of multi-phase SAR images, uses an adaptive Gaussian threshold algorithm to calculate the optimal segmentation threshold for each difference map, and then marks the candidate landslide areas by majority voting. At the same time, the digital elevation model (DEM) is integrated to calculate the slope. The slope values ​​of each point in the slope distribution map are used as terrain constraints to effectively filter out false change signals in the plain area and finally extract the landslide area. The method of the present invention has been verified by experimental results and shows significant advantages in detecting landslides under complex terrain conditions. It can successfully identify most landslide events with an overall accuracy of 90.7%, showing good robustness and providing reliable technical support for post-earthquake disaster assessment and emergency response. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flow chart of the landslide extraction method based on SAR image change detection of the present invention;

[0036] Figure 2 The SAR images before and after the landslide preprocessed in this invention

[0037] Figure 3 The landslide extraction result of the ratio method in this invention

[0038] Figure 4 The landslide extraction result of the logarithmic ratio method in this invention

[0039] Figure 5 The landslide extraction result of the mean ratio method in this invention is

[0040] Figure 6 DEM and slope distribution of the experimental area in this invention

[0041] Figure 7 This is the landslide result extracted by the majority voting method in this invention. DETAILED DESCRIPTION

[0042] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as limiting the present invention. In addition, the terms "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, features defined as "first", "second", etc. may explicitly or implicitly include one or more of the features. In the description of the present invention, unless otherwise specified, "multiple" means two or more. In the description of the present invention, it should be noted that, unless otherwise expressly specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense. For example, they may refer to fixed connections, detachable connections, or integral connections; mechanical connections or electrical connections; direct connections or indirect connections through an intermediate medium; and internal communication between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0043] According to the prior art known to the inventors of the present invention, SAR, as a coherent imaging system, often produces speckle noise during image formation due to the coherent superposition of waves emitted by numerous randomly distributed scatterers. This noise interferes with the expression of the scattering characteristics of ground objects, significantly affecting the interpretation and recognition of SAR images. Prior to landslide extraction, this study performed a series of preprocessing procedures on the images, including trajectory correction, radiation calibration, filtering, terrain correction, decibel processing, and image registration, to improve image quality and ensure the accuracy of subsequent analysis. SAR is an active microwave imaging technology whose images can reflect the physical and geometric properties of the surface. Grayscale, as a basic characteristic, directly corresponds to the backscatter coefficient of the ground object, and the backscatter intensity is closely related to the dielectric constant, surface roughness, and geometric structure of the ground object. SAR can obtain structural information on different imaging surfaces by controlling the polarization of the signals in the transmission and reception paths. Compared to single-polarization SAR, dual-polarization data can provide richer ground feature features. Changes in SAR images primarily stem from changes in the amplitude and phase of targets in the study area, which in turn cause changes in the electromagnetic scattering coefficient and, consequently, in the grayscale of the SAR image. Random factors such as atmospheric conditions, sensor accuracy, angle of incidence, and surface humidity can significantly interfere with SAR image change detection, masking the true target change signal. Therefore, the core challenge of SAR image change detection lies in effectively distinguishing true changes caused by the target itself from spurious changes caused by these random factors.

[0044] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0045] Example 1

[0046] 1.1 Landslide information extraction process

[0047] This embodiment proposes a landslide extraction method based on SAR image change detection. The landslide feature information extraction process is as follows: Figure 1 As shown in the figure. First, Sentinel-1SAR images of the Luding earthquake experimental area before and after the landslide were preprocessed, including orbit correction, radiometric calibration, filtering, terrain correction, decibel processing, and image registration to ensure the spatial geometric consistency and numerical comparability of the multi-temporal data. Based on the VV and HV polarization backscatter intensity variation characteristics of the images, the ratio method, mean ratio method, and logarithmic ratio method were selected. Six difference maps were constructed using these two polarization data. The adaptive Gaussian threshold algorithm was used to solve the optimal segmentation threshold. The six difference maps were then binarized to preliminarily extract candidate landslide areas.

[0048] In order to reduce the errors that may be generated by a single method and make full use of the complementary advantages of VV and HV polarization data, this embodiment adopts the majority voting method to fuse the segmentation results of the three difference maps. If a pixel is judged to be a landslide in at least half of the difference maps, it is marked as a candidate landslide area. The slope distribution is further calculated in combination with DEM data, and candidate areas with slope interference are eliminated to solve the confusion problem between landforms such as fields and buildings and landslide features, and finally the landslide area and non-landslide area are divided. Finally, based on the optical images of the same period as the verification benchmark, the overall accuracy (OA) of the extraction results is quantitatively evaluated by the confusion matrix. The verification results show that this method significantly improves the accuracy and practicality of landslide identification, and provides reliable technical support for the rapid detection of landslides in strong earthquake areas.

[0049] 1.2 SAR Image Change Detection Method

[0050] Specifically, the common idea of ​​constructing a difference map based on pixel information is to use difference operations or ratio operations to calculate the dynamic difference of corresponding pixel points. The core of change detection is to maximize the difference between the information characteristics of the changed area and the information characteristics of the non-changed area. The multi-phase SAR intensity information images before and after the landslide event are respectively extracted using ratio operations, mean ratio operations and logarithmic ratio operations. Difference map generation refers to the use of specific algorithms or technical means to generate an image presented in grayscale form, which preliminarily reflects the changes in remote sensing images, where pixels with large grayscale values ​​correspond to areas where significant changes have occurred, and pixels with low grayscale values ​​correspond to areas where no significant changes have occurred. In the study of the Luding earthquake landslide experimental area, when generating difference maps and performing change detection on the acquired two-phase SAR images, a large amount of multiplicative speckle noise was introduced into the image due to the SAR imaging mechanism, which seriously interfered with subsequent image processing and change extraction. In order to quantify surface change information and landslide range, the two-phase SAR images need to be mathematically defined, and the expression is as follows:

[0051] I before ={I before (i,j),1<<i<<M,1<<j<<N},

[0052] I after ={I after (i,j),1< <i<<M,1<<j<<N},

[0053] Among them, I before and I after is the mean value of the intensity information of the multi-temporal SAR images before and after the landslide after mean polarization, M and N are I before and I after The number of pixels in the length and width directions, i represents the i-th pixel in the length direction, and j represents the j-th pixel in the width direction.

[0054] 1.2.1 Ratio method

[0055] In remote sensing image change detection, the image ratio method calculates the grayscale value ratio of corresponding pixels between images of different time phases to generate a difference image. Then, by selecting an appropriate threshold, regions of change are identified. If no change has occurred, the grayscale value ratio of the corresponding pixels in that region should be close to 1. However, for regions of change, the grayscale value ratio of the corresponding pixels deviates significantly from 1. While this method mitigates the effects of calibration errors to some extent, it can easily miss subtle changes and is sensitive to multiplicative speckle noise. Its advantages include a simple principle, intuitive and easy-to-use calculations, low computational complexity, and suitability for processing large-scale data. However, its disadvantages are that it struggles to account for all factors (e.g., susceptibility to interference from noise, shadows, overlap, and perspective contraction). Image registration errors can directly affect the ratio results, leading to false or missed detections. The threshold setting significantly influences the results and usually requires empirical or experimental determination. This method lacks universal applicability and has limited accuracy in complex scenarios, requiring integration with other methods for improved performance.

[0056] The calculation formula is:

[0057]

[0058] 1.2.2 Logarithmic Ratio Method

[0059] The logarithmic ratio method introduces logarithmic transformation on the basis of the traditional image ratio method. It can transform the multiplicative coherent speckle noise model into an additive noise model, compress the dynamic range of the image, enhance the contrast between the changing and non-changing classes, make weak changes more obvious, and effectively reduce the impact of speckle noise. At the same time, it performs nonlinear compression on the image's change range, making the background (non-changing) area smoother and the distribution of pixel values ​​more uniform, reducing the impact of extreme values ​​on the results, and making the change detection results more stable. This method can effectively suppress noise, but the pixel values ​​in the edge area are easily blurred. Compared with the simple ratio method, the amount of calculation is slightly increased. Although the logarithmic transformation can alleviate multiplicative noise, additive noise may still affect the detection accuracy. It is suitable for scenes with a large dynamic range and obvious noise. The calculation formula is as follows:

[0060]

[0061] 1.2.3 Mean Ratio Method

[0062] The mean ratio method utilizes the neighborhood information of pixels. By calculating the mean of the local area, it makes the foreground (change) area of ​​the difference map closer to the actual situation, can smooth the noise, reduce the impact of noise on the change detection results, and has stronger robustness. It has a good suppression effect on single noise points, because the comparison is no longer a single pixel point, but the mean of the area where the pixel point is located. Due to the introduction of neighborhood information, the boundary between the foreground area and the background area becomes blurred. The mean ratio method will highlight the changes in the local area, can detect subtle local changes in the image, and improve the missed detection rate or false alarm rate. The size of the local area (window size) needs to be adjusted according to the specific scene. Improper selection will affect the detection effect. The calculation formula is:

[0063]

[0064] 1.3 Majority Voting

[0065] Majority voting is an effective method for optimizing decision-making processes. It typically aggregates independent decision results from multiple modalities and makes the final decision based on the most frequently occurring category or outcome. This method is simple and easy to implement, eliminating the need to assume independence between information sources. It integrates information from multiple sources through the principle of majority rule, thereby improving the robustness and generalization of decisions. In radar remote sensing image processing, information from different modalities can be represented by pixel values. This representation not only extracts key data features but also effectively reduces redundant information, thereby improving model performance and generalization capabilities.

[0066] In order to obtain the binary landslide candidate area, we k (i,j)(k∈{R(i,j),LR(i,j(,MR(i,j)})) application adaptive

[0067]

[0068] Among them, B k (i, j) is the binarization result of pixel (i, j) in the corresponding difference map (1 represents the changed area, 0 represents the non-changed area), T k is the optimal segmentation threshold of the corresponding difference map determined by the adaptive Gaussian threshold method.

[0069] For a given pixel (i, j), the decision formula of the majority voting method is as follows:

[0070]

[0071] Among them: B MV(i, j) represents the final voting result of pixel (i, j), N is the total number of difference maps, and if at least half of the difference maps at pixel (i, j) are classified as a change area (i.e., 1), the final voting result is 1, otherwise it is 0; if the final voting result is 1, the pixel (i, j) is marked as a candidate landslide area.

[0072] Example 2

[0073] This embodiment provides a practical application example of the landslide extraction method based on SAR image change detection according to Embodiment 1.

[0074] 2.1 Experimental Data

[0075] The SAR (Synthetic Aperture Radar) imagery used in this example is Sentinel-1A data, as shown in Table 1. The data is sourced directly from the European Space Agency (ESA) Copernicus open platform (https: / / scihub.copernicus.eu). The Sentinel-1 satellite, successfully launched in 2014, is the first satellite equipped with a C-band synthetic aperture radar. A dual-satellite constellation consisting of Sentinel-1A and Sentinel-1B is currently in orbit. Single-look complex (SLC) images were acquired in interferometric wide-swath (IW) mode, with a swath width of 250 kilometers and a spatial resolution of 5 m × 20 m.

[0076] Table 1 Sentinel-1 Interferometric Wide Width (IW) Mode Product Information

[0077]

[0078] Sentinel-1 imagery was acquired on September 2, 2022 (before the earthquake) and September 7, 2022 (after the earthquake), providing two views of dual-polarization data, including VV (vertical transmit and vertical receive) and VH (vertical transmit and horizontal receive). Radar remote sensing technology uses radar echo signals to observe the Earth around the clock, acquiring reliable and stable high-spatial-resolution SAR imagery. This allows for the interpretation and identification of ground features, providing efficient data services for priority applications such as land cover classification, sea ice applications, and emergency management.

[0079] In the Luding earthquake landslide extraction experiment, the areas near the major rivers such as the Dadu River and Moshi River, where the earthquake triggered many and serious collapse and landslide geological disasters, were selected as the research areas.

[0080] To verify the accuracy of the majority voting fusion decision-making method for identifying and extracting landslide results, this study constructed a high-confidence landslide validation sample set by analyzing Sentinel-2 post-earthquake images and visually interpreting them. Using interactive human-computer interpretation, combined with pre-landslide historical images and cross-validation with terrain features, 4029 ground-truth samples with clear landslide boundaries were finally annotated. The validation sample contained a total of 310,439 valid pixels, corresponding to a total landslide area of ​​27.07 km. 2 The largest single landslide area reached 0.36 km 2 .

[0081] 2.2 Experimental Results

[0082] The experiment of this embodiment sequentially studies the complete process of image preprocessing, difference map construction, adaptive threshold segmentation, fusion voting, and DEM-assisted filtering methods. First, the Sentinel-1 image is preprocessed to improve the image quality, correct image geometric distortion, reduce speckle noise, enhance image details, and reduce data redundancy, so as to obtain real surface information and accurately reflect surface features and changes. The preprocessed image is as follows: Figure 2 As shown in Figure 1, the image preprocessing process includes orbit correction, radiation calibration, filtering, terrain correction, decibel processing and image registration. Figure 2 (a) and (b) are the pre-processed VV and VH polarization images of Sentinel-1 before the landslide (September 2, 2022), respectively. They are used as the reference data. To ensure the temporal and spatial consistency of the images, the post-landslide image was selected on September 14, 2022. Figure 2 (c) and (d) are preprocessed Sentinel-1 VV and VH polarimetric images. In SAR imaging, dual-polarization technology can simultaneously acquire data in both horizontal and vertical polarization modes. The echo strengths of the same feature vary under different polarization modes, providing rich information for landslide identification. Because VH polarization signals are highly sensitive to vegetation structure, canopy changes, and surface roughness, they can more accurately capture changes in scattering characteristics caused by exposed rock, soil damage, or disturbed vegetation during landslides. VV polarization, on the other hand, primarily reflects surface moisture and has a weaker direct response to these geological disturbances.

[0083] Early difference map generation algorithms primarily relied on simple pixel-level interpolation methods, directly subtracting two SAR images pixel by pixel. However, the inherent multiplicative noise characteristics of SAR images differ fundamentally from those of optical images, making traditional interpolation methods ineffective when processing SAR images. Consequently, the results struggled to accurately represent actual surface changes. By comparing SAR images before and after a landslide, this study employed the following three methods to generate dual-polarization difference maps.

[0084] The ratio method generates a new image by dividing the grayscale values ​​of corresponding pixels in two images (such as formula 1). It can eliminate multiplicative errors such as terrain undulations and shadows, reduce the interference of noise on change detection, and highlight the differences in ground objects. According to the brightness distribution of different regions of the image, for each pixel in the difference map of the contrast value method, a 5×5 area is taken and the local threshold is calculated according to the Gaussian kernel weight. If the grayscale value of a pixel in the difference map is greater than the grayscale value of the area, it is determined to be a changed area; otherwise, it is a non-changed area. The threshold segmentation result is as follows: Figure 3 As shown in the figure, the backscatter coefficient of the landslide target is strong and the color is bright. The backscatter coefficient of the water area is small and all are dark. Figure 3 (a) VV polarization backscattering is better than Figure 3 (b) The VH polarization is stronger and the color is slightly brighter. Although the ratio method can extract the change region with high spatial consistency, the result image still contains discrete noise with the distribution characteristics of salt and pepper noise.

[0085] The logarithmic ratio method combines ratio operation and logarithmic transformation. According to Formula 2, the pixel values ​​of the two phase images are ratioed and then logarithmically transformed to convert multiplicative noise into additive noise. The nonlinear shrinkage characteristics of the logarithmic operation stretch the dark details of the image and compress the dynamic range of the highlight area, which significantly enhances the contrast between the changed and unchanged areas. The difference map is segmented by a local (5×5) adaptive threshold using the local statistical characteristics and Gaussian weighted smoothing threshold. The weight is determined by the Gaussian equation. The difference value is binarized according to the degree of deviation from the local threshold. Gaussian weighted smoothing preserves the spatial correlation of the neighborhood and suppresses isolated false alarms. The segmentation results are shown in the figure below. Figure 4 As shown in the figure, the landslide area has a significantly larger logarithmic ratio due to surface damage and exposed rock mass, showing a brighter tone; while the non-landslide area has lower backscattering and a lower logarithmic ratio. Figure 4 (a) VV polarization is sensitive to soil moisture and surface roughness, while Figure 4 (b) VH polarization is more sensitive to changes in vegetation scattering.

[0086] The mean ratio method is a method that generates a new image by calculating the neighborhood mean participation ratio of corresponding pixels in two images. In this paper, a local area window of 5×5 is selected according to the SAR image resolution and the characteristics of the target object, which can achieve a good balance between noise suppression and detail retention. Although a larger neighborhood (such as 7×7 or 9×9) can better suppress coherent speckle noise, it may cause detail loss and a large amount of calculation; a smaller neighborhood (such as 3×3) helps to retain image details, but has weaker noise resistance. According to the brightness distribution of different areas of the image, an adaptive Gaussian threshold is used for the mean ratio difference map, and the Gaussian weighted average of the landslide area in the 5×5 local neighborhood is calculated to determine the image threshold. The threshold is dynamically adjusted according to the characteristics of the local area to improve the separation accuracy of the changed area and the unchanged area. The threshold segmentation results are shown in the figure below. Figure 5 As shown in the figure. Due to the abnormally strong backscattering, its ratio to the surrounding mean is significantly increased, and it appears as a bright color in the mean ratio diagram; the backscattering value in the non-landslide area is low, and the ratio to the local mean is low, so it appears as a dark color in the figure. Figure 5 (a) VV polarization, Figure 5 (b) VH polarization is more sensitive to changes in surface details.

[0087] The three difference maps highlight the differences in how different algorithms represent surface deformation information, but the detection results all contain sporadic false alarms outside the landslide area. Pixels with large grayscale values ​​in the difference maps correspond to areas where significant ground changes have occurred, while pixels with low grayscale values ​​correspond to areas where no significant ground changes have occurred. Based on the brightness distribution of different image regions, an adaptive Gaussian threshold is used to detect local feature changes and improve threshold accuracy. The advantage of this method is that different optimal thresholds can be adaptively calculated for different regions in the same image, allowing for better results under varying brightness conditions. This effectively addresses the threshold selection challenge and yields the landslide information extraction results using three different threshold segmentation methods.

[0088] Figure 6 (a) is from the ALOS-12.5m resolution digital elevation model (DEM), which represents the elevation information of the earth's surface in digital form. Figure 6 (b) is the slope distribution calculated directly using DEM data. Slope is a key parameter in terrain analysis and is used to measure the steepness of surface units. It is usually expressed as the angle between the slope surface and the horizontal plane. The slope tangent (such as Formula 5) is the ratio of the vertical height change (Δh) of a point on the terrain surface to its horizontal distance (Δd).

[0089]

[0090] Combining the three difference maps with the slope distribution map reveals that landslides are often located in the upper and middle portions of slopes, revealing a correlation between topographic parameters and the spatial distribution of landslides. Furthermore, slope stability is significantly enhanced when the slope is below 15°. By calculating slope values ​​based on DEM data to apply topographic constraints, areas with slopes less than 15° are eliminated, effectively filtering out false change signals in plain areas and optimizing the identification of landslide areas.

[0091] Then, based on the multi-feature fusion strategy, the ratio method with poor accuracy was discarded, and the majority voting method was used to perform decision-level fusion on the preliminary segmentation results of the four groups of difference maps of VV polarization logarithm ratio and mean ratio and VH polarization logarithm ratio and mean ratio. A slope greater than 15° was used for terrain constraint. The results are as follows: Figure 7As shown in the figure. In the analysis of landslide extraction results from SAR image majority voting fusion decision-making, dual-polarization data can provide more information to distinguish landslides from other surface changes. The area before a landslide typically consists of bare soil or vegetation with low scattering power, appearing as darker tones in SAR images. After a landslide disaster, the surface structure is destroyed, forming a landslide body with rich texture information. Its surface exhibits diffuse reflectivity, resulting in a significant increase in the echo power received by the sensor. This is reflected as a noticeable brightness change in the SAR image, with white areas representing identified landslide information.

[0092] from Figure 2-Figure 4 It can be seen that the landslide extraction results from the three difference maps are highly consistent with the actual landslide distribution interpreted visually, demonstrating the strong detection capabilities of the adopted change detection and threshold segmentation methods. Each of the three dual-polarization SAR difference map construction methods has its own unique advantages, but also has disadvantages such as low change detection accuracy and narrow applicability due to the use of a single difference map. By combining majority voting and slope terrain constraints, a more accurate landslide extraction result was ultimately obtained. Visually, the majority voting method outperformed the three difference map extraction methods.

[0093] During the experiment, the majority voting fusion strategy perfectly addressed this drawback. Using DEM for slope constraints effectively prevented misidentification of non-landslide areas, such as bare ground and buildings, further reducing the false alarm rate. This method achieved high recognition accuracy in landslide area demarcation, providing a strong technical foundation for subsequent landslide monitoring and disaster warning applications. Experimental results indicate that slopes along major rivers and their tributaries, such as the Dadu and Moxi Rivers in the Luding earthquake landslide experimental area, are the primary sites of earthquake-triggered collapse and landslides.

[0094] The landslide results finally extracted by the multi-temporal change detection method in this paper are as follows Figure 7 As shown, landslides are primarily concentrated in the upper and middle portions of the slopes, with high-level landslides showing a dense distribution. Multiple landslide zones converge to form a planar expansion pattern, dominated by small to medium-sized, shallow rock slides. The highest density of landslide points is observed in Moxi Town, the Hailuogou area, and along the Dadu River in Detuo Town.

[0095] This paper evaluates change detection results using both subjective performance and objective quantitative metrics. The change detection map is compared with the corresponding reference change image. Table 2 systematically evaluates the detection performance of different difference image construction methods using a confusion matrix and quantitative metrics for overall accuracy. A true positive occurs when the model correctly predicts an actual landslide area as a landslide; a false positive occurs when the model incorrectly predicts a non-landslide area as a landslide; a false negative occurs when the model incorrectly predicts a landslide area as a non-landslide; and a true negative occurs when the model correctly predicts a non-landslide area as a non-landslide. Overall accuracy (OA) refers to the proportion of correctly classified pixels. It reflects the proportion of correctly identified pixels across all evaluated samples in both landslide and non-landslide areas and is used to comprehensively evaluate the model's overall classification performance. Higher overall accuracy indicates better overall classification performance for the detection method.

[0096] Table 2 Quantitative analysis indicators of landslide extraction results using different methods

[0097]

[0098] 2.3 Results Analysis

[0099] When extracting landslide information from SAR images, this paper uses three difference map generation algorithms: Figure 2-Figure 4 It can be seen that VV and VH dual-polarization SAR have different characteristics when identifying landslides, indicating that different polarizations have different sensitivities to surface features. The results of the three different difference maps have their own characteristics. Figure 2 The difference image constructed using the ratio method is affected by the multiplicative speckle noise inherent in SAR images. It is not only difficult to highlight the changed areas, but also lacks sensitivity to changes in weak scattering areas, resulting in the failure to effectively highlight some landslide boundaries with low backscattering coefficients (such as shallow sliding areas); the logarithmic ratio method converts multiplicative noise into additive noise through nonlinear transformation, thereby reducing the impact of speckle noise to a certain extent, but its change detection map still contains some scattered noise. At the same time, the logarithmic ratio method can compress the dynamic range of intensity values ​​and enhance the identifiability of small change information, but its change sensitivity may still be insufficient for weak scattering areas.

[0100] The neighborhood-averaged ratio operator incorporates local information, giving the difference map a strong smoothing effect during analysis. This reduces the impact of overall image intensity fluctuations and improves the sensitivity of detecting changes in weakly scattering areas, thereby better capturing landslide boundary information. While maintaining high spatial detail, the spatial smoothing effect reduces the efficiency of speckle noise suppression, leading to an increase in isolated noise points in non-landslide areas and a loss of some local detail. By fusing the complementary features of the logarithmic ratio method and the mean ratio method through majority voting, white speckle noise is reduced while better preserving areas of minor changes. This achieves a balance between noise suppression and detail preservation at the decision-making level, effectively improving change detection performance using a simple fusion strategy.

[0101] Table 2 summarizes the evaluation criteria and accuracy of the methods used to study data from the Luding earthquake landslide experimental area. Comparing the experimental results of three change difference map methods using different polarization feature evaluation indicators reveals that the landslide experimental results extracted by each method have their own unique advantages, but they also suffer from disadvantages such as low change detection accuracy and a narrow scope of applicability due to the use of a single difference map. Fusion using the majority voting method not only effectively suppresses multiplicative speckle noise in SAR images but also significantly reduces the incidence of false alarms, achieving even better performance in extracting landslide information, with an OA of 90.7%.

[0102] In the ratio method, the false positive rate is high, indicating that the model over-identifies landslide areas and a large number of negative samples are misclassified as positive, resulting in serious false alarms. Although the overall accuracy is close to 80%, the model's generalization ability is poor and its practical application value is limited. The logarithmic ratio improves noise suppression through nonlinear logarithmic transformation, significantly reducing the false positive rate and slightly increasing the false negative rate. This is due to the over-smoothing effect of the logarithmic operation on weak change signals. The OA rate exceeds 85%. The improvement in accuracy may be due to better negative example classification but insufficient positive example recognition. The majority voting method effectively integrates the complementary advantages of each method through decision-level fusion. Its false positive rate is simultaneously reduced to the lowest level, the OA value is maximized, and the overall classification accuracy is high, which is effective for SAR image change detection.

[0103] Experimental results show that in dual-polarization SAR, different polarization modes exhibit significant differences in the scattering mechanisms of the ground surface and vegetation, which in turn affects the accuracy of change detection methods. Cross-polarization (VH) primarily reflects volume scattering within the vegetation canopy and is highly sensitive to changes in vegetation density and structure. Co-polarization (VV) primarily reflects surface or roughness scattering and is more sensitive to soil moisture and bare soil exposure. In change detection, VH polarization generally more accurately identifies vegetation disturbances because it mitigates soil moisture interference. In experiments, the VH ratio method generally outperformed the VV ratio method. However, the logarithmic ratio method (VV) polarization enhances the signal dynamic range through nonlinear transformation, better capturing surface changes caused by landslides. The mean ratio method (VV) polarization captures bare soil exposure and moisture changes by incorporating the domain mean. Analysis of historical meteorological data indicates that multiple days of rainy weather during earthquakes significantly increase surface moisture. VV polarization is more sensitive to soil moisture changes. This environmental factor may further enhance the detection performance of VV based on logarithmic and mean ratio methods.

[0104] In summary, to effectively reduce speckle noise in SAR imagery and improve change detection performance, this study proposed a SAR image change detection voting method based on a multimodal information fusion strategy. This method utilizes multi-temporal dual-polarization SAR to generate difference images using three change detection operators: the ratio method, the mean ratio method, and the logarithmic ratio method. The resulting change detection results are then fused using a majority voting method. Furthermore, terrain constraint optimization is performed using slope information calculated from a DEM to improve landslide identification accuracy. This paper optimizes the difference image construction, speckle noise suppression, adaptive Gaussian threshold segmentation, and majority voting decision-making to enhance overall change detection accuracy. Experimental change detection experiments using the Luding earthquake landslide as an example demonstrate that this method effectively suppresses speckle noise, enhances change information extraction, and significantly improves change detection performance, minimizing false positives and increasing the average error rate (OA) to 90.7%. Future work aims to improve landslide detection accuracy while enhancing the method's automation and adaptability, providing more efficient and accurate technical support for post-earthquake disaster monitoring and emergency response.

[0105] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A landslide extraction method based on SAR image change detection, characterized in that: include: Acquire multi-temporal SAR images of the earthquake area before and after the landslide and perform preprocessing; For the VV and VH polarization images of multi-temporal SAR images before and after the landslide, the difference maps of the earthquake area were constructed using the ratio method, logarithmic ratio method and mean ratio method respectively. For each difference map, an adaptive Gaussian threshold segmentation method is applied to solve the optimal segmentation threshold, and the corresponding difference map is binarized using the optimal segmentation threshold to divide the earthquake area into a change area and a non-change area; For a given pixel, the majority voting method is used to fuse the segmentation results of each difference map to make a decision. If the pixel is classified as a change area in at least half of the difference maps, it is marked as a candidate landslide area. The slope is calculated based on the DEM data of the earthquake area obtained in advance to obtain the slope distribution map of the earthquake area for terrain constraint. The candidate landslide areas with slopes less than a preset threshold are eliminated to finally obtain the landslide area.

2. The landslide extraction method based on SAR image change detection according to claim 1 is characterized in that: The preprocessing includes orbit correction, radiation calibration, filtering, terrain correction, decibel processing and image registration.

3. The landslide extraction method based on SAR image change detection according to claim 1 is characterized in that: The difference map is constructed based on the multi-temporal SAR intensity information image obtained by mathematical definition of the multi-temporal SAR image. The mathematical definition expression is as follows: I before ={I before (i,j),1<<i<<M,1<<j<<N}, I after ={I after (i,j),1<<i<<M,1<<j<<N}, Among them, I before and I after is the mean of the multi-temporal SAR images before and after the landslide, M and N are I before and I after The number of pixels in the length and width directions, i represents the i-th pixel in the length direction, and j represents the j-th pixel in the width direction.

4. The landslide extraction method based on SAR image change detection according to claim 3 is characterized in that: The calculation formulas for constructing difference maps using the ratio method, logarithmic ratio method, and mean ratio method are: Where R represents the ratio, LR represents the log ratio, and MR represents the mean ratio.

5. The landslide extraction method based on SAR image change detection according to claim 4 is characterized in that: The formula for binarizing the corresponding difference map using the optimal segmentation threshold is: Among them, B k (i, j) is the binarization result of pixel (i, j) in the corresponding difference map (1 represents the changed area, 0 represents the non-changed area), T k is the optimal segmentation threshold of the corresponding difference map determined by the adaptive Gaussian threshold method.

6. The landslide extraction method based on SAR image change detection according to claim 5 is characterized in that: The formula for making a decision by using majority voting to fuse the segmentation results of each difference map is: Among them: B MV (i, j) represents the final voting result of pixel (i, j), N is the total number of difference maps, and if at least half of the difference maps at pixel (i, j) are classified as a change area (i.e., 1), the final voting result is 1, otherwise it is 0; if the final voting result is 1, the pixel (i, j) is marked as a candidate landslide area.

7. The landslide extraction method based on SAR image change detection according to claim 1 is characterized in that: The calculation formula for the slope tangent value tanα of a point on the terrain surface in the slope distribution map of the earthquake area is: Where Δh is the vertical height change of the point (Δh), and Δd is the horizontal distance change of the point.

8. The landslide extraction method based on SAR image change detection according to claim 1 is characterized in that: The preset threshold value for eliminating slopes less than the preset threshold value is 15°.

9. A landslide extraction system based on SAR image change detection, characterized in that: include: An acquisition module is used to acquire multi-phase SAR images of the earthquake area before and after the landslide and perform preprocessing; A construction module is used to construct a difference map of the earthquake area using the ratio method, logarithmic ratio method and mean ratio method for the VV and VH polarization images of the multi-phase SAR images before and after the landslide; a segmentation module for applying an adaptive Gaussian threshold segmentation method to each difference map to obtain an optimal segmentation threshold, and using the optimal segmentation threshold to perform binarization processing on the corresponding difference map to divide the earthquake area into a change area and a non-change area; The decision module is used to make a decision by fusing the segmentation results of each difference map using the majority voting method for a given pixel. If the pixel is classified as a change area in at least half of the difference maps, it is marked as a candidate landslide area. The extraction module is used to calculate the slope based on the pre-acquired DEM data of the earthquake area to obtain the slope distribution map of the earthquake area, and eliminate the candidate landslide areas with slopes less than a preset threshold to finally extract the landslide areas.

Images