Satellite-borne SAR building damage assessment method based on double-domain feature fusion

Abstract

The invention belongs to the technical field of remote sensing image processing and damage assessment, and particularly relates to a satellite-borne SAR building damage assessment method based on double-domain feature fusion, which comprises the following steps: acquiring same-orbit satellite-borne SAR image pairs in the same area before and after a disaster, and collecting detailed situations of damage events; processing the single-view complex image to generate pre-disaster and post-disaster SAR image data after preprocessing; a double-branch architecture is adopted, spatial domain features and frequency domain features are extracted, and a double-domain feature fusion network model is constructed; utilizing a double-domain feature fusion network model to extract and optimize a damaged area; based on the damage distribution diagram, the total damage area, the area of each partition and the proportion are calculated; performing damage grade division based on the back scattering coefficient variation before and after the disaster; and generating a damage assessment report. According to the method, accurate extraction, grading and quantitative analysis of the damage area can be realized, the accuracy and reliability of damage identification are improved, and the high timeliness requirement of disaster emergency response is met.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing and damage assessment technology, specifically relating to a spaceborne SAR building damage assessment method based on dual-domain feature fusion. Background Technology

[0002] Following disasters such as earthquakes, explosions, and fires, rapidly and accurately acquiring information on building damage is crucial for disaster assessment, rescue decisions, and post-disaster reconstruction. Traditional field surveys are not only inefficient and dangerous, but also struggle to quickly grasp the overall disaster situation on a macro scale. In recent years, with the development of remote sensing technology, using optical remote sensing imagery for change detection has become an important means of disaster monitoring. However, optical remote sensing is heavily reliant on lighting and weather conditions, and during disasters, effective data is often obscured by clouds and smoke, significantly limiting its timeliness.

[0003] As an active microwave remote sensing sensor, spaceborne SAR possesses all-weather, all-day Earth observation capabilities and can effectively penetrate clouds and dust, providing unique data advantages for disaster emergency response. Currently, research on using spaceborne SAR data for building damage assessment has made some progress, mainly relying on change detection techniques for pre- and post-disaster SAR imagery. These techniques can be broadly divided into two categories: one is based on backscattering intensity change detection, which identifies damaged areas by comparing significant changes in building scattering intensity before and after damage (such as the failure of the original secondary scattering mechanism due to structural collapse, resulting in random scattering); the other is based on interferometric coherence change detection, which uses the decoherence characteristics of ground object scatterers caused by damage events to identify changes.

[0004] However, existing assessment methods based on spaceborne SAR still have significant limitations. First, they rely on information from a single feature domain: whether based on intensity or coherence, these methods often only utilize features of SAR images in a certain aspect, failing to fully explore and integrate the rich information contained in SAR data. Intensity changes are easily affected by various factors such as imaging geometry and moisture content, leading to false alarms; while coherence changes are sensitive to surface disturbances, their results are easily affected by time decorrelation and baseline decorrelation, and they are insufficient in distinguishing subtle damage levels. Second, the assessment accuracy and automation are limited: existing methods mostly remain at a coarse "change / no change" binary detection level, or rely on empirical thresholds for simple classification, making it difficult to achieve refined damage level classification (such as severe damage, moderate damage, minor damage). At the same time, traditional methods have a low degree of automation, often requiring a large amount of manual intervention for interpretation, which is difficult to meet the high timeliness requirements in disaster emergency response.

[0005] Therefore, there is an urgent need in this field for an assessment method that can overcome the above-mentioned shortcomings. It should be able to make full use of the multidimensional information of SAR data to achieve automated, high-precision building damage identification and quantitative level assessment. Summary of the Invention

[0006] The purpose of this invention is to provide a spaceborne SAR building damage assessment method based on dual-domain feature fusion. This method utilizes pre-disaster and post-disaster co-orbit spaceborne SAR image pairs, and through dual-domain feature fusion and intelligent recognition technology, it achieves accurate extraction, classification, and quantitative analysis of damaged areas, improves the accuracy and reliability of damage identification, meets the high timeliness requirements of disaster emergency response, and can effectively solve the problems of low assessment accuracy, insufficient automation, and difficulty in achieving refined quantitative assessment caused by reliance on single feature information in existing technologies.

[0007] Technical solution to achieve the purpose of this invention:

[0008] A spaceborne SAR building damage assessment method based on dual-domain feature fusion includes:

[0009] Step 1, Data Acquisition: Acquire pairs of satellite-borne SAR images of the same area before and after the disaster, and collect detailed information on the damage event;

[0010] Step 2, Data Preprocessing: Radiometric calibration, registration, multi-view processing, and adaptive filtering are performed on single-view complex images to generate pre-disaster and post-disaster SAR image data.

[0011] Step 3, Feature Fusion and Model Construction: A dual-branch architecture is adopted to extract spatial domain features and frequency domain features, fuse the two domains, and construct a dual-domain feature fusion network model;

[0012] Step 4, Damage Area Extraction and Optimization: Using the constructed dual-domain feature fusion network model, the damage area is extracted and optimized;

[0013] Step 5, Two-dimensional quantitative assessment: Based on the damage distribution map, calculate the total damaged area, the area of ​​each zone, and its proportion;

[0014] Step 6, Damage Severity Classification: Based on the change in backscattering coefficient before and after the disaster, damage severity is classified.

[0015] Step 7: Output of comprehensive damage assessment report: Based on the pre-processed pre-disaster and post-disaster SAR image data, two-dimensional damage range distribution map, damage level classification raster map, damage area and damage level classification results, a damage assessment report is generated.

[0016] Further, step 1 includes:

[0017] Step 1.1: Acquire satellite-borne SAR image pairs of the same area before and after the disaster: Select data from satellite systems as the data source; Select interferometric wide swath mode or strip mode data with a spatial resolution better than 15 meters; The data format is single-look complex data, containing amplitude and phase information;

[0018] Step 1.2: Collect detailed information about the damage event, including: the time of the event, the geographical coordinates of the center point; the estimated impact range of the damage event; and preliminary assessment information on the degree of damage.

[0019] Further, step 2 includes:

[0020] Step 2.1, Radiometric calibration processing: Convert the original digital quantization value of each pixel in the acquired single-view complex image into backscattering coefficients;

[0021] Step 2.2, Fine Registration Processing: Using pre-disaster imagery as the reference image and post-disaster imagery as the registration image; coarse registration is performed based on precise satellite orbit data; a registration method based on cross-correlation algorithm is adopted, control points are uniformly selected on the reference image, and the optimal corresponding point position of each control point is searched on the image to be registered by calculating the normalized cross-correlation coefficient, until the registration error is better than the accuracy requirement of 0.2 pixels, thus obtaining corresponding point pairs; through the corresponding point pairs, a polynomial transformation model is constructed, and the image to be registered is resampled to ensure that each pixel is strictly aligned with the geometric position of the reference image, thus obtaining the registered pre-disaster and post-disaster images;

[0022] Step 2.3, SAR image multi-view processing: Based on the registration and processing of pre-disaster and post-disaster images, a specific number of multi-views is set, and multi-view processing is performed on the SAR images to generate multi-view processed images;

[0023] Step 2.4 Adaptive Filtering Processing: Based on multi-view processing, an adaptive filtering algorithm is used to enhance the image and generate a filtered image.

[0024] Further, step 3 includes:

[0025] Step 3.1 Spatial Domain Feature Extraction: The preprocessed SAR image feature map is passed through a convolutional layer to generate a new feature map; the new feature map is divided into three groups of features: horizontal central region features, global region features, and vertical central region features; the three groups of features are processed through a 3x3 convolutional layer to obtain three new features; the three new features are then fused element-wise to generate the final spatial domain fused features.

[0026] Step 3.2, Frequency Domain Feature Extraction: The preprocessed SAR image patch is processed by pooling layer to generate three parallel paths; each path's image patch is transformed by DCT to convert spatial information to the frequency domain, and an attention mechanism is used to filter features and generate multi-scale frequency feature vectors; the multi-scale feature vectors are fused into a unified frequency domain feature by concatenation or weighted summation.

[0027] Step 3.3, Dual-domain feature fusion: A dual-path attention mechanism is used to filter the frequency components of the frequency domain features, generating an information vector carrying the original frequency information and a function-normalized attention gating vector. The information vector and the attention gating vector are then adaptively weighted using the Hadamard product to obtain the optimized frequency domain features. The optimized frequency domain features are then deeply fused with the spatial domain features through a concatenation operation. The fused dual-domain features are then input into a fully connected layer to complete the classification decision, thus constructing the dual-domain feature fusion network model.

[0028] Further, step 4 includes:

[0029] Step 4.1, Damage Area Extraction: The pre-processed pre-disaster and post-disaster SAR images are input into the dual-domain feature fusion neural network model. Through end-to-end calculation, a pixel-level preliminary damage area binary map is output, generating a preliminary damage area map.

[0030] Step 4.2, Damage Area Optimization: Morphological filtering is applied to the extracted preliminary damage area for post-processing to generate an optimized damage area map;

[0031] Step 4.3, Damage Pixel Mapping: Based on geocoding information, each damaged pixel is mapped to the real geographic coordinate system to generate a damage distribution map with the actual coordinate system.

[0032] Furthermore, in step 5, the formula for calculating the total damaged area is:

[0033] S damage=N×R ground2

[0034] In the formula, Sdamage represents the total damaged area, N represents the total number of damaged pixels identified, and Rground2 represents the ground area represented by a single pixel;

[0035] The formula for calculating the percentage of damaged area is:

[0036] P damage=(S damage / S total)×100%

[0037] In the formula, Pdamage represents the percentage of damaged area, and Stotal represents the total area.

[0038] Further, step 6 includes:

[0039] Step 6.1, Damage Level Classification: Set threshold ranges for different damage levels, and classify the damage levels based on the change in backscattering coefficient Δσ° before and after the disaster, matching the threshold ranges for the damage levels.

[0040] Furthermore, the damage level classification in step 6.1 is specifically as follows:

[0041] When Δσ° < -10dB, it is judged as "severe damage", indicating that the building structure has completely collapsed or been destroyed, the original regular geometric shape and strong secondary scattering characteristics are basically lost, and the echo signal is drastically weakened.

[0042] When -10dB≤Δσ°<-5dB, it is judged as "moderate damage", indicating that the main structure of the building has been significantly damaged, but the overall outline is still intact and the echo intensity is significantly reduced.

[0043] When -5dB≤Δσ°<-2dB, it is judged as "minor damage", indicating that the scattering is slightly changed due to the increase in surface roughness caused by the damage to the building surface;

[0044] When the change is between -2dB and 2dB, it is judged as "basically intact", indicating that the scattering characteristics have not undergone significant changes beyond the normal fluctuation range.

[0045] Furthermore, step 6 also includes:

[0046] Step 6.2, Damage Level Determination: The threshold system judgment value is determined by adopting a regional statistical smoothing strategy; the threshold system is fine-tuned and verified regionally according to the main building structure types and typical SAR scattering response characteristics of different regions; a damage level classification raster map corresponding to the damage distribution map space is generated.

[0047] Further, step 7 includes:

[0048] Step 7.1, Damage Thematic Map Creation: Using the post-disaster SAR image preprocessed in Step 2 as the base map, the two-dimensional damage range distribution map generated in Step 4 is overlaid on it. The damage level classification raster map obtained in Step 6 is overlaid on the base map in a rendering manner. Color coding is used to visualize the spatial distribution of damage at different levels to generate a damage spatial distribution map. At the same time, the comparison map of pre-disaster and post-disaster remote sensing images is displayed side by side to form a damage thematic map.

[0049] Step 7.2, Damage Assessment Report Generation: Based on the damage area conversion in Step 5 and the damage level classification results in Step 6, a structured data table is automatically generated; the damage thematic map and data table are comprehensively interpreted to provide a descriptive summary of the overall spatial distribution pattern of the damage, the composition of the main damage levels, and the key damaged areas. Targeted follow-up action suggestions can be proposed based on preset rules or an expert knowledge base, generating textual analysis conclusions and recommendations; a damage assessment report is generated by integrating the damage thematic map, data table, textual analysis conclusions, and recommendations.

[0050] The beneficial technical effects of this invention are as follows:

[0051] 1. This invention provides a spaceborne SAR building damage assessment method based on dual-domain feature fusion, which has high accuracy and low false alarm rate: By fusing complementary information from the spatial domain (such as texture and edge) and frequency domain (such as global structure pattern) of spaceborne SAR images, a more comprehensive feature representation is constructed. It makes full use of the multidimensional information of the data, reduces false detection and missed detection. Compared with methods that only use single features such as intensity or coherence, it can more accurately distinguish between real damage and false changes caused by imaging noise, seasonal changes, etc., which significantly improves the accuracy and reliability of damage identification and enhances detection precision.

[0052] 2. The present invention provides a spaceborne SAR building damage assessment method based on dual-domain feature fusion, which is highly intelligent and efficient: it uses deep learning networks to realize end-to-end automated processing from feature extraction to change detection, from raw data to final assessment report, minimizing human intervention, meeting the high timeliness requirements of disaster emergency response, overcoming the dependence of traditional methods on manually designed features and set thresholds, greatly improving processing efficiency, and meeting the needs of large-scale and rapid disaster assessment.

[0053] 3. The present invention provides a spaceborne SAR building damage assessment method based on dual-domain feature fusion, which provides in-depth assessment results and strong practicality: the final output of this method not only includes the damage range, but also provides quantitative statistics based on physical area (such as damage area and percentage) and damage level classification based on scattering mechanism (such as severe, moderate, slight, and basically intact), realizing the leap from "whether it has changed" to "how much it has changed and how severe it is", achieving refined quantitative assessment. The assessment results are more refined and objective, and have higher decision support value.

[0054] 4. The present invention provides a spaceborne SAR building damage assessment method based on dual-domain feature fusion, which has strong anti-interference ability: based on spaceborne SAR data, it has the advantage of working in all weather and all time; at the same time, the deep learning model has a certain fault tolerance and generalization ability for the data, which further enhances the stability and reliability of the method in complex scenarios. Attached Figure Description

[0055] Figure 1 The flowchart of a spaceborne SAR building damage assessment method based on dual-domain feature fusion provided by the present invention is shown below.

[0056] Figure 2 This is a schematic diagram illustrating the extraction and fusion of spatial and frequency domains in a spaceborne SAR building damage assessment method based on dual-domain feature fusion provided by the present invention.

[0057] Figure 3 This is a schematic diagram of spatial domain feature extraction in a spaceborne SAR building damage assessment method based on dual-domain feature fusion provided by the present invention.

[0058] Figure 4 The image shown is a result of automatically identifying damaged buildings in the Port of Beirut using a spaceborne SAR building damage assessment method based on dual-domain feature fusion provided by this invention. In the image, A is the SAR image before damage, B is the SAR image after damage, and C is the optical image after damage. The red polygon represents the boundary of the identified damaged building. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0060] like Figure 1 As shown, this invention provides a method for assessing damage to spaceborne SAR buildings based on dual-domain feature fusion, specifically including the following steps:

[0061] Step 1: Data Acquisition: Acquire satellite-borne SAR image pairs of the same area before and after the disaster to gather detailed information on the damage event.

[0062] Step 1.1: Acquire satellite-borne SAR images of the same area before and after the disaster.

[0063] "Same orbit" means that two observations use essentially the same orbital parameters, with orbital deviations controlled within a preset range; data sources include, but are not limited to, Sentinel-1, TerraSAR-X, and Gaofen-3 satellite systems; data preferably uses C-band or X-band interferometric wide-swath or stripe mode data, with a spatial resolution better than 15 meters; the data format is single-look complex data (SLC), containing amplitude and phase information.

[0064] Step 1.2: Gather detailed information about the damage incident.

[0065] Detailed information about the damage event includes: the time of the event, the geographical coordinates (latitude and longitude) of the center point; an estimated impact range of the damage event, expressed in the form of polygon boundary coordinates; and preliminary assessment information on the degree of damage, including a description of the damage type (such as explosion, fire, earthquake, etc.).

[0066] Step 2, Data Preprocessing: Radiometric calibration, registration, multi-view processing, and adaptive filtering are performed on the single-view complex images to generate radiometrically normalized and geometrically aligned pre-disaster and post-disaster SAR image data.

[0067] Step 2.1: Radiometric calibration processing of single-view complex images

[0068] This step aims to convert the sensor's raw measurements into physically meaningful backscattering coefficients for ground features, ensuring the comparability of radiative intensity between images from different time phases. Specifically, calibration parameters (such as the calibration constant K) provided by the sensor manufacturer are used. For the input raw single-view complex image, the raw digital quantization value of each pixel is first extracted. Subsequently, the calibration formula is applied:

[0069] σ°=DN 2 / K

[0070] Here, DN represents the raw digital quantization value of the pixel, and K is a calibration constant specified by the satellite data provider, which is usually recorded in the data's metafile (such as a .xml file). Through this calculation, each pixel value is converted into a backscattering coefficient σ°, which directly reflects the radar scattering characteristics of ground targets, laying the foundation for radiometric consistency for subsequent quantitative change detection.

[0071] Step 2.2: Fine-grained registration and processing of pre-disaster and post-disaster images

[0072] To achieve pixel-level accurate comparison between pre-disaster and post-disaster images, sub-pixel-level fine registration is necessary. This step uses the geometrically stable pre-disaster image as a spatial reference to register the post-disaster image. First, coarse registration is performed, typically based on precise satellite orbit data, to eliminate overall translation and rotation caused by orbital deviations between images. Then, fine registration is performed using a cross-correlation algorithm. A large number of control points are uniformly selected on the reference image (pre-disaster image), and the optimal corresponding point position for each control point is searched on the image to be registered (post-disaster image) by calculating the normalized cross-correlation coefficient, until the registration error is better than the accuracy requirement of 0.2 pixels. Finally, using these precise corresponding point pairs, a multinomial transformation model is constructed to resample the image to be registered (commonly using bilinear or cubic convolution interpolation algorithms), ensuring that each pixel is strictly aligned geometrically with the reference image, guaranteeing the accuracy of subsequent change detection.

[0073] Step 2.3: SAR Image Multi-look Processing

[0074] Based on pre-disaster and post-disaster image registration processing, multi-look averaging is used to suppress inherent speckle noise in SAR images, improve the signal-to-noise ratio, and generate multi-look processed images. Processing is performed separately in the range and azimuth directions. A specific multi-look ratio is set according to the initial image resolution and the desired output resolution, for example, performing 2:1 multi-look processing in both the range and azimuth directions. This process essentially involves averaging multiple adjacent single pixels (single look) in the range and azimuth directions using non-overlapping or overlapping blocks to generate a new pixel value. For single-look complex data, this averaging operation is typically performed in the intensity domain. After multi-look processing, the equivalent number of views increases, speckle noise is significantly suppressed, and the image visual effect is smoother, but at the cost of a corresponding decrease in spatial resolution. Therefore, the selection of the multi-look ratio needs to strike a balance between noise suppression and detail preservation.

[0075] Step 2.4, Adaptive Filtering Process

[0076] Building upon multi-view processing, an adaptive filtering algorithm is employed to enhance the image quality and generate a multi-view processed image. This invention preferentially uses the Lee filtering algorithm because it can adaptively adjust the filtering intensity based on the statistical characteristics of local image regions. Specifically, a fixed-size sliding window (e.g., 7×7 pixels) traverses the entire image. For each pixel to be filtered at the center of the window, the algorithm calculates the mean and standard deviation within that local window. The filtering operation is based on a linear model, the core idea of ​​which is that the true backscattering coefficient of a pixel can be best estimated from the mean of that local region, and the deviation between the observed value and the mean is partly caused by the actual signal changes and partly by speckle noise. The filter weights are adaptively adjusted according to the relationship between local variance and global variance: in uniform regions (small variance), the mean is given greater weight for stronger smoothing to suppress noise; in edge or textured regions (large variance), the weight for retaining the original pixel values ​​is greater to better preserve edge and texture features. This step effectively smooths noise in uniform regions while preserving detailed information such as building edges to the greatest extent, providing a high-quality data foundation for subsequent feature extraction.

[0077] Step 3, Feature Fusion and Model Construction: A dual-branch architecture is adopted to extract spatial domain features and frequency domain features, and the two domains are fused to construct a dual-domain feature fusion network model.

[0078] like Figure 2As shown, a dual-domain feature fusion network for building damage in spaceborne SAR imagery is constructed. A dual-branch architecture is adopted. In the spatial domain branch, a deep feature extraction network is built using four-level cascaded multi-directional convolutional modules. In the frequency domain branch, three parallel paths are generated through pooling layers to preserve local details, focus on the overall contour, and define the baseline scale, respectively. After DCT transformation of the image patch for each path, an attention mechanism is used for feature selection to generate multi-scale frequency feature vectors.

[0079] Step 3.1 Spatial Domain Feature Extraction

[0080] A deep feature extraction network is constructed using four cascaded multi-directional convolutional modules, with each MDC module employing a feature grouping mechanism. For example... Figure 3 As shown, the preprocessed SAR image feature map (assumed to be 7×7) is first passed through a convolutional layer to generate a new feature map S with 15 channels. Then, these 15 channels are divided into three groups of features (5 channels per group), S = Sh + Sr + Sv:

[0081] Sh represents the horizontal middle region feature. To emphasize the center, the values ​​of the top two and bottom two rows of the feature map are set to 0, retaining only the effective information from the middle three rows (i.e., the horizontal strip region); Sr represents the global region feature, retaining the complete 7x7 region information and using a spatial attention mechanism to generate a global feature map to capture global context; Sv represents the vertical middle region feature. Similarly, to emphasize the center, the values ​​of the leftmost two and rightmost columns of the feature map are set to 0, retaining only the effective information from the middle three columns (i.e., the vertical strip region);

[0082] Finally, the three sets of features Sh, Sr, and Sv are each processed by an independent 3x3 convolutional layer to obtain three new sets of features S'h, S'r, and S'v. These three new sets of features are then fused element-wise to generate the final spatial domain fusion feature S', where S' = S'h + S'r + S'v. This design enables the network to adaptively model contextual information and central region features, preserving the necessary global context while significantly reducing the negative impact of edge noise on the classification decision of the central pixel, ultimately outputting a more robust feature representation.

[0083] Step 3.2: Frequency Domain Feature Extraction. This framework first performs multi-scale processing on the preprocessed SAR image patches using pooling layers, generating three parallel paths: Path 1 scales the image to a fine scale of 2×12×12 to retain more local details; Path 2 maintains a basic scale of 2×8×8 as a balancing benchmark; and Path 3 scales to a coarse scale of 2×6×6 to focus on the overall contour. Each path's image patch undergoes Discrete Cosine Transform (DCT) to convert spatial information to the frequency domain. Then, independent attention modules ("ON-OFF" switching mechanism) are used for feature selection, generating frequency feature vectors (Vf1, Vf2, Vf3) at corresponding scales. Finally, the multi-scale feature vectors are fused into a unified frequency domain feature Vf_multi through concatenation or weighted summation strategies. This preserves complementary information at different scales and strengthens key frequency components through the attention mechanism, significantly improving the model's ability to perceive complex changes.

[0084] Step 3.3: Dual-domain feature fusion, constructing a dual-domain feature fusion network model.

[0085] This step utilizes an attention-weighted feature fusion mechanism to achieve information complementarity between spatial domain features and frequency domain features.

[0086] Adaptive filtering of frequency components:

[0087] In the feature optimization stage, a dual-path attention mechanism is employed to filter frequency components in the frequency domain features. This mechanism uses two parallel fully connected layers to generate information vectors (i = W) carrying the original frequency information. i v+b i The attention gating vector (g = σ(Wgv + bg)) is normalized by the Sigmoid function. When the gating value is close to 1, the corresponding frequency component is preserved and enhanced; when the gating value is close to 0, the component is suppressed.

[0088] Attention weighting and feature optimization:

[0089] Adaptive feature weighting (V_f = i⊙g) is achieved by combining the information vector and the attention-gated vector through the Hadamard product. This dynamic weighting mechanism automatically strengthens frequency features useful for change detection tasks (such as low-frequency components characterizing ground structures) while weakening interference components that may contain noise (such as some high-frequency components), thus obtaining optimized frequency domain features.

[0090] Deep fusion of dual-domain features:

[0091] The optimized frequency domain features and spatial domain features are deeply fused through a splicing operation. The spatial domain features provide detailed texture information, while the frequency domain features contribute macroscopic spectral characteristics.

[0092] Classification decision output:

[0093] The fused dual-domain features are input into a fully connected layer for feature mapping and dimensional transformation. Finally, a classifier makes a classification decision, thus constructing a dual-domain feature fusion network model.

[0094] Step 4, Damaged Area Extraction and Optimization: Using the constructed dual-domain feature fusion network model, the damaged area is extracted and optimized.

[0095] Step 4.1: Extraction of damaged areas

[0096] The preprocessed pre-disaster and post-disaster SAR images are used as input to construct a dual-domain feature fusion neural network model. Through end-to-end computation, a pixel-level preliminary damage area binary map is output, in which the highlighted area (pixels with a value of 1) represents the identified potential damage range, thus generating a preliminary damage area map.

[0097] Step 4.2, Damage Area Optimization

[0098] Morphological filtering is applied to the extracted preliminary damaged areas for post-processing. A 3×3 pixel structuring element is used for opening operations to eliminate isolated noise points and false damaged areas with too small a size, thereby reducing the false detection rate. Then, a 5×5 pixel structuring element is used for closing operations to fill the small holes that may be generated inside the damaged areas due to complex scattering mechanisms or occlusion, and to connect the local fracture areas to ensure the integrity and smoothness of the damaged patches, generating an optimized damaged area map.

[0099] Step 4.3, Damaged Pixel Mapping

[0100] A damage distribution map with an actual coordinate system is generated using geocoding parameters.

[0101] This step first maps each damaged pixel to a real geographic coordinate system (such as WGS-84UTM projection) based on geocoding information, generating a damage distribution map with the actual coordinate system. The geocoding process utilizes precise satellite orbital parameters and a digital elevation model (DEM) to accurately calculate the ground area corresponding to each pixel using a slant-range-Doppler model or a corresponding geometric correction algorithm. Specifically, the ground resolution (R_ground) is determined by the slant-range resolution of the SAR system (ρr) and the local incident angle (θ), and its calculation formula is as follows:

[0102] R_ground = ρr / sinθ.

[0103] Step 5, Two-dimensional quantitative assessment: Based on the damage distribution map, calculate the total damaged area, the area of ​​each zone, and its proportion.

[0104] Based on the damage distribution map, the total damage area (Sdamage) is obtained by multiplying the total number of identified damaged pixels (N) by the ground area represented by a single pixel (Rground2), i.e.

[0105] S damage = N × R ground2.

[0106] For a specific target assessment area (such as an entire urban block or the area surrounding a specific infrastructure), calculate its total area (S total).

[0107] The percentage of damaged area (P damage) is calculated using the formula P damage = (S damage / S total) × 100%.

[0108] Step 6: Damage Severity Classification: Based on the change in backscattering coefficient before and after the disaster, damage severity is classified.

[0109] Step 6.1, Damage Level Classification

[0110] This step is based on the precise backscattering coefficient (σ°) image obtained after preprocessing. By calculating the change in backscattering coefficient of corresponding pixels in the pre-disaster and post-disaster images (Δσ°=σ°post-σ°pre), a refined quantitative assessment of the degree of damage is achieved. Based on the physical correspondence between SAR scattering mechanism and the degree of damage to building structure, a strict quantification threshold range is set for classification: when Δσ° < -10dB, it is judged as "severe damage", indicating that the building structure has completely collapsed or been destroyed, the original regular geometric shape and strong secondary scattering characteristics are basically lost, and the echo signal is drastically weakened; when -10dB ≤ Δσ° < -5dB, it is judged as "moderate damage", indicating that the main structure of the building (such as load-bearing walls, beams and columns) has been significantly damaged, but the overall outline is still present, and the echo intensity is significantly reduced; when -5dB ≤ Δσ° < -2dB, it is judged as "minor damage", which usually corresponds to the slight change in scattering caused by the increase in surface roughness due to damage to the building surface (such as the peeling of exterior wall decoration, broken glass curtain wall, etc.); when the change is between -2dB and 2dB, it is judged as "basically intact", meaning that the scattering characteristics have not undergone significant changes beyond the normal fluctuation range.

[0111] Step 6.2, Damage Level Assessment

[0112] To improve the accuracy and reliability of the classification results, a regional statistical smoothing strategy can be adopted to determine the threshold system judgment value. Specifically, the mean Δσ° within a local window centered on a pixel (e.g., 3x3 or 5x5) is used as the final judgment value for that central pixel, reducing speckle noise interference. Simultaneously, this threshold system can be regionally fine-tuned and verified based on the main building structure types (e.g., reinforced concrete frames, brick-concrete structures) and their typical SAR scattering response characteristics in different regions, ensuring a high degree of consistency between the classification results and the actual physical damage status. Finally, a damage level classification raster map corresponding to the damage distribution map space is generated, providing core quantitative classification information for the comprehensive assessment report.

[0113] Step 7: Output of Comprehensive Damage Assessment Report: Based on the preprocessed pre-disaster and post-disaster SAR image data, two-dimensional damage range distribution map, damage level classification raster map, damage area and damage level classification results, a damage assessment report is generated.

[0114] Step 7.1: Creating thematic damage maps

[0115] This step systematically integrates the processing results of the preceding steps to generate a comprehensive damage assessment report with a complete structure and detailed content. The report first focuses on spatial overlay analysis, generating a series of thematic maps: using the geometrically corrected post-disaster SAR amplitude image from step 2 as a base map, a vector-formatted two-dimensional damage extent distribution map (usually in ESRI Shapefile format) generated in step 4 is precisely overlaid on it. This vector layer clearly delineates the boundaries of the damaged area. Simultaneously, the damage level grading raster map (e.g., GeoTIFF format) obtained in step 6 is overlaid on the base map using semi-transparent color rendering. Intuitive color coding (e.g., red for severe damage, orange for moderate damage, yellow for minor damage, and green for largely intact) is used to visualize the spatial distribution of different damage levels, generating a damage spatial distribution map. Furthermore, the report includes a set of juxtaposed pre-disaster and post-disaster remote sensing image comparison maps, visually presenting the changes in the target area before and after the disaster through a roll-up or flashing display method. A thematic map of damage is created by combining the spatial distribution map of the damage with a comparison map of pre- and post-disaster remote sensing images.

[0116] Step 7.2: Damage Assessment Report Generation

[0117] The report provides precise quantitative statistical and analytical conclusions. Based on the damage area conversion in step 5 and the damage level classification results in step 6, a structured data table is automatically generated. This table details the physical area corresponding to each damage level, the percentage of the total assessed area, and the number of damaged patches, among other key quantitative indicators. Finally, the system automatically or semi-automatically generates textual analysis conclusions and recommendations. This textual section comprehensively interprets the information from the aforementioned charts and graphs, providing a descriptive summary of the overall spatial distribution pattern of the damage, the composition of the main damage levels, and the key damaged areas. It can also propose targeted follow-up action suggestions (such as priority rescue area assessment and preliminary loss estimation) based on preset rules or an expert knowledge base, generating textual analysis conclusions and recommendations. All charts, tables, and texts are ultimately integrated into a unified document (such as a PDF report) or a Geographic Information System (GIS) engineering package containing all original data and output files, generating a damage assessment report and ensuring the systematic nature, traceability, and direct applicability of the assessment results.

[0118] The damage assessment method for spaceborne SAR structures based on dual-domain feature fusion, provided by this invention, was used to identify damaged structures in the Port of Beirut. The automatically identified damaged structures in the Port of Beirut are shown in the image below. Figure 4 As shown, A is the SAR image before damage, B is the SAR image after damage, C is the optical image after damage, and the red polygon is the boundary of the damaged building.

[0119] The present invention has been described in detail above with reference to the accompanying drawings and embodiments. However, the present invention is not limited to the above embodiments, and various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention. All contents not described in detail in the present invention can be derived from existing technologies.

Claims

1. A method for assessing building damage based on dual-domain feature fusion in spaceborne SAR, characterized in that, include: Step 1, Data Acquisition: Acquire pairs of satellite-borne SAR images of the same area before and after the disaster, and collect detailed information on the damage event; Step 2, Data Preprocessing: Radiometric calibration, registration, multi-view processing, and adaptive filtering are performed on single-view complex images to generate pre-disaster and post-disaster SAR image data. Step 3, Feature Fusion and Model Construction: A dual-branch architecture is adopted to extract spatial domain features and frequency domain features, fuse the two domains, and construct a dual-domain feature fusion network model; Step 4, Damage Area Extraction and Optimization: Using the constructed dual-domain feature fusion network model, the damage area is extracted and optimized; Step 5, Two-dimensional quantitative assessment: Based on the damage distribution map, calculate the total damaged area, the area of ​​each zone, and its proportion; Step 6, Damage Severity Classification: Based on the change in backscattering coefficient before and after the disaster, damage severity is classified. Step 7: Output of comprehensive damage assessment report: Based on the pre-processed pre-disaster and post-disaster SAR image data, two-dimensional damage range distribution map, damage level classification raster map, damage area and damage level classification results, a damage assessment report is generated.

2. The spaceborne SAR building damage assessment method based on dual-domain feature fusion according to claim 1, characterized in that, Step 1 includes: Step 1.1: Acquire satellite-borne SAR image pairs of the same area before and after the disaster: Select data from satellite systems as the data source; Select interferometric wide swath mode or strip mode data with a spatial resolution better than 15 meters; The data format is single-look complex data, containing amplitude and phase information; Step 1.2: Collect detailed information about the damage event, including: the time of the event, the geographical coordinates of the center point; the estimated impact range of the damage event; and preliminary assessment information on the degree of damage.

3. The method for assessing damage to spaceborne SAR buildings based on dual-domain feature fusion according to claim 2, characterized in that, Step 2 includes: Step 2.1, Radiometric calibration processing: Convert the original digital quantization value of each pixel in the acquired single-view complex image into backscattering coefficients; Step 2.2, Fine Registration Processing: Using pre-disaster imagery as the reference image and post-disaster imagery as the registration image; coarse registration is performed based on precise satellite orbit data; a registration method based on cross-correlation algorithm is adopted, control points are uniformly selected on the reference image, and the optimal corresponding point position of each control point is searched on the image to be registered by calculating the normalized cross-correlation coefficient, until the registration error is better than the accuracy requirement of 0.2 pixels, thus obtaining corresponding point pairs; through the corresponding point pairs, a polynomial transformation model is constructed, and the image to be registered is resampled to ensure that each pixel is strictly aligned with the geometric position of the reference image, thus obtaining the registered pre-disaster and post-disaster images; Step 2.3, SAR image multi-view processing: Based on the registration and processing of pre-disaster and post-disaster images, a specific number of multi-views is set, and multi-view processing is performed on the SAR images to generate multi-view processed images; Step 2.4 Adaptive Filtering Processing: Based on multi-view processing, an adaptive filtering algorithm is used to enhance the image and generate a filtered image.

4. The spaceborne SAR building damage assessment method based on dual-domain feature fusion according to claim 3, characterized in that, Step 3 includes: Step 3.1 Spatial Domain Feature Extraction: The preprocessed SAR image feature map is passed through a convolutional layer to generate a new feature map; the new feature map is divided into three groups of features: horizontal central region features, global region features, and vertical central region features; the three groups of features are processed through a 3x3 convolutional layer to obtain three new features; the three new features are then fused element-wise to generate the final spatial domain fused features. Step 3.2, Frequency Domain Feature Extraction: The preprocessed SAR image patch is processed by pooling layer to generate three parallel paths; each path's image patch is transformed by DCT to convert spatial information to the frequency domain, and an attention mechanism is used to filter features and generate multi-scale frequency feature vectors; the multi-scale feature vectors are fused into a unified frequency domain feature by concatenation or weighted summation. Step 3.3, Dual-domain feature fusion: A dual-path attention mechanism is used to filter the frequency components of the frequency domain features, generating an information vector carrying the original frequency information and a function-normalized attention gating vector. The information vector and the attention gating vector are then adaptively weighted using the Hadamard product to obtain the optimized frequency domain features. The optimized frequency domain features are then deeply fused with the spatial domain features through a concatenation operation. The fused dual-domain features are then input into a fully connected layer to complete the classification decision, thus constructing the dual-domain feature fusion network model.

5. The spaceborne SAR building damage assessment method based on dual-domain feature fusion according to claim 4, characterized in that, Step 4 includes: Step 4.1, Damage Area Extraction: The pre-processed pre-disaster and post-disaster SAR images are input into the dual-domain feature fusion neural network model. Through end-to-end calculation, a pixel-level preliminary damage area binary map is output, generating a preliminary damage area map. Step 4.2, Damage Area Optimization: Morphological filtering is applied to the extracted preliminary damage area for post-processing to generate an optimized damage area map; Step 4.3, Damage Pixel Mapping: Based on geocoding information, each damaged pixel is mapped to the real geographic coordinate system to generate a damage distribution map with the actual coordinate system.

6. The spaceborne SAR building damage assessment method based on dual-domain feature fusion according to claim 5, characterized in that, In step 5, the formula for calculating the total damaged area is: S damage=N×R ground2 In the formula, Sdamage represents the total damaged area, N represents the total number of damaged pixels identified, and Rground2 represents the ground area represented by a single pixel; The formula for calculating the percentage of damaged area is: P damage=(S damage / S total)×100% In the formula, Pdamage represents the percentage of damaged area, and Stotal represents the total area.

7. The spaceborne SAR building damage assessment method based on dual-domain feature fusion according to claim 6, characterized in that, Step 6 includes: Step 6.1, Damage Level Classification: Set threshold ranges for different damage levels, and classify the damage levels based on the change in backscattering coefficient Δσ° before and after the disaster, matching the threshold ranges for the damage levels.

8. The spaceborne SAR building damage assessment method based on dual-domain feature fusion according to claim 7, characterized in that, The damage level classification in step 6.1 is as follows: When Δσ° < -10dB, it is judged as "severe damage", indicating that the building structure has completely collapsed or been destroyed, the original regular geometric shape and strong secondary scattering characteristics are basically lost, and the echo signal is drastically weakened. When -10dB≤Δσ°<-5dB, it is judged as "moderate damage", indicating that the main structure of the building has been significantly damaged, but the overall outline is still intact and the echo intensity is significantly reduced. When -5dB≤Δσ°<-2dB, it is judged as "minor damage", indicating that the scattering is slightly changed due to the increase in surface roughness caused by the damage to the building surface; When the change is between -2dB and 2dB, it is judged as "basically intact", indicating that the scattering characteristics have not undergone significant changes beyond the normal fluctuation range.

9. The spaceborne SAR building damage assessment method based on dual-domain feature fusion according to claim 8, characterized in that, Step 6 also includes: Step 6.2, Damage Level Determination: The threshold system judgment value is determined by adopting a regional statistical smoothing strategy; the threshold system is fine-tuned and verified regionally according to the main building structure types and typical SAR scattering response characteristics of different regions; a damage level classification raster map corresponding to the damage distribution map space is generated.

10. The method for assessing damage to spaceborne SAR buildings based on dual-domain feature fusion according to claim 9, characterized in that, Step 7 includes: Step 7.1, Damage Thematic Map Creation: Using the post-disaster SAR image preprocessed in Step 2 as the base map, the two-dimensional damage range distribution map generated in Step 4 is overlaid on it. The damage level classification raster map obtained in Step 6 is overlaid on the base map in a rendering manner. Color coding is used to visualize the spatial distribution of damage at different levels to generate a damage spatial distribution map. At the same time, the comparison map of pre-disaster and post-disaster remote sensing images is displayed side by side to form a damage thematic map. Step 7.2, Damage Assessment Report Generation: Based on the damage area conversion in Step 5 and the damage level classification results in Step 6, a structured data table is automatically generated; the damage thematic map and data table are comprehensively interpreted to provide a descriptive summary of the overall spatial distribution pattern of the damage, the composition of the main damage levels, and the key damaged areas. Targeted follow-up action suggestions can be proposed based on preset rules or an expert knowledge base, generating textual analysis conclusions and recommendations; a damage assessment report is generated by integrating the damage thematic map, data table, textual analysis conclusions, and recommendations.

Images