Highway water damage disaster rapid identification method and system based on remote sensing image analysis

Abstract

This invention discloses a rapid identification method and system for highway flood damage based on remote sensing image analysis, belonging to the field of remote sensing image analysis and disaster detection technology. The method includes: extracting change areas from pre- and post-disaster remote sensing images using a multi-scale differential twin network change detection module; identifying five types of flood damage—roadbed erosion, pavement collapse, slope landslide, retaining wall collapse, and bridge / culvert damage—using a cascaded attention-based flood damage semantic segmentation network; calculating the damage severity level by fusing morphological, spectral, and spatial features using a multi-dimensional damage severity collaborative assessment module; and determining traffic status and generating a repair priority report using a traffic impact adaptive quantification module. This invention achieves rapid identification, accurate classification, and comprehensive assessment of highway flood damage, providing technical support for post-disaster emergency rescue and traffic restoration decisions.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image analysis and disaster detection technology, specifically relating to a method and system for rapid identification of highway flood damage based on remote sensing image analysis. Background Technology

[0002] Frequent rainstorms and floods cause severe water damage to highway infrastructure, including roadbed washouts, road surface collapses, slope landslides, retaining wall collapses, and bridge and culvert damage. A rapid and accurate assessment of the overall damage along the entire route is crucial for developing emergency response and road clearing plans and making decisions on repair projects. Traditional ground-based inspection methods are inefficient, have limited coverage, and require significant manual labor, making them unsuitable for meeting the timeliness requirements of large-scale disaster emergency assessments.

[0003] In the prior art, Chinese invention patent application CN120355703A discloses a geological disaster detection method based on image processing. This method acquires initial pavement image data of the area to be detected and performs geometric correction to obtain pavement feature image data. It then uses the outer edge coefficients and pixel grayscale values ​​of the spots in the pavement feature image data to filter and remove all spots, using the remaining spots as the re-detection area. This re-detection area is then processed to obtain the crack areas of the pavement. While this method can detect pavement cracks, it has the following technical shortcomings: First, it only processes single-phase images and lacks the ability to compare and analyze pre- and post-disaster remote sensing images over time, making it unable to effectively distinguish between newly added damage caused by disasters and existing pavement defects. Second, the method uses a traditional spot outer edge coefficient and grayscale value filtering mechanism, which is insufficient for identifying large-scale structural damage such as roadbed erosion and slope collapse caused by water disasters. Third, this method fails to perform fine-grained classification of water damage types and fails to assess the degree of damage and traffic impact, thus failing to provide decision support for prioritizing post-disaster repairs.

[0004] In recent years, deep learning technology has made significant progress in the field of change detection in remote sensing images. Studies have shown that change detection methods based on Siamese network architecture can effectively extract differential features from dual-temporal images, and combining them with the global attention mechanism of Transformer can further enhance the modeling ability of long-distance spatial dependencies. However, existing methods still face challenges in identifying highway flood damage: on the one hand, flood damage exhibits complex multi-scale features, requiring feature extraction from different receptive fields, ranging from microscopic pavement cracks to macroscopic roadbed erosion; on the other hand, different types of flood damage show differences in spectral response and morphological features in remote sensing images, requiring sophisticated semantic segmentation networks for differentiation; furthermore, damage assessment and traffic impact analysis require comprehensive consideration of multi-dimensional features and the establishment of quantitative models.

[0005] Therefore, there is an urgent need to develop a technical solution that can make full use of multi-temporal remote sensing images to achieve rapid and accurate identification and comprehensive assessment of water damage disasters. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and system for rapid identification of highway flood damage based on remote sensing image analysis. By constructing a deeply coupled closed-loop collaborative architecture of four core modules—a multi-scale differential twin network change detection module, a cascaded attention-based flood damage semantic segmentation network, a multi-dimensional damage degree collaborative assessment module, and a traffic impact adaptive quantification module—it achieves rapid identification, accurate classification, degree assessment, and traffic impact analysis of highway flood damage.

[0007] The technical solution adopted in this invention is as follows:

[0008] A rapid identification method for highway flood damage based on remote sensing image analysis includes the following steps:

[0009] The system acquires post-disaster remote sensing images and pre-disaster archived baseline images of the affected road sections. After performing registration preprocessing on the dual-temporal images, the images are input into the multi-scale differential twin network change detection module. In the multi-scale differential twin network change detection module, a pyramid feature extractor with shared weights extracts multi-scale feature pyramids from the pre-disaster archived baseline images and the post-disaster remote sensing images. Feature difference maps are calculated at each scale level, and a mask of the changed area is obtained through multi-scale feature fusion and gated aggregation.

[0010] The changed area mask and the post-disaster remote sensing image are input into the cascaded attention-based water damage semantic segmentation network. The context enhancement features of the changed area are extracted by the multi-head cross-attention encoder, and the water damage type semantic segmentation map is generated by the adaptive category-aware decoder to identify five types of water damage: roadbed washout, road surface collapse, slope landslide, retaining wall collapse and bridge and culvert damage.

[0011] For each flood-damaged area, the multi-dimensional damage degree collaborative assessment module extracts morphological feature vectors, spectral anomaly index vectors, and spatial distribution feature vectors. Through feature cascade fusion and nonlinear mapping, a comprehensive damage degree score is calculated, and the damage degree is divided into three levels: mild, moderate, and severe based on an adaptive threshold.

[0012] The traffic impact adaptive quantification module calculates the traffic capacity impact coefficient based on the spatial occupancy ratio of the water-damaged area relative to the road cross section, the degree of damage, and the road grade attribute. It then determines the traffic status of the damaged road section as normal traffic, half-width traffic, single-lane traffic, or complete interruption, and generates a water damage disaster distribution map and a repair priority ranking report.

[0013] A rapid identification system for highway flood damage based on remote sensing image analysis includes:

[0014] The multi-scale differential twin network change detection module is used to receive post-disaster remote sensing images and pre-disaster archived baseline images. It extracts multi-scale feature pyramids from the pre-disaster archived baseline images and post-disaster remote sensing images through a pyramid feature extractor with shared weights. It calculates feature difference maps at each scale level and outputs a change area mask after multi-scale feature fusion and gated aggregation.

[0015] A cascaded attention-based water damage semantic segmentation network is used to receive a mask of the changed area and post-disaster remote sensing images. The contextual enhancement features of the changed area are extracted by a multi-head cross-attention encoder and then generated by an adaptive category-aware decoder to produce a water damage type semantic segmentation map.

[0016] The multi-dimensional damage assessment module is used to extract morphological feature vectors, spectral anomaly index vectors, and spatial distribution feature vectors for each water-damaged area. It calculates a comprehensive damage score and classifies the damage level by feature cascade fusion and nonlinear mapping.

[0017] The traffic impact adaptive quantification module is used to calculate the traffic capacity impact coefficient based on the spatial occupancy ratio of the flood-damaged area, the degree of damage, and the road grade attributes, determine the traffic status of the damaged road sections, and generate a flood disaster distribution map and a repair priority ranking report.

[0018] The beneficial effects of this invention are as follows:

[0019] First, this invention adopts a multi-scale differential twin network architecture, which simultaneously processes pre-disaster and post-disaster dual-temporal images through a pyramid feature extractor with shared weights. It calculates feature difference maps by combining a weighted fusion strategy of absolute value difference and cosine distance difference, which effectively overcomes the limitation of existing technologies that only process single-temporal images. It can accurately distinguish between new damage caused by disasters and existing road surface defects, and the F1 score for change detection reaches above 0.85.

[0020] Second, this invention constructs a cascaded attention-based semantic segmentation network for water damage, employs a multi-head cross-attention encoder to model the spatial dependency relationship between the changing region and the global context, and combines the dynamic convolution kernel generation mechanism of an adaptive category-aware decoder to achieve fine-grained identification of five types of water damage: roadbed erosion, road surface collapse, slope landslide, retaining wall collapse, and bridge and culvert damage, with an average classification accuracy of over 0.82.

[0021] Third, this invention designs a multi-dimensional damage degree collaborative assessment module that integrates information from three dimensions: morphological features, spectral anomaly index vector, and spatial distribution feature vector. It generates a comprehensive damage degree score through nonlinear mapping of a multi-layer fully connected network, and combines dynamic thresholds of K-means clustering to classify damage levels. The consistency rate between the assessment results and human interpretation reaches over 90%.

[0022] Fourth, this invention designs an adaptive quantification module for traffic impact, which comprehensively considers three factors: spatial occupancy ratio, degree of damage, and road grade to calculate the traffic capacity impact coefficient, realizes automatic determination of the traffic status of damaged road sections, provides a quantitative basis for prioritizing post-disaster repairs, and the processing time for a single remote sensing image does not exceed 30 seconds. Attached Figure Description

[0023] Fig. 1 This is a flowchart of the method for rapid identification of highway flood damage based on remote sensing image analysis according to the present invention;

[0024] Fig. 2 This is an architecture diagram of the rapid identification system for highway flood damage based on remote sensing image analysis according to the present invention. Detailed Implementation

[0025] Please refer to the attached document. Figs. 1-2 The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0026] like Fig. 1 As shown, the rapid identification method for highway flood damage based on remote sensing image analysis provided by this invention includes the following steps:

[0027] Step 1: Remote Sensing Image Acquisition and Registration Preprocessing. After a flood disaster, drones are urgently deployed or satellite remote sensing images are acquired to collect multispectral and visible light aerial images of the affected road sections as post-disaster remote sensing images. At the same time, pre-disaster baseline images of the same area are retrieved from the image archive database.

[0028] Registration preprocessing was performed on pre-disaster archived baseline images and post-disaster remote sensing images. The registration process included three stages: geometric correction, radiometric correction, and spatial registration. Geometric correction used a rational function model to eliminate sensor imaging geometric distortion. Radiometric correction used a relative radiometric normalization method to eliminate radiometric inconsistencies caused by differences in atmospheric conditions and solar altitude angle. Spatial registration employed a feature point matching method based on scale-invariant feature transformation, combined with a random sampling consensus algorithm to eliminate mismatched points, and achieved sub-pixel-level registration accuracy through affine transformation. The registered dual-temporal images have the same spatial resolution, coordinate system, and image range, providing reliable input data for subsequent change detection.

[0029] Step 2: Multi-scale differential twin network change detection. The registered pre-disaster archived reference image and post-disaster remote sensing image are input into the multi-scale differential twin network change detection module 1. This module adopts a twin network architecture with shared weights. The two branch networks have completely identical network structures and parameters, and process the pre-disaster archived reference image and the post-disaster remote sensing image respectively.

[0030] The pyramid feature extractor employs a four-level feature pyramid structure. The first-level feature map has a spatial resolution of 1 / 4 of the input image and 64 feature channels, primarily capturing local texture and edge information. The second-level feature map has a spatial resolution of 1 / 8 of the input image and 128 feature channels, capturing mid-scale structural features. The third-level feature map has a spatial resolution of 1 / 16 of the input image and 256 feature channels, capturing a larger range of contextual information. The fourth-level feature map has a spatial resolution of 1 / 32 of the input image and 512 feature channels, capturing global semantic features. Feature extraction at each level is implemented using residual convolutional blocks, comprising two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function.

[0031] The differences between pre-disaster and post-disaster feature maps are calculated at each scale level. The feature difference map is calculated using a weighted fusion strategy of absolute value difference and cosine distance difference, and the calculation formula is as follows: .in, For the first Feature difference map of the layer The pyramid level sequence number and The values ​​are 1, 2, 3, and 4. The baseline image for pre-disaster archiving was in the first Feature map of the layer For the post-disaster remote sensing images in the first Feature map of the layer The weighting coefficient for the absolute value difference is 0.6. The weighting coefficient for the cosine distance difference is 0.4. This indicates the element-wise absolute value operation. This represents L2 norm operations. Absolute value differencing directly reflects the magnitude of change in feature values ​​and is sensitive to regions of significant change. Cosine distance differencing measures changes in the direction of feature vectors and is sensitive to alterations in semantic information. The fusion of these two differencing methods can comprehensively capture multi-level changes caused by highway flooding disasters.

[0032] Multi-scale feature fusion employs a bidirectional feature pyramid network structure combining bottom-up and top-down approaches. First, 1×1 convolutions unify the feature difference maps of each level to the same number of channels. Then, high-level semantic information is passed down from top to bottom by adding the upsampled low-resolution features to the high-resolution features using bilinear interpolation. Simultaneously, detail information is passed down from bottom to top by fusing it with the low-resolution features after downsampling through convolutions with a stride of 2.

[0033] Gated aggregation employs a channel attention gating mechanism to adaptively weight and aggregate multi-scale feature difference maps. First, global average pooling is performed on the fused features at each level to obtain channel description vectors. Then, importance weights for each level are generated through a two-layer fully connected network. Finally, the weighted multi-scale features are concatenated along the channel dimension and compressed using a 1×1 convolution, then the resulting change region mask is output after passing through a Sigmoid activation function.

[0034] The change region mask is a binary image with the same spatial resolution as the input image, where pixels with a value of 1 indicate a detected change, and pixels with a value of 0 indicate no detected change. To eliminate isolated noise points, a morphological opening operation is performed on the change region mask using a 5×5 elliptical structuring element.

[0035] Step 3: Cascaded Attention-Based Semantic Segmentation of Flood Damage. The changed area mask and the post-disaster remote sensing image are input into the Cascaded Attention-Based Semantic Segmentation Network 2. The goal of this network is to perform fine-grained classification of pixels within the changed area and identify five types of flood damage: roadbed erosion, road surface collapse, slope landslide, retaining wall collapse, and bridge and culvert damage.

[0036] The multi-head cross-attention encoder consists of two parts: an encoding backbone network and a multi-head cross-attention module. The encoding backbone network uses a pre-trained residual network to extract local features of the changed regions, and the spatial resolution of the output feature map is 1 / 16 of the input. The multi-head cross-attention module establishes the spatial dependency between the features of the changed regions and the global features of the post-disaster remote sensing image.

[0037] The calculation process for multi-head cross-attention is as follows: First, the features of the changed region are flattened into a sequence form, and a query matrix Q is generated through linear transformation. Then, the global features of the post-disaster remote sensing image are flattened into a sequence form, and a key matrix K and a value matrix V are generated through linear transformation. The attention calculation formula is: .in, The query matrix is ​​obtained by linear transformation of the features of the changing region. The key matrix is ​​obtained by linear transformation of global features from post-disaster remote sensing images. The value matrix is ​​obtained by linear transformation of the global features of the post-disaster remote sensing image. The dimension of the key vector is 64. The spatial location encoding matrix is ​​used to inject the absolute position information of pixels. Represents the normalized exponential function, This represents the matrix transpose operation.

[0038] The multi-head attention mechanism evenly divides the feature channels into eight heads, each independently calculating attention weights. Different attention heads can focus on different spatial locations and feature patterns, enhancing the model's expressive power. The attention outputs of each head are concatenated and then linearly transformed to restore the original number of channels. The multi-head cross-attention module is stacked in four layers, with residual connections and layer normalization added between each layer.

[0039] The adaptive category-aware decoder employs a dynamic convolutional kernel generation mechanism. First, global average pooling compresses the encoder's output features into category prototype vectors. For each type of water damage, an independent convolutional kernel generation network is designed. This network takes the category prototype vector as input and outputs the corresponding category's segmentation convolutional kernel weights. The dynamically generated convolutional kernels are then depthwise separable convolutionally processed with the encoded features to obtain activation maps for each category. Finally, pixel-wise softmax normalization is used to generate semantic segmentation maps for each water damage type.

[0040] The water damage type semantic segmentation map is a six-channel probabilistic map containing the background and five water damage types. The category with the highest probability is taken as the classification result for each pixel. The classification result is multiplied by the change region mask, retaining only the water damage type label within the change region.

[0041] The types of water damage are defined as follows: roadbed erosion refers to the loss or hollowing of roadbed soil caused by flood erosion; pavement collapse refers to the sinking or collapse of pavement structure layers due to loss of support; slope landslide refers to the sliding or collapse of soil on road cut slopes or embankment slopes; retaining wall collapse refers to the collapse or severe deformation of highway retaining wall structures; bridge and culvert damage refers to the erosion or damage of bridge and culvert structures.

[0042] Step 4: Multi-dimensional damage degree collaborative assessment. For each water-damaged area identified in the semantic segmentation map of water damage type, the multi-dimensional damage degree collaborative assessment module 3 extracts three types of features and performs fusion assessment.

[0043] Morphological feature vector The extraction includes four components. Area of ​​the flood-damaged region. The perimeter of the flood-damaged area is calculated by counting the number of pixels within the damaged area and multiplying it by the square of the spatial resolution. The shape complexity is obtained by detecting the boundary pixels of the water-damaged area and calculating the boundary length. The roundness index is used, and the calculation formula is as follows: The value ranges from 0 to 1; the smaller the value, the more complex the shape. Aspect Ratio This represents the ratio of the longer side to the shorter side of the smallest bounding rectangle of the flood-damaged area. The morphological feature vector is represented as... .

[0044] Spectral Anomaly Index Vector The extraction includes two components: Normalized Difference Vegetation Index Change. The calculation formula is the NDVI of the post-disaster remote sensing image minus the NDVI of the pre-disaster archived reference image, where NDVI equals the near-infrared band minus the red band divided by the near-infrared band plus the red band. This applies when water damage causes vegetation destruction. The value is negative. Normalized Difference (NDD) water index change. The calculation formula is the NDWI of the post-disaster remote sensing image minus the NDWI of the pre-disaster archived reference image, where NDWI equals the green band minus the near-infrared band divided by the green band plus the near-infrared band. (Flood-inundated area) It shows a positive value. The spectral anomaly index vector is represented as... .

[0045] Spatial distribution feature vector The extraction includes two components: the distance from the centroid of the flood-damaged area to the road centerline. The lateral location of the flood damage reflects its location; the closer it is to the road centerline, the greater the impact on traffic. The spatial clustering of the flood-damaged area with adjacent flood-damaged areas. Using kernel density estimation, high clustering indicates that the concentrated distribution of multiple flood-damaged sites may lead to continuous flooding. The spatial distribution feature vector is represented as follows: .

[0046] The comprehensive damage score is calculated using a nonlinear mapping through a multi-layer fully connected network. First, the three types of feature vectors are concatenated into an 8-dimensional input vector. The first fully connected layer maps the 8-dimensional input to a 32-dimensional hidden representation and introduces nonlinearity through the ReLU activation function. The second fully connected layer maps the 32-dimensional hidden representation to 16 dimensions and uses the ReLU activation function. The third fully connected layer maps the 16 dimensions to a 1-dimensional scalar and normalizes it to the range of 0 to 1 using the Sigmoid activation function. The formula for calculating the comprehensive damage score is as follows: .in, A comprehensive damage score is given, with values ​​ranging from 0 to 1. The morphological feature vector includes the area, perimeter, shape complexity, and aspect ratio of the water-damaged area. The spectral anomaly index vector includes the normalized difference vegetation index change and the normalized difference water index change. The spatial distribution feature vector includes the distance from the centroid of the flood-damaged area to the road centerline and the spatial clustering of the flood-damaged area with its adjacent flood-damaged areas. This represents a vector concatenation operation. The first layer weight matrix has a size of 32×8. The second layer weight matrix has a size of 16×32. The third layer weight matrix has a size of 1×16. The first layer bias vector has a length of 32. The second layer bias vector has a length of 16. For the third layer bias scalar, To modify the linear unit activation function, its expression is: , The expression for the Sigmoid activation function is: .

[0047] The damage severity classification employs a dynamic threshold generation strategy based on K-means clustering. Comprehensive damage severity scores for all flood-damaged areas in the current scenario are collected, and the scores are clustered into three clusters using the K-means algorithm. The midpoint between the centers of adjacent clusters is used as the classification threshold. Areas with comprehensive damage severity scores below the first threshold are classified as slightly damaged, those between the two thresholds are classified as moderately damaged, and those above the second threshold are classified as severely damaged. This dynamic threshold strategy can adapt to the differences in damage severity distribution under different disaster scenarios, improving the rationality of the classification.

[0048] Step 5: Adaptive Quantification of Traffic Impact. The Adaptive Quantification of Traffic Impact module 4 comprehensively assesses the impact on traffic capacity based on the spatial geometric characteristics of the flood-damaged area, the severity of the damage, and the importance of the road.

[0049] Space occupancy ratio This reflects the proportion of the road cross-section occupied by the flood-damaged area. First, the road centerline is extracted from the road vector data, generating cross-sectional sampling lines perpendicular to the centerline at 10-meter intervals. For each cross-sectional sampling line, the proportion of the line segment length intersecting the flood-damaged area to the total width of the cross-section is calculated. The maximum proportion of all cross-sectional sampling lines within the flood-damaged area is taken as the spatial occupancy proportion of that flood-damaged area. The value ranges from 0 to 1, with a larger value indicating more severe encroachment of water damage on the road cross section.

[0050] Normalized values ​​of damage level Discrete damage levels are converted into continuous numerical values. Slight damage corresponds to a value of 0.3, indicating minor damage with limited impact on traffic. Moderate damage corresponds to a value of 0.6, indicating significant damage requiring speed reduction or vehicle type restrictions. Severe damage corresponds to a value of 1.0, indicating severe damage that may lead to traffic disruption.

[0051] Normalized weights of road grade attributes This reflects the varying importance of roads of different grades. Expressways, serving as long-distance rapid transport routes, have a weight of 1.0. National highways, serving as inter-regional trunk lines, have a weight of 0.8. Provincial highways, serving as intra-provincial connections, have a weight of 0.6. County roads, serving as intra-county passageways, have a weight of 0.4. Road grade information is obtained from the highway infrastructure database.

[0052] Traffic capacity impact coefficient The weighted summation method is used for calculation, and the formula is as follows: .

[0053] in, This is the traffic capacity impact coefficient. This represents the spatial proportion of the flood-damaged area relative to the road cross-section, with a value ranging from 0 to 1. These are normalized values ​​for the degree of damage, where 0.3 is for mild, 0.6 for moderate, and 1.0 for severe. The normalized weights for road grade attributes are as follows: expressway = 1.0, national highway = 0.8, provincial highway = 0.6, and county road = 0.4. , , These are weighting coefficients with values ​​of 0.5, 0.3, and 0.2, respectively. The spatial occupancy ratio has the highest weight, reflecting the decisive role of physical traffic conditions; the degree of damage is next, reflecting safety risk considerations; and the road grade has the lowest weight, serving as an adjustment factor.

[0054] The traffic status is determined using a segmented threshold rule. When the traffic capacity influence coefficient... When the value is less than 0.2, the traffic status is determined to be normal, indicating that the flood damage is minor and does not affect normal traffic. When the value is greater than or equal to 0.2 and less than 0.5, the traffic status is determined to be half-width traffic, indicating that one lane is damaged but the opposite lane is passable. When the value is greater than or equal to 0.5 and less than 0.8, the traffic status is determined to be single-lane traffic, meaning that only one lane is kept open for alternating traffic. When When the value is greater than or equal to 0.8, the traffic status is determined to be completely interrupted, indicating that the road is impassable and needs to be closed.

[0055] Step Six: Results Output and Report Generation. Based on the above analysis results, the system generates a flood damage distribution map and a repair priority ranking report.

[0056] The flood damage distribution map is presented in the form of a geographic information system layer and includes the following information: the highway network base map shows the road centerline, road name and road grade; flood-damaged areas are marked in polygon form, and different types of flood damage are distinguished by different colors; the degree of damage is represented by the shade of color, with darker colors indicating more severe damage; and the traffic status is marked by icons or text on the corresponding road sections.

[0057] The emergency repair priority ranking report lists information on each flood-damaged location in descending order of priority. The overall priority score is calculated by multiplying the traffic capacity impact coefficient by the road traffic flow. Completely interrupted locations on highways and national roads have the highest priority and require immediate repair. Flood-damaged locations at road network nodes have a higher priority than ordinary road sections. The estimated repair workload is based on the area of ​​the flood-damaged region and the degree of damage, providing a reference for resource allocation.

[0058] The invention establishes a deep coupling relationship at the parameter level and the state level among its four core modules, forming a complete processing pipeline.

[0059] The output change region mask of the multi-scale differential twin network change detection module 1 is used as the spatial constraint input of the cascaded attention-damaged semantic segmentation network 2 to limit the effective region of semantic segmentation and avoid misclassification in regions without change.

[0060] The semantic segmentation map of water damage type output by the cascaded attention water damage semantic segmentation network 2 provides the basis for region division for the multi-dimensional damage degree collaborative assessment module 3. Different water damage types are evaluated using differentiated feature weight combinations.

[0061] The damage level output by the multi-dimensional damage level collaborative assessment module 3 is directly used as the input parameter of the traffic impact adaptive quantification module 4 to participate in the calculation of the traffic capacity impact coefficient.

[0062] The evaluation results of the traffic impact adaptive quantification module 4 can be used to optimize the sensitivity threshold of change detection. When high-grade roads suffer severe water damage, the sensitivity of the change detection module for that area is increased to ensure that no critical damage information is missed.

[0063] The synergistic effect of the four modules is reflected in the following aspects: change detection provides precise spatial constraints for semantic segmentation, achieving complementary advantages; semantic segmentation provides category priors for severity assessment, achieving synergistic effects; severity assessment provides severity indicators for traffic quantification, achieving mutual promotion; and traffic quantification feedback optimizes the detection strategy, achieving adaptive adjustment. The overall technical solution achieves a non-linear effect gain of 1+1>2.

[0064] like Fig. 2 As shown, the present invention also provides a rapid identification system for highway flood damage based on remote sensing image analysis, comprising:

[0065] The multi-scale differential twin network change detection module 1 receives post-disaster remote sensing images and pre-disaster archived baseline images. It extracts multi-scale feature pyramids from both the pre-disaster archived baseline images and the post-disaster remote sensing images using a shared-weight pyramid feature extractor. Feature difference maps are calculated at each scale level, and a change region mask is output after multi-scale feature fusion and gated aggregation. This module is deployed on a GPU server and uses a deep learning inference engine for accelerated computation.

[0066] The Cascaded Attention-Based Water Damage Semantic Segmentation Network 2 receives a mask of the changed region and post-disaster remote sensing imagery. It extracts contextual enhancement features of the changed region through a multi-head cross-attention encoder, and generates a water damage type semantic segmentation map via an adaptive category-aware decoder. This module shares GPU computing resources with the change detection module and employs a pipelined parallel strategy to improve processing efficiency.

[0067] The multi-dimensional damage assessment module 3 extracts morphological feature vectors, spectral anomaly index vectors, and spatial distribution feature vectors for each flood-damaged area. It then calculates a comprehensive damage score and classifies the damage level through feature cascade fusion and nonlinear mapping. This module uses a CPU for feature calculation and a GPU for neural network inference.

[0068] The Traffic Impact Adaptive Quantification Module 4 is used to calculate the traffic capacity impact coefficient based on the spatial occupancy ratio of the flood-damaged area, the degree of damage, and the road grade attributes. It determines the traffic status of damaged road sections and generates a flood damage distribution map and a repair priority ranking report. This module integrates geographic information system functions and supports spatial analysis and visualization output.

[0069] The system also includes auxiliary components such as remote sensing image access interface, highway basic database interface, result visualization interface and report export module, forming a complete application system.

[0070] To verify the effectiveness of the technical solution of the present invention, experimental tests were conducted in multiple real-world water damage disaster scenarios.

[0071] The test data includes orthophotos taken by drones after a rainstorm and flood disaster in a province in July 2024, as well as pre-disaster archived satellite images of the same area. The spatial resolution of the images is 0.5 meters, covering approximately 200 kilometers of roads. 156 locations of water damage were manually marked, including 42 instances of roadbed erosion, 35 instances of road surface collapse, 38 instances of slope landslides, 25 instances of retaining wall collapse, and 16 instances of bridge and culvert damage.

[0072] Performance evaluation of change detection: The multi-scale differential Siamese network change detection module achieved an accuracy of 0.91, a recall of 0.87, and an F1 score of 0.89, which are significantly better than the F1 score of 0.72 of the traditional differential thresholding method and the F1 score of 0.81 of the single-scale Siamese network method.

[0073] Semantic segmentation performance evaluation: The cascaded attention-based water-damaged semantic segmentation network has an overall classification accuracy of 0.86, an average intersection-union ratio of 0.78, and an average accuracy of 0.84 for each category, which is better than the baseline model that does not use the cross-attention mechanism.

[0074] Damage assessment and verification: Compared with the on-site interpretation results of professional technicians, the consistency rate of the damage level assessment was 92%, of which the consistency rate was 95% for minor damage, 90% for moderate damage, and 91% for severe damage.

[0075] Traffic impact assessment verification: The accuracy rate of the traffic status assessment results compared with the actual road closure measures was 94%, with the accuracy rate of the complete traffic interruption assessment reaching 98%.

[0076] Processing efficiency assessment: On a server equipped with an NVIDIA RTX 4090 GPU, the entire process of processing a 5000×5000 pixel remote sensing image takes approximately 25 seconds, which meets the timeliness requirements for rapid emergency response after a disaster.

[0077] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A rapid identification method for highway flood damage based on remote sensing image analysis, characterized in that, Includes the following steps: The system acquires post-disaster remote sensing images and pre-disaster archived baseline images of the affected road sections. After performing registration preprocessing on the dual-temporal images, the images are input into the multi-scale differential twin network change detection module. In the multi-scale differential twin network change detection module, a pyramid feature extractor with shared weights extracts multi-scale feature pyramids from the pre-disaster archived baseline images and the post-disaster remote sensing images. Feature difference maps are calculated at each scale level, and a mask of the changed area is obtained through multi-scale feature fusion and gated aggregation. The changed area mask and the post-disaster remote sensing image are input into the cascaded attention-based water damage semantic segmentation network. The context enhancement features of the changed area are extracted by the multi-head cross-attention encoder, and the water damage type semantic segmentation map is generated by the adaptive category-aware decoder to identify five types of water damage: roadbed washout, road surface collapse, slope landslide, retaining wall collapse and bridge and culvert damage. For each flood-damaged area, the multi-dimensional damage degree collaborative assessment module extracts morphological feature vectors, spectral anomaly index vectors, and spatial distribution feature vectors. Through feature cascade fusion and nonlinear mapping, a comprehensive damage degree score is calculated, and the damage degree is divided into three levels: mild, moderate, and severe based on an adaptive threshold. The traffic impact adaptive quantification module calculates the traffic capacity impact coefficient based on the spatial occupancy ratio of the flood-damaged area relative to the road cross-section, the degree of damage, and the road grade attribute. It determines the traffic status of the damaged road section as normal traffic, half-width traffic, single-lane traffic, or complete interruption, and generates a flood damage distribution map and a repair priority ranking report. The spatial occupancy ratio is obtained by extracting the road centerline from the road vector data, generating cross-sectional sampling lines perpendicular to the road centerline, and calculating the proportion of the length of the line segment intersecting the flood-damaged area of ​​each cross-sectional sampling line to the total width of the cross-section. In the multi-dimensional damage degree collaborative assessment module, the morphological feature vector includes the area, perimeter, shape complexity, and aspect ratio of the water-damaged area; the spectral anomaly index vector includes the normalized difference vegetation index change and the normalized difference water index change; and the spatial distribution feature vector includes the distance from the centroid of the water-damaged area to the road centerline and the spatial clustering degree between the water-damaged area and adjacent water-damaged areas.

2. The method for rapid identification of highway flood damage based on remote sensing image analysis according to claim 1, characterized in that, In the multi-scale differential twin network change detection module, the pyramid feature extractor adopts a four-level feature pyramid structure, with the spatial resolution of the feature maps at each level being 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the input image, respectively. The calculation of the feature difference map adopts a weighted fusion strategy of absolute value difference and cosine distance difference. The gating aggregation adopts a channel attention gating mechanism to adaptively weight and aggregate the multi-scale feature difference map.

3. The method for rapid identification of highway flood damage based on remote sensing image analysis according to claim 2, characterized in that, In the calculation of the feature difference map, the weight coefficient of the absolute value difference is 0.6, and the weight coefficient of the cosine distance difference is 0.

4.

4. The method for rapid identification of highway flood damage based on remote sensing image analysis according to claim 1, characterized in that, In the cascaded attention-based water damage semantic segmentation network, the query matrix of the multi-head cross-attention encoder is obtained by linear transformation of the change region features, and the key matrix and value matrix are obtained by linear transformation of the global features of the post-disaster remote sensing image. The dimension of the key vector is 64, and the spatial location encoding matrix is ​​added to the attention calculation.

5. The method for rapid identification of highway flood damage based on remote sensing image analysis according to claim 1, characterized in that, The adaptive category-aware decoder employs a dynamic convolution kernel generation mechanism, which adaptively generates segmentation convolution kernel weights for each category based on the input features, enabling fine-grained identification of five types of water damage: roadbed erosion, road surface collapse, slope landslide, retaining wall collapse, and bridge and culvert damage.

6. The method for rapid identification of highway flood damage based on remote sensing image analysis according to claim 1, characterized in that, The adaptive threshold for classifying the degree of damage adopts a dynamic threshold generation strategy based on K-means clustering, which divides the comprehensive degree of damage score into three levels: mild damage, moderate damage, and severe damage.

7. The method for rapid identification of highway flood damage based on remote sensing image analysis according to claim 1, characterized in that, In the adaptive quantification module for traffic impact, the weight coefficient for the spatial occupancy ratio in the calculation of the traffic capacity impact coefficient is 0.5, the weight coefficient for the normalized value of the damage level is 0.3, and the weight coefficient for the normalized weight of the road grade attribute is 0.

2. Among the normalized values ​​of damage severity levels, the values ​​are 0.3 for mild, 0.6 for moderate, and 1.0 for severe. Among the normalized weights of road grade attributes, the values ​​are 1.0 for expressways, 0.8 for national highways, 0.6 for provincial highways, and 0.4 for county roads.

8. The method for rapid identification of highway flood damage based on remote sensing image analysis according to claim 1, characterized in that, The rules for determining the traffic status are as follows: when the traffic capacity influence coefficient is less than 0.2, it is determined to be normal traffic; when the traffic capacity influence coefficient is greater than or equal to 0.2 and less than 0.5, it is determined to be half-width traffic; when the traffic capacity influence coefficient is greater than or equal to 0.5 and less than 0.8, it is determined to be single-lane traffic; and when the traffic capacity influence coefficient is greater than or equal to 0.8, it is determined to be completely interrupted.

9. A rapid identification system for highway flood damage based on remote sensing image analysis, used to implement the rapid identification method for highway flood damage based on remote sensing image analysis as described in any one of claims 1-8, characterized in that, include: The multi-scale differential twin network change detection module is used to receive post-disaster remote sensing images and pre-disaster archived baseline images. It extracts multi-scale feature pyramids from the pre-disaster archived baseline images and post-disaster remote sensing images through a pyramid feature extractor with shared weights. It calculates feature difference maps at each scale level and outputs a change area mask after multi-scale feature fusion and gated aggregation. A cascaded attention-based water damage semantic segmentation network is used to receive a mask of the changed area and post-disaster remote sensing images. The contextual enhancement features of the changed area are extracted by a multi-head cross-attention encoder and then generated by an adaptive category-aware decoder to produce a water damage type semantic segmentation map. The multi-dimensional damage assessment module is used to extract morphological feature vectors, spectral anomaly index vectors, and spatial distribution feature vectors for each water-damaged area. It calculates a comprehensive damage score and classifies the damage level by feature cascade fusion and nonlinear mapping. The traffic impact adaptive quantification module is used to calculate the traffic capacity impact coefficient based on the spatial occupancy ratio of the flood-damaged area, the degree of damage, and the road grade attributes, determine the traffic status of the damaged road sections, and generate a flood disaster distribution map and a repair priority ranking report.

Images