Deep learning-based hazard-affected body damage assessment method and device

Abstract

The present disclosure relates to a kind of deep learning-based hazard-bearing body damage assessment method and device.Wherein, the method comprises: inputting post-disaster remote sensing image into pre-trained semantic segmentation model, to obtain the preliminary segmentation mask output by semantic segmentation model;The remote sensing image corresponding to preliminary segmentation mask is converted to the color space of pre-set, and the spectral feature component in the color space of the pixel region of at least one target category is extracted according to preliminary segmentation mask;The pixel corresponding to the spectral feature component that does not meet the spectral threshold condition is reclassified, to obtain the segmentation mask after optimization;Pixel-level statistical analysis is carried out on the segmentation mask after optimization, to obtain the pixel number and pixel proportion of each category hazard-bearing body;Based on the pixel number and proportion, assessment is carried out, to obtain hazard-bearing body damage assessment result.The present disclosure improves the reliability of disaster assessment.

Description

Technical Field

[0001] This disclosure relates to the field of disaster assessment technology, and in particular to a method and apparatus for assessing damage to disaster-bearing bodies based on deep learning. Background Technology

[0002] In related technologies, damage assessment of disaster-bearing structures is a crucial step in natural disaster emergency response, aiming to identify and quantify the extent of damage to critical disaster-bearing structures such as buildings, roads, and bridges. Traditional manual on-site investigations are inefficient and cannot meet the needs of large-scale emergency response. With the development of UAVs and deep learning technologies, intelligent assessment based on high-resolution remote sensing imagery is gradually becoming mainstream. Current traditional methods use deep convolutional neural networks to perform semantic segmentation of images to achieve automatic identification and classification of disaster-bearing structures. However, existing methods show a significant decrease in segmentation accuracy when faced with a large amount of debris obscuring the imagery or drastic changes in lighting; moreover, the identification results lack an effective post-verification mechanism, which is prone to misjudgment, such as misidentifying gray concrete pavement as water, thus affecting the reliability of disaster assessment. Summary of the Invention

[0003] To overcome the problems existing in related technologies, this disclosure provides a method and apparatus for assessing the damage of disaster-bearing bodies based on deep learning.

[0004] According to a first aspect of the present disclosure, a method for assessing damage to disaster-bearing structures based on deep learning is provided, comprising:

[0005] Acquire post-disaster remote sensing images, input the post-disaster remote sensing images into a pre-trained semantic segmentation model, and obtain a preliminary segmentation mask output by the semantic segmentation model; the preliminary segmentation mask includes multiple disaster-bearing body categories; The remote sensing image corresponding to the preliminary segmentation mask is converted to a preset color space, and the spectral feature components of at least one target category pixel region in the color space are extracted according to the preliminary segmentation mask. Pixels corresponding to spectral feature components that do not meet the spectral threshold conditions are reclassified to obtain an optimized segmentation mask. Pixel-level statistical analysis was performed on the optimized segmentation mask to obtain the number of pixels and the percentage of pixels for each type of disaster-bearing body; Based on the preset disaster level assessment rules, a comprehensive assessment of multiple indicators is conducted based on the number and proportion of pixels to obtain the disaster-bearing body damage assessment result.

[0006] According to a second aspect of the present disclosure, a disaster-bearing body damage assessment device based on deep learning is provided, comprising: The acquisition unit is used to acquire post-disaster remote sensing images, input the post-disaster remote sensing images into a pre-trained semantic segmentation model, and obtain a preliminary segmentation mask output by the semantic segmentation model; the preliminary segmentation mask includes multiple disaster-bearing body categories. The extraction unit is used to convert the remote sensing image corresponding to the preliminary segmentation mask to a preset color space, and extract the spectral feature components of at least one target category pixel region in the color space according to the preliminary segmentation mask. An optimization unit is used to reclassify pixels corresponding to spectral feature components that do not meet the spectral threshold conditions to obtain an optimized segmentation mask. The analysis unit is used to perform pixel-level statistical analysis on the optimized segmentation mask to obtain the number of pixels and the percentage of pixels for each type of disaster-bearing body. The assessment unit is used to conduct a comprehensive assessment of multiple indicators based on the number and proportion of pixels according to the preset disaster level assessment rules, and obtain the disaster-bearing body damage assessment result.

[0007] According to a third aspect of the present disclosure, an electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described in any one of the first aspects.

[0008] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the method as described in any one of the first aspects.

[0009] According to a fifth aspect of the present disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the method as described in any one of the first aspects.

[0010] The technical solutions provided by the embodiments of this disclosure can include the following beneficial effects: by acquiring post-disaster remote sensing images and inputting them into a pre-trained semantic segmentation model, a preliminary segmentation mask including multiple disaster-bearing body categories is obtained, realizing automatic identification of disaster-bearing body categories; by converting remote sensing images to a preset color space and extracting the spectral feature components of the target categories, a verification mechanism based on the spectral characteristics of ground objects is constructed; by reclassifying pixels that do not meet the spectral threshold conditions to obtain an optimized segmentation mask, the misjudgment caused by fragment occlusion or illumination changes in the semantic segmentation model is corrected, significantly improving the accuracy and robustness of the segmentation results; by performing pixel-level statistical analysis on the optimized segmentation mask, the number of pixels and the proportion of pixels in each category are obtained, providing a quantitative basis for disaster assessment; by performing multi-indicator comprehensive evaluation of the number and proportion of pixels based on preset evaluation rules, the damage assessment results are obtained, realizing the transformation from pixel-level data to decision-level information, and automatically outputting graded assessment results, providing scientific and objective decision support for emergency response.

[0011] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0013] Figure 1 This is a flowchart illustrating a disaster-bearing body damage assessment method based on deep learning, according to an exemplary embodiment.

[0014] Figure 2 This is a flowchart of the disaster-bearing body damage assessment method proposed in the embodiments of this disclosure.

[0015] Figure 3 This is a block diagram illustrating a disaster-bearing body damage assessment device based on deep learning, according to an exemplary embodiment.

[0016] Figure 4 This is a block diagram illustrating an apparatus for a deep learning-based disaster damage assessment method according to an exemplary embodiment.

[0017] Figure Labels 301 - Acquisition unit; 302 - Extraction unit; 303 - Optimization unit; 304 - Analysis unit; 305 - Evaluation unit; 400 - Device; 402 - Processing component; 404 - Memory; 406 - Power component; 408 - Multimedia component; 410 - Audio component; 412 - I / O interface; 414 - Sensor component; 416 - Communication component; 420 - Processor. Detailed Implementation

[0018] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0019] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the present disclosure. The singular forms “a” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise.

[0020] It should be understood that although the terms first, second, third, etc., may be used to describe various information in embodiments of this disclosure, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of embodiments of this disclosure, and similarly, second information may also be referred to as first information. Depending on the context, the words “if” and “suppose” as used herein may be interpreted as “when”, “when”, or “in response to a determination”.

[0021] Furthermore, various forms of processes shown in the embodiments of this disclosure can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and no limitation is imposed herein.

[0022] It should be noted that the collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in the technical solution disclosed herein all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0023] In related technologies, disaster-bearing body damage assessment is the process of identifying, classifying, and quantifying the degree of damage to key disaster-bearing bodies such as buildings, roads, and bridges under the influence of natural disasters such as earthquakes, floods, and typhoons. It widely serves emergency response, disaster relief, and loss assessment. With the evolution of UAV and deep learning technologies, intelligent disaster-bearing body damage assessment based on UAV remote sensing imagery is rapidly developing. Compared with manual on-site surveys, UAV remote sensing imagery is characterized by multi-view, high resolution, and rapid acquisition, and can provide large-scale disaster imagery support in the immediate aftermath of a disaster.

[0024] Current methods for assessing damage to disaster-bearing bodies are mainly divided into three categories: methods based on manual interpretation, methods based on traditional image processing, and methods based on deep learning. Methods based on manual interpretation rely on subjective assessments by experts based on remote sensing images or on-site photographs and their experience. While highly accurate, these methods are inefficient, costly, and difficult to support large-scale data processing. Methods based on traditional image processing extract features such as texture, color, and shape from remote sensing images and identify them according to preset rules or thresholds; however, their adaptability and generalization ability to complex scenes are limited.

[0025] To address the aforementioned issues, this disclosure provides a deep learning-based disaster-bearing body damage assessment method. By acquiring post-disaster remote sensing images and inputting them into a pre-trained semantic segmentation model, a preliminary segmentation mask encompassing multiple disaster-bearing body categories is obtained, enabling automatic identification of these categories. A verification mechanism based on the spectral characteristics of ground features is constructed by converting the remote sensing images to a preset color space and extracting the spectral feature components of the target categories. An optimized segmentation mask is obtained by reclassifying pixels that do not meet spectral threshold conditions, correcting misjudgments caused by debris occlusion or illumination changes in the semantic segmentation model, significantly improving the accuracy and robustness of the segmentation results. Pixel-level statistical analysis of the optimized segmentation mask yields the number and proportion of pixels in each category, providing a quantitative basis for disaster assessment. Finally, a comprehensive multi-indicator assessment of the number and proportion of pixels based on preset assessment rules is performed to obtain the damage assessment results, realizing the transformation from pixel-level data to decision-level information. This method can automatically output graded assessment results, providing scientific and objective decision support for emergency response.

[0026] Figure 1 This is a flowchart illustrating a deep learning-based disaster-bearing body damage assessment method according to an exemplary embodiment, such as... Figure 1 As shown, it should be noted that the deep learning-based disaster-bearing body damage assessment method of this disclosure is applied in a deep learning-based disaster-bearing body damage assessment device. For example... Figure 1 As shown, the method may include the following steps: Step 101: Acquire post-disaster remote sensing images and input them into a pre-trained semantic segmentation model to obtain the preliminary segmentation mask output by the semantic segmentation model.

[0027] The initial segmentation mask includes multiple disaster-bearing categories.

[0028] In this embodiment, high-resolution remote sensing images collected by UAVs after a disaster can be acquired. These images can be stored in BGR format and include information on typical disaster-bearing entities under the influence of natural disasters such as earthquakes, floods, and typhoons, such as buildings, roads, bridges, and water bodies. The acquired remote sensing images are input into a pre-trained semantic segmentation model, and the feature extraction capabilities of the deep learning model are used to classify each pixel in the image, thereby outputting a preliminary segmentation mask.

[0029] As an example, the preliminary segmentation mask mentioned above can be a two-dimensional matrix, where the value of each element represents the disaster-bearing category identifier of that pixel, such as 0 for background, 1 for water area, and 2 for undamaged building.

[0030] It is understandable that the disaster-bearing body category can include land cover type, as well as damage level information. For example, buildings can be further divided into subcategories such as undamaged, slightly damaged, severely damaged, and completely destroyed.

[0031] It should be noted that the automatic identification of disaster-bearing body categories through semantic segmentation models replaces the traditional manual interpretation method, significantly improving the efficiency and scalability of disaster assessment and laying the foundation for subsequent rapid assessment of large-scale disasters.

[0032] In some embodiments of this disclosure, the semantic segmentation model is the LRFormer model, and step 101 may specifically include the following sub-steps: Step a1: Input the post-disaster remote sensing image into the LRFormer (Low-Resolution Self-Attention) model so that the LRFormer model can downsample the input features to a fixed size for global context modeling through the low-resolution self-attention mechanism.

[0033] As an example, post-disaster remote sensing imagery is input into the encoder of the LRFormer model. This model uses LRSA (Low-Resolution Self-Attention) to downsample the input high-resolution feature map to a fixed low-resolution size. By reducing the spatial resolution of the feature map, the self-attention mechanism can model global contextual information with controllable computational complexity, capturing the dependencies between distant pixels in the image.

[0034] For example, in building damage assessment, global context can help the model understand the overall layout of the building complex and its relationship with the surrounding environment.

[0035] Understandably, the low-resolution self-attention mechanism significantly reduces computational overhead while ensuring the model's receptive field, enabling the model to achieve efficient inference on computing resources such as GPUs, thus meeting the speed requirements of emergency response scenarios.

[0036] Step a2 uses depthwise convolution to capture high-resolution spatial details.

[0037] In one embodiment, in the LRFormer model, 3×3 depthwise convolutions are used in parallel with the low-resolution self-attention mechanism to process the original high-resolution features, in order to preserve and capture spatial detail information in the image. The depthwise convolutions are performed independently on each input channel, effectively extracting local features such as edges, textures, and corners.

[0038] For post-disaster remote sensing imagery, detailed information such as the damaged edges of buildings, cracks in roads, and the distribution of debris is crucial for accurately determining the level of damage. High-resolution spatial details captured through deep convolution complement global contextual information captured by low-resolution self-attention mechanisms, together constructing a complete feature representation.

[0039] Understandably, this parallel design balances global context modeling with preservation of local details, improving the model's adaptability to complex disaster scenarios.

[0040] Step a3: Extract multi-level features with step sizes of 4, 8, 16 and 32, and use the encoder and decoder to perform multi-scale feature aggregation on the multi-level features to obtain the preliminary segmentation mask output by the LRFormer model, which includes multiple disaster-bearing body categories and their damage levels.

[0041] Specifically, the LRFormer model extracts four levels of features with step sizes of 4, 8, 16 and 32 through progressive downsampling in the encoder stage, denoted as F1, F2, F3 and F4.

[0042] Shallow features (such as F1 and F2) have high spatial resolution and include rich detail information; deep features (such as F3 and F4) have a large receptive field and include rich semantic information. In the decoder stage, the model can use progressive upsampling and feature fusion to gradually transfer deep semantic information to shallow layers, while simultaneously performing multi-scale aggregation of features from different levels.

[0043] Through the encoder-decoder structure, the model can comprehensively utilize the advantages of multi-level features to generate a segmentation mask with the same resolution as the input image. Each pixel in this initial segmentation mask is assigned a category label, which includes not only the type of land cover, but also the specific damage level of buildings (such as minor damage, severe damage, and complete destruction) and the traffic status of roads (such as unobstructed and blocked).

[0044] As an example, the model outputs 11 semantic categories, including background, water, undamaged buildings, slightly damaged buildings, severely damaged buildings, completely destroyed buildings, unobstructed roads, blocked roads, vehicles, trees, and pools.

[0045] It should be noted that multi-scale feature aggregation enables the model to simultaneously identify large-area continuous land features (such as water bodies and building complexes) and small-sized discrete targets (such as vehicles and trees), improving the comprehensiveness and accuracy of the segmentation results.

[0046] Step 102: Convert the remote sensing image corresponding to the preliminary segmentation mask to a preset color space, and extract the spectral feature components of at least one target category pixel region in the color space based on the preliminary segmentation mask.

[0047] In this embodiment of the disclosure, after obtaining the initial segmentation mask, a post-processing mechanism based on color space is introduced to further verify and correct the accuracy of the segmentation results.

[0048] As an example of a possible implementation, the original remote sensing image corresponding to the preliminary segmentation mask can be converted from the BGR color space to a preset color space in order to extract the spectral feature components of the ground features. Based on the pixel category information recorded in the preliminary segmentation mask, the target category (e.g., water area) that needs to be spectrally verified is determined, and the spectral feature components of the pixel area of ​​that category are extracted in the specified color space.

[0049] It should be noted that spectral feature components reflect the physical characteristics of ground features in a specific color space. Different ground features (such as water bodies and gray road surfaces) often show significant differences on the same spectral feature component. This difference can be used to perform secondary verification of semantic segmentation results.

[0050] In one embodiment, the spectral characteristic component may be a saturation component.

[0051] In some embodiments of this disclosure, the preset color space is the HSV (Hue, Saturation, Value) color space, and step 102 may specifically include the following sub-steps: Step b1: For each target category in at least one target category, determine the position of the pixels of the target category in the preliminary segmentation mask based on the preliminary segmentation mask.

[0052] As an example, the target category requiring spectral verification can be selected from the initial segmentation mask. In disaster assessment scenarios, water bodies are a key category that is easily confused with other features (especially gray concrete pavements), so water bodies can be used as the target category. For the target category of water bodies, the initial segmentation mask is traversed to find all pixel locations with a value of the water body category identifier (e.g., 1), and the row and column coordinates of these pixels are recorded.

[0053] It is understood that this disclosure is not limited to verification only of water areas. Depending on actual needs, other easily confused categories (such as damaged buildings and rubble piles) can also be set as target categories for spectral verification. By determining the location of the target category pixels, a spatial index is provided for subsequent spectral feature extraction.

[0054] Step b2: Extract the saturation component of the pixel corresponding to the location from the remote sensing image converted to HSV color space to obtain the spectral feature component of the pixel region of the target category in the color space.

[0055] Specifically, the original remote sensing image can be converted from the BGR color space to the HSV color space. The HSV color space consists of three components: hue, saturation, and value. The saturation component reflects the purity or vividness of the color. Real water bodies (such as rivers and stagnant water) typically exhibit high saturation (generally greater than 50), while the saturation of gray concrete pavement is close to 0, showing a significant difference in the saturation component between the two.

[0056] Based on the target category pixel location determined in step b1, the saturation value of the corresponding location is extracted from the converted HSV image to obtain the set of spectral feature components of the target category pixel region.

[0057] As an example, for the tens of thousands of pixels that have been initially segmented into water areas, their saturation values ​​will be extracted one by one to form a spectral feature dataset to be verified.

[0058] It should be noted that using saturation components as the verification basis utilizes the inherent physical spectral characteristics of ground objects, without relying on the feature learning results of deep learning models, and has good generalization ability and robustness.

[0059] In some embodiments of this disclosure, a saturation verification and correction method based on the HSV color space is provided. First, the original RGB image and its corresponding preliminary segmentation mask are obtained as input. The RGB image is then converted to the HSV color space using the `cv2.cvtColor()` function of OpenCV (Open Source Computer Vision Library). Subsequently, pixel regions initially identified as water bodies are extracted based on the water body category identifier in the segmentation mask, and the saturation values ​​of these pixels are obtained from the saturation channels of the HSV image. The core judgment logic of the scheme is based on the spectral difference in saturation between water bodies and gray ground features: real water bodies typically have high saturation (usually greater than 50), while the saturation of easily confused ground features such as gray concrete pavement approaches zero. Based on this, a saturation threshold of 20 is set, and pixels with a saturation below 20 are judged as gray misclassified pixels, and their category in the segmentation mask is re-labeled as background. After correction, an optimized segmentation mask is output. Experimental data show that the verification method is effective: before verification, there were 235,680 pixels in the water area, with a misclassification rate as high as 60%-80%; after verification, the number of pixels in the water area decreased to 8,520, and the misclassification rate decreased to 0-10%, effectively eliminating the interference of ground features such as gray road surfaces on water area recognition and significantly improving the accuracy of segmentation results.

[0060] Step 103: Reclassify the pixels corresponding to the spectral feature components that do not meet the spectral threshold conditions to obtain the optimized segmentation mask.

[0061] In this embodiment of the disclosure, after extracting the spectral feature components of the target category, pixels are filtered and reclassified according to preset spectral threshold conditions.

[0062] As an example, the spectral feature components of each pixel are compared with the corresponding threshold condition. If the threshold condition is not met, the pixel is determined to be a misclassification by the semantic segmentation model and is removed from the original category and reclassified to another category (such as background). After reclassification, the category label of the corresponding pixel in the initial segmentation mask is updated to obtain the optimized segmentation mask.

[0063] It should be noted that post-processing correction of the segmentation results using spectral threshold conditions can effectively eliminate misjudged pixels caused by factors such as changes in illumination, shadows, and debris occlusion, significantly improving the accuracy and reliability of the segmentation results.

[0064] In some embodiments of this disclosure, step 103 may specifically include the following steps: Step c1: For each target category in at least one target category, determine whether the spectral feature component corresponding to the target category meets the spectral threshold condition that matches the target category.

[0065] Specifically, for each target category that requires spectral verification, a matching spectral threshold condition is pre-set.

[0066] Taking the water area category as an example, based on the distribution characteristics of saturation in the HSV color space, the saturation threshold condition is set to "saturation below 20". For the saturation value of each water area pixel extracted in step b2, it is judged one by one: if the saturation value is ≥20, the pixel is considered to meet the spectral threshold condition and is judged as a correctly identified water area pixel; if the saturation value is <20, the pixel is considered to not meet the spectral threshold condition and is judged as a misjudged pixel.

[0067] It is understandable that different target categories can have different spectral threshold conditions and different spectral feature components. For example, for vegetation categories, threshold conditions can be set based on hue components; for shaded areas, threshold conditions can be set based on lightness components. Those skilled in the art can flexibly set corresponding spectral verification rules according to the spectral characteristics of specific land features.

[0068] Step c2: Remove pixels whose spectral feature components do not meet the corresponding spectral threshold conditions from the current target category and reclassify them to the preset category.

[0069] As an example, pixels with a saturation level below 20 are removed from the water category, and the category identifier of these pixels in the segmentation mask is modified to another preset category, typically the background category (category identifier 0).

[0070] It should be noted that reclassifying misclassified pixels to the background category is the most direct approach. However, in some embodiments, misclassified pixels can also be reclassified to other more relevant disaster-bearing body categories based on spectral characteristics. For example, gray road surface pixels misclassified as water can be reclassified to the road category. This reclassification operation not only eliminates misclassifications but may also supplement pixel information from other categories.

[0071] Step c3: Based on the results of the reclassification, update the initial segmentation mask to obtain the optimized segmentation mask.

[0072] Specifically, based on the modification of the category labels of misclassified pixels in step c2, the initial segmentation mask is updated as a whole. The category labels of all pixels that have passed spectral verification and reclassification are modified to the new values; while the category labels of other pixels that have not been verified or meet the threshold conditions remain unchanged.

[0073] After the update, an optimized segmentation mask is obtained. Compared to the initial segmentation mask, this optimized mask has higher accuracy in identifying target categories. For example, gray road surfaces that were initially misidentified as water areas are successfully removed, making the identification of water areas more realistic and reliable. Understandably, the optimized segmentation mask serves as the basis for subsequent statistical analysis and disaster assessment, and its quality directly affects the reliability of the final assessment results.

[0074] Step 104: Perform pixel-level statistical analysis on the optimized segmentation mask to obtain the number of pixels and pixel percentage of each type of disaster-bearing body.

[0075] In this embodiment of the disclosure, after obtaining the optimized segmentation mask, pixel-level statistical analysis is performed on it to convert the segmentation results into quantifiable statistical data.

[0076] As an example, the number of pixels appearing in the mask for each type of disaster-bearing body is counted, and the proportion of each type of pixel in the mask is calculated to obtain the pixel percentage of each type. These statistics objectively reflect the coverage and distribution of various types of disaster-bearing bodies in remote sensing imagery, providing a quantitative basis for subsequent disaster level assessment.

[0077] It should be noted that pixel ratio is an important indicator for measuring the area covered by ground features. Given the spatial resolution of the image, it can be further converted into the actual area (such as square meters), providing more intuitive data support for disaster loss assessment.

[0078] In some embodiments of this disclosure, step 104 may specifically include the following steps: Step d1: Count the number of pixels for each disaster-bearing category in the optimized segmentation mask.

[0079] Specifically, the optimized segmentation mask can be traversed to count the category identifier of each pixel.

[0080] As an example of a possible implementation, the number of pixels can be counted in 11 categories, including background, water area, undamaged buildings, slightly damaged buildings, severely damaged buildings, completely destroyed buildings, unobstructed roads, blocked roads, vehicles, trees, and pools. These pixel counts reflect the coverage scale of each type of disaster-bearing entity in the image. Understandably, for categories that include damage levels (such as buildings), the pixel count can also be used to assess the relative size of the damage extent.

[0081] Step d2: Calculate the total number of pixels in the optimized segmentation mask.

[0082] As an example of a possible implementation, the dimensions of the optimized segmentation mask are obtained, namely the height H and width W, and the total number of pixels is calculated as H×W. This total number of pixels represents the pixel scale of the entire remote sensing image and is the denominator for calculating the pixel percentage. As an example, for a 1920×1080 resolution image, the total number of pixels is 2,073,600.

[0083] Step d3: For each disaster-bearing body category, divide the number of pixels in the disaster-bearing body category by the total number of pixels to obtain the pixel percentage of the disaster-bearing body category.

[0084] Specifically, for each disaster-bearing category, its pixel count is divided by the total pixel count to obtain the pixel percentage of that category in the image.

[0085] As an example, if the number of pixels representing water areas is 622,080 and the total number of pixels is 2,073,600, then the water area percentage is 30%. These pixel percentages eliminate the influence of image size differences, making the assessment results of images at different scales comparable, and also facilitating direct comparison with preset threshold conditions. Understandably, pixel percentage is a core input parameter for subsequent disaster level assessments; for example, a water area percentage exceeding 30% may trigger a flood warning.

[0086] Step 105: Based on the preset disaster level assessment rules, a comprehensive assessment of multiple indicators is conducted based on the number and proportion of pixels to obtain the disaster-bearing body damage assessment results.

[0087] In this embodiment of the disclosure, after obtaining the number of pixels and the percentage of pixels of various disaster-bearing bodies, a comprehensive evaluation of multiple indicators is performed based on preset disaster level assessment rules.

[0088] As an example, the assessment rules can include multiple assessment indicators and their corresponding threshold ranges. Each indicator corresponds to a key disaster-bearing body category (such as blocked roads, damaged bridges, destroyed buildings, water bodies, etc.). The indicator value of each category can be determined based on the pixel proportion of each category. Then, the indicator value is compared with the preset threshold range to comprehensively determine the current disaster level (such as normal, warning, critical), and generate a damage assessment result that includes the disaster level and warning information.

[0089] It should be noted that multi-indicator comprehensive assessment can reflect the severity of a disaster from multiple dimensions, avoiding the one-sided judgment that may be caused by a single indicator, and providing more comprehensive information support for emergency decision-making.

[0090] In some embodiments of this disclosure, step 105 may specifically include the following steps: Step e1: Obtain multiple preset evaluation indicators and their corresponding threshold ranges.

[0091] The assessment indicators include at least one of the following: the percentage of blocked roads, the percentage of damaged bridges, the percentage of destroyed buildings, and the percentage of water area.

[0092] As an example, the assessment rules are shown in Table 1, including four assessment indicators: the percentage of blocked roads, the percentage of damaged bridges, the percentage of destroyed buildings, and the percentage of water area. Each indicator corresponds to a set of threshold ranges to define three levels: normal, warning, and critical. For example, for the percentage of destroyed buildings, the threshold for the critical level is >15%, the threshold for the warning level is 5%~15%, and the threshold for the normal level is ≤5%. It is understandable that the selection of assessment indicators and the setting of threshold ranges can be flexibly adjusted according to the type of disaster (e.g., earthquakes focus on buildings, floods focus on water areas) and regional characteristics.

[0093] Table 1 Examples of Disaster Level Assessment Rules

[0094] Step e2: Determine the index value of each evaluation index based on the number of pixels and the percentage of pixels of each type of disaster-bearing body.

[0095] In one embodiment, based on the number of pixels and the percentage of pixels obtained in step 104, category data corresponding to each evaluation index is extracted.

[0096] As an example, the percentage of blocked roads is taken from the pixel percentage of the "blocked roads" category in the segmentation mask; the percentage of damaged bridges is taken from the pixel percentage of the "damaged bridges" category in the segmentation mask (note that bridges may be included in the road category or be a separate category); the percentage of destroyed buildings is taken from the pixel percentage of the "completely destroyed buildings" category in the segmentation mask; and the percentage of water area is taken from the pixel percentage of the "water area" category in the segmentation mask (after merging the pool category). These percentage values ​​are used as the indicator values ​​for each evaluation metric in subsequent threshold comparisons.

[0097] Step e3: Compare each indicator value with its corresponding threshold range to obtain the comparison result, determine the disaster level based on the comparison result, and generate the disaster-bearing body damage assessment result based on the disaster level.

[0098] Specifically, the values ​​of each indicator determined in step e2 are compared one by one with the threshold ranges obtained in step e1.

[0099] As an example, for a case where 18% of buildings are destroyed, this is compared with the critical level threshold of ">15%", and it is found to meet the critical condition; for a case where 10% of the water area is destroyed, this is compared with the warning level threshold of "2%~30%", and it is found to meet the warning condition. Based on the comprehensive comparison results of multiple indicators, the final disaster level is determined according to the preset judgment logic.

[0100] Understandably, a comprehensive assessment of multiple indicators can be made using a priority principle (such as determining a situation as critical if any indicator reaches a critical level) or a weighted scoring method.

[0101] In some embodiments of this disclosure, a damage assessment result, including a level identifier, Chinese description, a list of warning information, and detailed statistical data, can be generated based on the determined disaster level and output to the user interface or stored as a report file.

[0102] In some embodiments of this disclosure, step e3 may specifically include the following sub-steps: Step e31: According to the preset priority of the assessment indicators from high to low, each indicator value is compared with its corresponding threshold range to obtain the comparison results, and the disaster level is determined based on the comparison results.

[0103] The disaster levels include critical level, early warning level, and normal level.

[0104] As an example, indicators such as the percentage of blocked roads, damaged bridges, destroyed buildings, and water area can be assigned different priorities according to their ability to represent the severity of the disaster.

[0105] As another example, we can first determine if any indicators meet the criteria for a critical level. If so, it is directly classified as a critical level. If no critical indicators exist, we continue to determine if any indicators meet the criteria for a warning level. If so, it is classified as a warning level. If all indicators are within the normal threshold range, it is classified as a normal level. This hierarchical judgment logic ensures that the most severe disaster situations can be identified and responded to first.

[0106] It should be noted that this priority determination rule conforms to the basic principle of "safety first, prevention foremost" in emergency response and can promptly capture high-risk disaster signals.

[0107] Step e32: Generate the damage assessment results of the disaster-bearing body based on the disaster level.

[0108] Specifically, based on the disaster level determined in step e31, a complete assessment of the damage to affected structures is generated. As an example, the assessment results may include the following information: disaster level identifiers (e.g., red "Critical", yellow "Warning", green "Normal"), a Chinese description of the level, a list of specific warning information (e.g., "Severe damage detected, 18% of buildings destroyed", "Large areas of water detected, potential flooding, 10% of water area"), and precise percentage statistics for each land cover type. This assessment result can be used to generate a visual report for emergency command personnel to review. It is understood that the output format of the assessment results can be flexibly adjusted according to the application scenario, providing a quantitative basis for emergency decision-making based on the disaster level conclusion derived from a comprehensive assessment of multiple indicators.

[0109] In this embodiment, the method may further include performing a temporal-dimensional statistical evaluation of the video data. For the input video stream, a frame interval sampling strategy is used to extract keyframes, for example, processing once every 5 frames by default. Steps 101 to 105 above are performed on each keyframe to obtain the disaster-bearing body damage assessment results for each keyframe. Simultaneously, the total number of video frames, the actual number of processed frames, the processing frame rate, and the total processing time are statistically analyzed to generate a video-dimensional evaluation report. It should be noted that the video processing capability allows this method to be applied not only to single remote sensing images but also to real-time or near-real-time analysis of UAV aerial video, providing technical support for dynamic disaster monitoring.

[0110] In one embodiment of this disclosure, the method may further include a visualization output step. Based on a preset category-color mapping relationship (e.g., cyan for water areas, red for completely destroyed buildings, and yellow-green for blocked roads), the optimized segmentation mask is converted into a color segmentation result image, visually displaying the spatial distribution of various disaster-bearing bodies. Simultaneously, an assessment visualization interface is generated, including a category statistics table, a pie chart of disaster type proportions, and a bar chart of disaster category statistics, facilitating emergency personnel to quickly grasp the disaster overview. Furthermore, an early warning management list is displayed, including original image thumbnails, segmentation result images, identification category label groups, and disaster level indicators, supporting full lifecycle management of early warning records (viewing, modifying, deleting, and exporting). Through visualization output, abstract data is transformed into an intuitive graphical interface, improving the system's usability and decision support capabilities.

[0111] In some embodiments of this disclosure, such as Figure 2As shown, the technical solution disclosed in this paper first processes the post-disaster remote sensing image through a semantic segmentation module. It then employs the PELRA-Net (Pyramid-Enhanced Low-Resolution Attention Network) inference framework for multi-scale feature extraction. This framework combines a low-resolution self-attention mechanism with encoder feature aggregation to output a preliminary segmentation mask. Subsequently, a post-processing correction stage is initiated, including converting the image from the BGR color space to the HSV color space, optimizing category merging for easily confused categories (such as pools and bodies of water), extracting pixels from bodies of water and thresholding and classifying their saturation, and reclassifying misclassified pixels to optimize the segmentation results. After correction, the optimized segmentation mask is used to calculate pixel proportions and perform statistical analysis to obtain quantitative data for various disaster-bearing bodies. Finally, based on the statistical results, a disaster level assessment is performed and an assessment report is output. Simultaneously, a color segmentation result map is generated for visualization, providing an intuitive basis for emergency decision-making. By introducing a spectral feature verification mechanism to post-process and correct the semantic segmentation results, the problem of misjudgment caused by changes in illumination or debris occlusion in a single semantic segmentation model under complex disaster scenarios is effectively solved, and the accuracy of disaster-bearing body identification and the reliability of damage assessment are significantly improved.

[0112] The deep learning-based disaster-bearing body damage assessment method proposed in this disclosure acquires post-disaster remote sensing images and inputs them into a pre-trained semantic segmentation model to obtain a preliminary segmentation mask including multiple disaster-bearing body categories, thus achieving automatic identification of disaster-bearing body categories. A verification mechanism based on the spectral characteristics of ground objects is constructed by converting the remote sensing images to a preset color space and extracting the spectral feature components of the target categories. An optimized segmentation mask is obtained by reclassifying pixels that do not meet the spectral threshold conditions, correcting misjudgments caused by fragment occlusion or illumination changes in the semantic segmentation model, and significantly improving the accuracy and robustness of the segmentation results. Pixel-level statistical analysis of the optimized segmentation mask yields the number and proportion of pixels in each category, providing a quantitative basis for disaster assessment. Damage assessment results are obtained by comprehensively evaluating the number and proportion of pixels based on preset assessment rules using multiple indicators, realizing the transformation from pixel-level data to decision-level information. This method can automatically output graded assessment results, providing scientific and objective decision support for emergency response.

[0113] Figure 3 This is a block diagram illustrating a disaster-bearing body damage assessment device based on deep learning, according to an exemplary embodiment. (Refer to...) Figure 3 The device includes an acquisition unit 301, an extraction unit 302, an optimization unit 303, an analysis unit 304, and an evaluation unit 305.

[0114] Among them, the acquisition unit 301 is used to acquire post-disaster remote sensing images, input the post-disaster remote sensing images into a pre-trained semantic segmentation model, and obtain the preliminary segmentation mask output by the semantic segmentation model; the preliminary segmentation mask includes multiple disaster-bearing body categories. Extraction unit 302 is used to convert the remote sensing image corresponding to the preliminary segmentation mask to a preset color space, and extract the spectral feature components of at least one target category pixel region in the color space according to the preliminary segmentation mask. The optimization unit 303 is used to reclassify the pixels corresponding to spectral feature components that do not meet the spectral threshold conditions to obtain the optimized segmentation mask. Analysis unit 304 is used to perform pixel-level statistical analysis on the optimized segmentation mask to obtain the number of pixels and the percentage of pixels for each type of disaster-bearing body. The assessment unit 305 is used to conduct a comprehensive assessment of multiple indicators based on the number and proportion of pixels according to the preset disaster level assessment rules, and obtain the disaster-bearing body damage assessment result.

[0115] In some embodiments of this disclosure, the semantic segmentation model is the LRFormer model, and the acquisition unit 301 can be specifically used for: The post-disaster remote sensing imagery is input into a pre-trained semantic segmentation model to obtain a preliminary segmentation mask output by the semantic segmentation model, including: The post-disaster remote sensing images are input into the LRFormer model, which uses a low-resolution self-attention mechanism to downsample the input features to a fixed size for global context modeling, uses deep convolution to capture high-resolution spatial details, and extracts multi-level features with strides of 4, 8, 16, and 32. The encoder and decoder are used to perform multi-scale feature aggregation on the multi-level features to obtain the preliminary segmentation mask output by the LRFormer model, which includes multiple disaster-bearing body categories and their damage levels.

[0116] In some embodiments of this disclosure, the preset color space is the HSV color space, and the extraction unit 302 can specifically be used for: For each target category in at least one target category, determine the position of the pixel of the target category in the preliminary segmentation mask based on the preliminary segmentation mask; Saturation components corresponding to the location of pixels are extracted from remote sensing images converted to HSV color space to obtain the spectral feature components of the pixel region of the target category in the color space.

[0117] In some embodiments of this disclosure, the optimization unit 303 may specifically be used for: For each target category in at least one target category, determine whether the spectral feature component corresponding to the target category meets the spectral threshold condition that matches the target category; Pixels whose spectral feature components do not meet the corresponding spectral threshold conditions are removed from the current target category and reclassified to the preset category; Based on the results of the reclassification, the initial segmentation mask is updated to obtain the optimized segmentation mask.

[0118] In some embodiments of this disclosure, the analysis unit 304 may specifically be used for: The number of pixels for each disaster-bearing category in the optimized segmentation mask is counted. Calculate the total number of pixels in the optimized segmentation mask; For each disaster-bearing body category, the number of pixels in the disaster-bearing body category is divided by the total number of pixels to obtain the pixel percentage of the disaster-bearing body category.

[0119] In some embodiments of this disclosure, the evaluation unit 305 may specifically be used for: Obtain multiple preset evaluation indicators and their corresponding threshold ranges; the evaluation indicators include at least one of the following: the percentage of blocked roads, the percentage of damaged bridges, the percentage of destroyed buildings, and the percentage of water area. The value of each evaluation indicator is determined based on the number of pixels and the percentage of pixels in each category of disaster-bearing bodies. Each indicator value is compared with its corresponding threshold range to obtain the comparison results. The disaster level is determined based on the comparison results, and the damage assessment results of the disaster-bearing body are generated based on the disaster level.

[0120] In some embodiments of this disclosure, the evaluation unit 305 may specifically be used for: According to the preset priority of the assessment indicators from high to low, each indicator value is compared with its corresponding threshold range to obtain the comparison results. The disaster level is determined based on the comparison results. The disaster level includes critical level, early warning level and normal level.

[0121] In some embodiments of this disclosure, the apparatus may further include a visualization unit, which may specifically be used for: Based on the preset mapping relationship between pixel categories and colors, the optimized segmentation mask is converted into a color segmentation result image; Generate an assessment visualization interface that includes category statistics tables, disaster type percentage charts, and disaster category statistics charts; The display includes a list of early warning management features such as original image thumbnails, segmentation results, identification category label groups, and disaster level indicators.

[0122] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0123] The deep learning-based disaster-bearing body damage assessment device proposed in this disclosure acquires post-disaster remote sensing images and inputs them into a pre-trained semantic segmentation model to obtain a preliminary segmentation mask including multiple disaster-bearing body categories, thus achieving automatic identification of disaster-bearing body categories. A verification mechanism based on the spectral characteristics of ground objects is constructed by converting the remote sensing images to a preset color space and extracting the spectral feature components of the target categories. An optimized segmentation mask is obtained by reclassifying pixels that do not meet the spectral threshold conditions, correcting misjudgments caused by fragment occlusion or illumination changes in the semantic segmentation model, and significantly improving the accuracy and robustness of the segmentation results. Pixel-level statistical analysis of the optimized segmentation mask yields the number and proportion of pixels in each category, providing a quantitative basis for disaster assessment. Damage assessment results are obtained by comprehensively evaluating the number and proportion of pixels based on preset assessment rules using multiple indicators, realizing the transformation from pixel-level data to decision-level information. This allows for the automatic output of graded assessment results, providing scientific and objective decision support for emergency response.

[0124] Figure 4 This is a block diagram illustrating an apparatus for a deep learning-based disaster damage assessment method according to an exemplary embodiment. For example, apparatus 400 may be an electronic device, such as a mobile phone, computer, digital broadcasting terminal, messaging device, tablet device, personal digital assistant, etc.

[0125] Reference Figure 4 The device 400 may include one or more of the following components: processing component 402, memory 404, power component 406, multimedia component 408, audio component 410, input / output I / O interface 412, sensor component 414, and communication component 416.

[0126] Processing component 402 typically controls the overall operation of device 400, such as operations associated with display, telephone calls, data communication, camera operation, and recording. Processing component 402 may include one or more processors 420 to execute instructions to perform all or part of the steps of the methods described above. Furthermore, processing component 402 may include one or more modules to facilitate interaction between processing component 402 and other components. For example, processing component 402 may include a multimedia module to facilitate interaction between multimedia component 408 and processing component 402.

[0127] Memory 404 is configured to store various types of data to support the operation of device 400. Examples of such data include instructions for any application or method operating on device 400, contact data, phonebook data, messages, pictures, videos, etc. Memory 404 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0128] The power supply component 406 provides power to the various components of the device 400. The power supply component 406 may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power to the device 400.

[0129] Multimedia component 408 includes a screen that provides an output interface between the device 400 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touchscreen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors may sense not only the boundaries of the touch or swipe action but also the duration and pressure associated with the touch or swipe operation. In some embodiments, multimedia component 408 includes a front-facing camera and / or a rear-facing camera. When the device 400 is in an operating mode, such as a shooting mode or a video mode, the front-facing camera and / or the rear-facing camera may receive external multimedia data. Each front-facing camera and rear-facing camera may be a fixed optical lens system or have focal length and optical zoom capabilities.

[0130] Audio component 410 is configured to output and / or input audio signals. For example, audio component 410 includes a microphone (MIC) configured to receive external audio signals when device 400 is in an operating mode, such as call mode, recording mode, and voice recognition mode. The received audio signals may be further stored in memory 404 or transmitted via communication component 416. In some embodiments, audio component 410 also includes a speaker for outputting audio signals.

[0131] I / O interface 412 provides an interface between processing component 402 and peripheral interface modules, such as keyboards, click wheels, buttons, etc. These buttons may include, but are not limited to, home buttons, volume buttons, power buttons, and lock buttons.

[0132] Sensor assembly 414 includes one or more sensors for providing status assessments of various aspects of device 400. For example, sensor assembly 414 may detect the on / off state of device 400, the relative positioning of components such as the display and keypad of device 400, changes in the position of device 400 or a component of device 400, the presence or absence of user contact with device 400, the orientation or acceleration / deceleration of device 400, and temperature changes of device 400. Sensor assembly 414 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. Sensor assembly 414 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, sensor assembly 414 may also include an accelerometer, a gyroscope, a magnetometer, a pressure sensor, or a temperature sensor.

[0133] Communication component 416 is configured to facilitate wired or wireless communication between device 400 and other devices. Device 400 can access wireless networks based on communication standards, such as WiFi, 2G, or 3G, or combinations thereof. In one exemplary embodiment, communication component 416 receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, communication component 416 also includes a near-field communication (NFC) module to facilitate short-range communication. For example, the NFC module may be implemented based on radio frequency identification (RFID) technology, Infrared Data Association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0134] In an exemplary embodiment, the apparatus 400 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the methods described above.

[0135] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 404 including instructions, which can be executed by a processor 420 of the device 400 to perform the above-described method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0136] In an exemplary embodiment, a computer program product is also provided, including a computer program that implements the above-described method when executed by a processor 420 of the device 400.

[0137] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This disclosure is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the appended claims.

[0138] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A disaster-bearing body damage assessment method based on deep learning, characterized in that, include: Acquire post-disaster remote sensing images, input the post-disaster remote sensing images into a pre-trained semantic segmentation model, and obtain a preliminary segmentation mask output by the semantic segmentation model; the preliminary segmentation mask includes multiple disaster-bearing body categories; The remote sensing image corresponding to the preliminary segmentation mask is converted to a preset color space, and the spectral feature components of at least one target category pixel region in the color space are extracted according to the preliminary segmentation mask. Pixels corresponding to spectral feature components that do not meet the spectral threshold conditions are reclassified to obtain the optimized segmentation mask. Pixel-level statistical analysis was performed on the optimized segmentation mask to obtain the number of pixels and the percentage of pixels for each type of disaster-bearing body; Based on the preset disaster level assessment rules, a comprehensive assessment of multiple indicators is conducted based on the number and proportion of pixels to obtain the disaster-bearing body damage assessment result.

2. The disaster-bearing body damage assessment method based on deep learning according to claim 1, characterized in that, The semantic segmentation model is a low-resolution self-attention network LRFormer model. The step of inputting the post-disaster remote sensing image into a pre-trained semantic segmentation model to obtain the preliminary segmentation mask output by the semantic segmentation model includes: The post-disaster remote sensing images are input into the LRFormer model, which uses a low-resolution self-attention mechanism to downsample the input features to a fixed size for global context modeling, uses deep convolution to capture high-resolution spatial details, and extracts multi-level features with strides of 4, 8, 16, and 32. The encoder and decoder are used to perform multi-scale feature aggregation on the multi-level features to obtain a preliminary segmentation mask output by the LRFormer model, which includes multiple disaster-bearing body categories and their damage levels.

3. The disaster-bearing body damage assessment method based on deep learning according to claim 1, characterized in that, The preset color space is the HSV color space; The step of extracting the spectral feature components of at least one target category pixel region in the color space based on the preliminary segmentation mask includes: For each target category in at least one target category, the position of the pixel of the target category in the preliminary segmentation mask is determined according to the preliminary segmentation mask; The saturation component of the pixel corresponding to the location is extracted from the remote sensing image converted to the HSV color space to obtain the spectral feature component of the pixel region of the target category in the color space.

4. The disaster-bearing body damage assessment method based on deep learning according to claim 1, characterized in that, The step of reclassifying pixels corresponding to spectral feature components that do not meet the spectral threshold condition to obtain an optimized segmentation mask includes: For each of the at least one target categories, determine whether the spectral feature component corresponding to the target category meets the spectral threshold condition that matches the target category; Pixels whose spectral feature components do not meet the corresponding spectral threshold conditions are removed from the current target category and reclassified to the preset category; Based on the results of the reclassification, the initial segmentation mask is updated to obtain the optimized segmentation mask.

5. The disaster-bearing body damage assessment method based on deep learning according to claim 1, characterized in that, The optimized segmentation mask is subjected to pixel-level statistical analysis to obtain the number of pixels and pixel proportion of each type of disaster-bearing body, including: The number of pixels for each disaster-bearing category in the optimized segmentation mask is counted. Calculate the total number of pixels in the optimized segmentation mask; For each disaster-bearing body category, the number of pixels in that category is divided by the total number of pixels to obtain the pixel percentage of that disaster-bearing body category.

6. The disaster-bearing body damage assessment method based on deep learning according to claim 1, characterized in that, The process of conducting a comprehensive multi-indicator assessment based on the number and proportion of pixels according to preset disaster level assessment rules to obtain the disaster-bearing body damage assessment result includes: Obtain multiple preset evaluation indicators and their corresponding threshold ranges; the evaluation indicators include at least one of the following: percentage of blocked roads, percentage of damaged bridges, percentage of destroyed buildings, and percentage of water area. The value of each evaluation indicator is determined based on the number of pixels and the percentage of pixels in each category of disaster-bearing bodies. Each indicator value is compared with its corresponding threshold range to obtain a comparison result. The disaster level is determined based on the comparison result, and the damage assessment result of the disaster-bearing body is generated based on the disaster level.

7. The disaster-bearing body damage assessment method based on deep learning according to claim 6, characterized in that, The step of comparing each indicator value with its corresponding threshold range to obtain a comparison result, and determining the disaster level based on the comparison result, includes: According to the preset priority of the evaluation indicators from high to low, each indicator value is compared with its corresponding threshold range to obtain the comparison result. The disaster level is determined based on the comparison result. The disaster level includes critical level, early warning level and normal level.

8. The disaster-bearing body damage assessment method based on deep learning according to claim 1, characterized in that, Also includes: Based on the preset mapping relationship between pixel category and color, the optimized segmentation mask is converted into a color segmentation result image; Generate an assessment visualization interface that includes category statistics tables, disaster type percentage charts, and disaster category statistics charts; The display includes a list of early warning management features such as original image thumbnails, segmentation results, identification category label groups, and disaster level indicators.

9. A disaster-bearing body damage assessment device based on deep learning, characterized in that, include: The acquisition unit is used to acquire post-disaster remote sensing images, input the post-disaster remote sensing images into a pre-trained semantic segmentation model, and obtain a preliminary segmentation mask output by the semantic segmentation model; the preliminary segmentation mask includes multiple disaster-bearing body categories. The extraction unit is used to convert the remote sensing image corresponding to the preliminary segmentation mask to a preset color space, and extract the spectral feature components of at least one target category pixel region in the color space according to the preliminary segmentation mask. The optimization unit is used to reclassify the pixels corresponding to spectral feature components that do not meet the spectral threshold conditions to obtain the optimized segmentation mask. The analysis unit is used to perform pixel-level statistical analysis on the optimized segmentation mask to obtain the number of pixels and the percentage of pixels for each type of disaster-bearing body. The assessment unit is used to conduct a comprehensive assessment of multiple indicators based on the number and proportion of pixels according to the preset disaster level assessment rules, and obtain the disaster-bearing body damage assessment result.

10. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as described in any one of claims 1 to 8.

Images