A vulnerability assessment method for marine hazard-bearing bodies based on SETR and geographic grid

Abstract

The present invention provides a method for evaluating the vulnerability of marine hazard-bearing bodies based on SETR and geographic grids, which belongs to the field of intelligent ocean data processing technology. The method comprises the following steps: establishing a sample library, constructing a semantic segmentation model of remote sensing images of marine hazard-bearing bodies based on a SETR network, constructing a grid spatiotemporal data set using geographic gridding technology, establishing a grid and spatial value density hazard-bearing body vulnerability assessment method, calculating a marine hazard-bearing body vulnerability index, and realizing the vulnerability assessment of marine hazard-bearing bodies in a region. In view of the difficulties in investigating and extracting massive marine hazard-bearing bodies and the lack of intelligent early warning and monitoring technical means, a marine hazard-bearing body vulnerability assessment model based on SETR and geographic grids is constructed using deep learning, geographic gridding, spatial value density, disaster science and other technologies, so as to realize intelligent and rapid evaluation of marine hazard-bearing bodies in a large range of scenarios, which is of great significance to the intelligent monitoring and early warning of disasters of marine hazard-bearing bodies under typhoon disasters and the refined management of disaster prevention and mitigation.

Description

Technical Field

[0001] The present invention relates to the technical field of smart ocean data processing, and in particular to a method for evaluating the vulnerability of marine disaster-prone bodies based on SETR and geographic grids. Background Art

[0002] In recent years, global marine disasters have occurred frequently, such as hurricanes, tsunamis, ocean acidification and sea level rise, which have had a significant impact on coastal areas. The background of the research on marine hazard-bearing body vulnerability assessment is to gain a deeper understanding of the vulnerability of the marine system to disasters, and to provide a scientific basis to support relevant decision-making and reduce the harm of marine disasters to human society. By evaluating the vulnerability of marine hazard-bearing bodies, we can gain a deeper understanding of the sensitivity and resistance of the natural characteristics, ecological functions and human activities of the marine system to marine disasters. This will help to take relevant measures before the disaster occurs to reduce the risk of disasters. The basis of the research on marine hazard-bearing body vulnerability assessment can provide scientific analysis and prediction capabilities and provide decision-making support for disaster prevention and mitigation. Continuous technological progress and research innovation will continue to promote the development of marine hazard-bearing body vulnerability assessment.

[0003] Due to the wide variety of marine hazard-prone objects, including houses, roads, street lights, communication facilities, seawalls, marine aquaculture facilities, coastal towns, etc., it is impossible to fully understand their characteristic data; and the vulnerability assessment of marine hazard-prone objects involves a wide range of data and a large amount of data, such as meteorology, hydrology, basic geography, humanities and economics, transportation, remote sensing and other massive spatiotemporal data. Traditional technical means are difficult to achieve fast and effective data organization and data analysis and processing. At present, it is difficult for marine management departments to evaluate and analyze marine hazard-prone objects and to carry out refined management of marine disaster prevention and mitigation. Therefore, it is necessary to design a vulnerability assessment method for marine hazard-prone objects based on SETR and geographic grids. Summary of the invention

[0004] The purpose of the present invention is to provide a method for evaluating the vulnerability of marine hazard-prone bodies based on SETR and geographic grids, so as to solve the technical problems of the existing difficulties in investigating and extracting massive marine hazard-prone bodies and the lack of intelligent early warning and monitoring technical means.

[0005] Deep learning is a powerful machine learning technology that can extract the key features of marine hazard-prone bodies and establish an accurate marine hazard-prone body extraction model by learning relevant data of marine hazard-prone bodies. Using deep learning algorithms, large-scale and rapid extraction of marine hazard-prone body targets from massive remote sensing data can be achieved, improving the efficiency and accuracy of assessment. Geographic grid is a method of dividing geographic information into grid cells of equal size. The vulnerability assessment method based on geographic grid can make comprehensive use of multi-source data, including ocean, meteorological, geographical, and human factors, and integrate the multi-dimensional information of marine hazard-prone bodies into the assessment model through deep learning algorithms to improve the comprehensiveness and comprehensiveness of the assessment. The geographic grid can be adjusted according to different assessment scales. The method based on deep learning and geographic grid can perform fine-grained assessments as needed, such as assessing the vulnerability of marine hazard-prone bodies in local areas, or conducting overall assessments in large areas. Such flexibility enables the assessment method to adapt to assessment needs at different scales.

[0006] Based on the results of vulnerability assessment, the disaster risks faced by the marine system are assessed, including disaster probability, potential losses, etc. Multi-criteria decision analysis, risk management framework and other methods are used to provide decision makers with suggestions on risk management and disaster response. The vulnerability assessment of marine hazard-bearing bodies involves knowledge from multiple disciplines and requires collaboration and comprehensive analysis by interdisciplinary experts. By integrating data, models and methods from different disciplines, the vulnerability of the marine system can be more comprehensively understood and assessed.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0008] A method for evaluating the vulnerability of marine hazard-bearing bodies based on SETR and geographic grids, the method comprising the following steps:

[0009] S1: Establish a sample library of semantic segmentation of marine hazard-prone bodies, build a semantic segmentation model of marine hazard-prone remote sensing images based on the SETR network, extract and classify marine hazard-prone bodies in remote sensing images, and calculate the geometric attributes of marine hazard-prone bodies through the statistical pixel method;

[0010] S2: Divide the space-time grid and construct the grid space-time data set. Use the geographic gridding technology to perform space-time correlation on the massive space-time data and organize it into a data set with grids as units.

[0011] S3: Use oblique photogrammetry technology to model the regional geographic environment, classify and extract geographic element information, establish a grid and spatial value density disaster-prone body vulnerability assessment method, build a marine disaster-prone body vulnerability geographic calculation model, and realize the quantitative calculation of the physical vulnerability of natural disasters;

[0012] S4: Calculate the vulnerability index of marine hazard-prone bodies and evaluate the vulnerability of marine hazard-prone bodies in the region based on the vulnerability index.

[0013] Furthermore, the specific process of constructing the semantic segmentation model of marine hazard-prone bodies for remote sensing images in step 1 is as follows:

[0014] (1) Serialize the input image into a one-dimensional vector, divide the image into image blocks of equal size, encode the spatial information of each image block to form the final sequence input, and adjust each image block to the form of a one-dimensional vector;

[0015] (2) Input the serialized one-dimensional vector into the transformer encoder to extract features. The transformer encoder consists of several transformer modules, each of which consists of several self-attention modules, a normalization layer, and several perceptron layers;

[0016] (3) The output of the transformer encoder is passed to the decoder, which is also composed of several transformer modules. The decoder combines global context information with local details by using self-attention and cross-attention mechanisms, maps the encoded vector back to the image block, performs pixel-level classification, and generates semantic segmentation prediction results.

[0017] Furthermore, in step 2, when the massive spatiotemporal data are temporally and spatially associated, geographic grid codes are generated. The specific process is as follows:

[0018] (1) Obtain the latitude, longitude, altitude and time nodes related to the data;

[0019] (2) Express the latitude and longitude in the form of whole degrees, whole minutes and seconds × 2048, and then convert it into a binary number; perform a bit-by-bit crossover operation to form a binary one-dimensional code, which constitutes the spatial code in the marine information grid code;

[0020] (3) Time nodes are encoded in the form of year, month, and day;

[0021] (4) Combining spatial coding and temporal coding to form an ontology code;

[0022] (5) The verification code is generated from the main code to uniformly form the sea area information grid code.

[0023] Furthermore, the specific process of constructing the geographic calculation model of marine hazard-bearing body vulnerability in step 3 is as follows:

[0024] (1) Construct a rating index system. Based on the geographical conditions of the demonstration area and the degree of exposure to marine disasters, select disaster-prone elements to establish a spatial value density evaluation index system, and determine the weight of each factor through expert scoring and normalization method;

[0025] (2) Establish a quantitative calculation model for the vulnerability of disaster-prone bodies. In the quantitative calculation model, the demonstration area is geographically divided into 10m grids. The demonstration area is divided according to the divided grids, and the value index and vulnerability index of each grid are calculated respectively, and the product is calculated:

[0026] V=E×D (1)

[0027] In the formula: V is the vulnerability index, E is the value index, and D is the vulnerability index;

[0028] (3) Value index calculation: The value Q of each land feature is calculated in each grid. The calculated results are normalized and then weighted summed. The calculation formula is as follows:

[0029]

[0030] In the formula, E is the value density, Q is the value of a certain land feature in a square kilometer, W is the land feature weight, i represents the i-th land feature, and n represents a total of n land features;

[0031] (4) Calculation of vulnerability index: The vulnerability index of an assessment unit is determined by two factors: the disaster risk level L of the assessment unit and the probability of damage to the ground objects P. L expresses the magnitude of the disaster risk of the assessment unit caused by the difference in ground elevation, and P is the statistical value of the damage probability of various ground objects in the entire assessment unit. The calculation formula is as follows:

[0032] D=L×P (3)

[0033] Where D is the vulnerability index, L is the disaster risk level, and P is the probability of ground damage, ranging from 0 to 1.

[0034] Furthermore, the calculation process of the vulnerability index of the marine hazard-bearing body in step 4 is:

[0035] (1) Based on the value index and vulnerability index of a single grid, calculate the vulnerability index of the marine hazard-bearing body in grid units;

[0036] (2) Summarize the results of all grids to form a vulnerability assessment of the marine hazard-bearing bodies in the entire region.

[0037] Furthermore, the calculation process of the vulnerability index of the marine hazard-bearing body in step 4 is:

[0038] (1) Based on the value index and vulnerability index of a single grid, calculate the vulnerability index of the marine hazard-bearing body in grid units;

[0039] (2) Summarize the results of all grids to form a vulnerability assessment of the marine hazard-bearing bodies in the entire region.

[0040] 6. The method for evaluating the vulnerability of marine hazard-bearing bodies based on SETR and geographic grid according to claim 2 is characterized in that the specific process of serializing the input image is:

[0041] Divide the image evenly The image blocks are , where H is the image height and W is the image width. Each image block is then flattened into a one-dimensional vector sequence, and the vectorized image block p is mapped to C dimension using linear mapping, that is, C is the size of the hidden channel, which is the size of the classification. Each pixel is positionally encoded to obtain p i , and then with the vector e i Add them together to get the final input sequence E = {e1+p1, e2+p2, ..., e L +p L}, the sequence length L is The final dimension that needs to be input to the transformer encoder is e1...e L Represents the vector corresponding to each image block, p1...p L Indicates the position information encoding corresponding to each image block.

[0042] Furthermore, the decoder performs pixel-level segmentation on the image, extracts a feature every 6 layers from the encoder, and then transfers the feature from The feature vector of The feature maps are processed by a 1×1, 3×3 and 3×3 three-layer convolutional network, where the first and third layers reduce the number of channels to half of the original, and the output Then use bilinear interpolation to upsample by 4 times to get Four The features of are then fused from top to bottom, and the top-level features and the three fused features are concatenated to obtain The size of the image is , an additional 3×3 convolution is applied after adding the function element by element, and finally a 4-fold upsampling and convolution is performed to obtain the final H×W×C feature map. The category probability of the pixel point is normalized using the softmax function to obtain the final semantic segmentation result.

[0043] Furthermore, the specific process of dividing the space-time grid in step S2 is:

[0044] The grid division space of the earth surface space adopts the coordinate space of longitude and latitude. In order to ensure the division of whole degrees, whole minutes and whole seconds of longitude and latitude, the earth space is expanded into the longitude and latitude space of integer powers of 2, which is defined as 512°×512°×512° grid. The 60′ space of each degree is expanded to 64′, and the 60″ space of each minute is expanded to 64″. It is divided into 32 levels step by step according to the octave method, and the minimum expression accuracy can be accurate to 1.5 cm. The part within the range of 180°×360° in the expanded longitude and latitude coordinate space is consistent with the actual geographic space, and the part beyond the range of 180°×360° has no actual geographical significance. When the elevation is fixed to the earth's surface height, the land and sea space-time information grid is used as the earth's plane grid. When the elevation value is not needed, the plane grid can be used.

[0045] Furthermore, the calculation process of ground object damage probability is:

[0046] The calculation method of the probability of damage to the ground objects in the evaluation unit is to determine the types of ground objects, ground object parameters and the area ratio of each type of ground objects in the unit cell network, and perform weighted summation. The calculation formula is as follows. After calculation, the result is normalized so that the value range of the ground object damage probability P is 0-1;

[0047]

[0048] In the formula, P is the probability of damage to the object in the evaluation unit, F is the probability of damage to a single object, and S is the area percentage of a single object in the evaluation unit. Since the evaluation unit is a kilometer grid, the area of ​​each grid is 1 km. 2 The actual area of ​​each feature is the percentage of its area in the evaluation unit.

[0049] Furthermore, the specific process of disaster risk level classification is as follows:

[0050] The calculation of disaster risk level takes kilometer grid as unit, calculates the average ground elevation within each kilometer grid, and divides it into 10 levels. The lower the elevation, the greater the disaster risk. According to historical marine disaster statistics, areas with elevations exceeding 180 meters will not be affected by marine disasters. The calculation of disaster risk level of land features takes the spatial object of land features as unit, and requires the superposition of several spatial layers for spatial analysis and calculation to determine the disaster risk level formed by the spatial location distribution.

[0051] The present invention has the following beneficial effects due to the adoption of the above technical solution:

[0052] (1) In terms of disaster-prone body extraction, the present invention realizes large-scale and rapid extraction of marine disaster-prone body data through remote sensing data and SETR network, which improves efficiency and saves costs compared with traditional manual delineation methods;

[0053] (2) In terms of marine hazard-prone body assessment, the correlation analysis of multi-source heterogeneous data is realized through the global geographic grid. The same spatiotemporal attributes of data of different formats and categories in the same grid are used to fully explore the spatiotemporal correlation and intrinsic value of various hazard factor data, reduce the limitations of knowledge on the vulnerability of hazard-prone bodies of different types, and improve the scientificity and reliability of the assessment results.

[0054] (3) In practice, deep learning models can provide a scientific basis for the formulation of emergency measures, which is conducive to guiding the practice of emergency management, minimizing the harm of disaster events, and improving the effectiveness and scientific nature of emergency management. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a flow chart of the marine disaster-bearing body evaluation method of the present invention;

[0056] Figure 2 Schematic diagram of the transformer-based encoder structure;

[0057] Figure 3 Schematic diagram of multi-layer feature fusion decoding;

[0058] Figure 4 Calculate the schematic diagram for the floor;

[0059] Figure 5 This is a schematic diagram for calculating the number of lanes on a traffic road;

[0060] Figure 6 It is a schematic diagram of the spatial and temporal geographic grid division of land and sea;

[0061] Figure 7 It is a diagram of the spatial-temporal grid coding structure of the entire land and sea area;

[0062] Figure 8 Coding effect map for coastline grid;

[0063] Fig. 9 This is the effect diagram of the island grid coding;

[0064] Fig.10 Schematic diagram of vulnerability assessment of marine hazard-prone bodies based on geographic grid. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and preferred embodiments. However, it should be noted that many details listed in the specification are only for the purpose of enabling the reader to have a thorough understanding of one or more aspects of the present invention, and these aspects of the present invention can be implemented even without these specific details.

[0066] like Figure 1As shown in FIG. 1 , a vulnerability assessment method for marine hazard-bearing bodies based on SETR and geographic grids includes the following steps:

[0067] S1. Establish a semantic segmentation sample library, set SETR network parameters, build a semantic segmentation model for remote sensing images of marine hazard-prone bodies, and realize the extraction, classification and geometric attribute calculation of marine hazard-prone bodies in remote sensing images. The specific implementation is as follows:

[0068] (1) Image serialization processing. Divide the image into The image blocks are then flattened into a one-dimensional vector sequence, and the vectorized image blocks p are mapped to C dimensions using linear mapping, that is, C is the size of the hidden channel, that is, the size of the classification; each pixel is positionally encoded to obtain p i , and then with the vector e i Add them together to get the final input sequence E = [e1+p1, e2+p2, ..., e L +p L}, the sequence length L is The final dimension that needs to be input to the transformer encoder is e1...e L Represents the vector corresponding to each image block, p1...p L Indicates the position information encoding corresponding to each image block.

[0069] (2) Transformer encoding. The one-dimensional sequence E is used as input, and the transformer is used as an encoder to learn features. Each transformer layer has a global receptive field. Each transformer layer consists of a multi-head self-attention (MSA), layer normalization (LN), and a multi-layer perceptron (MLP), for a total of 24 layers. Figure 2 shown.

[0070] The input to the self-attention module is a set of tuples (query, key, value), the input is The query, key, and value are calculated as follows: query = Z l-1 W Q , key = Z l-1 W K ,value=Z l-1 W V ,in is a learnable parameter and d is the dimension of (query, key, value).

[0071] Z l-1 Represents the features of the transformer layer l-1, and l represents the transformer layer l.

[0072] The self-attention module SA will do the following processing:

[0073]

[0074] SA() represents the operation of the self-attention module, Z l-1 Represents the features of the l-1th layer of the transformer. Softmax is a normalized exponential function commonly used in deep learning. is a learnable parameter and d is the dimension of (query, key, value).

[0075] The multi-head self-attention module MSA is an extension with m independent self-attention operations and projects their concatenated outputs:

[0076] MSA(Z l-1 )=[SA1(Z l-1 );SA2(Z l-1 );...;SA m (Z l-1 )]W O (5)

[0077] SA1...SA m Represents m self-attention module operations, and MSA() represents multi-head self-attention operations, which mainly realize the encoding of image features.

[0078] in d is usually set to

[0079] The output of MSA is transformed through a residual skip MLP block, and the layer output is:

[0080]

[0081] MLP means that the multi-layer perceptron performs feature convolution encoding on the image. The entire function completes the encoding of the image features by the l-th layer transformer.

[0082] Z l Represents the features of the lth layer of transformer.

[0083] Finally, the output of each layer of the Transformer encoder is formed {Z1 , Z 2 , ..., Z L}.

[0084] (3) Design a decoder to perform pixel-level segmentation on the image. Extract a feature from the encoder every 6 layers, and then extract the feature from The feature vector of The feature maps are processed by a three-layer convolutional network (1×1, 3×3, 3×3), where the first and third layers reduce the number of channels to half of the original, and the output Then use bilinear interpolation to upsample by 4 times to get These four The features of are then fused from top to bottom, and the top-level features and the three fused features are concatenated to obtain After adding features element by element, an additional 3×3 convolution is applied, and finally a 4-fold upsampling and convolution is performed to obtain the final H×W×C feature map. For the category probability of the pixel point, the softmax function is used for normalization to obtain the final semantic segmentation result. Figure 3 , Figure 4 shown.

[0085] (4) Based on the results of image semantic segmentation, the statistical pixel method is used to calculate the geometric properties of the marine hazard-prone body, such as the length, area, and circumference of the hazard-prone body.

[0086] S2. Divide the space-time grid and construct the grid data set. The land and sea space-time information grid takes the intersection of the prime meridian and the reference ellipsoid equatorial plane as the origin, and the longitude and latitude coordinate system and the earth reference ellipsoid follow the provisions of GJB 6304-2008.

[0087] (1) Spatial grid division. The grid division space of the earth surface space adopts the coordinate space of longitude and latitude. In order to ensure the division of whole degrees, whole minutes and whole seconds of longitude and latitude, the earth space is expanded into the longitude and latitude space of integer powers of 2, defined as 512°×512°×512° grid. The 60′ space of each degree is expanded to 64′, and the 60″ space of each minute is expanded to 64″. It is divided into 32 levels step by step according to the octave method (Table 1), and the minimum expression accuracy can be accurate to 1.5 cm. The part within the range of 180°×360° in the expanded longitude and latitude coordinate space is consistent with the actual geographic space, and the part beyond the range of 180°×360° has no actual geographical meaning. When the elevation is fixed to the height of the earth's surface, the Guangxi land and sea space-time information grid is the earth's plane grid. When the elevation value is not required, this plane grid can be used.

[0088] Table 1 Statistics of grid scales at each level

[0089]

[0090]

[0091] (2) Grid code. The global land-sea space-time grid code should be a characteristic combination code and consist of 33 characters. The first 32 characters are the main code and the last character is the check code. The main code should be arranged from left to right as follows: 24-bit marine information grid space code segment, 8-bit time code segment. For the given data latitude, longitude, altitude and time point, the global land-sea space-time grid code is generated according to the following steps.

[0092] 1) Obtain the latitude, longitude, altitude and time nodes related to the data;

[0093] 2) Express the latitude and longitude altitude in the form of whole degrees, whole minutes and seconds × 2048, and then convert it into a binary number. Perform a bit-by-bit crossover operation to form a binary one-dimensional code, which constitutes the spatial code in the global land and sea space-time grid code.

[0094] 3) Time nodes are encoded in the form of year, month and day.

[0095] 4) Combine spatial coding and temporal coding to form the ontology code.

[0096] 5) Generate the check code from the main code to uniformly form the space-time grid code for the entire land and sea.

[0097] For example, the latitude and longitude of the given data is 109.03°E, 21.56°N, and the height from the center of the earth is 6372 kilometers (surface height). First convert the height to the corresponding degree, and the calculation formula is: height (unit: degree) = height (unit: kilometer) / 40000 (equatorial circumference, unit: kilometer) × 360°, that is, 57.348°H. Convert the data represented by degrees into binary numbers. Taking longitude as an example, it is expressed in degrees, minutes, and seconds: 109°1′48.0″. Degrees and minutes are directly converted to binary numbers. The first digit of the whole degree represents the positive or negative value, a total of 9 digits, and the whole minute has 6 digits. They are: (001101101)2, (000001)2, and then the seconds and the fraction of the second are multiplied by 2048 (the minimum accuracy of the grid division) and converted to binary numbers, which is (1100000000000000)2. The longitude is converted to (0011011010000011100000000000000)2. The latitude and altitude can be obtained in the same way, which are (0000101011000011001000000000000)2 and

[0098] (000111001010100110100110011001100)2. The spatial code is formed by the bit-by-bit crossover method. The values ​​are taken in the order of taking one bit each in the longitude and latitude, and finally (000000100101011101110000111010001000001000110111101000011000000001001000000001001000)2 is formed, and the corresponding hexadecimal number is 025770E88237A18048048048. The time node of the data is represented as a time code in the format of year, month and day. For example, 20170101. For data with missing time, the default value of this segment is 00000000. Combine the space code and time code to form the main code, and generate the check code. Finally, the coding is completed.

[0099] (3) Constructing a grid data set. Using geographic grids as the temporal and spatial association, the marine environment data, marine carrier vulnerability data, and strong typhoon forecast data (such as wind intensity, wind direction, wind volume, and precipitation) are discretely loaded into each grid to form a grid data set.

[0100] S3. Construct a physical model for vulnerability assessment of marine disaster-prone bodies. Divide the entire space into evaluation units using grids, calculate the sum of the value of the features in each grid, and use it as the value index E (Economic value); at the same time, calculate the degree to which each grid unit is susceptible to disasters based on the type of marine disasters, and use it as the vulnerability index D (easy-damage Degree). The product of the two indices is the vulnerability index V (Vulnerability) of the evaluation unit.

[0101] (1) Constructing a rating index system. Ten types of land features were selected to establish a spatial value density evaluation index system, and the weight of each factor was determined through expert scoring, as shown in Table 2. Among them, buildings are both a type of land feature and reflect the distribution of people. In a typhoon storm surge disaster, buildings may not be destroyed, but people living in the buildings will be evacuated, resulting in economic losses, so they have a greater weight.

[0102] Table 2 Factors and weights of ground features

[0103]

[0104] (2) Calculation of value index. The calculation steps of value density E are as follows:

[0105] 1) Calculation of building value density. Building value density is the sum of the values ​​of all buildings within a unit area, calculated by the building area and unit price. The parameters introduced include the building floor area, number of floors, and prices of different types of buildings. The floor area and number of floors are extracted from the three-dimensional model of the city, and the building price refers to the city's economic statistics.

[0106]

[0107] Where S is the building floor area; L is the number of floors of the building; P is the building price per unit area (square meters); and i is each building within the grid of kilometers.

[0108] Buildings are divided into different types: sheds, blocks, ordinary residences, high-rise buildings, public facilities, etc. Different building types have different p values. The number of floors L of a building is calculated based on the building height. The floor height of ordinary residences and high-rise buildings is 3 meters, and the floor height of blocks and public venues is 3.5 meters. Figure 4 shown.

[0109] Table 3 Building classification table

[0110]

[0111] 2) Calculation of road value density. Roads are important infrastructure in cities and are easily damaged in typhoon storm surges. They are also important evaluation factors. Road information is easy to obtain and can be accurately quantitatively calculated.

[0112]

[0113] Where L is the length of the road; w is the number of lanes; p is the value weight, which is determined according to the road grade; and i is each highway within the kilometer grid.

[0114] Highway grades are generally divided into: expressways, first-class highways, second-class highways, and other-class highways. Different highway grades have different p values. According to relevant design standards for Chinese traffic roads, the width of each lane on urban roads is 3.5 meters. Based on this parameter and combined with the extracted road width information, the number of lanes on the main roads in the city can be calculated. Figure 5 shown.

[0115] Table 4 Road classification table

[0116]

[0117]

[0118] 3) Calculation of population factors. Population is an important factor in urban natural disaster vulnerability assessment. Many disaster loss assessment studies at home and abroad use the death toll as an indicator, but the death toll can only be obtained through post-disaster statistics and is not easy to use as a vulnerability assessment indicator. For vulnerability assessments that require quantitative calculations, population factors are mainly reflected in the spatial distribution of the population and the cost of large-scale evacuation of personnel in the face of disasters.

[0119] Q=N×P (9)

[0120] In the formula, N is the number of people to be evacuated, and P is the cost of evacuating one person. The value of P is related to factors such as the evacuation distance and the evacuation resettlement time, and requires specific analysis.

[0121] 4) Calculation of value density of other factors. Mainly includes the following categories:

[0122] a. Transportation facilities:

[0123] Q=N×P (10)

[0124] Among them, S is the floor area, L is the number of floors, and P is the unit price (yuan / square meter);

[0125] b. Linear water system:

[0126]

[0127] Among them, L is the length of the river, P is the value corresponding to different types of water quality;

[0128] c. Surface water system:

[0129]

[0130] Among them, S is the water area, P is the value corresponding to different types of water quality;

[0131] d. Vegetation:

[0132]

[0133] Among them, S is the vegetation area, P is the value corresponding to different vegetation types;

[0134] e. Pipeline:

[0135]

[0136] Among them, L is the length of pipeline, P is the value corresponding to different types of pipelines;

[0137] f. Tourist attractions:

[0138]

[0139] Among them, S is the area of ​​scenic spot, P is the value corresponding to different types of scenic spots;

[0140] g. Dam:

[0141]

[0142] Among them, L is the length of the dam, H is the height of the dam, and P is the cost of the dam per unit volume.

[0143] (3) Calculation of vulnerability index, including the classification of disaster risk level L and the calculation of ground damage probability P.

[0144] 1) Classification of disaster risk levels. The calculation of L is based on kilometer grids. The average ground elevation in each kilometer grid is calculated and divided into 10 levels. The lower the elevation, the greater the disaster risk. According to historical marine disaster statistics, areas with elevations exceeding 180 meters are generally not affected by marine disasters. The calculation of disaster risk levels for various types of land features is based on the spatial objects of land features (such as a single building or a section of road). Multiple spatial layers need to be superimposed for spatial analysis and calculation to determine the disaster risk level due to spatial location distribution.

[0145] Table 5 Hazard levels of spatial grid cells

[0146]

[0147] 2) Calculation of the probability of damage to ground objects. The calculation method of the probability of damage to ground objects in the evaluation unit is to determine the types of ground objects in the unit grid, the main parameters of the ground objects, and the area ratio of each type of ground objects, and perform weighted summation. The calculation formula is as follows. After calculation, the results are normalized so that the value range of P is (0-1).

[0148]

[0149] In the formula, P is the probability of damage to the object in the evaluation unit, F is the probability of damage to a single object, and S is the area percentage of a single object in the evaluation unit. Since the evaluation unit is a kilometer grid, the area of ​​each grid is 1 km 2 The actual area of ​​each feature is the percentage of its area in the evaluation unit.

[0150] The damage probability F of various types of ground objects is shown in Table 6. Buildings are the most important ground object factor, and different building types, floor heights, and construction years have different vulnerabilities.

[0151] Table 6 Damage probability of various types of ground objects

[0152]

[0153] 3) Calculate the vulnerability index according to formula (3).

[0154] S4. Calculate the vulnerability index of marine hazard-bearing bodies. Combine DEM data, marine hazard-bearing body data, and storm surge numerical simulation results to perform mixed calculations of spatial position and attribute values. Calculate the flooding depth of each grid based on the difference between tide level and ground elevation. Calculate the value index and vulnerability index of various elements in the grid using the grid as a unit. Calculate the vulnerability index of each grid based on formula (1). Then summarize the results of all grids to form a vulnerability assessment of the marine hazard-bearing bodies in the entire region.

[0155] In view of the difficulties in investigating and extracting massive marine hazard-prone bodies and the lack of intelligent early warning and monitoring technology, this embodiment proposes and implements a marine hazard-prone body vulnerability assessment method based on SETR and geographic grid evaluation, and constructs a marine hazard-prone body vulnerability assessment model by combining deep learning, disaster science, geography, and disaster-related oceanography. Through experimental verification, this method can be used for the study of hazard-prone body response mechanism, hazard-prone body assessment, and disaster early warning, providing a basis for refined management of disaster prevention and mitigation when strong typhoons come.

[0156] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for assessing the vulnerability of marine hazard-bearing bodies based on SETR and geographic grids, characterized by: The method comprises the following steps: S1: Establish a sample library of semantic segmentation of marine hazard-prone bodies, build a semantic segmentation model of marine hazard-prone remote sensing images based on the SETR network, extract and classify marine hazard-prone bodies in remote sensing images, and calculate the geometric attributes of marine hazard-prone bodies through the statistical pixel method; S2: Divide the space-time grid and construct the grid space-time data set. Use the geographic gridding technology to perform space-time correlation on the massive space-time data and organize it into a data set with grids as units. S3: Use oblique photogrammetry technology to model the regional geographic environment, classify and extract geographic element information, establish a grid and spatial value density disaster-prone body vulnerability assessment method, build a marine disaster-prone body vulnerability geographic calculation model, and realize the quantitative calculation of the physical vulnerability of natural disasters; S4: Calculate the vulnerability index of marine hazard-bearing bodies and evaluate the vulnerability of marine hazard-bearing bodies in the region based on the vulnerability index; The specific process of constructing the semantic segmentation model of marine hazard-prone bodies for remote sensing images in step 1 is as follows: (1) Serialize the input image into a one-dimensional vector, divide the image into image blocks of equal size, encode the spatial information of each image block to form the final sequence input, and adjust each image block to the form of a one-dimensional vector; (2) The serialized one-dimensional vector is input into the transformer encoder to extract features. The transformer encoder consists of several transformer modules, and each transformer layer has a global receptive field. Each module consists of several multi-head self-attention modules, a normalization layer, and several perceptron layers, for a total of 24 layers. (3) The output of the transformer encoder is passed to the decoder, which is also composed of several transformer modules. The decoder combines global context information with local details by using self-attention and cross-attention mechanisms, maps the encoded vector back to the image block, performs pixel-level classification, and generates semantic segmentation prediction results; In step 2, geographic grid codes are generated when massive spatiotemporal data are temporally and spatially associated. The specific process is as follows: (1) Obtain the latitude, longitude, altitude and time nodes related to the data; (2) Express the latitude and longitude in the form of whole degrees, whole minutes and seconds × 2048, and then convert it into a binary number; perform a bit-by-bit crossover operation to form a binary one-dimensional code, which constitutes the spatial code in the marine information grid code; (3) Time nodes are encoded in the form of year, month, and day; (4) Combining spatial coding and temporal coding to form an ontology code; (5) Generate a check code from the main code to uniformly form the sea area information grid code; The specific process of constructing the geographic calculation model of marine hazard-bearing body vulnerability in step 3 is as follows: (1) Construct a rating index system. Based on the geographical conditions of the demonstration area and the degree of exposure to marine disasters, select disaster-prone elements to establish a spatial value density evaluation index system, and determine the weight of each factor through scoring and normalization method; (2) Establish a quantitative calculation model for the vulnerability of disaster-prone bodies. In the quantitative calculation model, the demonstration area is geographically divided into 10m grids. The demonstration area is divided according to the divided grids, and the value index and vulnerability index of each grid are calculated respectively, and the product is calculated: V=E×D(1) In the formula: V is the vulnerability index, E is the value index, and D is the vulnerability index; (3) Value index calculation: the value of each land feature is calculated in each grid. The calculated results are normalized and then weighted summed. The calculation formula is as follows: In the formula, E is the value density, Q is the value of a certain land feature in a square kilometer, W is the land feature weight, i represents the i-th land feature, and n represents a total of n land features; (4) Calculation of vulnerability index: The vulnerability index of an assessment unit is determined by two factors: the disaster risk level L of the assessment unit and the probability of damage to the ground objects P. L expresses the magnitude of the disaster risk of the assessment unit caused by the difference in ground elevation, and P is the statistical value of the damage probability of various ground objects in the entire assessment unit. The calculation formula is as follows: D=L×P(3) In the formula, D is the vulnerability index, L is the disaster risk level, and P is the probability of ground damage, with a value range of 0 to 1; The calculation process of the vulnerability index of the marine hazard-bearing body in step 4 is: (1) Based on the value index and vulnerability index of a single grid, calculate the vulnerability index of the marine hazard-bearing body in grid units; (2) Summarize the results of all grids to form a vulnerability assessment of the marine hazard-bearing bodies in the entire region; The specific process of input image serialization is: Divide the image evenly The image blocks are then flattened into a one-dimensional vector sequence, and the vectorized image blocks p are mapped to C dimensions using linear mapping, i.e., f:p→e∈R C , C is the size of the hidden channel, is the size of the classification, and each pixel is positionally encoded to obtain p j , and then with the vector e k Add them together to get the final input sequence E = {e1+p1, e2+p2, ..., e L +p L }, j represents the jth position code, k represents the kth vector, and the sequence length L is The final dimension that needs to be input to the transformer encoder is e1...e L Represents the vector corresponding to each image block, p1...p L Indicates the position information encoding corresponding to each image block; The decoder performs pixel-level segmentation on the image, extracts a feature from the encoder every 6 layers, and then The feature vector of The feature maps are processed by a 1×1, 3×3 and 3×3 three-layer convolutional network, where the first and third layers reduce the number of channels to half of the original, and the output Then use bilinear interpolation to upsample by 4 times to get Four The features of are then fused in a top-down manner, and the top-level features and the three fused features are concatenated to obtain The size of the H×W×C feature map is obtained by applying an additional 3×3 convolution after adding features element by element, and then upsampling and convolution by 4 times to obtain the final H×W×C feature map. The category probability of the pixel points is normalized using the softmax function to obtain the final semantic segmentation result. The specific process of dividing the space-time grid in step S2 is: The space of the earth surface space grid is divided into the coordinate space of longitude and latitude. In order to ensure the division of the whole degree, whole minute and whole second of the longitude and latitude, the earth space is expanded into the longitude and latitude space of the integer power of 2, which is defined as 512°×512°×512° grid. The 60' space of each degree is expanded to 64', and the 60" space of each minute is expanded to 64". It is divided into 32 levels step by step according to the octave method. The minimum expression accuracy can be accurate to 1.5 cm. The part within the range of 180°×360° in the expanded longitude and latitude coordinate space is consistent with the actual geographic space. The part beyond the range of 180°×360° has no actual geographic meaning. When the elevation is fixed to the earth's surface height, the land and sea space-time information grid is used as the earth's plane grid. When the elevation value is not needed, the plane grid can be used. The calculation process of ground object damage probability is as follows: The calculation method of the probability of damage to the ground objects in the evaluation unit is to determine the types of ground objects, ground object parameters and the area ratio of each type of ground objects in the unit cell network, and perform weighted summation. The calculation formula is as follows. After calculation, the result is normalized so that the value range of the ground object damage probability P is 0-1; In the formula, P is the probability of damage to the object in the evaluation unit, F is the probability of damage to a single object, and S is the area percentage of a single object in the evaluation unit. Since the evaluation unit is a kilometer grid, the area of ​​each grid is 1 km. 2 , the actual area of ​​each feature is the percentage of its area in the evaluation unit; The specific process of disaster risk classification is as follows: The calculation of disaster risk level takes kilometer grid as unit, calculates the average ground elevation within each kilometer grid, and divides it into 10 levels. The lower the elevation, the greater the disaster risk. According to historical marine disaster statistics, areas with elevations exceeding 180 meters will not be affected by marine disasters. The calculation of disaster risk level of land features takes the spatial object of land features as unit, and requires the superposition of several spatial layers for spatial analysis and calculation to determine the disaster risk level formed by the spatial location distribution.

Images