Space-time dimension disaster intensity assessment and visualization method for urban flood prevention construction

Abstract

The invention discloses a space-time dimension disaster intensity evaluation and visualization method for urban flood prevention construction, and the method comprises the steps: obtaining an original flood detection data set of an urban region, and expanding the data set; constructing and training a global sensing network model for cross-modal change detection flood extraction; acquiring a detection result of the multi-temporal urban flood inundation area by using the trained model through a threshold segmentation method; counting the multi-temporal flood loss, and performing weighted calculation on the loss of the urban flood inundation area in each temporal to obtain the disaster intensity in the time dimension; the land use type of each region of the current city is analyzed, and the disaster intensity is evaluated according to the spatial dimension; and generating a flood assessment map according to the disaster intensity of the time dimension and the space dimension, carrying out intersection operation on the flood assessment map and the current urban flood inundation area, obtaining a comprehensive flood assessment map, and generating a thermodynamic diagram. According to the method, the multi-temporal urban flood inundation area is accurately identified, and accurate and effective guidance is provided for post-disaster flood prevention construction of the urban area.

Description

Technical Field

[0001] The present invention relates to a method for evaluating and visualizing the disaster-affected intensity in the spatio-temporal dimension for urban flood control and disaster reduction construction, and belongs to the technical field of flood monitoring. Background Art

[0002] In recent years, due to the dense urban population, the loss of life and property caused by flood disasters every year has been particularly serious. Therefore, accurately identifying the flood inundation area is crucial for post-disaster construction. Traditional flood monitoring methods mainly rely on manual data collection. This method not only has a limited coverage area, but also appears to be too high in terms of human and time costs, making it difficult to meet the demand for accurate delineation of flood areas. Remote sensing technology has now become the primary tool for flood disaster monitoring due to its convenient data acquisition, accurate surface observation, and wide coverage.

[0003] However, in urban areas, due to the large ground inhomogeneity, such as buildings, roads, and green belts, the boundaries of flood areas become complex and irregular, increasing the difficulty of identification. In addition, previous flood disaster monitoring studies have mainly focused on the identification of flood inundation ranges, lacking in-depth analysis of the differences in flood losses in urban areas over previous floods and the losses of different land use types. This limitation fails to effectively evaluate the losses of cities and residents in terms of life and property, making it difficult to assess the impact of flood disasters on society and the economy. Summary of the Invention

[0004] Object of the Invention: Aiming at the problems and deficiencies of the existing technology, the present invention provides a method for evaluating and visualizing the disaster-affected intensity in the spatio-temporal dimension for urban flood control and disaster reduction construction. The invention constructs a global perception network based on cross-modal change detection to achieve high-precision identification of multi-temporal urban flood inundation areas. On the basis of obtaining multi-temporal urban flood inundation areas, a comprehensive flood evaluation is carried out in combination with the disaster-affected intensity in the spatio-temporal dimension. The flood evaluation results are intuitively presented using a heat map, providing effective guidance for post-disaster urban planning and flood control project design.

[0005] Technical Solution: A method for evaluating and visualizing the disaster-affected intensity in the spatio-temporal dimension for urban flood control and disaster reduction construction, comprising the following steps:

[0006] S1: Obtain the original flood detection data set of the urban area, preprocess the data set, label the urban areas affected by floods, and use data augmentation methods to expand the data set;

[0007] S2: Construct a global perception network model for cross-modal change detection flood extraction, and train the model through the data set;

[0008] S3: Through the threshold segmentation method, use the trained global perception network model for cross-modal change detection flood extraction to obtain the detection results of multi-temporal urban flood inundation areas;

[0009] S4: Statistically calculate the multi-temporal flood losses by weighted calculation of the losses in the flood inundation areas of each time phase of the city to obtain the disaster-affected intensity in the time dimension.

[0010] S5: Analyze the land use types of each area in the current city and evaluate the disaster-affected intensity in the spatial dimension.

[0011] S6: Generate a flood assessment map based on the disaster-affected intensities in the time dimension and the spatial dimension, and perform an intersection operation with the current urban flood inundation area to obtain a comprehensive flood assessment map.

[0012] S7: Generate a heat map based on the comprehensive flood assessment map and use it as a basis for guiding post-disaster construction.

[0013] In the step S1, SAR images and multi-spectral remote sensing images for original flood detection in urban areas are obtained using Sentinel-1 and Sentinel-2 satellite data to construct a data set. The remote sensing images are preprocessed using radiometric calibration, geometric correction, atmospheric correction, image enhancement, filtering denoising, and image cropping. The flood areas in the remote sensing images are labeled using an image annotation tool, and the labeled data is converted into a binary map that distinguishes between flood and non-flood areas through a script. The image processing library functions are called to perform data augmentation operations on the remote sensing images and their corresponding binary map labels to increase the diversity of the data and obtain an extended data set.

[0014] The global perception network model includes a multi-scale feature extractor, a global feature refinement module (GFRM), a gated feature fusion module (GFM), and a symbiotic relationship learning module (SRLM). The step S2 specifically includes:

[0015] S21: Use the multi-scale feature extractor to extract features from the flood detection data set to obtain the multi-scale global-local features of the dual-temporal remote sensing images.

[0016] S22: Use the global feature refinement module (GFRM) to explicitly model and refine the effective information in the context relationship of the multi-scale global-local features of the dual-temporal remote sensing images to enhance the network's context information learning and feature effective extraction capabilities.

[0017] S23: Fuse the refined features of the dual-temporal through the gated feature fusion module (GFM) to obtain the fused features.

[0018] S24: Input the refined features of the first phase and the obtained fused features into the Symbiotic Relationship Learning Module (SRLM) according to steps S22 and S23, learn the symbiotic relationship between the scene and the foreground, and focus on the foreground area, so as to extract features that enhance the foreground and reduce false alarms.

[0019] S25: Again, fuse the dual-phase features through the Gated Feature Fusion Module (GFM), and under a series of convolutional layer operations, learn the change results in the dual-phase remote sensing images.

[0020] In the above-mentioned step S21, the multi-scale feature extractor T of the first phase 1 adopts ResNet, and the input is the multi-spectral remote sensing image; the multi-scale feature extractor T of the second phase 2 adopts EfficientNet, and the input is the SAR image.

[0021] In the above-mentioned step S22, the Global Feature Refinement Module (GFRM) improves the Convolutional Block Attention Module (CBAM), and this module is composed of a channel attention branch and a spatial attention branch. Extract low-level detail information, refine the high-level upsampled features through deformable convolution, and obtain better boundary results.

[0022] The calculation steps of GFRM are as follows:

[0023] A1: Pass f ci =MLP c (MaxPool s (f i ), AvgPool s (f i ) to globally pool the input feature f i in the spatial dimension respectively through MaxPool s and AvgPool s . After learning the relationship between channel features by MLP c and adding them for fusion, obtain f ci . Among them, MaxPool s is the maximum pooling in the spatial dimension, AvgPool s is the average pooling in the spatial dimension, MLP c is the multi-layer perceptron in the channel dimension, and f ci is the output channel feature;

[0024] A2: Pass f si =MLP s (MaxPool c (f i ), AvgPool c (f i))The input feature f i is respectively passed through MaxPool c and AvgPool c to perform global pooling in the channel dimension. After the MLP s learns the relationship between channel features, they are added and fused to obtain f si . Among them, MaxPool c is the maximum pooling in the channel dimension, AvgPool c is the average pooling in the channel dimension, MLP s is the multi-layer perceptron in the spatial dimension, and f si is the output spatial feature;

[0025] A3: Through G CAM , G SAM = Split(MLP(concat(f ci , f si ))) to merge the feature f ci that has learned the relationship between channel features and the feature f si that has learned the relationship between spatial features, and learn the relationship between channel and spatial features. Among them, concat represents the merging of spatio-temporal features, MLP is used to learn the relationship between channel and spatial features, Split re-divides the channel and spatial features, and G CAM represents the channel feature that has learned the spatial-channel relationship, and G SAM represents the spatial feature that has learned the spatial-channel relationship;

[0026] A4: Through multiply and fuse the feature weights in the spatial and channel dimensions with the original features to further refine the features in the channel and space; among them, σ represents the Sigmoid activation function, represents matrix multiplication, represents feature fusion addition, and f mi represents the output feature map after channel and spatial weighting;

[0027] A5: To obtain complex edge features, deformable convolution is used for effective edge recognition, which is defined as follows:

[0028] △p k , △m′ k = Conv(g i-1 )

[0029]

[0030] Among them, K is the sampling number of the convolution kernel. w k and p kDenote the weights and sampling grids in ordinary convolution. △p k and △m' k are the sampling offsets and scalar masks learned through ordinary convolution, and g i is the feature learned in the deformable convolution for the i-th one, where i ∈ {1, 2, 3, 4}, and when i = 1, g 0 = g 1 .

[0031] In step S23, the gated feature fusion module (GFM) fuses the refined features of two time phases to enhance the feature representation. Fusing two data sources reduces the limitations of a single data source and provides a more stable and reliable output. The calculation steps of GFM are as follows:

[0032]

[0033] Among them, and respectively represent the features of the first time phase and the second time phase learned in GFRM. Concat represents feature merging. Conv 1×1 represents performing a 1×1 convolution on the feature map. σ represents the Sigmoid activation function. represents matrix multiplication, represents the output feature map that fuses two data sources.

[0034] In step S24, the symbiotic relationship learning module (SRLM) learns the symbiotic relationship between the foreground and the background by expanding the receptive field and enhancing foreground perception, thereby effectively associating the context related to the foreground and focusing the network's attention on the foreground region.

[0035] The function of SRLM can be defined as:

[0036] B1: Continuously use dilated convolution to continuously increase the receptive field of the feature map and learn the symbiotic relationship between the foreground and the background. The specific method is:

[0037] DC j (r) = Dropout(R(B N (D j (r))))

[0038]

[0039] Among them, j ∈ {1, 2, 3, 4}, and g a represents the feature learned in GFRM. D i (r) represents the j-th layer of dilated convolution with a dilation rate of r, where the dilation rates are 1, 3, 5, and 3 respectively. B NBatch normalization is denoted as. ReLU activation function is denoted as R. Dropout is denoted as denotes a convolutional operation. f a denotes the output feature.

[0040] B2: Refine the features learned for the foreground and background co-occurrence relationship of f a by fusing with g a and f a to further refine the features. The steps are as follows:

[0041]

[0042] where denotes element-wise addition. denotes performing a 1×1 convolution on the feature map with the number of channels halved. denotes performing a 1×1 convolution on the feature map with the number of channels becoming 1. R denotes the ReLU activation function. σ denotes the Sigmoid activation function. f a denotes the input feature. g a denotes the features learned in GFRM. x n is the refined feature.

[0043] In step S25, fuse the dual-temporal features, and the result is denoted as X. Obtain the change features through Upsample(Conv 1×1 (Conv 3×3 (Conv 3×3 (X)))). Where X is the dual-temporal remote sensing image features learned through GFM. Upsample is the upsampling operation. Conv 3×3 denotes performing a 3×3 convolution on the feature map with the number of channels halved each time. Conv 1×1 denotes performing a 1×1 convolution on the feature map with the number of channels reduced to 1. Use the joint loss function L Total = L Bce + α·L Dice and Adam as the optimization method, where α is a tuning parameter used to balance the binary cross-entropy loss L Bce and the Dice loss L Dice between the weights.

[0044] In step S3, the threshold segmentation method is as follows:

[0045] S31: Input multiple historical flood remote sensing image data into the trained network model to obtain single-channel output features;

[0046] S32: Use the Sigmoid function and threshold Θ to determine the change region for the output features of a single channel and use it as the detection result.

[0047] In step S4, use Obtain the normalized loss weight for the k-th flood disaster. The obtained loss weight is used as the disaster intensity in the time dimension. Among them, L k is the computable loss caused by the k-th flood disaster. n is the total number of historical floods. λ is a non-negative real number used to control the decay rate of the exponential decay function. w k represents the k-th output weight.

[0048] In step S5, obtain various land use types from the current urban planning data, count the average losses suffered by various land use types in historical flood events, multiply by the spatial resolution of the flood remote sensing image, and calculate the average inundation loss threshold l m,n , where m is the m-th flood disaster and n is the n-th region. l m,n is used as the disaster intensity in the spatial dimension.

[0049] In step S6, obtaining the comprehensive flood assessment map includes:

[0050] S61: Considering the disaster intensities in the time dimension and the spatial dimension, more attention can be paid to multiple floods with larger losses in time, and the loss differences brought by current various land use types can be considered in space. Finally, through multi-temporal flood superposition, a comprehensive loss assessment is carried out on the flood inundation areas of each time phase of the city. The specific assessment expression is:

[0051]

[0052] Among them, t represents the total number of floods counted. x (m,n) represents the flood inundation detection result of the n-th region of the m-th flood disaster after being converted by a set threshold, where 0 represents not inundated and 1 represents inundated. w m represents the disaster intensity of the m-th flood. l m,n represents the disaster intensity of the land use type to which the n-th region of the m-th flood disaster belongs. S (n) represents the comprehensive disaster intensity of the n-th region.

[0053] S62: In order to provide guidance for flood prevention and construction after this flood disaster, an intersection operation is performed using the current flood inundation detection result and the comprehensive loss assessment, and the comprehensive loss assessment result of the current flood inundation area obtained is used as the assessment map.

[0054] In step S7, use an image processing library to save or display the assessment result as a heat map to visualize the assessment result of the flood disaster intensity in the urban area.

[0055] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0056] 1. By utilizing the global feature refinement module, the context features of remote sensing images are effectively learned.

[0057] 2. The gated feature fusion module combines multi-spectral and SAR remote sensing image data sources to overcome the limitations brought by relying on a single data source, thereby providing more stable and reliable results.

[0058] 3. The symbiotic relationship learning module uses continuous dilated convolutions to increase the receptive field and learn the relationship between the foreground and the background, improving the detection accuracy of urban irregular flood inundation areas.

[0059] 4. Starting from effectively reducing the impact of floods on society and economy, the flood disaster intensity is evaluated from multiple aspects in the time dimension and the space dimension, and guidance is provided for the construction of flood prevention and control in urban areas after disasters. Brief description of the drawings

[0060] Figure 1 is the flowchart of the method of the embodiment of the present invention;

[0061] Figure 2 is the framework diagram of the global perception network for cross-modal change detection and flood extraction of the embodiment of the present invention;

[0062] Figure 3 is the schematic diagram of the global feature refinement module of the embodiment of the present invention;

[0063] Figure 4 is the schematic diagram of the gated feature fusion module of the embodiment of the present invention;

[0064] Figure 5 is the schematic diagram of the symbiotic relationship learning module of the embodiment of the present invention. Detailed implementation manners

[0065] The following further clarifies the present invention in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. After reading the present invention, various equivalent modifications made by those skilled in the art to the present invention fall within the scope defined by the appended claims of this application.

[0066] Embodiment 1:

[0067] As Figure 1 shown, a method for evaluating and visualizing the disaster intensity in the time and space dimensions for urban flood prevention and control construction in this embodiment includes the following steps:

[0068] Step 1: Use Sentinel-1 and Sentinel-2 satellite data to obtain SAR images and multispectral remote sensing images for original flood detection in urban areas, and construct a dataset. Perform preprocessing operations on the obtained remote sensing images, such as radiometric calibration, geometric correction, atmospheric correction, image enhancement, and filtering for noise reduction. Crop the processed remote sensing images to a size of 512×512 pixels and use an image annotation tool to label the flood areas in the remote sensing images, and convert the annotation data into a binary map that distinguishes flood areas from non-flood areas through a script. Call the functions of the image processing library to perform necessary data augmentation operations on the remote sensing images and their corresponding binary map labels in the dataset to increase data diversity and obtain an extended dataset. In this embodiment, the training set is 70%, the validation set is 20%, and the test set is 10%.

[0069] Step 2: Construct a global perception network model for cross-modal change detection flood extraction, as Figure 2 shown, and use the dataset to train the model.

[0070] Specifically, the above Step 2 includes the following steps:

[0071] Step 21: Use a multi-scale feature extractor to extract features from the flood change water body dataset in Step 1 to obtain multi-scale global-local features of the dual-temporal remote sensing images, where the multi-scale feature extractor T 1 for the first temporal phase adopts ResNet and the input is the multispectral remote sensing image; the multi-scale feature extractor T 2 for the second temporal phase adopts EfficientNet and the input is the SAR image.

[0072] Step 22: After constructing the backbone network, use the global feature refinement module (GFRM) to explicitly model and refine the effective information in the context relationship to enhance the network's context information learning and feature effective extraction capabilities, as Figure 3 shown.

[0073] The global feature refinement module (GFRM) can be defined as:

[0074] A1: Pass the input feature f ci through f c =MLP s (MaxPool i )(f s ), AvgPool i )(f i ) to perform global pooling on the input feature f s in the spatial dimension through MaxPool s and AvgPool c respectively, and add and fuse the learned relationships between channel features by MLPci . Among them, MaxPool s is the maximum pooling in the spatial dimension, and AvgPool s is the average pooling in the spatial dimension. MLP c is the multi-layer perceptron in the channel dimension, and f ci is the output channel feature;

[0075] A2: Through f si = MLP s (MaxPool c (f i ), AvgPool c (f i )) the input feature f i is respectively globally pooled in the channel dimension through MaxPool c and AvgPool c . After the relationship between channel features is learned by MLP s and added and fused, f si is obtained. Among them, MaxPool c is the maximum pooling in the channel dimension, AvgPool c is the average pooling in the channel dimension, MLP s is the multi-layer perceptron in the spatial dimension, and f si is the output spatial feature;

[0076] A3: Through G CAM , G SAM = Split(MLP(concat(f ci , f si ))) the feature f ci that has learned the relationship between channel features and the feature f si that has learned the relationship between spatial features are merged and the relationship between channel and spatial features is learned. Among them, concat represents the merger of spatio-temporal features, MLP is used to learn the relationship between channel and spatial features, Split re-divides the channel and spatial features, and G CAM represents the channel feature that has learned the spatial-channel relationship, and G SAM represents the spatial feature that has learned the spatial-channel relationship;

[0077] A4: Through the feature weights in the spatial and channel dimensions are multiplied and fused with the original features, and the features are further refined in the channel and space; among them, σ represents the Sigmoid activation function, represents matrix multiplication, represents feature fusion addition, and f miRepresents the output feature map after channel and spatial weighting;

[0078] A5: To obtain complex edge features, deformable convolution is used for effective edge recognition, which is defined as follows:

[0079] △p k ,△m′ k =Conv(g i-1 )

[0080]

[0081] where K is the number of samplings of the convolution kernel. w k and p k represent the weights and sampling grids in ordinary convolution. △p k and △m' k are the sampling offsets and scalar masks learned through ordinary convolution, and g i is the feature learned in the i-th deformable convolution, i ∈ {1, 2, 3, 4}, and when i = 1, g 0 =g 1 .

[0082] Step 23: Use the gated feature fusion module (GFM) to fuse the refined features of the two time phases for feature enhancement. Fusing two data sources reduces the limitations of a single data source and provides a more stable and reliable output, as Figure 4 shown.

[0083] The gated feature fusion module (GFM) can be defined as:

[0084]

[0085] where, and represent the features of the first time phase and the second time phase learned in the GFRM respectively. concat represents feature merging. Conv 1×1 represents performing a 1×1 convolution on the feature map. σ represents the Sigmoid activation function. represents matrix multiplication, represents the output feature map that fuses the two data sources.

[0086] Step 24: The symbiotic relationship learning module (SRLM) explores the symbiotic relationship between the foreground and the background by expanding the receptive field and enhancing the perception of the foreground. This method helps the network effectively link the foreground-related context and focus the attention on the foreground area, as Figure 5 shown.

[0087] The function of the symbiotic relationship learning module (SRLM) can be defined as:

[0088] B1: Continuously use dilated convolution to continuously increase the receptive field of the feature map and learn the symbiotic relationship between the foreground and the background. The specific method is as follows:

[0089] DC j (r) = Droopout(R(B N (D j (r))))

[0090]

[0091] where j ∈ {1, 2, 3, 4}, g a represents the features learned in GFRM. D i (r) represents the j-th layer of dilated convolution with a dilation rate of r, where the dilation rates are 1, 3, 5, and 3 respectively. B N represents batch normalization. R represents the ReLU activation function. Dropout represents random dropout. represents the convolution operation. f a represents the output features.

[0092] B2: Refine the features f that have learned the symbiotic relationship between the foreground and the background, and further refine the features through the fusion with g a and f a and f a as follows:

[0093]

[0094] where, represents element-wise addition. represents performing a 1×1 convolution on the feature map to halve the number of channels. represents performing a 1×1 convolution on the feature map to change the number of channels to 1. R represents the ReLU activation function. σ represents the Sigmoid activation function. f a represents the input features. g a represents the features learned in GFRM. x n is the refined feature.

[0095] Step 25: Fuse the dual-temporal features, and the result is denoted as X. Obtain the change features through Upsample(Conv 1×1 (Conv 3×3 (Conv 3×3 (X)))). Where X is the dual-temporal remote sensing image features learned through GFM. Upsample is the upsampling operation. Conv 3×3 represents performing a 3×3 convolution on the feature map to halve the number of channels each time. Conv 1×1Indicates a 1×1 convolution on the feature map, reducing the number of channels to 1. Use a combined loss function \(L\) composed of binary cross-entropy loss and Dice loss Total = \(L\) Bce + \(\alpha\cdot L\) Dice and Adam as the optimization method, where \(\alpha\) is a tuning parameter used to balance the weights between the binary cross-entropy loss \(L\) Bce and the Dice loss \(L\) Dice and is set to 0.7 in this embodiment.

[0096] Step 3: Through the threshold segmentation method, use the globally aware network trained for cross-modal change detection flood extraction to obtain the detection results of multi-temporal urban flood inundation areas. Specifically:

[0097] C1: Input multiple historical flood remote sensing image data into the trained network model to obtain single-channel output features. In this embodiment, the flood period from May to September each year in the past 10 years is used as the key data extraction period to capture flood events during this time.

[0098] C2: Use the Sigmoid function and threshold \(\Theta\) to determine the change area for the single-channel output features and use it as the detection result.

[0099] Step 4: Statistically calculate the multi-temporal flood losses, and use to obtain the normalized loss weight of the \(k\)th flood disaster. The obtained loss weight is used as the disaster intensity in the time dimension. Among them, \(L\) k is the computable loss caused by the \(k\)th flood disaster. \(n\) is the total number of historical floods. \(\lambda\) is a non-negative real number used to control the decay rate of the exponential decay function and is set to 0.1 in this embodiment. \(w\) k represents the \(k\)th output weight.

[0100] Step 5: Obtain various land use types from the current urban planning materials, statistically calculate the average losses suffered by various land use types in historical flood events, and multiply by the spatial resolution of the flood remote sensing images to calculate the average inundation loss threshold \(l\) m,n , where \(m\) is the \(m\)th flood disaster and \(n\) is the \(n\)th area. \(l\) m,n is used as the disaster intensity in the spatial dimension. The average inundation losses suffered by various land use types in this example are shown in Table 1.

[0101] Table 1 Average inundation losses suffered by various land use types

[0102]

[0103] Step 6: Generate a flood assessment map based on the disaster intensity in the time dimension and the space dimension, and perform an intersection operation with the current urban flood inundation area to obtain a comprehensive flood assessment map. Specifically, obtaining the comprehensive flood assessment map includes:

[0104] D1: Considering the disaster intensity in the time dimension and the space dimension, more attention can be paid to multiple floods with greater losses in terms of time, and the loss differences brought by current land use types can be considered in terms of space. Finally, through multi-temporal flood superposition, a comprehensive loss assessment is carried out on the urban flood inundation areas in each time phase. The specific assessment expression is:

[0105]

[0106] Among them, t represents the total number of floods counted. x (m,n) represents the flood inundation detection result of the nth area in the mth flood disaster after being converted by a set threshold, where 0 represents non-inundated and 1 represents inundated. w m represents the disaster intensity of the mth flood. l m,n represents the disaster intensity of the land use type to which the nth area in the mth flood disaster belongs. S (n) represents the comprehensive disaster intensity of the nth area.

[0107] D2: In order to provide guidance for flood prevention and control construction after this flood disaster, an intersection operation is performed using the current flood inundation detection result and the comprehensive loss assessment, and the comprehensive loss assessment result of the current flood inundation area obtained is used as the assessment map.

[0108] Step 7: Generate a heat map based on the comprehensive flood assessment map, and use an image processing library to save or display the assessment result as a heat map to visualize the assessment result of the urban flood disaster intensity.

Claims

1. A method for assessing and visualizing the disaster intensity in spatiotemporal dimensions for urban flood control construction, characterized in that: The steps include: S1: Obtain the original flood detection dataset in urban areas, preprocess the dataset, mark the urban areas affected by floods, and expand the dataset using data augmentation methods; S2: Build a global perception network model for cross-modal change detection flood extraction and train the model with the dataset; S3: Through the threshold segmentation method, the global perception network model trained for cross-modal change detection and flood extraction is used to obtain the detection results of multi-temporal urban flood inundation areas; S4: Calculate the multi-temporal flood losses by weighting the losses of the flood-inundated areas of the cities in each phase to obtain the disaster intensity in the time dimension; S5: Analyze the land use types in each area of ​​the current city and evaluate the disaster intensity according to the spatial dimension; S6: Generate a flood assessment map based on the disaster intensity in the time dimension and the space dimension, and perform an intersection operation with the current urban flood inundation area to obtain a comprehensive flood assessment map; S7: Generate a heat map based on the comprehensive flood assessment map.

2. The method for evaluating and visualizing disaster intensity in spatiotemporal dimensions for urban flood control construction according to claim 1 is characterized in that: In the step S1, SAR images and multispectral remote sensing images for original flood detection in urban areas are obtained to construct a data set; remote sensing images are preprocessed using radiation calibration, geometric correction, atmospheric correction, image enhancement, filtering denoising, and image cropping; flood areas in remote sensing images are annotated using image annotation tools, and the annotated data are converted into binary images that distinguish flooded areas from non-flooded areas through a script; image processing library functions are called to perform data enhancement operations on remote sensing images and their corresponding binary image labels to obtain an expanded data set.

3. The method for evaluating and visualizing the disaster intensity in the spatiotemporal dimension for urban flood control construction according to claim 1 is characterized in that: The global perception network model includes a multi-scale feature extractor, a global feature refinement module, a gated feature fusion module and a symbiotic relationship learning module. Step S2 specifically includes: S21: Use a multi-scale feature extractor to extract features from the flood detection dataset and obtain multi-scale global-local features of the dual-temporal remote sensing image; S22: using a global feature refinement module to explicitly model the effective information in the context relationship of the multi-scale global-local features of the dual-temporal remote sensing image and refine the features; S23: fusing the bi-phase refined features through a gated feature fusion module to obtain fused features; S24: according to step S22 and step S23, the refined features of the first phase and the obtained fusion features are respectively transmitted to the symbiotic relationship learning module, the symbiotic relationship between the scene and the foreground is learned, and the attention is focused on the foreground area; S25: The dual-phase features are fused again through the gated feature fusion module and the changes in the dual-phase remote sensing image are learned through a series of convolutional layer operations.

4. The method for evaluating and visualizing the disaster intensity in the spatiotemporal dimension for urban flood control construction according to claim 3 is characterized in that: In step S21, the multi-scale feature extractor T1 of the first phase adopts ResNet, and the input is the multi-spectral remote sensing image; the multi-scale feature extractor T2 of the second phase adopts EfficientNet, and the input is the SAR image.

5. The method for evaluating and visualizing disaster intensity in spatiotemporal dimensions for urban flood control construction according to claim 3 is characterized in that: In step S22, the global feature refinement module includes a channel attention branch and a spatial attention branch; extracting low-level detail information, and refining high-level upsampled features through deformable convolution to obtain better boundary results; The calculation steps of the global feature refinement module are as follows: A1: Through f ci =MLP c (MaxPool s (f i ),AvgPool s (f i )) Input feature f i Through MaxPool s and AvgPool s Global pooling is performed in the spatial dimension, by MLP c After learning the relationship between channel features, we add and fuse them to get f ci ; Among them, MaxPool s It is the maximum pooling in the spatial dimension, AvgPool s is the average pooling in the spatial dimension, MLP c is a multi-layer perceptron in the channel dimension, f ci is the channel feature of the output; A2: Through f si =MLP s (MaxPool c (f i ),AvgPool c (f i )) Input feature f i Through MaxPool c and AvgPool c Global pooling is performed in the channel dimension, by MLP s After learning the relationship between channel features, we add and fuse them to get f si ; Among them, MaxPool c It is the maximum pooling in the channel dimension, AvgPool c is the average pooling in the channel dimension, MLP s is a multi-layer perceptron in the spatial dimension, f si is the spatial feature of the output; A3: Through G CAM ,G SAM =Split(MLP(concat(f ci ,f si ))) The feature f that learns the relationship between channel features ci and features f that learn the relationship between spatial features si Merge and learn the relationship between channel and spatial features; where concat represents the merging of spatiotemporal features, MLP is used to learn the relationship between channel and spatial features, Split re-divides channel and spatial features, and G CAM Indicates the channel features of the learned spatial channel relationship, G SAM Indicates that the spatial features of the spatial channel relationship have been learned; A4: Pass The feature weights in the spatial and channel dimensions are multiplied and fused with the original features to further refine the features in the channel and space; where σ represents the Sigmoid activation function, represents matrix multiplication, represents feature fusion addition, f mi Represents the output feature map after channel and spatial weighting; A5: In order to obtain complex edge features, deformable convolution is used for effective edge recognition, which is defined as follows: Δp k ,Δm′ k =Conv(g i-1 ) Among them, K is the number of samples of the convolution kernel; w k and p k represents the weight and sampling grid in ordinary convolution; Δp k and Δm′ k is the sampling offset and scalar mask learned by ordinary convolution, g i is the i-th feature learned in the deformable convolution, i∈{1,2,3,4}, and when i=1, g0=g1.

6. The method for evaluating and visualizing disaster intensity in spatiotemporal dimensions for urban flood control construction according to claim 3 is characterized in that: In step S23, the gated feature fusion module fuses the bi-phase refined features to enhance feature representation; the calculation steps of the gated feature fusion module are as follows: in, and They represent the first phase and the second phase features learned in the global feature refinement module respectively; concat represents feature merging; Conv 1×1 Indicates 1×1 convolution of the feature map; σ indicates the Sigmoid activation function; represents matrix multiplication, Represents the output feature map of the fusion of two data sources.

7. The method for evaluating and visualizing disaster intensity in spatiotemporal dimensions for urban flood control construction according to claim 3 is characterized in that: In step S24, the symbiotic relationship learning module learns the symbiotic relationship between the foreground and the background by expanding the receptive field and enhancing the foreground perception, thereby effectively associating the foreground-related context and focusing the network's attention on the foreground area; The function of the symbiotic relationship learning module is defined as: B1: Continuously use dilated convolution to continuously increase the receptive field of the feature map and learn the symbiotic relationship between foreground and background; the specific method is: DC j (r)=Dropout(R(B N (D j (r)))) f a =g a οDC1(r)οDC2(r)οDC3(r)οDC4(r) Among them, j∈{1,2,3,4}, g a Denotes the features learned in GFRM; D i (r) represents the jth layer of atrous convolution based on the dilation rate r, where the dilation rates are 1, 3, 5, and 3 respectively; B N represents batch normalization; R represents the ReLU activation function. Dropout represents random dropout; ° represents the convolution operation; f a represents the output features; B2: f that learns the symbiotic relationship between foreground and background a The features are refined by comparing with g a and f a The fusion of further refines the features; proceed as follows: in, Represents element-by-element addition. It means that 1×1 convolution is performed on the feature map, and the number of channels is halved; Indicates that a 1×1 convolution is performed on the feature map, and the number of channels becomes 1; R represents the ReLU activation function; σ represents the Sigmoid activation function; f a represents the input feature; g a represents the features learned in GFRM; x n It is the refined feature.

8. The method for evaluating and visualizing disaster intensity in spatiotemporal dimensions for urban flood control construction according to claim 3 is characterized in that: In step S25, the dual-phase features are fused, and the result is represented by X; by Upsample (Conv 1 ×1 (Conv 3×3 (Conv 3×3 (X))))Get the change characteristics; Where X is the dual-temporal remote sensing image feature learned by GFM; Upsample is the upsampling operation; Conv 3×3 Indicates that a 3×3 convolution is performed on the feature map, and the number of channels is halved each time; Conv 1×1 Indicates that a 1×1 convolution is performed on the feature map, and the number of channels is reduced to 1; a joint loss function L based on binary cross entropy loss and Dice loss is used Total =L Bce +α·L Dice and Adam as the optimization method, where α is a tuning parameter used to balance the binary cross entropy loss L Bce and Dice loss L Dice The weight between .

9. The method for evaluating and visualizing disaster intensity in spatiotemporal dimensions for urban flood control construction according to claim 1 is characterized in that: In step S3, the threshold segmentation method is: S31: Input multiple historical flood remote sensing image data into the trained network model to obtain the output features of a single channel; S32: Use the Sigmoid function and threshold Θ to determine the change area of ​​the output feature of the single channel and use it as the detection result. In step S4, use Get the normalized loss weight of the kth flood disaster. The obtained loss weight is used as the disaster intensity in the time dimension. Among them, L k is the calculable loss caused by the kth flood disaster. n is the total number of historical floods. λ is a non-negative real number used to control the decay speed of the exponential decay function. w k represents the kth output weight.

10. The method for evaluating and visualizing disaster intensity in spatiotemporal dimensions for urban flood control construction according to claim 1, characterized in that: In step S5, the land use types are obtained from the current urban planning data, the average losses of various land use types in historical flood events are calculated, and the average flood loss threshold l is calculated by multiplying it by the spatial resolution of the flood remote sensing image. m,n , where m is the mth flood disaster, n is the nth region; l m,n As a kind of disaster intensity in spatial dimension; In step S6, obtaining a comprehensive flood assessment map includes: S61: Considering the disaster intensity in both time and space dimensions, a comprehensive loss assessment is conducted on the flood-inundated areas of cities in each time phase through the superposition of multi-temporal floods. The specific assessment expression is: Where t represents the total number of floods counted; x (m,n) represents the flood inundation detection result of the nth area in the mth flood disaster after the set threshold conversion, where 0 represents no inundation and 1 represents inundation; w m represents the intensity of the flood disaster at the mth event; l m,n S represents the disaster intensity of the land use type in the nth area of ​​the mth flood disaster; (n) It represents the comprehensive disaster intensity of the nth region; S62: performing an intersection operation using the current flood and waterlogging detection result and the comprehensive loss assessment, and using the obtained comprehensive loss assessment result of the current flood and waterlogging inundation area as an assessment map; In step S7, the evaluation results are saved or displayed as a heat map using an image processing library to visualize the evaluation results of the flood damage intensity in the urban area.

Images