Flood impact area extraction method and system based on convolution and state space model

Abstract

The application discloses a flood influence range extraction method and system based on convolution and state space model, and belongs to the technical field of remote sensing image processing and disaster monitoring; the method comprises the following steps: acquiring remote sensing image data of a region to be detected, and performing pretreatment; inputting the pretreated remote sensing image into a multi-stage and multi-branch hierarchical feature extraction encoder; utilizing channel attention and spatial attention mechanisms of the hierarchical feature extraction encoder to dynamically weight and fuse local and global features of different scales; inputting the fused multi-scale feature maps into a decoder for upsampling, and finally outputting a pixel-level binary segmentation mask of the flood influence range. The application has the advantages of high calculation efficiency, accurate edge positioning and strong anti-interference capability, and is suitable for actual businesses such as emergency monitoring, disaster assessment and post-disaster reconstruction.

Description

Technical Field

[0001] This invention relates to the fields of remote sensing image processing and disaster monitoring technology, specifically to a method for extracting the impact range of floods based on convolution and state space models, and also to a system for extracting the impact range of floods based on convolution and state space models. Background Technology

[0002] Floods are a type of natural disaster that occurs frequently, has a wide impact, and is characterized by its suddenness. Quickly and accurately determining the extent of flood inundation is crucial for emergency response, risk assessment, and post-disaster reconstruction. Remote sensing imagery, with its advantages of wide coverage and timely delivery, has become an important data source for flood monitoring.

[0003] Existing methods for flood extent extraction mainly include traditional image analysis methods and deep learning methods. Traditional methods, such as threshold segmentation, exponential methods, and water ratio methods, are simple to implement but are easily affected by factors such as water body obfuscation, shadows, topographic relief, and vegetation moisture, resulting in poor adaptability. Deep learning methods have been widely used in recent years. Among them, Convolutional Neural Networks (CNNs) can effectively extract local texture and boundary features, but due to their limited receptive field, they struggle to capture the overall distribution structure of flood areas on a large spatial scale. While Transformer models possess global modeling capabilities, their computational complexity increases quadratically with image size, making them difficult to adapt to high-resolution remote sensing imagery scenes. Furthermore, they lack sufficient local inductive bias, which can easily lead to unstable flood boundary predictions.

[0004] State-space models (SSMs), particularly selective state-space models like Mamba, have emerged in recent years and can process long-sequence data with linear time complexity, demonstrating significant advantages in long-distance dependency modeling. However, the application of SSMs in remote sensing imagery is still limited, and how to effectively combine them with convolutional features to balance local detail and global semantic modeling remains a pressing issue for the industry.

[0005] In summary, existing technologies still have the following shortcomings in the automatic extraction of flood impact areas:

[0006] (1) It is difficult to simultaneously take into account both local edge feature extraction and global region dependency modeling;

[0007] (2) High-resolution remote sensing images have high computational complexity and insufficient real-time performance;

[0008] (3) The feature fusion methods are limited, and it is impossible to fully aggregate multi-scale and multi-type features;

[0009] (4) It has poor robustness in complex terrain backgrounds (such as farmland, wetlands, and shade).

[0010] Therefore, it is necessary to design a flood impact range extraction method that can effectively integrate convolutional features and state space modeling capabilities to improve extraction accuracy and computational efficiency. Summary of the Invention

[0011] The purpose of this invention is to provide a method and system for extracting the impact range of floods based on convolution and state space models. By multi-level collaborative fusion of convolution features and state space modeling features, it achieves high-precision, robust and efficient automatic identification of flood-inundated areas, solving the problems of insufficient local detail extraction, poor long-distance dependency modeling, high computational complexity and insufficient robustness in existing technologies.

[0012] To achieve the above functions, this invention designs a flood impact range extraction method based on convolution and state-space models, executing the following steps S1-S5 to complete the extraction of the flood impact range:

[0013] Step S1: Acquire images of the target area and perform preprocessing to obtain standardized image blocks;

[0014] Step S2: Construct a hierarchical encoder with parallel local feature extraction branches and global context modeling branches. Standardized image blocks are input into the local feature extraction branches and global context modeling branches of the encoder, respectively. The local feature extraction branch is based on a convolutional neural network and outputs local features. The global context modeling branch is based on the Mamba state space model and adopts a dual-gated parallel architecture and gated fusion to output global features.

[0015] Step S3: Construct an adaptive feature fusion module to generate attention features based on channel attention and spatial attention; if the current stage is the first run, generate the fusion features of the current stage based on local features, global features, and attention features; otherwise, generate the fusion features of the current stage based on local features, global features, attention features, and the fusion features of the previous stage.

[0016] Step S4: Construct a decoder, perform upsampling and fusion of the fusion features output from each stage of the encoder, map the fusion features back to the spatial resolution of the standardized image blocks, and obtain the flood prediction map.

[0017] Step S5: For the flood prediction map, generate a pixel-level probability map, and perform thresholding and binary segmentation mask generation. The binary segmentation mask represents whether each pixel in the image is affected by the flood, and is used as the extracted flood impact range for output.

[0018] As a preferred technical solution of the present invention, the specific steps of step S1 are as follows:

[0019] Step S1.1: Acquire images of the target area, in the form of high-resolution optical remote sensing images or synthetic aperture radar images;

[0020] Step S1.2: Perform pixel-level registration on the image of the target area;

[0021] Step S1.3: For each pixel in the image of the target region, normalize the pixel value to the [0,1] range;

[0022] Step S1.4: The image of the target area is cropped into image blocks of a preset fixed size to obtain standardized image blocks.

[0023] As a preferred technical solution of the present invention, the specific steps of step S2 are as follows:

[0024] Step S2.1: Construct the input tensor from normalized image patches Input tensor The dimensions are (B,C,H,W), where B is the batch size, C is the number of channels, and H and W are the height and width of the standardized image block, respectively.

[0025] Step S2.2: Convert the input tensor Input local feature extraction branch, for input tensor The processing procedure is as follows:

[0026] ;

[0027] In the formula, This represents the local features output at the current stage. This represents a convolution operation with a kernel size of n. for Depth-separable convolution, for convolution, This represents the activation function. Representation layer normalization;

[0028] Step S2.3: Convert the input tensor Input the main path in the global context modeling branch, input tensor First, it undergoes hierarchical normalization and... Convolution, followed by the SS2D mechanism on the input tensor in the form of a two-dimensional image feature map. Perform a four-directional scan along the spatial directions of upper left to lower right, lower right to upper left, upper right to lower left, and lower left to upper right, respectively, and flatten it into a one-dimensional sequence; for each element at time t in the sequence... We use discretized state-space equations for recursive modeling, as shown in the following equation:

[0029] ;

[0030] ;

[0031] in, and These are the state parameters after zero-order hold-out discretization. , For continuous time parameters, Let be the hidden state at time t. Let be the hidden state at time t-1. The output at time t, To output the mapping matrix; The time step parameter is a learnable parameter. To and A dimensionally consistent identity matrix;

[0032] The output sequences generated by the four-way scanning are merged and spatially recombined to generate main path features with a global receptive field. ;

[0033] Step S2.4: Convert the input tensor Input the auxiliary path in the global context modeling branch, input tensor After a Convolutional layers, batch normalization layers, and the SiLU nonlinear activation function generate auxiliary path features. ;

[0034] Step S2.5: Analyze the main road features and auxiliary road characteristics Channel shuffling operations are performed separately, followed by gating fusion through feature multiplication; the gating fused features are then processed by the end-point... The convolution performs a final adjustment of the number of channels and is then compared with the input tensor. Perform residual connections to output the global features of the current stage. :

[0035] ;

[0036] In the formula, express convolution, This indicates a channel mixing operation. This indicates feature multiplication.

[0037] As a preferred technical solution of the present invention, the specific steps of step S3 are as follows:

[0038] Step S3.1: Based on local features Global features By utilizing channel attention and spatial attention, combined with bilinear pooling, attention features are generated. As shown in the following formula:

[0039] ;

[0040] In the formula, This is a bilinear pooling operation. and These are channel attention operations and spatial attention operations, respectively.

[0041] Step S3.2: If this is the first run, execute step S3.2.1; otherwise, execute step S3.2.2.

[0042] Step S3.2.1: Transfer local features Global features After splicing, dimensionality reduction and nonlinear transformation are performed to generate spliced ​​features. :

[0043] ;

[0044] in, For channel splicing operations;

[0045] splicing features and attention characteristics By splicing and merging, the fusion characteristics of the current stage are obtained. :

[0046] ;

[0047] In the formula, This represents a 3×3 convolution;

[0048] Step S3.2.2: Transfer local features Global features The integration characteristics of the previous stage After splicing, dimensionality reduction and nonlinear transformation are performed to generate spliced ​​features. :

[0049] ;

[0050] in, For splicing operations;

[0051] splicing features Attention characteristics The integration characteristics of the previous stage By splicing and merging, the fusion characteristics of the current stage are obtained. :

[0052] ;

[0053] In the formula, This represents a 3×3 convolution;

[0054] Step S3.3: Generate fusion features It is divided into two paths: one path outputs to the next stage as input; the other path is sent to the corresponding stage level of the decoder through a jump connection.

[0055] As a preferred technical solution of the present invention: In step S4, the fusion features output by the encoder in the i-th stage are... The decoder outputs the decoding features in stage i as follows: The decoder targets decoding features The specific processing steps are as follows:

[0056] Step S4.1: Decoding features of stage i Perform a 2x upsampling operation:

[0057] ;

[0058] In the formula, This is the upsampling processing function. This represents the decoded features after upsampling;

[0059] Step S4.2: Upsample the decoded features Fusion features from the corresponding stage of the encoder Perform splicing along the channel dimension:

[0060] ;

[0061] In the formula, Indicates splicing characteristics, For splicing operations;

[0062] Step S4.3: Add splicing features The features are then processed through a convolutional layer to obtain the next level of decoding features. :

[0063] ;

[0064] In the formula, It is an activation function. It is batch normalization. This represents a 3×3 convolution;

[0065] Step S4.4: Repeat steps S4.1-S4.3 until a decoded feature map with the same resolution as the input normalized image patch is obtained. This is output as a flood forecast map.

[0066] As a preferred embodiment of the present invention, the specific steps of step S5 are as follows:

[0067] Step S5.1: Decode the feature map output in step S4. Enter to Convolutional layers map multi-channel features to single-channel logistic values ​​L:

[0068] ;

[0069] In the formula, This represents a 1×1 convolution;

[0070] Step S5.2: Based on the logistic value L, generate a pixel-level probability map using the Sigmoid activation function:

[0071] ;

[0072] in, Representing coordinates in a probability graph The probability that a pixel at a given location is identified as part of a flood. It is an activation function. Representing coordinates The logical value at the location;

[0073] Step S5.3: Set the segmentation threshold Thresholding is performed on the probability map P, and the final binary segmentation mask M is output:

[0074] ;

[0075] In the formula, Represents coordinates in the image The pixels at that location were affected by the flood. Represents coordinates in the image The pixels at that location were not affected by the flood;

[0076] The binary segmentation mask M is used as the output to extract the flood impact range.

[0077] This invention also designs a flood impact range extraction system based on convolution and state space models, including a data acquisition module, a storage module, and a processing module, to implement the aforementioned flood impact range extraction method based on convolution and state space models.

[0078] The data acquisition module is used to acquire images of the target area and convert them into digital signals for transmission to the processing module;

[0079] The storage module is used to store the program instructions, algorithm model weights, historical and real-time flood forecast maps, and flood impact range required for system operation;

[0080] The processing module is used to call the instructions and parameters in the storage module to execute the flood impact range extraction method based on convolution and state space model.

[0081] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0082] (1) Convolutional branches ensure detailed representation, and the state space model realizes long-range dependency modeling. The two are dynamically coupled in the fusion module, so that the model has both spatial detail sensitivity and overall structure perception.

[0083] (2) The state space model adopts the SS2D mechanism, which maintains linear complexity and can model a wide range of contexts without significant increase in computational overhead.

[0084] (3) By using the residual-attention dual-path fusion strategy, interference signals from shadow areas, bare land, water bodies and other land features are effectively suppressed, thereby improving the accuracy and stability of flood disaster extraction.

[0085] (4) The encoder-decoder structure is easy to deploy on platforms such as servers and edge computing devices, and is suitable for emergency response and disaster monitoring scenarios. Attached Figure Description

[0086] Figure 1 This is a flowchart of a flood impact range extraction method based on convolution and state space model provided in an embodiment of the present invention;

[0087] Figure 2 This is a schematic diagram of an encoder-decoder structure provided according to an embodiment of the present invention;

[0088] Figure 3 This is a schematic diagram of the structure of the local feature extraction branch and the global context modeling branch provided in an embodiment of the present invention;

[0089] Figure 4 This is a flowchart illustrating the adaptive feature fusion module provided according to an embodiment of the present invention;

[0090] Figure 5 This is a schematic diagram of the structure of the adaptive feature fusion module provided in an embodiment of the present invention;

[0091] Figure 6 This is a schematic diagram of a flood impact range extraction system based on convolution and state space models provided in an embodiment of the present invention. Detailed Implementation

[0092] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0093] The flood impact range extraction method based on convolution and state-space models provided in this invention embodiment is described in reference to... Figure 1 Perform the following steps S1-S5 to complete the extraction of the flood impact area:

[0094] Step S1: Acquire images of the target area and perform preprocessing to obtain standardized image blocks;

[0095] The specific steps of step S1 are as follows:

[0096] Step S1.1: Acquire images of the target area, in the form of high-resolution optical remote sensing images or synthetic aperture radar (SAR) images; in this embodiment, the data source comes from Sentinel-1 / 2, Gaofen series satellites or UAV aerial photography.

[0097] Step S1.2: Ensure that the image has accurate geographic coordinates. If multi-temporal data is used, perform pixel-level registration for the image of the target area.

[0098] Step S1.3: For each pixel in the image of the target area, normalize the pixel value to the [0,1] range to eliminate the influence of atmosphere and lighting;

[0099] Step S1.4: Considering the limitations of computer video memory, the large-format remote sensing image needs to be cropped into image blocks of fixed size. The image of the target area is cropped into image blocks of a preset fixed size to obtain standardized image blocks. In this embodiment, the standardized image block is... or Pixels, and set an overlap rate of 25% to eliminate edge effects during stitching.

[0100] Step S2: Construct a hierarchical encoder with parallel local feature extraction branches and global context modeling branches. Standardized image blocks are input into the local feature extraction branches and global context modeling branches of the encoder, respectively. The local feature extraction branch is based on a convolutional neural network (CNN) and outputs local features. The global context modeling branch is based on the Mamba state space model (SSM) and uses gated fusion to output global features.

[0101] Reference Figure 2 The encoder adopts a hierarchical architecture, consisting of four cascaded stages; with increasing stage hierarchy... As the image deepens, the spatial resolution of the feature map decreases step by step (to 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original image, respectively), while the channel dimension increases step by step.

[0102] The specific steps of step S2 are as follows:

[0103] Step S2.1: Construct the input tensor from normalized image patches Input tensor The dimensions are (B,C,H,W), where B is the batch size, C is the number of channels, and H and W are the height and width of the standardized image block, respectively.

[0104] Step S2.2: Convert the input tensor Input local feature extraction branch, refer to Figure 3 Using convolutional neural networks to The local feature extraction branch first processes the data through a... Depthwise separable convolutions extract spatial details, followed by layer normalization (LN) and two... Convolution (including GELU activation function) enhances nonlinear representation and focuses on extracting high-frequency local details such as flood boundaries and water texture.

[0105] Local feature extraction branch on input tensor The processing procedure is as follows:

[0106] ;

[0107] In the formula, This represents the local features output at the current stage. This represents a convolution operation with a kernel size of n. for Depth-separable convolution, for The convolution (i.e., linear projection). This represents the activation function. Representation layer normalization;

[0108] Step S2.3: Global Context Modeling Branch utilizes the Mamba architecture to... To maintain local coherence of features while establishing global relationships, this branch employs a dual-gated parallel architecture. The specific processing logic is as follows:

[0109] Reference Figure 3 , input tensor Input the main path in the global context modeling branch, input tensor First, it undergoes hierarchical normalization and... Convolution, followed by the SS2D mechanism on the input tensor in the form of a two-dimensional image feature map. Perform a four-directional scan along the spatial directions of upper left to lower right, lower right to upper left, upper right to lower left, and lower left to upper right, respectively, and flatten it into a one-dimensional sequence; for each element at time t in the sequence... The time step element here refers to a feature vector in a one-dimensional sequence, which is derived from one or more pixels in the original two-dimensional image. It is recursively modeled using a discretized state-space equation, as shown in the following formula:

[0110] ;

[0111] ;

[0112] in, and These are the state parameters after zero-order hold (ZOH) discretization. , For continuous time parameters, convert the continuous time parameters , Convert to discrete parameters , Its purpose is to enable the discretized model to still approximate the global dependency modeling capability of a continuous system, while also being compatible with batch training. Let be the hidden state at time t. Let be the hidden state at time t-1. The output at time t, This is the output mapping matrix, used to map the hidden states to the output features of the current time step.

[0113] The and They represent the time steps respectively. Under the influence of continuous time parameters and The matrix after scaling; It is a learnable time step parameter that can adapt to the rhythm of different sequences (such as the size of patches in flash flood images). For matrix An identity matrix with consistent dimensions.

[0114] The continuous time parameter The state transition matrix in a continuous-time state-space model represents the dynamic characteristics of the hidden states as they change over time; the continuous-time parameters This is the input mapping matrix in the continuous-time state-space model, used to characterize the influence of input features on hidden state updates.

[0115] The output sequences generated by the four-way scanning are merged and spatially recombined to generate main path features with a global receptive field. The Global Receptive Field (GRF) refers to the field that the SSM mechanism can recursively compute through sequences, enabling the output features to... Each point in the image contains contextual information about the entire input image, making it suitable for capturing long-range dependencies such as the overall view of a river or a large flooded area.

[0116] Step S2.4: Convert the input tensor The auxiliary path in the global context modeling branch is used to compensate for the shortcomings of the state-space model in local inductive bias; input tensor After a Convolutional layers, batch normalization layers, and the SiLU nonlinear activation function generate auxiliary path features rich in local location information. ;

[0117] Step S2.5: In order to efficiently integrate the above two features, the main feature is... and auxiliary road characteristics Channel shuffling operations are performed separately, followed by gating fusion through feature multiplication; the gating fused features are then processed by the end-point... The convolution performs a final adjustment of the number of channels and is then compared with the input tensor. Perform residual connections to output the global features of the current stage. :

[0118] ;

[0119] In the formula, express convolution, This indicates a channel mixing operation. This indicates feature multiplication.

[0120] Step S3: Construct an adaptive feature fusion module (CMF Block, CNN-Mamba-Fuse), which generates attention features based on channel attention and spatial attention; if the current stage is the first run, generate the fusion features of the current stage based on local features, global features, and attention features; otherwise, generate the fusion features of the current stage based on local features, global features, attention features, and the fusion features of the previous stage.

[0121] The adaptive feature fusion module, acting as the convergence node in this stage, does not perform independent feature extraction but instead receives three input signals: local features of the current stage. Global characteristics at the current stage and the fusion features output from the previous stage (First stage without fusion features) The adaptive feature fusion module uses channel attention and spatial attention mechanisms to calculate weights, and combines them with bilinear pooling to fuse the multi-source features into the final output features for the current stage. .

[0122] Reference Figure 4 , Figure 5 The specific steps of step S3 are as follows:

[0123] Step S3.1: Based on local features Global features By utilizing channel attention (CA) and spatial attention (SA), combined with bilinear pooling to capture second-order feature interactions, attention features are generated. As shown in the following formula:

[0124] ;

[0125] In the formula, This is a bilinear pooling operation. and These are channel attention operations and spatial attention operations, respectively.

[0126] Step S3.2: If this is the first run, execute step S3.2.1; otherwise, execute step S3.2.2.

[0127] Step S3.2.1: Transfer local features Global features After splicing, dimensionality reduction and nonlinear transformation are performed to generate spliced ​​features. :

[0128] ;

[0129] in, For channel splicing operations;

[0130] splicing features and attention characteristics By splicing and merging, the fusion characteristics of the current stage are obtained. :

[0131] ;

[0132] In the formula, This represents a 3×3 convolution;

[0133] Step S3.2.2: Transfer local features Global features The integration characteristics of the previous stage After splicing, dimensionality reduction and nonlinear transformation are performed to generate spliced ​​features. :

[0134] ;

[0135] in, For splicing operations;

[0136] splicing features Attention characteristics The integration characteristics of the previous stage By splicing and merging, the fusion characteristics of the current stage are obtained. :

[0137] ;

[0138] In the formula, This represents a 3×3 convolution;

[0139] Step S3.3: Generate fusion features It is divided into two paths: one path outputs to the next stage as input; the other path is sent to the corresponding stage level of the decoder through a jump connection.

[0140] Step S4: Construct a decoder, perform upsampling and fusion of the fusion features output from each stage of the encoder, map the fusion features back to the spatial resolution of the standardized image blocks, and obtain the flood prediction map.

[0141] In step S4, refer to Figure 2 The decoder employs a bottom-up, step-by-step recovery strategy, targeting the fused features output by the encoder in stage i. The decoder outputs the decoding features in stage i as follows: The decoder targets decoding features The specific processing steps are as follows:

[0142] Step S4.1: Decoding features of stage i Perform a 2x upsampling operation to make its spatial resolution match that of the previous level of encoded features. Maintain consistency:

[0143] ;

[0144] In the formula, This is the upsampling processing function. This represents the decoded features after upsampling;

[0145] Step S4.2: Upsample the decoded features Fusion features from the corresponding stage of the encoder By concatenating along the channel dimension, this operation combines deep, strong semantic information with shallow, fine-grained edge information:

[0146] ;

[0147] In the formula, Indicates splicing characteristics, For splicing operations;

[0148] Step S4.3: Add splicing features The features are processed through a convolutional layer to fuse them and adjust the number of channels, resulting in the decoded features for the next stage. :

[0149] ;

[0150] In the formula, It is an activation function. It is batch normalization. This represents a 3×3 convolution;

[0151] Step S4.4: Repeat steps S4.1-S4.3 until a decoded feature map with the same resolution as the input normalized image patch is obtained. This is output as a flood forecast map.

[0152] Step S5: For the flood prediction map, generate a pixel-level probability map, and perform thresholding and binary segmentation mask generation. The binary segmentation mask represents whether each pixel in the image is affected by the flood, and is used as the extracted flood impact range for output.

[0153] The specific steps of step S5 are as follows:

[0154] Step S5.1: Decode the feature map output in step S4. Enter to Convolutional layers map multi-channel features to single-channel logistic values ​​L:

[0155] ;

[0156] In the formula, This represents a 1×1 convolution;

[0157] Step S5.2: Based on the logistic value L, generate a pixel-level probability map using the Sigmoid activation function:

[0158] ;

[0159] in, Representing coordinates in a probability graph The probability that a pixel at a given location is identified as part of a flood. It is an activation function. Representing coordinates The logical value at the location;

[0160] Step S5.3: Set the segmentation threshold In the embodiment, the following settings are provided: Thresholding is performed on the probability map P, and the final binary segmentation mask M is output:

[0161] ;

[0162] In the formula, Represents coordinates in the image The pixels at that location were affected by the flood. Represents coordinates in the image The pixels at that location were not affected by the flood;

[0163] The binary segmentation mask M is used as the output to extract the flood impact range.

[0164] This invention also provides a flood impact range extraction system based on convolution and state-space models, referring to... Figure 6 It includes a data acquisition module, a storage module, and a processing module to implement the flood impact range extraction method based on convolution and state space models.

[0165] The data acquisition module is used to acquire images of the target area and convert them into digital signals for transmission to the processing module;

[0166] The data acquisition module includes multi-source data interfaces, supporting access to satellite ground receiving stations (for acquiring remote sensing satellite imagery) and UAV image transmission links;

[0167] For flood emergency scenarios, the built-in image filtering unit automatically reads image metadata and removes optical images with cloud coverage exceeding 20%, or automatically performs speckle filtering preprocessing on synthetic aperture radar images.

[0168] The data acquisition module connects to the processing module via a gigabit Ethernet port or a USB 3.0 high-speed interface to ensure real-time transmission of high-resolution images;

[0169] The storage module is used to store the program instructions, algorithm model weights, historical and real-time flood forecast maps, and flood impact range required for system operation;

[0170] The storage module includes a program storage area, a model parameter area, and a data cache area, among which:

[0171] The program storage area stores preprocessing algorithm libraries (such as GDAL), deep learning inference engines (such as TensorRT or ONNX Runtime), and compiled code;

[0172] The model parameter area is stored in the model weight file (e.g., .pth or .engine format) that has been pre-trained on the flood disaster image dataset. The model weight file contains the convolution kernel weights of the local feature extraction branch, the discretized state parameters of the global context modeling branch, and the attention weights of the adaptive feature fusion module.

[0173] The data cache area uses a high-speed NVMe SSD solid-state drive to temporarily store standardized image blocks after slicing and the generated binary segmentation mask, ensuring that the I / O read and write speed is not less than 20MB / s;

[0174] The processing module is used to call the instructions and parameters in the storage module to execute the flood impact range extraction method based on convolution and state space model.

[0175] The hardware architecture of the processing module includes a CPU and a GPU. The CPU (such as an Intel Xeon or ARM architecture processor) is responsible for data preprocessing (geometric correction, slicing) and logic control, while the GPU (such as an NVIDIA RTX series or Jetson edge computing module) is responsible for parallel computing.

[0176] The execution logic of the processing module includes a loading phase, a parallel inference phase, and a post-processing and output phase, wherein:

[0177] During the loading phase, when the processing module is initialized, the model parameters in the storage module are loaded into the GPU memory;

[0178] During the parallel inference phase, for each standardized image patch in the input, the processing module utilizes the GPU's CUDA cores to perform convolution operations in the local feature extraction branch and SS2D sequence scanning operations in the global context modeling branch in parallel; subsequently, feature fusion is completed in the adaptive feature fusion module; finally, the flood prediction map is generated through the decoder.

[0179] In the post-processing and output stage, the CPU stitches together the flood forecast map output by the GPU to restore it into a complete geographic coordinate image, and performs thresholding processing to generate the final flood impact range vector file or binary segmentation mask.

[0180] The processing module also has an interactive interface, which can output the processing results to a large visualization screen via HDMI interface, or send them to the emergency command center via 5G / 4G communication module.

[0181] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for extracting the impact range of floods based on convolution and state-space models, characterized in that, Perform the following steps S1-S5 to complete the extraction of the flood impact area: Step S1: Acquire images of the target area and perform preprocessing to obtain standardized image blocks; Step S2: Construct a hierarchical encoder with parallel local feature extraction branches and global context modeling branches. Standardized image blocks are input into the local feature extraction branches and global context modeling branches of the encoder, respectively. The local feature extraction branch is based on a convolutional neural network and outputs local features. The global context modeling branch is based on the Mamba state space model and adopts a dual-gated parallel architecture and gated fusion to output global features. The specific steps of step S2 are as follows: Step S2.1: Construct the input tensor from normalized image patches Input tensor The dimensions are (B,C,H,W), where B is the batch size, C is the number of channels, and H and W are the height and width of the standardized image block, respectively. Step S2.2: Convert the input tensor Input local feature extraction branch, for input tensor The processing procedure is as follows: ； In the formula, This represents the local features output at the current stage. This represents a convolution operation with a kernel size of n. for Depth-separable convolution, for convolution, This represents the activation function. Representation layer normalization; Step S2.3: Convert the input tensor Input the main path in the global context modeling branch, input tensor First, it undergoes hierarchical normalization and... Convolution, followed by the SS2D mechanism on the input tensor in the form of a two-dimensional image feature map. Perform a four-directional scan along the spatial directions of upper left to lower right, lower right to upper left, upper right to lower left, and lower left to upper right, respectively, and flatten it into a one-dimensional sequence; for each element at time t in the sequence... We use discretized state-space equations for recursive modeling, as shown in the following equation: ； ； in, and These are the state parameters after zero-order hold-out discretization. , For continuous time parameters, Let be the hidden state at time t. Let be the hidden state at time t-1. The output at time t, To output the mapping matrix; The time step parameter is a learnable parameter. To and A dimensionally consistent identity matrix; The output sequences generated by the four-way scanning are merged and spatially recombined to generate main path features with a global receptive field. ; Step S2.4: Convert the input tensor Input the auxiliary path in the global context modeling branch, input tensor After a Convolutional layers, batch normalization layers, and the SiLU nonlinear activation function generate auxiliary path features. ; Step S2.5: Analyze the main road features and auxiliary road characteristics Channel shuffling operations are performed separately, followed by gating fusion through feature multiplication; the gating fused features are then processed by the end-point... The convolution performs a final adjustment of the number of channels and is then compared with the input tensor. Perform residual connections to output the global features of the current stage. : ； In the formula, express convolution, This indicates a channel mixing operation. Indicates feature multiplication; Step S3: Construct an adaptive feature fusion module to generate attention features based on channel attention and spatial attention; if the current stage is the first run, generate the fusion features of the current stage based on local features, global features, and attention features; otherwise, generate the fusion features of the current stage based on local features, global features, attention features, and the fusion features of the previous stage. Step S4: Construct a decoder, perform upsampling and fusion of the fusion features output from each stage of the encoder, map the fusion features back to the spatial resolution of the standardized image blocks, and obtain the flood prediction map. Step S5: For the flood prediction map, generate a pixel-level probability map, and perform thresholding and binary segmentation mask generation. The binary segmentation mask represents whether each pixel in the image is affected by the flood, and is used as the extracted flood impact range for output.

2. The method for extracting the impact range of floods based on convolution and state-space models according to claim 1, characterized in that, The specific steps of step S1 are as follows: Step S1.1: Acquire images of the target area, in the form of high-resolution optical remote sensing images or synthetic aperture radar images; Step S1.2: Perform pixel-level registration on the image of the target area; Step S1.3: For each pixel in the image of the target region, normalize the pixel value to the [0,1] range; Step S1.4: The image of the target area is cropped into image blocks of a preset fixed size to obtain standardized image blocks.

3. The method for extracting the impact range of floods based on convolution and state-space models according to claim 2, characterized in that, The specific steps of step S3 are as follows: Step S3.1: Based on local features Global features By utilizing channel attention and spatial attention, combined with bilinear pooling, attention features are generated. As shown in the following formula: ； In the formula, This is a bilinear pooling operation. and These are channel attention operations and spatial attention operations, respectively. Step S3.2: If this is the first run, execute step S3.2.1; otherwise, execute step S3.2.

2. Step S3.2.1: Transfer local features Global features After splicing, dimensionality reduction and nonlinear transformation are performed to generate spliced ​​features. : ； in, For channel splicing operations; splicing features and attention characteristics By splicing and merging, the fusion characteristics of the current stage are obtained. : ； In the formula, This represents a 3×3 convolution; Step S3.2.2: Transfer local features Global features The integration characteristics of the previous stage After splicing, dimensionality reduction and nonlinear transformation are performed to generate spliced ​​features. : ； in, For splicing operations; splicing features Attention characteristics The integration characteristics of the previous stage By splicing and merging, the fusion characteristics of the current stage are obtained. : ； In the formula, This represents a 3×3 convolution; Step S3.3: Generate fusion features It is divided into two paths: one path outputs to the next stage as input; the other path is sent to the corresponding stage level of the decoder through a jump connection.

4. The method for extracting the flood impact range based on convolution and state-space models according to claim 3, characterized in that, In step S4, the fused features output by the encoder in stage i are... The decoder outputs the decoding features in stage i as follows: The decoder targets decoding features The specific processing steps are as follows: Step S4.1: Decoding features of stage i Perform a 2x upsampling operation: ； In the formula, This is the upsampling processing function. This represents the decoded features after upsampling; Step S4.2: Upsample the decoded features Fusion features from the corresponding stage of the encoder Perform splicing along the channel dimension: ； In the formula, Indicates splicing characteristics, For splicing operations; Step S4.3: Add splicing features The features are then processed through a convolutional layer to obtain the decoding features for the next stage. : ； In the formula, It is an activation function. It is batch normalization. This represents a 3×3 convolution; Step S4.4: Repeat steps S4.1-S4.3 until a decoded feature map with the same resolution as the input normalized image patch is obtained. This is output as a flood forecast map.

5. The method for extracting the flood impact range based on convolution and state-space models according to claim 4, characterized in that, The specific steps of step S5 are as follows: Step S5.1: Decode the feature map output in step S4. Enter to Convolutional layers map multi-channel features to single-channel logistic values ​​L: ； In the formula, This represents a 1×1 convolution; Step S5.2: Based on the logistic value L, generate a pixel-level probability map using the Sigmoid activation function: ； in, Representing coordinates in a probability graph The probability that a pixel at a given location is identified as part of a flood. It is an activation function. Representing coordinates The logical value at the location; Step S5.3: Set the segmentation threshold Thresholding is performed on the probability map P, and the final binary segmentation mask M is output: ； In the formula, Represents coordinates in the image The pixels at that location were affected by the flood. Represents coordinates in the image The pixels at that location were not affected by the flood; The binary segmentation mask M is used as the output to extract the flood impact range.

6. A flood impact range extraction system based on convolution and state-space models, characterized in that, The system includes a data acquisition module, a storage module, and a processing module to implement the flood impact range extraction method based on convolution and state-space models as described in any one of claims 1-5. The data acquisition module is used to acquire images of the target area and convert them into digital signals for transmission to the processing module; The storage module is used to store the program instructions, algorithm model weights, historical and real-time flood forecast maps, and flood impact range required for system operation; The processing module is used to call the instructions and parameters in the storage module to execute the flood impact range extraction method based on convolution and state space model.

Images