A method for extracting urban flood inundation range based on remote sensing images
By combining optical remote sensing images and radar images, super-resolution convolutional neural network processing is used to solve the problem of insufficient timeliness and monitoring range in flood disaster monitoring, and efficient urban flood range monitoring is achieved.
Patent Information
- Application Number
- CN202411179155.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-27
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-08-27
AI Technical Summary
The existing technology cannot meet the timeliness and monitoring range requirements of flood disaster monitoring at the same time. The single use of polar orbit satellites or radars cannot fully meet the timeliness and effective monitoring range requirements of flood disaster monitoring work.
Combining optical remote sensing images with low spatial resolution and high timeliness and radar images with high spatial resolution and low temporal resolution, super-resolution processing is used to obtain a higher resolution after reconstruction Urban flooding range.
It has achieved urban flood range images during the disaster-stricken period that meet both time and spatial resolution, making up for the defects of long polar-orbit satellite revisiting cycle and low spatial resolution of optical satellites, and meeting the timeliness requirements of flood disaster monitoring.
Smart Images

Figure CN118691981B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to image processing technology, and in particular to a method for extracting urban flood inundation range based on remote sensing images. Background Art
[0002] Flood disasters are characterized by large affected areas, strong suddenness, and high degree of destruction, which pose a serious threat to people's lives and property. In recent years, affected by global climate change, urban flood disasters have occurred frequently, and have become one of the important natural disasters affecting urban development. It is of great significance to carry out urban flood disaster monitoring and formulate urban flood emergency plans.
[0003] The focus of flood disaster monitoring is the accurate identification of surface water information. For real-time statistics of flood-inundated areas, traditional methods mainly include manual field inspections, hydrological monitoring stations, drones, remote sensing technology, etc. The use of traditional manual survey methods to monitor flood disasters has the disadvantages of long cycles, dangers, and large resource consumption; satellite remote sensing technology has the advantages of wide coverage, high timeliness, and large amount of information, which can make up for the shortcomings of traditional technical means and has been widely used in flood disaster monitoring and assessment, water resources investigation and other fields.
[0004] The research on flood disaster monitoring based on polar-orbiting satellite data is relatively mature, and with the help of high-spatial-resolution images, it is possible to initially extract the flood inundation range. However, flood disasters generally occur when heavy rainfall events occur. Optical image data are blocked by clouds, making it difficult to obtain information about the underlying surface at the time of the disaster, limiting its ability to comprehensively monitor flood inundation conditions. Usually, a water body index model is used to extract small water bodies, but high-resolution satellite images generally only have visible light and near-infrared bands. The water body index model has data requirements for mid-infrared and short-wave infrared bands, and needs to be used in conjunction with low-resolution satellites to extract the flood inundation range.
[0005] In addition, the revisit period of polar-orbiting satellites is often several days, which may cause the lack of monitoring of urban flooding and inundation events, and cannot fully meet the timeliness and continuity requirements of flood disaster monitoring. Although synthetic aperture radar is not affected by clouds and precipitation and can obtain all-weather satellite observation data, its disadvantage is that the scanning range cannot cover all flooded areas, resulting in a limited scope of flood monitoring work.
[0006] Therefore, the use of polar-orbiting satellites or radars alone cannot fully meet the requirements of flood disaster monitoring for timeliness and effective monitoring range. Summary of the invention
[0007] Purpose of the invention: In view of the above problems, the purpose of the present invention is to provide a method for extracting urban flood inundation range based on remote sensing images.
[0008] Technical solution: A method for extracting urban flood inundation range based on remote sensing images of the present invention comprises:
[0009] The remote sensing images of Fengyun-4 satellite were obtained and preprocessed, and the preprocessed remote sensing images were divided into thin cloud areas and thick cloud areas using convolutional neural networks;
[0010] Homomorphic filtering method is used to remove thin clouds in thin cloud areas;
[0011] The replacement method is used to remove thick clouds in thick cloud areas;
[0012] Extract the water area in the remote sensing image after removing the clouds and obtain the flooded area of the first city;
[0013] Obtain radar images of the same area and time, use the maximum inter-class variance method to extract the water area in the radar image, and obtain the flooded range of the second city;
[0014] The first city inundation range and the second city inundation range are used to train a super-resolution convolutional neural network, and the trained super-resolution convolutional neural network is used to obtain a reconstructed higher-resolution urban flood inundation range.
[0015] Furthermore, the homomorphic filtering method is used to remove thin clouds in the thin cloud area, including:
[0016] The remote sensing image corresponding to the thin cloud area is decomposed into multiple single-band remote sensing images, and the thin clouds are removed from each band remote sensing image.
[0017] Furthermore, thin clouds are removed from each band of remote sensing images separately, including:
[0018] f s (x,y) = f i (x,y) × f r (x,y) ,
[0019] In the formula, the two-dimensional function form f s (x,y) Indicates the remote sensing image corresponding to the thin cloud area obtained, x , y Represents the position coordinates of each pixel on the two-dimensional remote sensing image. f i (x,y) Represents the incident component contained in the acquired remote sensing image, also known as the thin cloud illumination component. f r (x,y) Represents the reflection component contained in the acquired remote sensing image, also known as the ground reflection component;
[0020] The remote sensing image corresponding to the thin cloud area is logarithmically transformed to separate the thin cloud illumination component and the ground reflection component of the remote sensing image. The mathematical expression is:
[0021] Inf s ( x , y ) = Inf i ( x , y ) + Inf r ( x , y ) ,
[0022] The logarithmic transformation results Inf s ( x , y ) and then perform Fourier transform to reduce the thin cloud illumination component. The mathematical expression is:
[0023] F{Inf s (x, y)} = F{Inf i (x, y)} + F{Inf r (x, y)} ,
[0024] make S(u,v) = F{Inf s (x, y)} , I(u,v) = F{Inf i (x, y)} , R(u,v) = F{Inf r (x, y)} , then:
[0025] S(u,v) = I(u,v) + R(u,v) ,
[0026] The result of Fourier transformation is filtered through a high-pass filter. S(u,v) Low frequency components I(u,v) Suppression and high frequency components R(u,v) The mathematical expression for extraction is:
[0027] p(u,v) = S(u,v) × H(u,v) = I(u,v) × H(u,v) + R(u,v) × H(u,v) ,
[0028] In the formula, H(u,v) is the transfer function of the high-pass filter;
[0029] The result after high pass filteringp(u,v) Perform inverse Fourier transform, the mathematical expression is:
[0030] p(x, y) = F -1 {p(u,v)} = F -1 {I(u,v) × H(x, y)} + F -1 {R(u,v) × H(x, y)} ,
[0031] make i(x, y) = F -1 {I(u,v) × H(x, y)} , r(x, y) = F -1 {R(u,v) × H(x, y)} ,
[0032] Then we have: p(x, y) = i(x, y) + r(x, y) , the result after inverse Fourier transformation p(x, y) Perform exponential transformation, the mathematical expression is:
[0033] g(x, y) = exp{i(x, y)} + exp{r(x, y)} ,
[0034] make i 0 (x, y) = exp{i(x, y)} , r 0 (x, y) = exp{r(x, y)} ,in i 0 (x, y) represents the final thin cloud lighting component, r 0 (x, y) is the ground reflection component, then the remote sensing image expression after removing thin clouds using homomorphic filtering is:
[0035] g(x, y) = i 0 (x, y) + r 0 (x, y) .
[0036] Further, removing thick clouds on thick cloud areas by using a replacement method includes:
[0037] Obtain cloud-free remote sensing images and clouded remote sensing images taken by the Fengyun-4 satellite at different times or with different sensors in the same area, and perform geocoding to ensure that the coordinate systems of the two remote sensing images are consistent. On the registered remote sensing image, replace the cloud-blocked area of the clouded remote sensing image with the corresponding area of the cloud-free remote sensing image.
[0038] Furthermore, the water area in the remote sensing image after cloud removal is extracted, and the flooded range of the first city is obtained, including:
[0039] Based on the remote sensing image data after cloud removal, the normalized difference water index is calculated using the green band and the red band MNDWI , the expression is:
[0040] ,
[0041] In the formula, represents the reflectance of the green band, Represents the reflectivity in the mid-infrared band;
[0042] Will be greater than the threshold MNDWI 0 The remote sensing image of the target water body is divided into the target water body, otherwise it is divided into the background, and the remote sensing image area corresponding to the target water body is taken as the first urban flooding range;
[0043] Calculate the water area according to the target water body, the expression is:
[0044] ,
[0045] In the formula, P w and P total Respectively represent the number of water pixels and the number of all pixels in the study area. A w and A total They represent the water area in the study area and the total image area of the entire study area respectively.
[0046] Furthermore, the maximum inter-class variance method is used to extract the water area in the radar image, and the second city flooding range is obtained, including:
[0047] Convert the radar image to a grayscale image and normalize it, mapping the pixel values to the range of 0 to 255;
[0048] Count the number of pixels at each gray level in the grayscale image to form a grayscale histogram;
[0049] Calculate the total number of all pixels in the grayscale histogram. The expression is:
[0050] N = Σ( n i ),
[0051] In the formula, n i The grayscale in the grayscale histogram is i The total number of pixels, i The range is 0 to 255;
[0052] Calculate the probability of each gray level, the expression is:
[0053] p i = n i / N ,
[0054] Based on pixel value threshold t 0 The pixels in the grayscale histogram are divided into two categories, where the pixel value is less than t 0 The pixel set with is recorded as class A, otherwise it is recorded as class B;
[0055] Calculate the weight of class A separately p A (t) and the weight of class B p B (t) , the expressions are:
[0056] ,
[0057] ,
[0058] Calculate the grayscale mean of class A respectively m A (t) and the grayscale mean of class B m B (t) , the expressions are:
[0059] ,
[0060] ,
[0061] Calculate the grayscale mean of the entire image. The expression is:
[0062] ,
[0063] Calculate the between-class variance, the expression is:
[0064] ,
[0065] Iterates through all possible pixel value thresholds from 0 to 255 t 0 , calculate the inter-class variance σ² corresponding to each pixel value threshold, and find the pixel value threshold that maximizes the inter-class variance σ² t * ;
[0066] Use maximum pixel value threshold t *Binarize the grayscale image of the radar image, that is, t * Pixels with values greater than or equal to 0 are set to represent the background. t * The pixel is set to 255 to represent the target water body, and the radar image area corresponding to the target water body is used as the second city flooding range;
[0067] Calculate the water body area based on the target water body.
[0068] Further, using the flood range of the first city and the flood range of the second city to train a super-resolution convolutional neural network includes:
[0069] The flooded area of the first city is used as the input item of the super-resolution convolutional neural network, and the flooded area of the second city is used as the output item of the super-resolution convolutional neural network, and the super-resolution convolutional neural network is trained.
[0070] Furthermore, the urban flood inundation range reconstructed with higher resolution using the trained super-resolution convolutional neural network includes:
[0071] The Fengyun-4 satellite remote sensing images are input into the trained super-resolution convolutional neural network to obtain the reconstructed higher-resolution urban flooding range.
[0072] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:
[0073] The present invention combines optical remote sensing images with low spatial resolution and high timeliness with radar images with high spatial resolution and low temporal resolution, and uses a super-resolution convolutional neural network to perform super-resolution processing using a deep learning method to obtain images of urban flood ranges during the disaster period that can simultaneously meet temporal and spatial resolutions;
[0074] It meets the timeliness requirements of flood disaster monitoring work, and can effectively make up for the defects of long revisit period of polar orbit satellites and low spatial resolution of optical satellites in flood disaster monitoring work. It has good economy, practicality and flexibility. BRIEF DESCRIPTION OF THE DRAWINGS
[0075] Figure 1 This is a flowchart of a method for extracting urban flood inundation range based on remote sensing images;
[0076] Figure 2 This is a working principle diagram of the super-resolution convolutional neural network;
[0077] Figure 3 An overview of the super-resolution convolutional neural network. DETAILED DESCRIPTION
[0078] In order to make the objectives, technical solutions and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments.
[0079] The method for extracting urban flood inundation range based on remote sensing images described in this embodiment is as follows: Figure 1 As shown, including:
[0080] The remote sensing images of Fengyun-4 satellite were obtained and preprocessed, and the preprocessed remote sensing images were divided into thin cloud areas and thick cloud areas using convolutional neural networks;
[0081] Homomorphic filtering method is used to remove thin clouds in thin cloud areas;
[0082] The replacement method is used to remove thick clouds in thick cloud areas;
[0083] Extract the water area in the remote sensing image after removing the clouds, and obtain the flooded area of the first city;
[0084] Obtain radar images of the same area and time, use the maximum inter-class variance method to extract the water area in the radar image, and obtain the flooded range of the second city;
[0085] The first city inundation range and the second city inundation range are used to train a super-resolution convolutional neural network, and the trained super-resolution convolutional neural network is used to obtain a reconstructed higher-resolution urban flood inundation range.
[0086] In this example, the FY4A AGRI data of the National Satellite Meteorological Center is used as a remote sensing image, which has the advantages of wide coverage, high timeliness, and large amount of information, but it only has visible light and near-infrared bands, and cannot fully meet the dependence of the water index model on the mid-infrared and short-wave infrared bands. Taking the FY4A AGRI satellite sensor as an example, the resolution is 0.5 to 4km, so the satellite remote sensing image has the defect of insufficient spatial resolution. The spatial resolution of radar images is high, and the accurate extraction of flood inundation range can be achieved, but the revisit period is long. Taking the Sentinel-2 radar as an example, its revisit period is five days, which cannot meet the timeliness requirements of flood disaster monitoring. In order to make up for the lack of spatial resolution of the optical image of the FY4A satellite, the present invention proposes to combine it with the Sentinel-2 radar image with high spatial resolution but low temporal resolution, and use the deep learning method to perform super-resolution processing to obtain an image of the urban flood range in the disaster period that can simultaneously meet the temporal and spatial resolutions.
[0087] In one example, the preprocessing of the Fengyun-4 satellite remote sensing image includes normalization, cropping and other preprocessing operations. Then, the remote sensing images of thin clouds and thick clouds are classified through the convolutional neural network CNN. The preprocessed remote sensing image is input into the CNN. First, the input image is convolved through multiple convolution kernels to extract the local features of the image; then the feature map output by the convolution layer is pooled to reduce the dimension of the feature map and reduce the amount of calculation; the feature vector output by the pooling layer is input into the fully connected layer, and the classification of thin clouds and thick clouds is realized through the connection between multiple neurons and the action of the activation function; the classification result is output to determine whether the area in the input remote sensing image is thin cloud or thick cloud. It should be noted that before using the remote sensing image for classification, the convolutional neural network needs to be trained with the training set data so that it has the ability to classify thin clouds and thick clouds.
[0088] In one example, removing thin clouds in a thin cloud area using a homomorphic filtering method includes:
[0089] The remote sensing image corresponding to the thin cloud area is decomposed into multiple single-band remote sensing images, and the thin clouds are removed from each band remote sensing image.
[0090] Furthermore, thin clouds are removed from each band of remote sensing images separately, including:
[0091] The mathematical model of the remote sensing image corresponding to the thin cloud area is:
[0092] f s (x,y) = f i (x,y) × f r (x,y) ,
[0093] In the formula, the two-dimensional function form f s (x,y) Indicates the remote sensing image corresponding to the thin cloud area obtained. The value is a positive scalar value. x , y Represents the position coordinates of each pixel on the two-dimensional remote sensing image. f i (x,y) It indicates the incident component contained in the acquired remote sensing image. The incident component refers to the total amount of light source incident on the target area. It mainly includes the spectrum characteristics of the low-frequency part. The spatial range changes slowly. It is also called the thin cloud illumination component. f r (x,y)Represents the reflection component contained in the acquired remote sensing image. The reflection component refers to the total amount of light radiated by the ground objects in the target area, mainly the spectral characteristics of the high-frequency part, which contains the detailed information of the remote sensing image and is also called the ground reflection component. f r (x,y) Should satisfy 0< f r (x,y) <1. The process of removing thin clouds from remote sensing images is to remove f i (x,y) , seek f r (x,y) process.
[0094] To process cloud information, we must first convert the degradation model of thin clouds into the frequency domain. The remote sensing image corresponding to the thin cloud area is logarithmically transformed to separate the thin cloud illumination component and the ground reflection component of the remote sensing image. The mathematical expression is:
[0095] Inf s ( x , y ) = Inf i ( x , y ) + Inf r ( x , y ) ,
[0096] In order to reduce the influence of thin clouds, the illumination component of thin clouds can be reduced in the frequency domain. f i ( x , y ) . The result after logarithmic transformation Inf s ( x , y ) and then perform Fourier transform to reduce the thin cloud illumination component. The mathematical expression is:
[0097] F{Inf s (x, y)} = F{Inf i (x, y)} + F{Inf r (x, y)} ,
[0098] makeS(u,v) = F{Inf s (x, y)} , I(u,v) = F{Inf i (x, y)} , R(u,v) = F{Inf r (x, y)} , then:
[0099] S(u,v) = I(u,v) + R(u,v) ,
[0100] The result of Fourier transformation is filtered through a high-pass filter. S(u,v) Low frequency components I(u,v) Suppression and high frequency components R(u,v) The mathematical expression for extraction is:
[0101] p(u,v) = S(u,v) × H(u,v) = I(u,v) × H(u,v) + R(u,v) × H(u,v) ,
[0102] In the formula, H(u,v) is the transfer function of the high-pass filter;
[0103] The result after high pass filtering p(u,v) Perform inverse Fourier transform, the mathematical expression is:
[0104] p(x, y) = F -1 {p(u,v)} = F -1 {I(u,v) × H(x, y)} + F -1 {R(u,v) × H(x, y)} ,
[0105] make i(x, y) = F -1 {I(u,v) × H(x, y)} , r(x, y) = F -1 {R(u,v) × H(x, y)} , then:
[0106] p(x, y) = i(x, y) + r(x, y) ,
[0107] The result after inverse Fourier transformation p(x, y) Perform exponential transformation, the mathematical expression is:
[0108] g(x, y) = exp{i(x, y)} + exp{r(x, y)} ,
[0109] make i 0 (x, y)=exp {i(x, y)} , r 0 (x, y)=exp {r(x, y)} ,in i 0 (x, y) represents the thin cloud illumination component obtained after the above processing,r 0 (x, y) is the ground reflection component, then the remote sensing image expression after removing thin clouds using homomorphic filtering is:
[0110] g(x, y) = i 0 (x, y)+ r 0 (x, y) .
[0111] In one example, removing thick clouds from a thick cloud region using a replacement method includes:
[0112] Obtain cloud-free remote sensing images and clouded remote sensing images taken in the same area at different times or with different sensors, and perform geocoding to ensure that the coordinate systems of the two remote sensing images are consistent. On the registered remote sensing image, replace the cloud-blocked area of the clouded remote sensing image with the corresponding area of the cloud-free remote sensing image.
[0113] In one example, the water area in the remote sensing image after cloud removal is extracted to obtain the flooded area of the first city, including:
[0114] Based on the remote sensing image data after cloud removal, the normalized difference water index is calculated using the green band and the red band MNDWI , the expression is:
[0115] ,
[0116] In the formula, represents the reflectance of the green band, Represents the reflectivity in the mid-infrared band;
[0117] Will be greater than the threshold MNDWI 0 The remote sensing image of the target water body is divided into the target water body, otherwise it is divided into the background, and the remote sensing image area corresponding to the target water body is taken as the first urban flooding range;
[0118] Calculate the water area according to the target water body, the expression is:
[0119] ,
[0120] In the formula, P w and P total Respectively represent the number of water pixels and the number of all pixels in the study area. A w and A total They represent the water area in the study area and the total image area of the entire study area respectively.
[0121] In one example, the maximum inter-class variance method is used to extract the water area in the radar image, and the second city flooding range is obtained:
[0122] Convert the radar image to a grayscale image and normalize it, mapping the pixel values to the range of 0 to 255;
[0123] Count the number of pixels at each gray level in the grayscale image to form a grayscale histogram;
[0124] Calculate the total number of all pixels in the grayscale histogram. The expression is:
[0125] N = Σ( n i ),
[0126] In the formula, n i The grayscale in the grayscale histogram is i The total number of pixels, i The range is 0 to 255;
[0127] Calculate the probability of each gray level, the expression is:
[0128] p i = n i / N ,
[0129] Based on pixel value threshold t 0 The pixels in the grayscale histogram are divided into two categories, where the pixel value is less than t 0 The pixel set with is recorded as class A, otherwise it is recorded as class B;
[0130] Calculate the weight of class A separately p A (t) and the weight of class B p B (t) , the expressions are:
[0131] ,
[0132] ,
[0133] Calculate the grayscale mean of class A respectively m A (t) and the grayscale mean of class B m B (t) , the expressions are:
[0134] ,
[0135] ,
[0136] Calculate the grayscale mean of the entire image. The expression is:
[0137] ,
[0138] Calculate the between-class variance, the expression is:
[0139] ,
[0140] Iterates through all possible pixel value thresholds from 0 to 255 t 0 , calculate the inter-class variance σ² corresponding to each pixel value threshold, and find the pixel value threshold that maximizes the inter-class variance σ² t * ;
[0141] Use maximum pixel value threshold t * Binarize the grayscale image of the radar image, that is, t * Pixels with values greater than or equal to 0 are set to represent the background. t * The pixel is set to 255 to represent the target water body, and the radar image area corresponding to the target water body is used as the second city flooding range;
[0142] Calculate the water area according to the target water body, the expression is:
[0143] ,
[0144] In the formula, P w and P total Respectively represent the number of water pixels and the number of all pixels in the study area. A w and A total They represent the water area in the study area and the total image area of the entire study area respectively.
[0145] In one example, training a super-resolution convolutional neural network using a first city flooding range and a second city flooding range includes:
[0146] The flooded area of the first city is used as the input item of the super-resolution convolutional neural network, and the flooded area of the second city is used as the output item of the super-resolution convolutional neural network, and the super-resolution convolutional neural network is trained.
[0147] Based on the super-resolution convolutional neural network (SRCNN), low-resolution optical remote sensing image super-resolution is achieved. The implementation process of this method includes:
[0148] A low-resolution urban flooding range image extracted from the Fengyun-4 satellite remote sensing image is recorded as image Y. First, the bicubic interpolation method is used to enlarge image Y to the required size for preprocessing. The purpose of using SRCNN is to recover an image F(Y) from image Y that is as similar as possible to the ground truth high-resolution image X. For the sake of convenience, Y is called a low-resolution image, which has the same size as X:
[0149] (1) Patch extraction and representation:
[0150] Extract overlapping patches from the low-resolution image Y and represent each patch as a high-dimensional vector. These high-dimensional vectors include a set of feature maps, the number of which is equal to the dimension of the vector;
[0151] (2) Nonlinear mapping:
[0152] Each high-dimensional vector is nonlinearly mapped to another high-dimensional vector. Each mapped vector is conceptually a representation of a high-resolution patch. These vectors constitute another set of feature maps.
[0153] (3) Reconstruction:
[0154] The above high-resolution patch representations are aggregated to generate the final high-resolution image F(Y). This image F(Y) is similar to the ground truth high-resolution image X.
[0155] The urban flood inundation range reconstructed with higher resolution using the trained super-resolution convolutional neural network includes:
[0156] The Fengyun-4 satellite remote sensing images are input into the trained super-resolution convolutional neural network to obtain the reconstructed higher-resolution urban flooding range.
[0157] Specifically, Figures 2 to 3 As shown in the figure, the process of using SRCNN to achieve super-resolution of low-resolution optical remote sensing images is as follows:
[0158] (1) Assume that a low-resolution patch of size f1 × f1 is extracted from image Y, and the sparse coding solver projects the low-resolution patch onto a low-resolution dictionary. If the size of the low-resolution dictionary is n1, it is equivalent to applying n1 linear filters (f1×f1) to the input image, that is, convolving the image with a set of filters. The first layer is represented by operation F1, and the expression is:
[0159] F1(Y) = max (0,W1 * Y + B1),
[0160] Where W1 and B1 represent the corresponding filter and bias respectively, and * represents the convolution operation.
[0161] (2) Then, the sparse coding solver will iteratively process the n1 coefficients. The output of the sparse coding solver is n2 coefficients. In the case of sparse coding, n2 = n1. The n2 coefficients are the representation of the high-resolution patch. The second layer of nonlinear mapping is represented by F2, which is expressed as:
[0162] F2(Y) = max (0, W2 * F1(Y) + B2),
[0163] Where W2 and B2 represent the corresponding filter and bias respectively;
[0164] The n2 coefficients after sparse coding are projected onto another high-resolution dictionary to generate a high-resolution patch. Averaging the overlapping high-resolution patches is equivalent to performing linear convolution on the n2 feature maps. The third layer operation F3 defines a convolution layer to generate the final high-resolution image, expressed as:
[0165] F3(Y) = W3 * F2(Y) + B3,
[0166] Where W3 and B3 represent the corresponding filter and bias respectively.
[0167] (3) Training: Learning the end-to-end mapping function F requires estimating the network parameters θ = {W1; W2; W3; B1; B2; B3}, minimizing the loss between the reconstructed image F(Y; θ) and the corresponding ground truth high-resolution image X. Given a set of high-resolution images {Xi} and their corresponding low-resolution images {Yi}, the mean squared error MSE is used as the loss function, expressed as:
[0168] ,
[0169] In the formula, n is the number of training samples. Using MSE as the loss function is beneficial to obtain a higher peak signal-to-noise ratio PSNR and determine the mapping vector that can obtain the minimum MSE.
[0170] (4) Achieving super-resolution: Input the urban flooding range image with lower spatial resolution into SRCNN to obtain a reconstructed urban flooding range image with higher resolution.
Claims
1. A method for extracting urban flood inundation range based on remote sensing images, characterized in that: include: The remote sensing images of Fengyun-4 satellite were obtained and preprocessed, and the preprocessed remote sensing images were divided into thin cloud areas and thick cloud areas using convolutional neural networks. Homomorphic filtering method is used to remove thin clouds in thin cloud areas; The replacement method is used to remove thick clouds in thick cloud areas; Extract the water area in the remote sensing image after removing the clouds and obtain the flooded area of the first city; Obtain radar images of the same area and time, use the maximum inter-class variance method to extract the water area in the radar image, and obtain the flooded range of the second city; The first city flooding range and the second city flooding range are used to train a super-resolution convolutional neural network, and the trained super-resolution convolutional neural network is used to obtain a reconstructed higher-resolution urban flooding range; Extract the water area in the remote sensing image after removing the cloud, and obtain the flooded range of the first city, including: Based on the remote sensing image data after cloud removal, the normalized difference water index is calculated using the green band and the red band MNDWI , the expression is: , In the formula, represents the reflectance of the green band, Represents the reflectivity in the mid-infrared band; Will be greater than the threshold MNDWI 0 The remote sensing image of the target water body is divided into the target water body, otherwise it is divided into the background, and the remote sensing image area corresponding to the target water body is taken as the first urban flooding range; Calculate the water area according to the target water body, the expression is: , In the formula, P w and P total Respectively represent the number of water pixels and the number of all pixels in the study area. A w and A total They represent the water area in the study area and the total image area of the entire study area respectively; The maximum inter-class variance method is used to extract the water area in the radar image, and the second city flooding range is obtained, including: Convert the radar image to a grayscale image and normalize it, mapping the pixel values to the range of 0 to 255; Count the number of pixels at each gray level in the grayscale image to form a grayscale histogram; Calculate the total number of all pixels in the grayscale histogram. The expression is: N = Σ( n i ), In the formula, n i The grayscale in the grayscale histogram is i The total number of pixels, i The range is 0 to 255; Calculate the probability of each gray level, the expression is: p i = n i / N , Based on pixel value threshold t 0 The pixels in the grayscale histogram are divided into two categories, where the pixel value is less than t 0 The pixel set with is recorded as class A, otherwise it is recorded as class B; Calculate the weight of class A separately p A (t) and the weight of class B p B (t) , the expressions are: , , Calculate the grayscale mean of class A respectively m A (t) and the grayscale mean of class B m B (t) , the expressions are: , , Calculate the grayscale mean of the entire image. The expression is: , Calculate the between-class variance, the expression is: , Iterates through all possible pixel value thresholds from 0 to 255 t 0 , calculate the inter-class variance σ² corresponding to each pixel value threshold, and find the pixel value threshold that maximizes the inter-class variance σ² t * ; Use maximum pixel value threshold t * Binarize the grayscale image of the radar image, that is, t * Pixels with values greater than or equal to 0 are set to represent the background. t * The pixel is set to 255 to represent the target water body, and the radar image area corresponding to the target water body is used as the second city flooding range; Calculate the water body area based on the target water body.
2. The method for extracting urban flood inundation range based on remote sensing images according to claim 1 is characterized in that: The homomorphic filtering method is used to remove thin clouds in the thin cloud area, including: The remote sensing image corresponding to the thin cloud area is decomposed into multiple single-band remote sensing images, and the thin clouds are removed from each band remote sensing image.
3. The method for extracting urban flood inundation range based on remote sensing images according to claim 2 is characterized in that: Removing thin clouds from each band of remote sensing images includes: The mathematical model of the remote sensing image corresponding to the thin cloud area is: f s (x,y)= f i (x,y)×f r (x,y) , In the formula, the two-dimensional function form f s (x,y) Indicates the remote sensing image corresponding to the thin cloud area obtained, x , y Represents the position coordinates of each pixel on the two-dimensional remote sensing image. f i (x,y) Represents the incident component contained in the acquired remote sensing image, also known as the thin cloud illumination component. f r (x,y) Represents the reflection component contained in the acquired remote sensing image, also known as the ground reflection component; The remote sensing image corresponding to the thin cloud area is logarithmically transformed to separate the thin cloud illumination component and the ground reflection component of the remote sensing image. The mathematical expression is: Inf s ( x , y ) = Inf i ( x , y ) + Inf r ( x , y ) , The logarithmic transformation results Inf s ( x , y ) and then perform Fourier transform to reduce the thin cloud illumination component. The mathematical expression is: F{Inf s (x, y)} = F{Inf i (x, y)}+ F{Inf r (x, y)} , make S(u,v) = F{Inf s (x, y)} , I(u,v)= F{Inf i (x, y)} , R(u,v)= F{Inf r (x, y)} , then: S(u,v) = I(u,v) + R(u,v) , The result of Fourier transformation is filtered through a high-pass filter. S(u,v) Low frequency components I(u,v) Suppression and high frequency components R(u,v) The mathematical expression for extraction is: p(u,v)=S(u,v) ×H(u,v)=I(u,v) ×H(u,v)+ R(u,v) ×H(u,v) , In the formula, H(u,v) is the transfer function of the high-pass filter; The result after high pass filtering p(u,v) Perform inverse Fourier transform, the mathematical expression is: p(x, y) = F -1 {p(u,v)} = F -1 {I(u,v)× H(x, y)}+ F -1 {R(u,v)×H(x, y)} , make i(x, y) = F -1 {I(u,v)× H(x, y)} , r(x, y)=F -1 {R(u,v)×H(x, y)} , then: p(x, y) = i(x, y) + r(x, y) , The result after inverse Fourier transformation p(x, y) Perform exponential transformation, the mathematical expression is: g(x, y) =exp {i(x, y)} + exp {r(x, y)} , make i 0 (x, y)=exp {i(x, y)} , r 0 (x, y)=exp {r(x, y)} ,in i 0 (x, y) represents the final thin cloud lighting component, r 0 (x, y) is the ground reflection component, then the remote sensing image expression after removing thin clouds using homomorphic filtering is: g(x, y) = i 0 (x, y)+ r 0 (x, y) 。 4. The method for extracting urban flood inundation range based on remote sensing images according to claim 1 is characterized in that: The replacement method to remove thick clouds in thick cloud areas includes: Obtain cloud-free remote sensing images and clouded remote sensing images taken by the Fengyun-4 satellite at different times or with different sensors in the same area, and perform geocoding to ensure that the coordinate systems of the two remote sensing images are consistent. On the registered remote sensing image, replace the cloud-blocked area of the clouded remote sensing image with the corresponding area of the cloud-free remote sensing image.
5. The method for extracting urban flood inundation range based on remote sensing images according to claim 1 is characterized in that: Training a super-resolution convolutional neural network using the first city flooding range and the second city flooding range includes: The flooded area of the first city is used as the input item of the super-resolution convolutional neural network, and the flooded area of the second city is used as the output item of the super-resolution convolutional neural network, and the super-resolution convolutional neural network is trained.
6. The method for extracting urban flood inundation range based on remote sensing images according to claim 5 is characterized in that: The urban flood inundation range reconstructed with higher resolution using the trained super-resolution convolutional neural network includes: The Fengyun-4 satellite remote sensing images are input into the trained super-resolution convolutional neural network to obtain the reconstructed higher-resolution urban flooding range.
