Flood disaster water extraction method, device, equipment and storage medium

Abstract

The present invention provides a method and device, equipment and storage medium for extracting water bodies in flood disasters, wherein the method comprises: obtaining a SAR image of a water body to be extracted, wherein the SAR image of the water body to be extracted comprises an image composed of a dual polarization band and a spectral characteristic band of a flood disaster SAR image; inputting the SAR image of the water body to be extracted into a preset flood extraction convolutional neural network to obtain a distribution map of water body information in flood disasters; wherein the preset flood extraction convolutional neural network is used to perform residual convolution and attention mechanism correction on the SAR image of the water body to be extracted, and perform hole convolution on the corrected SAR image of the water body to be extracted to obtain the distribution map of water body information in flood disasters. The present invention can achieve the purpose of extracting water body information based on SAR images of flood disasters with high precision and high accuracy, and effectively improves the extraction precision and accuracy of water bodies in flood disasters.

Description

Technical Field

[0001] The invention relates to the technical field of natural disaster remote sensing, and in particular to a method and device for extracting water bodies from flood disasters, equipment and a storage medium. Background Art

[0002] Flood disasters are one of the most serious natural disasters in the world, causing a large number of casualties and economic losses around the world. Especially in my country, flood disasters occur frequently, seriously threatening the development of my country's national economy and the safety of people's lives and property. Therefore, timely and accurate acquisition of the spatiotemporal distribution of flood inundation is of great significance for relevant departments to formulate disaster prevention and relief plans and post-disaster loss assessment.

[0003] In traditional technology, synthetic aperture radar (SAR) and deep learning models are usually combined to extract water bodies from optical images of flood disasters. Initially, convolutional neural networks were used to extract water bodies. Then, urban surface water bodies in high-resolution remote sensing images were extracted based on convolutional neural network models, and higher water body extraction accuracy was achieved. Later, a multi-scale lake water body extraction network model was used to extract small lake water bodies.

[0004] However, since the research on traditional technology for extracting water bodies mainly focuses on the optical images of flood disasters, and water bodies are difficult to distinguish from shadows, small water bodies and rivers are difficult to extract completely, and the extraction accuracy is not high, the extraction accuracy and precision of flood disaster water bodies are not high. Summary of the invention

[0005] The present invention provides a method and device, equipment and storage medium for extracting water bodies caused by flood disasters, so as to solve the defects of low extraction precision and accuracy of water bodies caused by flood disasters in the prior art, and achieve the purpose of extracting water body information with high precision and high accuracy based on SAR images of flood disasters.

[0006] The present invention provides a method for extracting flood disaster water, comprising:

[0007] Acquire a SAR image of a water body to be extracted, wherein the SAR image of the water body to be extracted includes an image composed of a dual polarization band and a spectral characteristic band of a flood disaster SAR image;

[0008] Inputting the SAR image of the water body to be extracted into a preset flood extraction convolutional neural network to obtain a distribution map of flood disaster water body information;

[0009] Among them, the preset flood extraction convolutional neural network is used to perform residual convolution and attention mechanism correction on the SAR image of the water body to be extracted, and obtain the flood disaster water body information distribution map after performing void convolution on the corrected SAR image of the water body to be extracted.

[0010] According to a method for extracting water bodies caused by flood disasters provided by the present invention, the preset flood extraction convolutional neural network includes a preset residual neural network and a preset hole convolutional network, and the SAR image of the water body to be extracted is input into the preset flood extraction convolutional neural network to obtain a distribution map of water body information caused by flood disasters, including:

[0011] The SAR image of the water body to be extracted is input into a preset residual neural network for residual convolution and attention mechanism correction, so as to obtain a first water body information distribution map after precise correction of the edge of the flood disaster water body;

[0012] The first water body information distribution map is input into a preset void convolutional network for change processing of a preset expansion scale, so as to obtain a complete flood disaster water body information distribution map with flood disaster water body edge extraction.

[0013] According to a flood disaster water body extraction method provided by the present invention, the preset residual neural network includes P residual sub-networks, and each of the residual sub-networks is respectively connected to an attention mechanism correction sub-network, the SAR image of the water body to be extracted is input into the preset residual neural network for residual convolution and attention mechanism correction, and a first water body information distribution map after the flood disaster water body edge is accurately corrected is obtained, including:

[0014] Inputting the SAR image to be extracted from the water body into the qth residual subnetwork for convolution, batch normalization and linear sorting to obtain the qth residual convolution image;

[0015] Inputting the qth residual convolution image into the qth attention mechanism correction subnetwork for channel correction and spatial correction processing to obtain the qth corrected image;

[0016] Input the qth corrected image into the q+1th residual subnetwork for convolution, batch normalization and linear sorting to obtain the q+1th residual convolution image;

[0017] Input the q+1th residual convolution image into the q+1th attention mechanism correction subnetwork for channel correction and spatial correction processing to obtain the q+1th corrected image;

[0018] Let the value of q be increased by 1, and repeat the step of inputting the q+1th residual convolution image into the q+1th attention mechanism correction subnetwork for channel correction and spatial correction processing until the Pth corrected image is obtained, and the Pth corrected image is the first water body information distribution map after the precise correction of the edge of the flood disaster water body; q∈[1,…,P], q+1∈[2,…,P].

[0019] According to a flood disaster water body extraction method provided by the present invention, when the preset void convolution network includes Q different expansion rates, the first water body information distribution map is input into the preset void convolution network for change processing of the preset expansion scale to obtain a flood disaster water body information distribution map with a complete flood disaster water body edge extraction, including:

[0020] Inputting the first water body information distribution map into a preset atrous convolutional network for processing at different expansion scales, and obtaining Q small water body feature information maps at different expansion scales;

[0021] Performing fusion processing on the small water body feature information map to obtain a fused image;

[0022] The fused image is decoded corresponding to the residual convolution and attention mechanism correction to obtain a complete flood disaster water body information distribution map with flood disaster water body edge extraction.

[0023] According to a flood disaster water body extraction method provided by the present invention, the training process of the preset flood extraction convolutional neural network includes:

[0024] Acquire a training sample image set; the training sample image set includes training sample images, and the training sample images are SAR images obtained after sample annotation, regular grid cropping and data enhancement processing are performed on the SAR image of the water body to be extracted;

[0025] The preset initial flood extraction convolutional neural network is trained according to the training sample image set to obtain the preset flood extraction convolutional neural network.

[0026] According to a flood disaster water body extraction method provided by the present invention, the preset initial flood extraction convolutional neural network is trained according to the training sample image set to obtain the preset flood extraction convolutional neural network, including:

[0027] Iteratively training a preset initial flood extraction convolutional neural network according to the training sample image set to obtain an intermediate flood extraction convolutional neural network obtained after this round of iterative training;

[0028] Determine whether the cumulative number of training rounds corresponding to the current round of iterative training reaches a preset round number threshold;

[0029] If the accumulated number of training rounds reaches the preset round number threshold, the intermediate flood extraction convolutional neural network obtained after the current round of iterative training is used as the preset flood extraction convolutional neural network;

[0030] If the accumulated number of training rounds does not reach the preset round number threshold, the intermediate flood extraction convolutional neural network is trained to obtain the preset flood extraction convolutional neural network.

[0031] According to a flood disaster water body extraction method provided by the present invention, the process of determining the preset round number threshold comprises:

[0032] Dividing the training sample image set into training samples and verification samples according to a preset ratio;

[0033] Training a preset initial flood extraction convolutional neural network according to the training samples to obtain an intermediate flood extraction convolutional neural network obtained after a preset number of rounds of training;

[0034] Verifying the intermediate flood extraction convolutional neural network according to the verification sample to obtain a value of an evaluation index of the intermediate flood extraction convolutional neural network;

[0035] Determine whether the value of the evaluation index reaches a preset standard value;

[0036] If the value of the evaluation index reaches the preset standard value, the accumulated number of training rounds corresponding to the intermediate flood extraction convolutional neural network is used as the preset round number threshold;

[0037] If the value of the evaluation index does not reach the preset standard value, the intermediate flood extraction convolutional neural network is trained to obtain the preset round number threshold.

[0038] The present invention also provides a flood disaster water extraction device, comprising:

[0039] An acquisition module is used to acquire a SAR image of a water body to be extracted, wherein the SAR image of the water body to be extracted includes an image composed of a dual polarization band and a spectral characteristic band of a flood disaster SAR image;

[0040] A determination module is used to input the SAR image of the water body to be extracted into a preset flood extraction convolutional neural network to obtain a distribution map of flood disaster water body information;

[0041] Among them, the preset flood extraction convolutional neural network is used to perform residual convolution and attention mechanism correction on the SAR image of the water body to be extracted, and obtain the flood disaster water body information distribution map after performing void convolution on the corrected SAR image of the water body to be extracted.

[0042] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of any of the above-mentioned methods for extracting water bodies from flood disasters are implemented.

[0043] The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of any of the above-mentioned methods for extracting flood disaster water bodies are implemented.

[0044] The present invention provides a method and device, equipment and storage medium for extracting water bodies in flood disasters, wherein the method for extracting water bodies in flood disasters first uses an image composed of a dual polarization band and a spectral characteristic band of a flood disaster SAR image as a SAR image of a water body to be extracted, and then obtains a distribution map of water body information in flood disasters by inputting the SAR image of the water body to be extracted into a preset flood extraction convolutional neural network for water body extraction. Since the preset flood extraction convolutional neural network is used to perform residual convolution and attention mechanism correction on the SAR image of the water body to be extracted, and to perform hole convolution on the corrected SAR image of the water body to be extracted to obtain a distribution map of water body information in flood disasters, it is possible to achieve the purpose of accurately correcting and reliably extracting the water body edge with obvious features of the SAR image of the water body to be extracted, and further capturing the feature information of the water body edge with more details and unclear features, so that the water body information in the obtained distribution map of water body information in flood disasters is richer and more complete, thereby effectively improving the extraction precision and accuracy of water bodies in flood disasters. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0046] Figure 1 It is a schematic diagram of the process of extracting flood disaster water provided by the present invention;

[0047] Figure 2 It is a schematic diagram of a flood disaster water body information distribution map obtained based on the method of the present invention provided by the present invention;

[0048] Figure 3 It is a schematic diagram of a process of performing convolution, batch normalization and linear arrangement processing on each residual sub-network provided by the present invention;

[0049] Figure 4 It is a schematic diagram of the flow of each attention mechanism correction sub-network provided by the present invention performing channel correction and space correction processing;

[0050] Figure 5 It is a schematic diagram of a preset dilated convolution network provided by the present invention performing dilated convolutions with different dilation rates;

[0051] Figure 6 It is a schematic diagram of the accuracy inspection area extracted by the method of the present invention provided by the present invention;

[0052] Figure 7 It is a schematic structural diagram of a flood disaster water extraction device provided by the present invention;

[0053] Figure 8 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] Flood disasters are one of the most serious natural disasters in the world, causing a large number of casualties and economic losses around the world. According to the "China Flood and Drought Bulletin" issued by the Ministry of Water Resources, between 1998 and 2019, flood disasters caused an average annual direct economic loss of 168.288 billion yuan and 1,288 deaths in my country; it can be seen that flood disasters have caused huge losses to my country's economic development and people's lives. Therefore, timely and accurate acquisition of the spatiotemporal distribution range of flood inundation is of great significance for relevant departments to formulate disaster prevention and relief plans and post-disaster loss assessments.

[0056] At present, flood monitoring mainly includes hydrological station monitoring method and remote sensing monitoring method. Among them, the hydrological station monitoring method has high accuracy, but it has shortcomings such as small number of stations and inability to monitor spatial distribution range; remote sensing technology has the advantages of short revisit time and ability to monitor spatial distribution range, and it has gradually become the main method for flood monitoring. However, since floods are usually accompanied by bad weather, traditional optical sensors are affected by clouds and rain, and it is difficult to provide cloud-free high-quality optical images. On the contrary, SAR has the ability to monitor all day and all weather, and it is not easily affected by bad weather such as clouds and rain, so it plays an increasingly important role in flood monitoring and extraction.

[0057] The flood monitoring and extraction methods based on SAR data mainly include threshold method, object-oriented method, active contour method and data fusion method. However, due to the shortcomings of SAR images such as speckle noise and uneven grayscale distribution, the traditional SAR data flood monitoring method cannot meet the requirements well in large-area flood monitoring when faced with large data volumes and complex flood inundation scenes. In this context, combining SAR data with intelligent water body extraction algorithms has gradually become a research hotspot.

[0058] Furthermore, deep learning algorithms are mainly used in tasks such as semantic segmentation, target detection and image classification in the field of remote sensing, and have achieved good results. This is because deep learning methods avoid complex feature selection processes and can directly obtain features from original images, so they have strong applicability in the process of extracting information from multi-band remote sensing images. In 2015, Long et al. proposed a semantic segmentation method based on the fully convolutional network FCN, which solved the problem of image segmentation at the semantic level and classified the image at the pixel level; FCN is the first convolutional neural network used for semantic segmentation, but the network has shortcomings such as overly simple upsampling operations and loss of detail information. Later, Ronneberger et al. improved the FCN network and proposed a U-Net network with an encoding and decoding structure. The network can integrate low-resolution and high-resolution features, which greatly improves the accuracy of image segmentation. Since then, scholars have proposed many classic semantic segmentation models to improve the accuracy and performance of semantic segmentation, such as Deeplab v3, UNet++, HRNet and other models.

[0059] In recent years, deep learning models have gradually been applied to water body extraction from remote sensing images. For example, Chen Qian et al. used convolutional neural networks to extract water bodies. By comparing with traditional extraction methods such as NDWI, they proved the effectiveness of deep learning methods in extracting water bodies. Chen et al. designed a new convolutional neural network model to extract urban surface water bodies from high-resolution remote sensing images and achieved high water body extraction accuracy. Wang et al. proposed a multi-scale lake water body extraction network model that can effectively extract small lake water bodies.

[0060] However, existing research on water body extraction using deep learning models mainly focuses on optical images, while research on SAR images is relatively rare. In addition, there are some shortcomings, such as difficulty in distinguishing water bodies from shadows, difficulty in completely extracting small water bodies and rivers, and low extraction accuracy. Similar to optical remote sensing images, the key to flood monitoring using SAR images also lies in the identification and extraction of water body information.

[0061] Based on the above problems, the present invention provides a method for extracting water bodies caused by flood disasters. The execution subject of the method can be a flood disaster water body extraction device. The flood disaster water body extraction device can be implemented as part or all of a terminal device through software, hardware, or a combination of software and hardware. Optionally, the terminal device can be a personal computer (PC), a portable device, a laptop, a smart phone, a tablet computer, a portable wearable device, and other electronic devices, such as a tablet computer, a mobile phone, etc. The present invention does not limit the specific form of the terminal device.

[0062] It should be noted that the execution subject of the following method embodiment may be part or all of the above terminal device. The following method embodiment is described by taking the execution subject as the terminal device as an example.

[0063] Figure 1 The schematic diagram of the process of extracting flood water provided by the present invention is as follows: Figure 1 As shown, the flood disaster water extraction method comprises the following steps:

[0064] Step S110: Acquire a SAR image of a water body to be extracted, wherein the SAR image of the water body to be extracted includes an image composed of a dual-polarization band and a spectral characteristic band of a flood disaster SAR image.

[0065] Specifically, when the terminal device acquires the SAR image of the water body to be extracted, it can first acquire the SAR image for flood disasters, and use the SNAP software to perform orbit correction, thermal noise removal, radiation calibration, coherent speckle filtering, terrain correction, decibelization, cropping and mosaicking and other pre-processing operations on the SAR image to obtain the backscattering distribution map of the SAR image, and then further acquire the dual-polarization bands of the backscattering distribution map, which may include the VH band and the VV band. Then, Python programming is used to obtain the spectral feature SDWI band derived from the polarization of the VH band and the VV band. Finally, the three-band image composed of the VH band, the VV band and the SDWI band is used as the SAR image of the water body to be extracted.

[0066] It should be noted that the electric field vector of the energy pulse emitted by the radar can be polarized in the vertical or horizontal plane. Moreover, regardless of the wavelength, the radar signal can transmit a horizontal (H) or vertical (V) electric field vector, and receive a horizontal (H) or vertical (V) or both return signals. Therefore, the radar remote sensing system for SAR radar usually adopts four polarization modes: HH, VV, HV and VH. HH and VV are co-polarizations, and HV and VH are anisotropic (or cross) polarizations.

[0067] Step S120: input the water body SAR image to be extracted into a preset flood extraction convolutional neural network to obtain a flood disaster water body information distribution map.

[0068] Among them, the preset flood extraction convolutional neural network is used to perform residual convolution and attention mechanism correction on the SAR image of the water body to be extracted, and obtain the flood disaster water body information distribution map after performing void convolution on the corrected SAR image of the water body to be extracted.

[0069] Specifically, when the terminal device obtains the SAR image of the water body to be extracted, it can first extract the water body information with obvious features in the SAR image of the water body to be extracted, such as a wide lake surface, a small pond, etc., according to the pre-set feature extraction algorithm and feature expansion algorithm, and then further extract more water body information with less obvious features on the basis of ensuring the extraction accuracy and extraction precision, such as small rivers with more details, and tiny ditches that are difficult to identify, so as to obtain a complete distribution map of flood disaster water body information. When the SAR image of the water body to be extracted is determined based on the Sentinel-1SAR image of July 26, 2020 during the flood disaster, the following can be obtained: Figure 2 The distribution map of water body information for flood disasters is shown.

[0070] It should be noted that in order to improve the accuracy and precision of extraction, the feature extraction algorithm used in the present invention includes a deep residual network combined with an attention mechanism algorithm. This is because the deep residual network can reduce the complexity of the network model and improve the processing speed of the network model, and the attention mechanism can not only extract more semantic features, but also enable the terminal device to have the function of ignoring irrelevant information and focusing on key information like the human brain. Therefore, the present invention uses a feature extraction algorithm to accurately and accurately extract all water body information with obvious water features in the SAR image to be extracted, such as wide lakes, small water bodies, ponds and other water body information with obvious features that are easily submerged during floods. Optionally, the deep residual network can be a deep residual network (Deep residual network, ResNet) 18, and the attention mechanism can be a concurrent spatial compression and channel excitation (Concurrent Spatial and Squeeze and Channel Excitation, Scse) attention mechanism.

[0071] The method for extracting water bodies in flood disasters provided by the present invention first uses an image composed of a dual polarization band and a spectral characteristic band of a flood disaster SAR image as a SAR image of water bodies to be extracted, and then obtains a distribution map of water body information in flood disasters by inputting the SAR image of water bodies to be extracted into a preset flood extraction convolutional neural network for water body extraction. Since the preset flood extraction convolutional neural network is used to perform residual convolution and attention mechanism correction on the SAR image of water bodies to be extracted, and to perform hole convolution on the corrected SAR image of water bodies to be extracted to obtain a distribution map of water body information in flood disasters, it is possible to achieve the purpose of accurately correcting and reliably extracting the characteristic edge of water bodies with obvious characteristics in the SAR image of water bodies to be extracted, and further capturing the characteristic information of water body edges with more details and unclear characteristics, so that the water body information in the obtained distribution map of water body information in flood disasters is richer and more complete, thereby effectively improving the extraction precision and accuracy of water bodies in flood disasters.

[0072] Based on the defects of low extraction precision and accuracy of water bodies in flood disasters in the prior art, the present invention proposes a flood extraction convolutional neural network (FENet) based on SAR data to extract water body information in flood disasters. The FENet network is composed of a coding correction network and a hole convolution network, wherein the coding correction network adopts the Resnet18 network and the residual subnetwork in the Resnet18 network is connected to the attention mechanism correction subnetwork, so as to realize the input of the feature image obtained after coding and correction into the hole convolution network with different expansion rates, so that it can extract deep semantic features at different scales, and then further process the image after the hole convolution to the original size through decoding corresponding to the aforementioned coding correction process. In addition, when the attention mechanism correction subnetwork adopted by the FENet network is the Scse attention mechanism, the accuracy of the model prediction can be further improved.

[0073] Therefore, when the preset flood extraction convolutional neural network includes a preset residual neural network and a preset hole convolutional network, step S120 can be implemented by the following process:

[0074] Firstly, the SAR image of the water body to be extracted is input into a preset residual neural network for residual convolution and attention mechanism correction, so as to obtain a first water body information distribution map after precise correction of the edge of the flood disaster water body; then, the first water body information distribution map is input into a preset hole convolution network for change processing of the preset expansion scale, so as to obtain a complete flood disaster water body information distribution map with the edge of the flood disaster water body extracted.

[0075] Specifically, when the terminal device obtains the SAR image of the water body to be extracted, it can first input the SAR image of the water body to be extracted into a preset residual neural network for residual convolution and attention correction, that is, the preset residual neural network first performs residual convolution on the SAR image of the water body to be extracted, and then performs attention correction on the SAR image after the residual convolution, so as to obtain a first water body information distribution map after the precise correction of the edge of the water body caused by flood disasters. The first water body information distribution map includes water body information with obvious characteristics and water body information after the water body edge is accurately corrected.

[0076] Furthermore, when the terminal device determines that the preset residual neural network has outputted the first water body information distribution map, the first water body information distribution map can be further input into the preset atrous convolutional neural network for dilation processing with multiple different dilation rates, so that more details of the water body to be extracted from the SAR image and the characteristic information of the water body edge with unclear features can be captured, thereby making the water body information distribution map of flood disasters richer and more comprehensive.

[0077] The method for extracting water bodies caused by flood disasters provided by the present invention extracts water body information with obvious water body characteristics through a preset residual neural network with residual convolution and attention mechanism correction functions, and extracts more small water body information with unclear water body characteristics and easy to be ignored through a preset void convolution neural network with different expansion processing functions, thereby improving the precision and accuracy and reliability of extracting water body information caused by flood disasters.

[0078] Optionally, when the preset residual neural network includes P residual sub-networks, and each residual sub-network is respectively connected to an attention mechanism correction sub-network, the SAR image of the water body to be extracted is input into the preset residual neural network for residual convolution and attention mechanism correction, and the specific process of obtaining the first water body information distribution map after the precise correction of the edge of the flood disaster water body includes:

[0079] Inputting the SAR image to be extracted from the water body into the qth residual subnetwork for convolution, batch normalization and linear sorting to obtain the qth residual convolution image;

[0080] Inputting the qth residual convolution image into the qth attention mechanism correction subnetwork for channel correction and spatial correction processing to obtain the qth corrected image;

[0081] Input the qth corrected image into the q+1th residual subnetwork for convolution, batch normalization and linear sorting to obtain the q+1th residual convolution image;

[0082] Input the q+1th residual convolution image into the q+1th attention mechanism correction subnetwork for channel correction and spatial correction processing to obtain the q+1th corrected image;

[0083] Let the value of q be increased by 1, and repeat the step of inputting the q+1th residual convolution image into the q+1th attention mechanism correction subnetwork for channel correction and spatial correction processing until the Pth corrected image is obtained, and the Pth corrected image is the first water body information distribution map after the precise correction of the edge of the flood disaster water body; q∈[1,…,P], q+1∈[2,…,P].

[0084] Specifically, when the preset residual neural network is a Resnet18 network including 4 residual subnetworks and each attention mechanism correction subnetwork represents the Scse attention mechanism, the terminal device can first input the acquired SAR image of the water body to be extracted into the first residual subnetwork for convolution, batch normalization and linear sorting to obtain the first residual convolution image, and then input the first residual convolution image into the first attention mechanism correction subnetwork for channel correction and spatial correction to obtain the first corrected image; then input the first corrected image into the second residual subnetwork for convolution, batch normalization and linear sorting to obtain the second residual convolution image, and input the second residual convolution image into the second attention mechanism correction subnetwork for channel correction and spatial correction to obtain. the second corrected image; the second corrected image is then input into the third residual subnetwork for convolution, batch normalization and linear sorting to obtain the third residual convolution image, the third residual convolution image is input into the third attention mechanism correction subnetwork for channel correction and spatial correction to obtain the third corrected image; the third corrected image is then input into the fourth residual subnetwork for convolution, batch normalization and linear sorting to obtain the fourth residual convolution image, the fourth residual convolution image is input into the fourth attention mechanism correction subnetwork for channel correction and spatial correction to obtain the fourth corrected image. The fourth corrected image obtained at this time is the first water body information distribution map after the precise correction of the edge of the water body caused by the flood disaster, and the first water body information distribution map is also a feature image obtained after P steps (for example, 4 steps) of encoding correction.

[0085] In the actual processing process, each residual sub-network can be Figure 3 The process of performing convolution, batch normalization, and linear sorting can be done as shown in Figure 3 As shown in (a) of the figure, two convolutions, two batch normalizations, and two linear sorting processes are performed and then directly added to the image x to obtain the corrected image. Figure 3 As shown in (b), two convolutions, two batch normalizations, and two linear reorganizations are performed and then added to the image x after one convolution to obtain a corrected image. In the figure, x is the SAR image of the water body to be extracted or any corrected image from the 1st to the P-1th corrected image, y is any corrected image from the 1st to the Pth corrected image, and Relu represents linear reorganization.

[0086] In addition, each attention correction subnetwork can be Figure 4As shown, channel correction and spatial correction processing are performed on each residual convolution image, and the size of each residual convolution image is (C, H, W), C represents the color channel of each residual convolution image, H represents the height of each residual convolution image, and W represents the width of each residual convolution image; wherein, the channel correction processing is to first perform global average pooling processing on the residual convolution image, and then use the relu function and the sigmoid function to activate the first feature image obtained after two convolutions, and finally use the channel multiplication method to perform channel correction to obtain the first image after channel correction; spatial correction is to first perform convolution operation on the residual convolution image, and then use the sigmoid function to activate it, and finally use the spatial multiplication method to perform spatial information correction to obtain the second image after spatial correction; further, the first image after channel correction and the second image after spatial correction are combined by channel addition to obtain a corrected image after channel correction and spatial correction, and its size is the same as the corresponding residual convolution image size. In the present invention, each attention mechanism correction sub-network is added after each residual sub-network, without changing the dimension size of each residual convolution image, thereby achieving the purpose of accurately correcting the water body edge in each residual convolution image.

[0087] It should be noted that although the commonly used deep residual networks include Resnet18, Resnet34, Resnet50, Resnet101 and Resnet152, in order to reduce the complexity of the network model and speed up the processing speed of the network model, the present invention adopts the Resnet18 network and selects the 4 residual sub-networks in the Resnet18 network to achieve the purpose of extracting water body information with more obvious water body characteristics, which greatly improves the use efficiency of the Resnet18 network, and when each residual sub-network is connected to the Scse attention mechanism, it can further improve the accuracy of model prediction.

[0088] The flood disaster water body extraction method provided by the present invention realizes the purpose of quickly and accurately extracting water body information with more obvious features by means of P residual sub-networks and each residual sub-network is connected to an attention mechanism correction sub-network. In this way, the prediction accuracy of the network model can be greatly improved on the basis of accurately correcting the water body edge.

[0089] Optionally, when the preset atrous convolution network includes Q different dilation rates, the specific process of inputting the first water body information distribution map into the preset atrous convolution network for change processing of the preset dilation scale to obtain a complete flood disaster water body information distribution map by extracting the edge of the flood disaster water body may include:

[0090] The first water body information distribution map is input into a preset atrous convolutional network for processing at different expansion scales to obtain Q small water body feature information maps at different expansion scales; the small water body feature information maps are fused to obtain a fused image; the fused image is decoded corresponding to the residual convolution and attention mechanism correction to obtain a complete flood disaster water body information distribution map with flood disaster water body edge extraction.

[0091] Specifically, when the preset atrous convolution network includes four different dilation rates, the terminal device can input the first water body information distribution map after the edge of the flood disaster water body is accurately corrected into the preset atrous convolution network for atrous convolution processing with different dilation rates, thereby obtaining feature images of four different dilation scales, and then further perform fusion processing (such as channel addition processing) on ​​the feature images of the four different dilation scales to obtain a fused image, and finally perform a P-step decoding operation on the fused image corresponding to the aforementioned P-step encoding correction, so as to obtain a flood disaster water body information distribution map with the same size as the original water body to be extracted SAR image, which includes water body information with relatively obvious features and water body information with more details and less obvious features, such as Figure 5 The following is a schematic diagram of dilated convolution with different expansion rates. Figure 5 In the figure, (a) is a hole and schematic diagram when the dilation rate is 1, (b) is a schematic diagram of the dilated convolution when the dilation rate is 2, and (c) is a schematic diagram of the dilated convolution when the dilated convolution is 4, where the convolution kernel with a dilation rate of 1 can be regarded as a standard convolution kernel. Since dilated convolutions with different dilation rates have different receptive fields, different receptive fields are very important for distinguishing mountain shadows and water bodies with similar spectral characteristics. However, if the dilation rate is set too large, it will cause too much feature information of small water bodies to be lost. Therefore, the present invention combines dilated convolutions with dilation rates of 1, 2, 4, and 8 to extract semantic features at different scales, thereby achieving a more complete purpose of extracting water body edges by appropriately reducing the dilation rate to capture more feature information of small water bodies and rivers.

[0092] The flood disaster water body extraction method provided by the present invention achieves the purpose of capturing more water body information with smaller features or less obvious water body features while ensuring that water body information with more obvious features in SAR images including flood disaster areas can be accurately and precisely extracted, by performing atrous convolution processing of different expansion scales on the first water body information distribution map and fusing and decoding the feature images corresponding to each expansion scale, thereby greatly improving the extraction accuracy and efficiency of flood disaster water bodies.

[0093] Optionally, in order to improve the flexibility of training the initial flood extraction convolutional neural network and improve the reliability of the preset flood extraction convolutional neural network, the training process of the preset flood extraction convolutional neural network may include:

[0094] Acquire a training sample image set; the training sample image set includes training sample images, and the training sample images are SAR images obtained after sample annotation, regular grid cropping and data enhancement processing of the SAR images of the water body to be extracted; train a preset initial flood extraction convolutional neural network according to the training sample image set to obtain the preset flood extraction convolutional neural network.

[0095] Specifically, the terminal device trains a preset initial flood extraction convolutional neural network, which can also be considered as a process of training a preset initial FENet network. The initial FENet network is composed of an initial Resnet18 network and an initial hole convolutional network.

[0096] In addition, before training the initial FENet network, a training sample image set can be obtained from the SAR image to be extracted from the water body. The process of obtaining the training image set includes:

[0097] First, two typical areas are cropped from the SAR image of the water body to be extracted as training images and test images respectively. There is no intersection between the two typical areas and each typical area satisfies the following conditions: it contains different types of water bodies, such as calm open water bodies, water bodies that submerge objects, small river water bodies, etc. The more water body types, the better; the larger the size of each typical area, the better.

[0098] Then, the training images and test images are respectively subjected to refined sample annotation processing to obtain training label images corresponding to the training images and test label images corresponding to the test images. The refined sample annotation processing process includes: combining the existing high-resolution optical remote sensing image data as an aid, and based on the researchers' experience, marking the boundaries of suspected water bodies in the training images and test images, respectively, to obtain training label images and test label images.

[0099] Then, the training images and training label images are cropped with regular grids and data enhanced to obtain a training sample image set including M training sample images; and the test images and test label images are cropped with regular grids and data enhanced to obtain a test sample image set including N test sample images. Data enhancement includes horizontal flipping, vertical flipping and diagonal mirroring operations on the images and corresponding label images. Data enhancement is to obtain sufficient samples to avoid overfitting during network training.

[0100] Taking the test image as an example, the regular grid cropping is explained, and the process includes: cropping each sample from left to right and from top to bottom in the test image according to the preset sample size, and there is an overlap area between adjacent samples in the cropped samples, and the size of the overlap area is preset (for example, 20). For the cropping of samples near the boundary of rows or columns, at this time, the last row is cropped from top to bottom (or from bottom to top) to obtain a sample of the preset sample size, and the last column is also cropped from left to right (or from right to left) to obtain a sample of the preset sample size. For such samples, the overlap area between them and adjacent samples is smaller than the preset size (for example, smaller than 20).

[0101] Finally, the initial FENet network is trained using the training sample image set until the trained network reaches the preset network accuracy, and the preset flood extraction convolutional neural network can be obtained.

[0102] The method for extracting water bodies caused by flood disasters provided by the present invention trains an initial flood extraction convolutional neural network by using a training sample image set obtained from SAR images of water bodies to be extracted. The training sample image set for training the initial flood extraction convolutional neural network is obtained from SAR images of water bodies to be extracted, and the SAR images of water bodies to be extracted include dual-polarization bands and spectral characteristic bands. Therefore, using the training sample image set to train the initial flood extraction convolutional neural network can improve network accuracy and network reliability, and lay a foundation for subsequent rapid and accurate extraction of water body information.

[0103] Optionally, the specific process of training the preset initial flood extraction convolutional neural network according to the training sample image set to obtain the preset flood extraction convolutional neural network may include:

[0104] Iteratively train the preset initial flood extraction convolutional neural network according to the training sample image set to obtain an intermediate flood extraction convolutional neural network obtained after this round of iterative training; determine whether the cumulative number of training rounds corresponding to the current round of iterative training reaches a preset round number threshold; if the cumulative number of training rounds reaches the preset round number threshold, use the intermediate flood extraction convolutional neural network obtained after the current round of iterative training as the preset flood extraction convolutional neural network; if the cumulative number of training rounds does not reach the preset round number threshold, train the intermediate flood extraction convolutional neural network to obtain the preset flood extraction convolutional neural network.

[0105] Specifically, the terminal device uses M training sample images to train the initial flood extraction convolutional neural network. After each training sample image trains the FENet network once, it is called a training. After all M training sample images train the FENet network once, it is called a round of training. After the terminal device uses M training sample images for training for a preset number of rounds, it can determine whether the cumulative number of training rounds after this round of training reaches the preset number of rounds threshold. When it is determined that the cumulative number of training rounds reaches the preset number of rounds threshold, the training is stopped, and the network corresponding to the time when the training is stopped is tested using a test sample image set including N test sample images. When the values ​​of the five indicators of precision (Precision), recall (Recall), F1 index (F1), water body intersection over Union (IoU) and mean intersection over Union (mIoU) are all the highest, it is determined that the network corresponding to the time when the training is stopped converges, that is, the preset flood extraction convolutional neural network is obtained; on the contrary, when it is determined that the cumulative number of training rounds does not reach the preset number of rounds threshold, the preset number of rounds of training of the network obtained after this round of training is continued until the preset number of rounds threshold is reached to obtain the preset flood extraction convolutional neural network. The calculation formulas for the five indicators are:

[0106]

[0107]

[0108]

[0109]

[0110]

[0111] Among them, n represents the number of water body categories predicted by the network, TP represents the number of real water body pixels predicted by the network as water bodies, FP represents the number of real non-water body pixels predicted by the network as water bodies, FN represents the number of real water body pixels predicted by the network as non-water bodies, Precision represents precision, Recall represents recall, F1 represents F1 index, mIoU represents mean intersection over union, and IoU represents water body intersection over union; the F1 index comprehensively considers precision and recall, and the higher the value, the better the model extraction effect; mean intersection over union mIoU comprehensively considers the model extraction intersection over union of water bodies and the intersection over union of non-water bodies, and the higher the value, the better the model extraction effect; both F1 and mIoU are comprehensive evaluation indicators for measuring network models.

[0112] It should be noted that since training the FENet network once for each training sample image is called one training, and training the FENet network once for M training sample images is called one round of training, when the process of training each training sample image once is known, the process of training the FENet network once for M training sample images is also known.

[0113] The process of training each training sample image once includes: the training sample image enters the initial flood extraction convolutional neural network and first performs P-step coding correction. When P takes a value of 4 and the initial flood extraction convolutional neural network includes 4 initial residual sub-networks and each initial residual sub-network is respectively connected to an initial Scse attention mechanism correction sub-network, the training sample image is first subjected to the first step of residual convolution and attention correction processing to obtain a first corrected coded image; the first corrected coded image is subjected to the second step of residual convolution and attention correction processing to obtain a second corrected coded image; the second corrected coded image is subjected to the third step of residual convolution and attention correction processing to obtain a third corrected coded image; the third corrected coded image is subjected to the fourth step of residual convolution and attention correction processing to obtain a fourth corrected coded image. At this time, the fourth coded image can be called a feature image obtained after encoding and correction.

[0114] Then, the feature image is subjected to dilated convolution processing. Q different dilation rates (preferably 4 dilation rates, such as 1, 2, 4, and 8) can be selected. After the feature image is subjected to dilated convolution processing, Q feature images with different dilation scales are obtained. Each feature image with different dilation scale is then fused (such as channel addition). Subsequently, the fused image is subjected to P-step decoding corresponding to the aforementioned P-step encoding correction to obtain a decoded image with the same size as the original training sample image and after training the initial flood extraction convolutional neural network once.

[0115] The flood disaster water body extraction method provided by the present invention achieves the purpose of quickly and efficiently obtaining a preset flood extraction convolutional neural network by training a preset initial flood extraction convolutional neural network with a training sample image set to determine whether the cumulative number of rounds reaches a preset round threshold, which not only ensures the network accuracy, but also greatly reduces the training complexity and speeds up the network training speed.

[0116] Optionally, the process of determining the preset round number threshold may include:

[0117] The training sample image set is divided into training samples and verification samples according to a preset ratio; the preset initial flood extraction convolutional neural network is trained according to the training samples to obtain an intermediate flood extraction convolutional neural network obtained after a preset number of rounds of training; the intermediate flood extraction convolutional neural network is verified according to the verification samples to obtain the value of the evaluation index of the intermediate flood extraction convolutional neural network; it is determined whether the value of the evaluation index reaches the preset standard value; if the value of the evaluation index reaches the preset standard value, the accumulated number of training rounds corresponding to the intermediate flood extraction convolutional neural network is used as the preset round number threshold; if the value of the evaluation index does not reach the preset standard value, the intermediate flood extraction convolutional neural network is trained to obtain the preset round number threshold.

[0118] Among them, the preset standard value can be used to characterize that the values ​​of other parameters such as the loss value and accuracy of the network are sufficient to indicate that the network has reached the convergence requirement.

[0119] Specifically, the terminal device determines whether the network converges after the current round of training based on the comparison of the cumulative number of training rounds after the current round of training with the preset round number threshold. Therefore, the specific value of the preset round number threshold is particularly important.

[0120] The process of determining the preset round number threshold may be to divide the M training sample images into training samples and verification samples according to a preset ratio. The preset ratio may be 6:4 or 7:3, or other ratios, as long as the number of training samples is greater than the number of verification samples.

[0121] After the terminal device uses the training samples to train the preset initial flood extraction convolutional neural network for a preset number of rounds, it can obtain the intermediate flood extraction convolutional neural network obtained after the preset number of rounds of training, and use the verification samples to evaluate the intermediate flood extraction convolutional neural network to obtain the value of the evaluation index of the intermediate flood extraction convolutional neural network, and the value of the evaluation index may include the value of the loss function, the network accuracy, etc.; when it is determined that the value of the evaluation index of the intermediate flood extraction convolutional neural network reaches the preset standard value, the cumulative number of training rounds corresponding to the intermediate flood extraction convolutional neural network is used as the preset round threshold; when it is determined that the value of the evaluation index of the intermediate flood extraction convolutional neural network does not reach the preset standard value, the training samples can be used to train the intermediate flood extraction convolutional neural network again until the value of the evaluation index of the trained network after verification using the verification samples reaches the preset standard value.

[0122] The flood disaster water body extraction method provided by the present invention achieves the purpose of accurately and reliably determining a preset round number threshold by first using training samples in a training sample image set to train the network for a certain number of rounds, and then using verification samples in the training sample image set to verify whether the network meets the accuracy requirements, thereby providing a strong basis for subsequently quickly and accurately obtaining a preset flood extraction convolutional neural network.

[0123] In the actual processing process, in order to verify the effectiveness of the method of the present invention, it is compared and analyzed with the traditional flood monitoring method based on SAR images. Figure 6 The two typical areas shown in the figure have a water body proportion of 33.4% in area 1 and 5.8% in area 2. The real water body information is obtained by visual interpretation with the assistance of high-resolution optical remote sensing images. The real water body information marked manually is compared with the water body information obtained by the algorithm. A variety of evaluation indicators are used to quantitatively evaluate the accuracy of water body information extraction, so as to verify the method proposed in the present invention. Specifically, the recall rate, precision rate, false alarm rate and Evaluation Indicators Four indicators are used to test the effectiveness and reliability of the method proposed in the present invention. The calculation formulas of the four indicators are as follows:

[0124]

[0125]

[0126] F a =1-P (8)

[0127]

[0128] Among them, P(T W ) represents the real water pixel interpreted by human vision, P(A W ) represents the water body pixels extracted by the method; R represents the recall rate of the extraction method, which indicates the degree to which the extracted water body range is close to the real water body, and the higher the value, the better; P represents the precision rate of the extraction method, which indicates the accuracy of the extracted water body, and the higher the value, the better; F a represents the false alarm rate of the extraction method; It represents a comprehensive evaluation index of the extraction method, which comprehensively considers the recall rate and precision rate. The higher the value, the better the method effect. The three methods include the method of the present invention, the Otsu method (also known as the Otsu global threshold method) and the object-oriented method, and the extraction accuracy comparison of the three methods is shown in Table 1.

[0129] Table 1

[0130]

[0131]

[0132] based on Figure 6 As shown in Table 1, the proportion of water bodies in the selected comparison area is different, and the advantages and disadvantages of the method of the present invention and the traditional method are tested from the perspective of different proportions of water bodies. The results show that when the proportion of water bodies is high, the recall rate and precision rate of the Otsu global threshold method are high, and the water body extraction effect is good; when the proportion of water bodies is low, the water body extraction accuracy of the Otsu global threshold method is the worst, and the accuracy of water body extraction is the worst; the object-oriented method has certain advantages over the Otsu global threshold method in the extraction of small water bodies, but the extraction accuracy is not as good as the preset flood extraction convolutional neural network in the method of the present invention; for the extraction of water body information in a large range of flood disasters, the object-oriented method has the worst extraction accuracy. Among the three methods, the preset flood extraction convolutional neural network in the method of the present invention has the best effect and the highest accuracy. For the extraction of large-scale macro water body information, the Otsu global threshold method is difficult to find a suitable threshold to extract water body information, and the segmentation rules and classification rules of the object-oriented method require a lot of prior knowledge. Through precision comparison analysis, it is proved that the preset flood extraction convolutional neural network in the method of the present invention has higher accuracy and better effect than the traditional flood monitoring method.

[0133] The flood disaster water body extraction device provided by the present invention is described below. The flood disaster water body extraction device described below and the flood disaster water body extraction method described above can correspond to each other.

[0134] Figure 7 An example of a flood disaster water extraction device is shown. Figure 7 As shown, the flood disaster water body extraction device 700 includes: an acquisition module 710, used to acquire a SAR image of a water body to be extracted, wherein the SAR image of the water body to be extracted includes an image composed of a dual-polarization band and a spectral characteristic band of the flood disaster SAR image; a determination module 720, used to input the SAR image of the water body to be extracted into a preset flood extraction convolutional neural network to obtain a flood disaster water body information distribution map; wherein the preset flood extraction convolutional neural network is used to perform residual convolution and attention mechanism correction on the SAR image of the water body to be extracted, and perform hole convolution on the corrected SAR image of the water body to be extracted to obtain the flood disaster water body information distribution map.

[0135] Optionally, the determination module 720 can be specifically used to input the SAR image of the water body to be extracted into a preset residual neural network for residual convolution and attention mechanism correction, so as to obtain a first water body information distribution map after precise correction of the edge of the flood disaster water body; input the first water body information distribution map into a preset hole convolution network for change processing of a preset expansion scale, so as to obtain a complete flood disaster water body information distribution map with the edge of the flood disaster water body extracted.

[0136] Optionally, the determination module 720 can also be specifically used to input the SAR image of the water body to be extracted into the qth residual subnetwork for convolution, batch normalization and linear sorting to obtain the qth residual convolution image; input the qth residual convolution image into the qth attention mechanism correction subnetwork for channel correction and spatial correction to obtain the qth corrected image; input the qth corrected image into the q+1th residual subnetwork for convolution, batch normalization and linear sorting to obtain the q+1th residual convolution image; input the qth corrected image into the q+1th residual subnetwork for convolution, batch normalization and linear sorting to obtain the q+1th residual convolution image; The q+1 residual convolution images are input into the q+1th attention mechanism correction subnetwork for channel correction and space correction processing to obtain the q+1th corrected image; let the value of q be increased by 1, and repeat the step of inputting the q+1th residual convolution image into the q+1th attention mechanism correction subnetwork for channel correction and space correction processing until the Pth corrected image is obtained, and the Pth corrected image is the first water body information distribution map after the precise correction of the edge of the water body for flood disasters; q∈[1,…,P],q+1∈[2,…,P].

[0137] Optionally, the determination module 720 can also be used to input the first water body information distribution map into a preset hole convolution network for processing at different expansion scales to obtain Q small water body feature information maps with different expansion scales; fuse the small water body feature information map to obtain a fused image; and decode the fused image corresponding to the residual convolution and attention mechanism correction to obtain a complete flood disaster water body information distribution map with flood disaster water body edge extraction.

[0138] Optionally, the determination module 720 can also be specifically used to obtain a training sample image set; the training sample image set includes training sample images, and the training sample images are SAR images obtained after sample annotation, regular grid cropping and data enhancement processing of the SAR images of the water body to be extracted; according to the training sample image set, a preset initial flood extraction convolutional neural network is trained to obtain the preset flood extraction convolutional neural network.

[0139] Optionally, the determination module 720 can also be specifically used to iteratively train the preset initial flood extraction convolutional neural network according to the training sample image set to obtain an intermediate flood extraction convolutional neural network obtained after this round of iterative training; determine whether the cumulative number of training rounds corresponding to the current round of iterative training reaches a preset round number threshold; if the cumulative number of training rounds reaches the preset round number threshold, the intermediate flood extraction convolutional neural network obtained after the current round of iterative training is used as the preset flood extraction convolutional neural network; if the cumulative number of training rounds does not reach the preset round number threshold, the intermediate flood extraction convolutional neural network is trained to obtain the preset flood extraction convolutional neural network.

[0140] Optionally, the determination module 720 can also be specifically used to divide the training sample image set into training samples and verification samples according to a preset ratio; train the preset initial flood extraction convolutional neural network according to the training samples, and obtain the intermediate flood extraction convolutional neural network obtained after a preset number of rounds of training; verify the intermediate flood extraction convolutional neural network according to the verification sample to obtain the value of the evaluation index of the intermediate flood extraction convolutional neural network; determine whether the value of the evaluation index reaches the preset standard value; if the value of the evaluation index reaches the preset standard value, then use the cumulative number of training rounds corresponding to the intermediate flood extraction convolutional neural network as the preset round threshold; if the value of the evaluation index does not reach the preset standard value, then train the intermediate flood extraction convolutional neural network to obtain the preset round threshold.

[0141] Figure 8 An example of a physical structure diagram of an electronic device is shown in FIG. Figure 8 As shown, the electronic device 800 may include: a processor 810, a communication interface 820, a memory 830 and a communication bus 840, wherein the processor 810, the communication interface 820 and the memory 830 communicate with each other through the communication bus 840. The processor 810 may call the logic instructions in the memory 830 to execute the flood disaster water body extraction method, the method comprising: obtaining a SAR image of a water body to be extracted, wherein the SAR image of the water body to be extracted includes an image composed of a dual polarization band and a spectral characteristic band of the flood disaster SAR image; inputting the SAR image of the water body to be extracted into a preset flood extraction convolutional neural network to obtain a flood disaster water body information distribution map; wherein the preset flood extraction convolutional neural network is used to perform residual convolution and attention mechanism correction on the SAR image of the water body to be extracted, and perform hole convolution on the corrected SAR image of the water body to be extracted to obtain the flood disaster water body information distribution map.

[0142] In addition, the logic instructions in the above-mentioned memory 830 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0143] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the flood disaster water body extraction method provided by the above-mentioned methods, and the method includes: obtaining a SAR image of a water body to be extracted, wherein the SAR image of the water body to be extracted includes an image composed of a dual-polarization band and a spectral characteristic band of the flood disaster SAR image; inputting the SAR image of the water body to be extracted into a preset flood extraction convolutional neural network to obtain a flood disaster water body information distribution map; wherein the preset flood extraction convolutional neural network is used to perform residual convolution and attention mechanism correction on the SAR image of the water body to be extracted, and performing hole convolution on the corrected SAR image of the water body to be extracted to obtain the flood disaster water body information distribution map.

[0144] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to execute the above-mentioned flood disaster water body extraction methods, the method comprising: obtaining a SAR image of a water body to be extracted, the SAR image of the water body to be extracted comprising an image composed of a dual-polarization band and a spectral characteristic band of the flood disaster SAR image; inputting the SAR image of the water body to be extracted into a preset flood extraction convolutional neural network to obtain a flood disaster water body information distribution map; wherein the preset flood extraction convolutional neural network is used to perform residual convolution and attention mechanism correction on the SAR image of the water body to be extracted, and performing hole convolution on the corrected SAR image of the water body to be extracted to obtain the flood disaster water body information distribution map.

[0145] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0146] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0147] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for extracting water bodies from flood disasters, characterized in that: include: Acquire a SAR image of a water body to be extracted, wherein the SAR image of the water body to be extracted includes an image composed of a dual polarization band and a spectral characteristic band of a flood disaster SAR image; Inputting the SAR image of the water body to be extracted into a preset flood extraction convolutional neural network to obtain a distribution map of flood disaster water body information; The preset flood extraction convolutional neural network is used to perform residual convolution and attention mechanism correction on the SAR image of the water body to be extracted, and to perform hole convolution on the corrected SAR image of the water body to be extracted to obtain the flood disaster water body information distribution map; The preset flood extraction convolutional neural network includes a preset residual neural network and a preset hole convolutional network. The SAR image of the water body to be extracted is input into the preset flood extraction convolutional neural network to obtain a flood disaster water body information distribution map, including: The SAR image of the water body to be extracted is input into a preset residual neural network for residual convolution and attention mechanism correction, so as to obtain a first water body information distribution map after precise correction of the edge of the flood disaster water body; Inputting the first water body information distribution map into a preset dilated convolutional network to perform a change process of a preset expansion scale, thereby obtaining a flood disaster water body information distribution map with complete flood disaster water body edge extraction; The preset residual neural network includes P residual sub-networks, and each of the residual sub-networks is respectively connected to an attention mechanism correction sub-network, the SAR image of the water body to be extracted is input into the preset residual neural network for residual convolution and attention mechanism correction, and a first water body information distribution map after precise correction of the edge of the flood disaster water body is obtained, including: Inputting the SAR image to be extracted from the water body into the qth residual subnetwork for convolution, batch normalization and linear sorting to obtain the qth residual convolution image; Inputting the qth residual convolution image into the qth attention mechanism correction subnetwork for channel correction and spatial correction processing to obtain the qth corrected image; Input the qth corrected image into the q+1th residual subnetwork for convolution, batch normalization and linear sorting to obtain the q+1th residual convolution image; Input the q+1th residual convolution image into the q+1th attention mechanism correction subnetwork for channel correction and spatial correction processing to obtain the q+1th corrected image; Let the value of q be increased by 1, and repeatedly execute the step of inputting the q+1th residual convolution image into the q+1th attention mechanism correction subnetwork for channel correction and spatial correction processing until the Pth corrected image is obtained, and the Pth corrected image is the first water body information distribution map after the edge of the flood disaster water body is precisely corrected; , ; When the preset hole convolution network includes Q different expansion rates, the first water body information distribution map is input into the preset hole convolution network for change processing of the preset expansion scale to obtain a complete flood disaster water body information distribution map with a flood disaster water body edge extraction, including: Inputting the first water body information distribution map into a preset atrous convolutional network for processing at different expansion scales, and obtaining Q small water body feature information maps at different expansion scales; Performing fusion processing on the small water body feature information map to obtain a fused image; The fused image is decoded corresponding to the residual convolution and attention mechanism correction to obtain a complete flood disaster water body information distribution map with flood disaster water body edge extraction.

2. The method for extracting flood water according to claim 1, characterized in that: The training process of the preset flood extraction convolutional neural network includes: Acquire a training sample image set; the training sample image set includes training sample images, and the training sample images are SAR images obtained after sample annotation, regular grid cropping and data enhancement processing are performed on the SAR image of the water body to be extracted; The preset initial flood extraction convolutional neural network is trained according to the training sample image set to obtain the preset flood extraction convolutional neural network.

3. The method for extracting flood water according to claim 2, characterized in that: The step of training a preset initial flood extraction convolutional neural network according to the training sample image set to obtain the preset flood extraction convolutional neural network includes: Iteratively training a preset initial flood extraction convolutional neural network according to the training sample image set to obtain an intermediate flood extraction convolutional neural network obtained after this round of iterative training; Determine whether the cumulative number of training rounds corresponding to the current round of iterative training reaches a preset round number threshold; If the accumulated number of training rounds reaches the preset round number threshold, the intermediate flood extraction convolutional neural network obtained after the current round of iterative training is used as the preset flood extraction convolutional neural network; If the accumulated number of training rounds does not reach the preset round number threshold, the intermediate flood extraction convolutional neural network is trained to obtain the preset flood extraction convolutional neural network.

4. The method for extracting flood water according to claim 3, characterized in that: The process of determining the preset round number threshold includes: Dividing the training sample image set into training samples and verification samples according to a preset ratio; Training a preset initial flood extraction convolutional neural network according to the training samples to obtain an intermediate flood extraction convolutional neural network obtained after a preset number of rounds of training; Verifying the intermediate flood extraction convolutional neural network according to the verification sample to obtain a value of an evaluation index of the intermediate flood extraction convolutional neural network; Determine whether the value of the evaluation index reaches a preset standard value; If the value of the evaluation index reaches the preset standard value, the accumulated number of training rounds corresponding to the intermediate flood extraction convolutional neural network is used as the preset round number threshold; If the value of the evaluation index does not reach the preset standard value, the intermediate flood extraction convolutional neural network is trained to obtain the preset round number threshold.

5. A flood disaster water body extraction device for executing the flood disaster water body extraction method according to any one of claims 1 to 4, characterized in that: include: An acquisition module is used to acquire a SAR image of a water body to be extracted, wherein the SAR image of the water body to be extracted includes an image composed of a dual polarization band and a spectral characteristic band of a flood disaster SAR image; A determination module is used to input the SAR image of the water body to be extracted into a preset flood extraction convolutional neural network to obtain a distribution map of flood disaster water body information; Among them, the preset flood extraction convolutional neural network is used to perform residual convolution and attention mechanism correction on the SAR image of the water body to be extracted, and obtain the flood disaster water body information distribution map after performing void convolution on the corrected SAR image of the water body to be extracted.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the flood disaster water body extraction method according to any one of claims 1 to 4 are implemented.

7. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the flood disaster water body extraction method according to any one of claims 1 to 4 are implemented.

Images