A method, device, equipment and storage medium for extracting flood monitoring information

Abstract

The present invention provides a flood monitoring information extraction method, device, equipment and storage medium, and relates to the field of water conservancy application technology. The method comprises the following steps: inputting an image to be predicted into an image semantic segmentation network model to obtain a water body information distribution map of the image to be predicted; wherein the image semantic segmentation network model is trained based on a training set and a verification set corresponding to a sample image, the training set and the verification set are obtained by dividing a data set obtained by superimposing and combining a first polarization image, a second polarization image and a spectral feature image, the first polarization image and the second polarization image are obtained after polarization processing of the sample image, and the spectral feature image is derived from the first polarization image and the second polarization image. The present invention can extract flood monitoring information more accurately, and at the same time is applicable to a wider range of occasions and has better promotion value.

Description

Technical Field

[0001] The present invention relates to the field of water conservancy application technology, and in particular to a flood monitoring information extraction method, device and storage medium. Background Art

[0002] As one of the natural disasters that seriously threaten human survival and development, floods have the characteristics of rapid occurrence, high frequency, wide impact range, and inundation of farmland, residential areas and other areas. They endanger social stability and unity, seriously affect people's lives and safety, and cause a lot of economic losses. According to the "China Flood and Drought Bulletin" published by the Ministry of Water Resources, floods have brought great challenges to my country's economic development. In order to reduce the losses caused by floods to the people and the country, it is of great significance to monitor floods in a timely and accurate manner and obtain the spatiotemporal distribution information of floods.

[0003] The traditional flood monitoring method mainly estimates flood information through fixed hydrological stations. However, this method is inefficient and cannot accurately grasp the spatiotemporal distribution information of floods. It is difficult to play a huge role in the actual disaster reduction process. Remote sensing technology has the advantages of wide coverage and short revisit time, and it has gradually become the main method for flood monitoring. However, since floods are usually accompanied by bad weather, optical sensors are affected by clouds and rainfall, and it is generally difficult to provide cloud-free and high-quality optical images. On the contrary, Synthetic Aperture Radar (SAR) has become the primary technical means for monitoring floods due to its all-day and all-weather working ability, which is not affected by clouds, rain, and weather. The traditional flood monitoring information extraction methods based on SAR images can be mainly divided into two types, namely threshold method and object-oriented method. Among them, the threshold method has fast calculation speed, simple principle, small calculation amount, and is the most widely used. Commonly used threshold methods include Otsu method, global threshold method, local threshold method, exponential method, etc. However, the threshold method is less effective when processing large-area remote sensing images, and it only considers the information of the pixel itself, ignoring the features of the object texture, terrain and color, which makes the water extraction results relatively discrete, and small rivers cannot be accurately extracted. The object-oriented method uses the image object as the basic extraction unit, not only considering the spectral characteristics of the object, but also considering its shape characteristics and texture characteristics, so the consistency of the extraction results is high, and it is not easy to have the "salt and pepper effect" caused by image speckle noise, but it is greatly affected by the segmentation scale, and the classification rules of image objects require a lot of experience. Therefore, the above limits the large-scale promotion and application of the object-oriented method.

[0004] Modern flood emergency response requires high-precision and timely flood monitoring methods. However, the water body information extracted by traditional flood monitoring information extraction methods is difficult to meet current requirements. Therefore, it is of great significance to develop more accurate flood monitoring information extraction methods and apply them in flood monitoring to obtain more accurate and timely flood monitoring. Summary of the invention

[0005] The present invention provides a flood monitoring information extraction method, device and storage medium, which are used to solve the defects of the traditional flood monitoring information extraction method in the prior art that cannot meet the requirements of accuracy and timeliness, and realize more accurate extraction of flood monitoring information. At the same time, it is applicable to a wider range of occasions and has better promotion value.

[0006] The present invention provides a method for extracting flood monitoring information, comprising the following steps:

[0007] Inputting the image to be predicted into the image semantic segmentation network model to obtain a water body information distribution map of the image to be predicted;

[0008] Among them, the image semantic segmentation network model is trained based on the training set and the verification set corresponding to the sample image. The training set and the verification set are obtained by dividing the data set obtained by superimposing and combining the first polarization image, the second polarization image and the spectral feature image. The first polarization image and the second polarization image are obtained after polarization processing of the sample image, and the spectral feature image is derived from the first polarization image and the second polarization image.

[0009] According to the flood monitoring information extraction method provided by the present invention, the image semantic segmentation network model is trained by the following steps:

[0010] Performing multi-polarization processing on the sample images respectively to obtain the first polarization image and the second polarization image;

[0011] Performing derivative processing on the first polarization image and the second polarization image to obtain a spectral characteristic image;

[0012] The data set is obtained by superimposing and combining the first polarization image, the second polarization image and the spectral feature image;

[0013] Dividing the data set to obtain the training set and the validation set;

[0014] Obtain corresponding training recognition results and verification recognition results according to the training set and the verification set;

[0015] The training set and the validation set are used as input data for training, the training recognition structure corresponding to the training set is used as the label of the training set, and the validation recognition result corresponding to the validation set is used as the label of the validation set. Deep learning is used for training to obtain an image semantic segmentation network model for generating the water body information distribution map of the image to be predicted.

[0016] According to the flood monitoring information extraction method provided by the present invention, obtaining corresponding training recognition results and verification recognition results according to the training set and the verification set also includes the following steps:

[0017] Regular net pruning and data enhancement are performed on the training set, the validation set, and the labels corresponding to the training set and the validation set; wherein the data enhancement includes horizontal flipping, vertical flipping, and diagonal mirroring operations.

[0018] According to the flood monitoring information extraction method provided by the present invention, before performing multi-polarization processing on the sample images to obtain the first polarization image and the second polarization image, the method further includes the following steps:

[0019] The sample image is preprocessed; wherein the preprocessing includes one or a combination of orbit correction, thermal noise removal, radiation calibration, coherent speckle filtering, terrain correction, decibelization, cropping and mosaicking.

[0020] According to the flood monitoring information extraction method provided by the present invention, the multi-polarization processing includes vertical transmission vertical reception polarization processing and vertical transmission horizontal reception polarization processing.

[0021] According to the flood monitoring information extraction method provided by the present invention, the image semantic segmentation network model is a U-Net network model.

[0022] According to the flood monitoring information extraction method provided by the present invention, the loss function of the image semantic segmentation network model adopts an edge-ignoring cross entropy function.

[0023] The present invention also provides a flood monitoring device, comprising:

[0024] An information extraction module is used to input the image to be predicted into the image semantic segmentation network model to obtain a water body information distribution map of the image to be predicted;

[0025] Among them, the image semantic segmentation network model is trained based on the training set and the verification set corresponding to the sample image. The training set and the verification set are obtained by dividing the data set obtained by superimposing and combining the first polarization image, the second polarization image and the spectral feature image. The first polarization image and the second polarization image are obtained after polarization processing of the sample image, and the spectral feature image is derived from the first polarization image and the second polarization image.

[0026] 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 flood monitoring information extraction methods described above are implemented.

[0027] 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 flood monitoring information extraction methods described above are implemented.

[0028] The present invention provides a flood monitoring information extraction method, device and storage medium. By using an improved image semantic segmentation network model, the water body information distribution map extracted from the image to be predicted can be more accurate, and the image to be predicted can use a SAR image with a higher resolution, thereby making the flood monitoring results subsequently obtained based on the water body information distribution map more accurate and more timely, and can monitor floods more timely and accurately. At the same time, it is applicable to a wider range of occasions and has better promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] 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.

[0030] Figure 1 It is a flow chart of the flood monitoring information extraction method provided by the present invention;

[0031] Figure 2 It is one of the specific flow diagrams of step S100 in the flood monitoring information extraction method provided by the present invention;

[0032] Figure 3 This is the second specific flow chart of step S100 in the flood monitoring information extraction method provided by the present invention;

[0033] Figure 4A schematic diagram of the network structure of the U-Net network model in the flood monitoring information extraction method provided by the present invention;

[0034] Figure 5 A schematic diagram of the process flow during inspection in the flood monitoring information extraction method provided by the present invention;

[0035] Figure 6 An image result diagram of the test in the flood monitoring information extraction method provided by the present invention;

[0036] Figure 7 An image map used in the accuracy test area when comparing the flood monitoring information extraction method provided by the present invention with other methods;

[0037] Figure 8 It is a schematic diagram of the flow of the flood monitoring information extraction device provided by the present invention;

[0038] Fig. 9 It is one of the specific structural schematic diagrams of the information extraction module in the flood monitoring information extraction device provided by the present invention;

[0039] Fig.10 This is the second specific structural diagram of the information extraction module in the flood monitoring information extraction device provided by the present invention;

[0040] Fig.11 It is a structural schematic diagram of the electronic device provided by the present invention. DETAILED DESCRIPTION

[0041] 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.

[0042] Combine the following Figure 1 The flood monitoring information extraction method of the present invention is described, and the flood monitoring information extraction method comprises the following steps:

[0043] S100, input the image to be predicted into the image semantic segmentation network model to obtain the water body information distribution map of the image to be predicted. The image semantic segmentation network model is trained based on the training set and the validation set corresponding to the sample image, the training set and the validation set are obtained by dividing the data set obtained by superimposing and combining the first polarization image, the second polarization image and the spectral feature image, the first polarization image and the second polarization image are obtained after polarization processing of the sample image, and the spectral feature image is derived from the first polarization image and the second polarization image.

[0044] In step S100, the image semantic segmentation network model adopts the U-Net network model.

[0045] The traditional method of extracting water information from optical images based on convolutional neural networks has low computational efficiency and high computer storage overhead because it needs to classify each pixel in the image. In addition, the classification performance and extraction accuracy are limited. In this context, Long et al. proposed a semantic segmentation method based on a fully convolutional network (FCN) in 2015 to classify images at the pixel level, thereby solving the problem of image segmentation at the semantic level. Subsequently, Ronneberger et al. proposed a U-shaped symmetric network model, namely the U-Net network model, which can integrate low-dimensional and high-dimensional features to greatly improve the accuracy of image semantic segmentation. The structure of the classic U-Net network model is based on the expansion and modification of the fully convolutional network. The network structure is clear and elegant in a U shape, and is currently one of the most popular image segmentation models.

[0046] Among the current flood monitoring information extraction methods based on SAR images, most methods are developed for medium and low resolution SAR images with large-scale applications, and are not completely applicable to high-resolution SAR images and the extraction of flood spatiotemporal distribution information that requires very accurate results.

[0047] In recent years, deep learning has been widely used in the field of remote sensing water body extraction due to its powerful feature extraction capabilities. However, existing research on water body extraction combined with deep learning mainly focuses on optical images, and there is relatively little research on SAR images. Due to the randomly distributed speckle noise on SAR images and the complex flood inundation scenes, this hinders the identification and monitoring of floods based on deep learning of SAR images. At the same time, SAR images are usually dual-polarization data with relatively few polarization bands. It is difficult to distinguish water bodies from shadows based on the backscattering coefficient, texture, and contour of the ground objects alone, resulting in low water body extraction accuracy. Therefore, applying deep learning to water body information extraction requires improving the existing deep learning network model to make it suitable for high-resolution SAR images.

[0048] In step S100, an improved U-Net network model is used in order to adapt to high-resolution SAR images, thereby making the water body information extracted from the high-resolution SAR images more in line with the requirements of flood monitoring.

[0049] Thereafter, the flood monitoring result may be obtained according to the water body information distribution map obtained in step S100.

[0050] Since the improved U-Net network model is used in step S100, the water body information distribution map obtained from the image to be predicted can be extracted with higher accuracy, and the image to be predicted can use a SAR image with higher resolution, thereby making the flood monitoring results obtained according to the water body information distribution map obtained in step S100 more accurate and more timely.

[0051] The flood monitoring information extraction method of the present invention, by using an improved image semantic segmentation network model, makes the extraction accuracy of the water body information distribution map obtained from the image to be predicted higher, and the image to be predicted can use a SAR image with higher resolution, thereby making the flood monitoring results subsequently obtained based on the water body information distribution map more accurate and more timely, and can monitor floods more timely and accurately, while also being applicable to a wider range of occasions and having better promotion value.

[0052] Combine the following Figure 2 The flood monitoring information extraction method of the present invention is described, and the image semantic segmentation network model is trained by the following steps:

[0053] S120, performing multi-polarization processing on the sample images respectively to obtain a first polarization image and a second polarization image.

[0054] In step S120, the multi-polarization processing includes vertical transmission vertical reception (VV) polarization processing and vertical transmission horizontal reception (VH) polarization processing. SAR remote sensing systems commonly use four polarization modes, namely horizontal transmission horizontal reception (HH) polarization processing, vertical transmission vertical reception (VV) polarization processing, horizontal transmission vertical reception (HV) polarization processing and vertical transmission horizontal reception (VH) polarization processing. The electric field vector of the energy pulse emitted by the radar can be polarized in the vertical or horizontal plane. Regardless of the wavelength, the radar signal can transmit horizontal (H) or vertical (V) electric field vectors, and receive horizontal (H) or vertical (V) or both return signals. Among them, HH and VV are co-polarized, and HV and VH are inverse (cross) polarizations. The first polarization image is the polarization image after VV polarization processing, and the second polarization image is the polarization image after VH polarization processing.

[0055] S130, performing derivative processing on the first polarization image and the second polarization image to obtain a spectral characteristic image.

[0056] In step S130 , Python programming can be used to obtain polarization-derived spectral feature SDWI data, ie, spectral feature images, after VV and VH processing.

[0057] Considering the limited polarization data of SAR images and the insufficient richness of model feature learning, spectral features are introduced in step S130 to solve the problem of insufficient model features. The spectral feature SDWI data derived from the dual polarization data and polarization data of the SAR image are superimposed and combined into a new image, and the superimposed new image is used as the data set of the U-Net network model. The spectral feature SDWI data is obtained by the SAR image water body information extraction method proposed by Jia Shichao et al. in 2018, see formula (1), formula (1) is as follows:

[0058] K SDWI =ln(10*VV*VH) (1)

[0059] In formula (1), K SDWI Represents the result value of band operation, VV and VH represent the dual-polarization data of SAR image, where VV is the polarization data obtained after vertical transmission and vertical reception polarization processing, and VH is the polarization data obtained after vertical transmission and horizontal reception polarization processing. The spectral feature SDWI data uses dual-polarization data multiplication to enhance water body characteristics and weaken soil and vegetation characteristics, achieving better water body information extraction results.

[0060] S140, superimposing and combining the first polarization image, the second polarization image and the spectral characteristic image to obtain a data set, superimposing VH polarization data, VV polarization data and spectral characteristic SDWI data to form a new three-band image.

[0061] In step S140, the dual-polarization data and the spectral features of the SAR image SDWI data are combined to construct a new deep learning data set, which can solve the problem of insufficient model features and improve the accuracy of water body information extraction in the flood monitoring process to a certain extent.

[0062] S150, dividing the data set into a training set, a validation set and a test set.

[0063] S160: Obtain corresponding training recognition results and verification recognition results according to the training set and the verification set.

[0064] S170. Use the training set and the validation set as input data for training, use the training recognition structure corresponding to the training set as the label of the training set, use the validation recognition result corresponding to the validation set as the label of the validation set, and use deep learning to perform training to obtain an image semantic segmentation network model for generating a water body information distribution map of the image to be predicted.

[0065] According to the flood monitoring information extraction method provided by the present invention, obtaining corresponding training recognition results and verification recognition results according to the training set and the verification set also includes the following steps:

[0066] Step S160 also includes the following steps:

[0067] Regular net pruning and data enhancement are performed on the training set, validation set, and the labels corresponding to the training set and validation set, where the data enhancement includes horizontal flipping, vertical flipping, and diagonal mirroring operations.

[0068] Data enhancement is to obtain sufficient samples to avoid overfitting during model training.

[0069] Combine the following Figure 3 The flood monitoring information extraction method of the present invention is described, and before step S120, the following steps are also included:

[0070] S110, preprocessing the sample image to obtain a high-resolution dual-polarization backscatter distribution map, wherein the preprocessing includes one or a combination of orbit correction, thermal noise removal, radiation calibration, coherent speckle filtering, terrain correction, decibelization, cropping and mosaicking.

[0071] In this embodiment, the loss function of the image semantic segmentation network model adopts the ignore-edges Categorical Cross Entropy (IELoss), see formula (2), and formula (2) is as follows:.

[0072]

[0073] In formula (2), r represents the ratio of the number of pixels in the selected region to the total number of pixels in the image, N represents the total number of pixels in the sample image, and M represents the total number of categories. It is represented as a tensor consisting of the total number of categories after the true category label is processed by one-hot encoding. A tensor representing the pixel category probability values ​​obtained by forward propagation of the sample.

[0074] IELoss ignores the edge area of ​​the sample image when calculating the loss value. The calculation area is no longer the entire image area, but the middle area, which can avoid the overfitting phenomenon caused by insufficient pixel features at the edge of the image and improve the classification accuracy on the test set to a certain extent. Therefore, the structure of the improved U-Net network model suitable for flood monitoring is shown in Figure 4. Its network structure consists of two parts. Figure 4 The left half consists of convolution and downsampling operations for feature extraction. The network inputs a 3-channel 256×256 resolution image. The feature map obtained after 3×3 convolution is activated using the ReLU function, and then downsampled using the maximum pooling method. Figure 4 The right half consists of convolution and upsampling operations to restore the dimension of the feature map. The upsampling operation is introduced to increase the resolution of the image representation, thereby compensating for the loss of spatial resolution. The network finally uses a 1×1 convolution to output the result map.

[0075] The flood monitoring information extraction method of the present invention applies the improved U-Net network model to the Poyang Lake flood disaster from July to August 2020 to test the effect and recognition accuracy of the flood monitoring information extraction method after being applied to flood monitoring. Please refer to Figure 5 and Figure 6 shown.

[0076] First, the Sentinel-1 SAR image of July 26, 2020, during the flood disaster, was obtained. The Sentinel-1 SAR image was preprocessed using the Sentinel data processing software SNAP to obtain the 10-meter resolution dual-polarization backscatter distribution map of Poyang Lake. After that, the preprocessed Sentinel-1 dual-polarization image was obtained, and the spectral features SDWI derived from VV and VH polarization were obtained using Python programming, and VH, VV and SDWI were combined into a new three-band image. Typical areas of the new image with resolutions of 4467×4365 and 2681×2585 were cropped as training images and test images, respectively. Then, the training images and test images were finely labeled, and then the images and corresponding label images were regularly grid cropped and data enhanced. Finally, 1155 training samples, 289 validation samples, and 132 test samples with a resolution of 256×256 were obtained. After multiple iterative experiments, suitable hyperparameters were finally obtained. In this embodiment, the number of iterations was set to 30, the batch size was set to 2, and Adam was selected as the optimizer. The initial learning rate was set to 10-4 for model training. The model finally converged after iterative training, and achieved classification accuracies of 99.65%, 99.25% and 98.99% on the training set, validation set and test set, respectively. The model was used to predict the entire Sentinel-1 image on July 26, 2020, so as to obtain the water body information distribution map of the entire study area, as shown in the figure below. Figure 6 shown.

[0077] In order to test the flood monitoring information extraction method provided by the present invention, a comparative analysis is conducted between it and the traditional flood monitoring method based on SAR images. Two typical areas ( Figure 7 ), of which the water bodies in area 1 account for 33.4% and the water bodies in area 2 account for 5.8%. With the assistance of high-resolution optical remote sensing images, the real water body information is visually interpreted, and the artificially marked real water body information is compared with the water body information obtained by the algorithm. In this embodiment, the flood monitoring information extraction method provided by the present invention is tested from four indicators: recall rate, precision rate, false alarm rate and balanced F score (F1 Score, F1), that is, a variety of evaluation indicators are used to quantitatively evaluate the accuracy of water body information extraction, see formula (3) to formula (6) for details.

[0078]

[0079]

[0080] F a =1-P (5)

[0081]

[0082] In formula (3) to formula (6), P(T W ) represents the real water pixel interpreted by human vision, P(A W ) represents the water body pixel extracted by the algorithm. R represents the recall rate of the extraction algorithm. The recall rate represents the degree to which the extracted water body range is close to the real water body. The higher the value, the better. P represents the precision rate of the extraction algorithm. The precision rate represents the accuracy of the extracted water body. The higher the value, the better. F a It represents the false alarm rate of the extraction algorithm. F1 represents the comprehensive evaluation index of the extraction algorithm, which comprehensively considers the recall rate and precision rate. The higher the value, the better the algorithm effect. The comparison of the extraction accuracy of the three methods is shown in Table 1.

[0083] Table 1 Comparison of extraction accuracy of three methods

[0084]

[0085] The selected comparison areas have different water body proportions, and the advantages and disadvantages of the flood monitoring information extraction method provided by the present invention and the traditional method are tested from the perspective of different water body proportions. The results show that the Otsu global threshold method has a high recall rate and precision rate when the water body proportion is high, and the water body extraction effect is better. When the water body proportion is low, the Otsu global threshold method has the worst water body extraction accuracy and the worst accuracy of extracted water bodies; 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 improved U-Net network model. For the extraction of water body information in large-scale flood disasters, the object-oriented method has the worst extraction accuracy; among the three methods, the improved U-Net network model 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 flood monitoring information extraction method provided by the present invention has higher accuracy and better effect than the traditional flood monitoring information extraction method.

[0086] The flood monitoring information extraction device provided by the present invention is described below. The flood monitoring information extraction device described below and the flood monitoring information extraction method described above can be referred to each other.

[0087] Combine the following Figure 8 The flood monitoring information extraction device of the present invention is described, and the flood monitoring information extraction device comprises:

[0088] The information extraction module 100 is used to input the image to be predicted into the image semantic segmentation network model to obtain the water body information distribution map of the image to be predicted. The image semantic segmentation network model is trained based on the training set and the verification set corresponding to the sample image. The training set and the verification set are obtained by dividing the data set obtained by superimposing and combining the first polarization image, the second polarization image and the spectral feature image. The first polarization image and the second polarization image are obtained after polarization processing of the sample image, and the spectral feature image is derived from the first polarization image and the second polarization image.

[0089] In the information extraction module 100, the image semantic segmentation network model adopts the U-Net network model.

[0090] The traditional optical image water body information extraction device based on convolutional neural network needs to classify each pixel in the image, which makes its computational efficiency low, the computer storage overhead is large, and the classification performance and extraction accuracy are limited. In this context, Long et al. proposed a semantic segmentation device based on a fully convolutional network (FCN) in 2015 to classify images at the pixel level, thereby solving the image segmentation problem at the semantic level. Subsequently, Ronneberger et al. proposed a U-shaped symmetric network model, namely the U-Net network model, which can integrate low-dimensional and high-dimensional features to greatly improve the accuracy of image semantic segmentation. The structure of the classic U-Net network model is based on the expansion and modification of the fully convolutional network. The network structure is clear and elegant in U-shape, and it is one of the most popular image segmentation models at present.

[0091] At present, most of the flood monitoring information extraction devices based on SAR images are developed for medium and low resolution SAR images with wide range of applications, and are not completely applicable to high resolution SAR images and the extraction of flood spatiotemporal distribution information that requires very accurate results.

[0092] In recent years, deep learning has been widely used in the field of remote sensing water body extraction due to its powerful feature extraction capabilities. However, existing research on water body extraction combined with deep learning mainly focuses on optical images, and there is relatively little research on SAR images. Due to the randomly distributed speckle noise on SAR images and the complex flood inundation scenes, this hinders the identification and monitoring of floods based on deep learning of SAR images. At the same time, SAR images are usually dual-polarization data with relatively few polarization bands. It is difficult to distinguish water bodies from shadows based on the backscattering coefficient, texture, and contour of the ground objects alone, resulting in low water body extraction accuracy. Therefore, applying deep learning to water body information extraction requires improving the existing deep learning network model to make it suitable for high-resolution SAR images.

[0093] In the information extraction module 100, an improved U-Net network model is used in order to adapt to high-resolution SAR images, thereby making the water body information extracted from the high-resolution SAR images more in line with the requirements of flood monitoring.

[0094] Thereafter, the flood monitoring result may be obtained based on the water body information distribution map obtained by the information extraction module 100 .

[0095] Since the information extraction module 100 uses an improved U-Net network model, the water body information distribution map obtained from the image to be predicted can be extracted with higher accuracy, and the image to be predicted can use a SAR image with higher resolution, thereby making the flood monitoring results obtained according to the water body information distribution map obtained by the information extraction module 100 more accurate and more timely.

[0096] The flood monitoring information extraction device of the present invention, by using an improved image semantic segmentation network model, makes the extraction accuracy of the water body information distribution map obtained from the image to be predicted higher, and the image to be predicted can use a SAR image with higher resolution, thereby making the flood monitoring results subsequently obtained based on the water body information distribution map more accurate and more timely, and can monitor floods more timely and accurately, while also being applicable to a wider range of occasions and having better promotion value.

[0097] Combine the following Fig. 9 The flood monitoring information extraction device of the present invention is described. The image semantic segmentation network model is obtained by training the following modules:

[0098] The multi-polarization processing unit 120 is used to perform multi-polarization processing on the sample images to obtain a first polarization image and a second polarization image.

[0099] In the multi-polarization processing unit 120, the multi-polarization processing includes vertical transmission vertical reception (VV) polarization processing and vertical transmission horizontal reception (VH) polarization processing. SAR remote sensing systems commonly use four polarization modes, namely horizontal transmission horizontal reception (HH) polarization processing, vertical transmission vertical reception (VV) polarization processing, horizontal transmission vertical reception (HV) polarization processing and vertical transmission horizontal reception (VH) polarization processing. The electric field vector of the energy pulse emitted by the radar can be polarized in the vertical or horizontal plane. 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. Among them, HH and VV are co-polarized, and HV and VH are inverse (cross) polarized. The first polarization image is the polarization image after VV polarization processing, and the second polarization image is the polarization image after VH polarization processing.

[0100] The derivative processing unit 130 is used to perform derivative processing on the first polarization image and the second polarization image to obtain a spectral characteristic image.

[0101] In the derivative processing unit 130 , Python programming can be used to obtain the polarization-derived spectral feature SDWI data, ie, the spectral feature image, after VV and VH processing.

[0102] Considering the limited polarization data of SAR images and the insufficient richness of model feature learning, spectral features are introduced in step S130 to solve the problem of insufficient model features. The spectral feature SDWI data derived from the dual polarization data and polarization data of the SAR image are superimposed and combined into a new image, and the superimposed new image is used as the data set of the U-Net network model. The spectral feature SDWI data is obtained by the SAR image water body information extraction device proposed by Jia Shichao et al. in 2018.

[0103] The data set acquisition unit 140 is used to obtain a data set by superimposing the first polarization image, the second polarization image and the spectral characteristic image, and to superimpose VH polarization data, VV polarization data and spectral characteristic SDWI data to form a new three-band image.

[0104] In the data set acquisition unit 140, the dual-polarization data spectral feature SDWI data of the SAR image are combined to construct a new deep learning data set, which can solve the problem of insufficient model features and improve the accuracy of water body information extraction in the flood monitoring process to a certain extent.

[0105] The partitioning unit 150 is used to partition the data set into a training set, a validation set and a test set.

[0106] The marking unit 160 is used to obtain corresponding training recognition results and verification recognition results according to the training set and the verification set.

[0107] The training unit 170 is used to use the training set and the validation set as input data for training, use the training recognition structure corresponding to the training set as the label of the training set, and use the validation recognition result corresponding to the validation set as the label of the validation set, and adopt deep learning to perform training to obtain an image semantic segmentation network model for generating a water body information distribution map of the image to be predicted.

[0108] According to the flood monitoring information extraction device provided by the present invention, the flood monitoring information extraction device also includes:

[0109] The marking unit 160 also includes the following steps:

[0110] Regular net pruning and data enhancement are performed on the training set, validation set, and the labels corresponding to the training set and validation set, where the data enhancement includes horizontal flipping, vertical flipping, and diagonal mirroring operations.

[0111] Data enhancement is to obtain sufficient samples to avoid overfitting during model training.

[0112] Combine the following Fig.10 The flood monitoring information extraction device of the present invention is described, and the steps before the multi-polarization processing unit 120 also include the following steps:

[0113] The preprocessing unit 110 is used to preprocess the sample image to obtain a high-resolution dual-polarization backscatter distribution map, wherein the preprocessing includes one or a combination of orbit correction, thermal noise removal, radiation calibration, coherent speckle filtering, terrain correction, decibelization, cropping and mosaicking.

[0114] Fig.11 An example of a physical structure diagram of an electronic device is shown in FIG. Fig.11 As shown, the electronic device may include: a processor 810, a communications interface 820, a memory 830 and a communication bus 840, wherein the processor 810, the communications interface 820 and the memory 830 communicate with each other via the communication bus 840. The processor 810 may call the logic instructions in the memory 830 to execute the flood monitoring information extraction method, which includes the following steps:

[0115] S100, inputting the image to be predicted into the image semantic segmentation network model to obtain a water body information distribution map of the image to be predicted;

[0116] Among them, the image semantic segmentation network model is trained based on the training set and the verification set corresponding to the sample image. The training set and the verification set are obtained by dividing the data set obtained by superimposing and combining the first polarization image, the second polarization image and the spectral feature image. The first polarization image and the second polarization image are obtained after polarization processing of the sample image, and the spectral feature image is derived from the first polarization image and the second polarization image.

[0117] 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.

[0118] On the other hand, the present invention also provides a computer program product, the computer program product includes a computer program stored on a non-transitory computer-readable storage medium, the computer program includes program instructions, when the program instructions are executed by a computer, the computer can execute the flood monitoring information extraction method provided by the above methods, the method comprising the following steps:

[0119] S100, inputting the image to be predicted into the image semantic segmentation network model to obtain a water body information distribution map of the image to be predicted;

[0120] Among them, the image semantic segmentation network model is trained based on the training set and the verification set corresponding to the sample image. The training set and the verification set are obtained by dividing the data set obtained by superimposing and combining the first polarization image, the second polarization image and the spectral feature image. The first polarization image and the second polarization image are obtained after polarization processing of the sample image, and the spectral feature image is derived from the first polarization image and the second polarization image.

[0121] In another aspect, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to perform the flood monitoring information extraction method provided above, the method comprising the following steps:

[0122] S100, inputting the image to be predicted into the image semantic segmentation network model to obtain a water body information distribution map of the image to be predicted;

[0123] Among them, the image semantic segmentation network model is trained based on the training set and the verification set corresponding to the sample image. The training set and the verification set are obtained by dividing the data set obtained by superimposing and combining the first polarization image, the second polarization image and the spectral feature image. The first polarization image and the second polarization image are obtained after polarization processing of the sample image, and the spectral feature image is derived from the first polarization image and the second polarization image.

[0124] 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.

[0125] 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, 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.

[0126] 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 flood monitoring information extraction method, characterized in that: The following steps are involved: Inputting the image to be predicted into the image semantic segmentation network model to obtain a water body information distribution map of the image to be predicted; The image semantic segmentation network model is obtained by training based on a training set and a validation set corresponding to a sample image. The training set and the validation set are obtained by dividing a data set obtained by superimposing and combining a first polarization image, a second polarization image, and a spectral feature image. The first polarization image and the second polarization image are obtained by polarization processing of the sample image, and the spectral feature image is derived from the first polarization image and the second polarization image. The image semantic segmentation network model is trained by the following steps: Performing multi-polarization processing on the sample images respectively to obtain the first polarization image and the second polarization image; Performing derivative processing on the first polarization image and the second polarization image to obtain a spectral characteristic image; The data set is obtained by superimposing and combining the first polarization image, the second polarization image and the spectral feature image; Dividing the data set to obtain the training set and the validation set; Obtain corresponding training recognition results and verification recognition results according to the training set and the verification set; Using the training set and the validation set as input data for training, using the training recognition result corresponding to the training set as the label of the training set, and using the validation recognition result corresponding to the validation set as the label of the validation set, training is performed using a deep learning method to obtain an image semantic segmentation network model for generating the water body information distribution map of the image to be predicted; The multi-polarization processing includes vertical transmission vertical reception polarization processing and vertical transmission horizontal reception polarization processing; The image semantic segmentation network model is a U-Net network model; The loss function of the image semantic segmentation network model adopts an edge-ignoring cross entropy function.

2. The method for extracting flood monitoring information according to claim 1, characterized in that: Obtaining corresponding training recognition results and verification recognition results according to the training set and the verification set also includes the following steps: Regular net pruning and data enhancement are performed on the training set, the validation set, and the labels corresponding to the training set and the validation set; wherein the data enhancement includes horizontal flipping, vertical flipping, and diagonal mirroring operations.

3. The method for extracting flood monitoring information according to claim 1, characterized in that: The method further includes the following steps before performing multi-polarization processing on the sample images to obtain the first polarization image and the second polarization image: The sample image is preprocessed; wherein the preprocessing includes one or a combination of orbit correction, thermal noise removal, radiation calibration, coherent speckle filtering, terrain correction, decibelization, cropping and mosaicking.

4. A flood monitoring device, characterized in that: include: An information extraction module (100) is used to input the image to be predicted into the image semantic segmentation network model to obtain a water body information distribution map of the image to be predicted; The image semantic segmentation network model is obtained by training based on a training set and a validation set corresponding to a sample image. The training set and the validation set are obtained by dividing a data set obtained by superimposing and combining a first polarization image, a second polarization image, and a spectral feature image. The first polarization image and the second polarization image are obtained by polarization processing of the sample image, and the spectral feature image is derived from the first polarization image and the second polarization image. The image semantic segmentation network model is trained by the following steps: Performing multi-polarization processing on the sample images respectively to obtain the first polarization image and the second polarization image; Performing derivative processing on the first polarization image and the second polarization image to obtain a spectral characteristic image; The data set is obtained by superimposing and combining the first polarization image, the second polarization image and the spectral feature image; Dividing the data set to obtain the training set and the validation set; Obtain corresponding training recognition results and verification recognition results according to the training set and the verification set; Using the training set and the validation set as input data for training, using the training recognition result corresponding to the training set as the label of the training set, and using the validation recognition result corresponding to the validation set as the label of the validation set, training is performed using a deep learning method to obtain an image semantic segmentation network model for generating the water body information distribution map of the image to be predicted; The multi-polarization processing includes vertical transmission vertical reception polarization processing and vertical transmission horizontal reception polarization processing; The image semantic segmentation network model is a U-Net network model; The loss function of the image semantic segmentation network model adopts an edge-ignoring cross entropy function.

5. 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 the flood monitoring information extraction method according to any one of claims 1 to 3 are implemented.

6. 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 monitoring information extraction method according to any one of claims 1 to 3 are implemented.

Images