Lake water body extraction method and system based on deep learning, storage medium and electronic equipment

Abstract

The invention discloses a lake water body extraction method and system based on deep learning, a storage medium and electronic equipment, and relates to the technical field of remote sensing image processing. The method comprises the following steps: constructing a multi-level deep learning water body extraction model by utilizing polarization characteristic differences between a lake water body and other ground features in polarization decomposition; on the basis, accurate extraction of water boundary information is further realized through a feature learning and optimization process. According to the method, a new thought is provided for obtaining long-time series lake continuous observation data, and data support is provided for further developing seasonal change analysis, rapid flood season early warning, disaster assessment and the like of a frozen circle lake water body.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing image processing, and in particular to a method for extracting lake water bodies. Background Art

[0002] Freshwater resources on Earth only account for 3% of the total water reserves. As an important freshwater reserve, lake water bodies not only affect the regulation and storage of normal water resources in their seasonal changes, but also play an important role in flood control. These fluctuations will further affect various aspects of the lake basin ecosystem, including the habitat and reproduction of wetland organisms. These impacts have expanded the scope of lake basin ecosystem processes and highlighted the importance of lakes in the balance of water resources and ecological balance on Earth. In particular, in the past 50 years, satellite remote sensing has detected rapid melting and retreat of glaciers on the Qinghai-Tibet Plateau, triggering changes in the spatiotemporal distribution of lake water resources and water cycle processes in the region. The evolution of the hydrological environment will undoubtedly lead to changes in the ecological environment. Therefore, it is particularly important to monitor the seasonal changes of lake water bodies.

[0003] Traditional hydrological field observations are arduous and time-consuming, especially in the observation range of cryosphere lakes, which are obviously limited. It is even more difficult to obtain large-scale water body information at the basin scale. The rapid development of remote sensing technology in recent decades has the advantages of wide coverage and short revisit period, making it possible to monitor large-scale and continuous lake water bodies. With the continuous improvement of satellite remote sensing data acquisition capabilities, lake water body monitoring technology based on remote sensing observations has become the focus of current research. The cryosphere is composed of frozen water bodies, including lakes, rivers, sea ice, glaciers, ice sheets, snow and permafrost. Therefore, any qualitative or quantitative changes in the physical properties and range of the cryosphere may affect global atmospheric circulation, ocean and temperature, sea level and ocean current patterns.

[0004] Due to the difficulty in obtaining on-site data of mountain glaciers and glacial lakes, satellite remote sensing provides an effective observation platform for remote sensing of cryosphere lakes on a global and regional scale, and further captures hydrological variables including water boundaries, water depth, water level, water temperature and runoff estimation through corresponding data analysis algorithms. For example, by utilizing the absorption of near-infrared and infrared rays by water bodies and the rapid attenuation of radar waves in water, the accurate location of water body boundaries can be obtained using optical remote sensing and SAR images.

[0005] Optical remote sensing uses the difference in spectral reflectance between water bodies and other landforms to calculate the normalized difference index of reflectance in the infrared and near-infrared bands, and then uses threshold screening, feature selection and other methods to extract lake water information. This is a widely used method. However, optical data is affected by factors such as cloudy and rainy weather, which makes it difficult to obtain image data and brings challenges to subsequent data processing. In particular, for monitoring lake water bodies during the flood season and in the cryosphere, continuous observation data is often required to monitor short-term changes in lake water bodies. Optical data is affected by factors such as cloudy and rainy weather and is difficult to meet the above requirements.

[0006] Therefore, how to reduce or avoid the influence of factors such as cloudy and rainy weather, reduce the difficulty of acquiring water body image data, and improve the accuracy of water body boundary information extraction is a difficult problem that needs to be solved urgently when using optical remote sensing technology to obtain water body information. Summary of the invention

[0007] Based on synthetic aperture radar data and using a multi-level deep learning model, the present invention provides a lake water extraction method, system, storage medium and electronic device based on deep learning, which solves the problem that the optical data of lake water cannot be continuously monitored due to the influence of cloudy and rainy weather, and realizes the accurate extraction of lake water boundary information; the present invention can obtain long-term series of lake water data, and provide data support for further seasonal change analysis of cryosphere lake water, rapid flood warning, disaster assessment, etc. It is specifically achieved through the following technologies.

[0008] A lake water body extraction method based on deep learning includes the following steps:

[0009] Preprocess the SAR images of the lake water body and the surrounding area to obtain the preprocessed covariance matrix;

[0010] Perform polarization decomposition on the preprocessed covariance matrix to generate a polarization decomposition matrix;

[0011] Preliminarily mark the water body area and the non-water body area of ​​the SAR image through visual interpretation, cut the polarization decomposition matrix into a number of samples, divide and form a training set and a test set;

[0012] Constructing an original deep learning water body extraction model, and training using the training set and the test set to obtain an optimal deep learning water body extraction model;

[0013] The SAR image of the lake water body to be tested is input into the optimal deep learning water body extraction model, and the extraction result of the lake water body to be tested is output.

[0014] The present invention uses the image data of Sentinel-1 radar satellite, which has a certain degree of penetration and can effectively identify the problem of lake water freezing, thereby accurately extracting the water area. In addition, the radar satellite can also obtain accurate water body information without being affected by the atmosphere such as clouds and rain, thereby improving the extraction accuracy of lake water bodies.

[0015] Furthermore, the process of preprocessing the SAR images of the lake water body and the surrounding area includes: performing orbit correction, visual interpretation, radiation calibration and strip merging on the SAR images of the lake water body and the surrounding area to obtain a scattering matrix of the complete strip; converting the scattering matrix into the original covariance matrix, performing multi-view, geometric correction and terrain correction to obtain a preprocessed covariance matrix.

[0016] Furthermore, the method of visual interpretation of SAR images is as follows: the observers preliminarily mark the water and non-water areas on the SAR images, and then select some of these areas as training samples, and select the water sample set in combination with the field data.

[0017] Furthermore, the method of converting the scattering matrix into the original covariance matrix is:

[0018] The expression of the scattering matrix S is shown in Formula I;

[0019] Formula I;

[0020] in, , , and They correspond to different scattering matrix components and satisfy ;

[0021] The calculation method of the covariance matrix C is shown in Formula II;

[0022] [C]= "> S HH 2 2 "> S HH S HV * "> S HH S VV * 2 "> S HV S HH * 2 "> S HV 2 2 "> S HV S VV * "> S VV S HH * 2 "> S VV S VH * "> S VV 2 Formula II;

[0023] Here, * represents complex conjugation.

[0024] Furthermore, the visual interpretation method is as follows: the observer preliminarily marks the water body and non-water body areas on the SAR image, randomly crops the corresponding polarization decomposition matrix into a number of samples of 512×512 size, and divides the samples into a training set and a test set.

[0025] Further, the original deep learning water body extraction model is constructed by several layers of convolutional network structures; the convolutional network structure includes convolution units, and the convolutional network structures other than the bottom layer of the convolutional network structure also include convolution compensation units and attention mechanisms; further, the deep learning water body extraction model is constructed by several layers of convolutional network structures; the convolutional network structure includes convolution units, and the convolutional network structures other than the bottom layer of the convolutional network structure also include convolution compensation units and attention mechanisms;

[0026] The training method of the original deep learning water body extraction model is: inputting the training set into the convolution unit of the original deep learning water body extraction model, performing feature extraction in the convolution feature kernel function of the first layer, and obtaining the initial convolution output feature of the first layer;

[0027] The convolution compensation unit is used to process the initial convolution output features of the first layer, and the adjacent pixels at the water body boundary are included in the convolution processing of the first layer; the adjacent pixel size of the first layer is 16, and the center offset position is 10;

[0028] A corresponding weight coefficient is set for the water body feature of each layer of convolutional units as the weight coefficient of the attention mechanism; the attention mechanism of the first layer is used to assign a weight to each latitude of the initial convolution output feature of the first layer to obtain the final output feature of the first layer of convolution;

[0029] The initial convolution output features of the first layer are downsampled and input into the convolution unit of the second layer. The features are extracted in the convolution feature kernel function of the second layer, and then processed by the convolution compensation unit and the attention mechanism in turn to obtain the final output features of the second layer. Similarly, the final output features of the third and fourth layers are obtained respectively.

[0030] The initial convolution output features of the fourth layer are directly downsampled and directly processed by the fifth layer convolution unit as the final output features of the fifth layer;

[0031] The final output features of the fifth layer are deconvolved and then fused with the final output features of the fourth layer to obtain the fused data of the fourth layer; the final output features of the fifth layer convolutional network structure are upsampled through the attention mechanism of the fourth layer to obtain the upsampled data of the fifth layer; the fused data of the fourth layer and the upsampled data of the fifth layer are deconvolved together to obtain the deconvolved data of the fourth layer;

[0032] The deconvolution data of the fourth layer is fused with the final output features of the third layer to obtain the fused data of the third layer; the deconvolution data of the fourth layer is upsampled through the attention mechanism of the third layer to obtain the upsampled data of the fourth layer; the fused data of the third layer and the upsampled data of the fourth layer are deconvolved together to obtain the deconvolution data of the third layer; and so on, the deconvolution data of the first layer is finally obtained; the convolution kernel function of each layer during deconvolution corresponds to the convolution kernel function parameters of the layer during convolution;

[0033] The deconvolution numbers of the first layer corresponding to the training set and the test set are respectively output as a probability value between 0 and 1, and the probability value is used to determine whether the deconvolution data of the first layer belongs to a water body or a non-water body, that is, to obtain the final water body extraction result;

[0034] The test set is input into the deep learning water body extraction model to obtain the output result, that is, the classification result of water body and non-water body; the cross information entropy loss function is used, and the classification accuracy of water body and non-water body in the output result is calculated; according to the normalized loss function and the classification accuracy of water body and non-water body, the optimal deep learning water body extraction model is judged and output.

[0035] Furthermore, when the normalized loss function is less than 0.005 and the classification accuracy of water bodies and non-water bodies is greater than 95, the training of the deep learning water body extraction model is completed and the output is used as the optimal deep learning water body extraction model; otherwise, the parameters at this time are used as the initialization parameters of the model and re-trained in the deep learning water body extraction model.

[0036] The present invention also provides a lake water body extraction system based on deep learning, and the lake water body extraction system based on deep learning includes:

[0037] A preprocessing module is used to preprocess the SAR images of the lake water body and the surrounding area to obtain a preprocessed covariance matrix;

[0038] A model building module is used to perform polarization decomposition on the preprocessed covariance matrix to obtain a polarization decomposition matrix; to obtain a training set and a test set based on the polarization decomposition matrix; to build an original deep learning water body extraction model; to train and test the original deep learning water body extraction model using the training set and the test set to obtain an optimal deep learning water body extraction model; and to output the extraction results of the SAR image of the lake water body to be tested.

[0039] The present invention also provides a computer-readable storage medium on which a computer program is stored, and the computer program is used to enable a computer to execute the above-mentioned lake water extraction method based on deep learning.

[0040] 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. When the processor executes the computer program, the above-mentioned lake water extraction method based on deep learning is implemented.

[0041] Compared with the prior art, the present invention is beneficial in that: the present invention utilizes the polarization feature differences between lake water and other landforms in polarization decomposition to construct a multi-level deep learning water extraction model; on this basis, the accurate extraction of water boundary information is further achieved through feature learning and optimization processes. The present invention provides a new idea for obtaining long-term series of continuous observation data of lakes, and provides data support for further seasonal change analysis of cryosphere lake water, rapid flood warning, disaster assessment, etc., and has obvious advantages and innovations. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is an overall flow chart of the lake area extraction method provided by the present invention.

[0043] Figure 2 This is the structural diagram of the deep learning water body extraction model in the lake area extraction method.

[0044] Figure 3 This is the monthly average result of the Selin Co Lake area in 2020 based on Sentinel-1 polarimetric data in the following specific implementation method.

[0045] Figure 4 It is the time series of the Sentinel-1 estimation results of the deep learning water extraction model of the present invention and the lake extraction area of ​​Sentinel-2 in the same period.

[0046] Figure 5 Correlation analysis plots of lake area extracted for Sentinel-1 and Sentinel-2.

[0047] Figure 6 This is an example of the extraction results of the area of ​​the Selin Co Lake from Sentinel-1 and Sentinel-2. DETAILED DESCRIPTION

[0048] The technical scheme of the present invention will be described clearly and completely below in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than 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.

[0049] This embodiment takes the water body extraction of Sentinel-1 polarimetric SAR image as an example. The polarimetric SAR lake water body extraction method used is as follows: Figure 1 As shown, it specifically includes the following steps one to five.

[0050] Step 1: Data preparation

[0051] The data used in this example is from the Selin Co Lake in the Tibet Autonomous Region, which has an area of ​​2395 km 2 , as the cryosphere lake test object. According to the data acquisition time of the sampling site, two Sentinel-1 multispectral optical remote sensing data with good imaging quality were downloaded, which are the original polarimetric SAR single-look complex data (SLC, SingleLook Complex) acquired from January to December 2020. The DEM (Digital Elevation Model) data used is derived from ASTER DEM.

[0052] Step 2: Data preprocessing

[0053] Sentinel-1 SAR images are multi-strip data in dual-polarization form. Before feature extraction and input into the deep learning model, the SAR image data needs to be preprocessed, multiple strips need to be merged, and then converted into a covariance matrix form that has been calibrated by radiation and terrain correction.

[0054] like Figure 1 As shown, the preprocessing process includes performing orbit correction, radiation calibration, and strip merging on the SAR image to obtain a scattering matrix of the complete strip; converting the scattering matrix into an original covariance matrix, performing multi-viewing, geometric correction, and terrain correction on the original covariance matrix to obtain a preprocessed covariance matrix.

[0055] 1. Track correction.

[0056] In order to eliminate the geometric position change (phase deviation) caused by the real-time orbit parameter error during Sentinel-1 imaging and improve the positioning accuracy of the water body boundary, it is necessary to perform accurate orbit correction on the acquired SAR images. The orbit correction process is implemented using SNAP software.

[0057] 2. Radiation calibration.

[0058] When the radar receives the reflected signal (echo signal) from the lake and the surrounding area, the refraction of the atmosphere will easily cause the echo signal to attenuate and change the polarization direction. The relative position change between the satellite and the ground object will easily cause the center frequency to change. The change in ambient temperature will easily cause the system parameters to change. In order to eliminate the above influences and obtain the real ground object reflection signal, the data needs to be calibrated by radiation.

[0059] Firstly, the ground control point data in the SAR image data is extracted and used as the reference data for absolute calibration. On this basis, a linear regression equation is established. The amplitude and phase components of the SLC data are corrected respectively to eliminate the scattering errors caused by atmospheric disturbances and the relative motion of the target and sensor, and the result after absolute radiation calibration is obtained.

[0060] 3. Strip merging.

[0061] Sentinel-1 data is usually in the form of multiple strips, for example, the IW (Interferometric Wide swath) mode contains three strips, and the EW (Extra Wide swath) mode contains 5 strips. After completing the radiation calibration, it is necessary to merge multiple strips and remove the pulse bands (i.e., the signal-free parts) in adjacent strips to obtain complete SAR image data, i.e., the polarization scattering matrix S of the complete strip, as shown in the following formula I.

[0062] Formula I

[0063] In the above formula II, , , and They correspond to different scattering matrix components and satisfy .

[0064] 4. Generate the covariance matrix.

[0065] The obtained polarization scattering matrix is ​​converted into the original covariance matrix C using the following formula II.

[0066] [C]= "> S HH 2 2 "> S HH S HV * "> S HH S VV * 2 "> S HV S HH * 2 "> S HV 2 2 "> S HV S VV * "> S VV S HH * 2 "> S VV S VH * "> S VV 2 Formula II

[0067] In formula II, * represents complex conjugation.

[0068] It should be noted that since Sentinel-1 is a dual-polarization mode, taking HH / HV as an example, the original covariance matrix is ​​expressed as follows:

[0069] Formula III

[0070] 5. Multi-view processing.

[0071] The SAR imaging process uses synthetic aperture technology to improve the resolution in azimuth, resulting in obvious coherent speckle noise in the image. This embodiment uses multi-view processing and azimuth averaging to reduce the impact of coherent speckle noise, while the downsampling process also greatly reduces the amount of subsequent data.

[0072] The multi-view processing method is to spatially downsample the original covariance matrix of the image and perform spatial averaging on the downsampled data. In the multi-view processing process, the equivalent number of views is 4.

[0073] 6. Geometry Correction

[0074] In order to eliminate perspective contraction, overlap and shadow caused by slant range imaging during radar imaging, it is necessary to perform geometric correction on the multi-view processed data. Common geometric correction methods include: range-Doppler, ellipsoid projection or collinear equation method.

[0075] 7. Terrain correction.

[0076] Due to the geometric and radiometric distortions caused by perspective contraction, overlap, shadow, etc. during radar slant range imaging, it is necessary to perform terrain correction on the geometrically corrected data.

[0077] Terrain correction uses digital elevation model (DEM) data to geocode the elements in the covariance matrix and give them real geographic coordinate information. At the same time, it uses the radiation distortion caused by terrain geometry in DEM data to correct the amplitude and phase of the covariance matrix to eliminate the influence of terrain.

[0078] In this embodiment, the DEM data used is SRTM (Shuttle Radar Topography Mission), and the terrain correction method is the range-Doppler method.

[0079] Step 3: Obtain polarization decomposition matrix

[0080] Water bodies are mainly surface scattered, and other objects (vegetation, buildings, etc.) are mostly dihedral and volume scattered. In order to accurately describe the scattering difference between lake water and other objects, this embodiment uses the Freeman decomposition method based on the terrain-corrected covariance matrix obtained in step 2, and uses the polarization feature difference between the two (water bodies and their surrounding objects) to obtain a polarization decomposition matrix (i.e., Freeman decomposition matrix).

[0081] The polarization decomposition matrix is ​​used to describe three different types of scatterers (surface scattering, dihedral scattering and volume scattering components) to characterize the scattering characteristics of water bodies.

[0082] Visual interpretation refers to the use of manual methods to preliminarily mark the approximate areas of water bodies and non-water bodies on SAR images, and then select some of these areas as training samples.

[0083] After obtaining the polarization decomposition matrix, the polarization decomposition matrix is ​​cropped into multiple 512×512 samples through visual interpretation to construct training sets and test sets; the labeled data corresponding to these samples are obtained through visual interpretation, and 1 is used to represent water bodies and 0 is used to represent non-water bodies.

[0084] The re-entry period of Sentinel-1 data is 6 days, and about 60 images of the same orbit can be obtained in a year. This implementation method uses data from 2015 to 2019 to select samples, with a total of 300 images, and about 15 samples are selected from each image. Finally, a total of 4179 samples were selected, and then 3643 of them were randomly selected for model training, and the rest were used for testing. Therefore, the training set of this implementation method contains 3643 samples of size 512×512, and the test set contains 536 samples of size 512×512.

[0085] Step 4: Build and train a deep learning water extraction model

[0086] In this step, a deep learning water body extraction model based on multi-layer convolution is constructed and trained.

[0087] 1. Construction of deep learning water extraction model

[0088] like Figure 2 As shown in the figure, the original deep learning water extraction model constructed in this step contains a total of five layers of downsampling convolutional network structures, each of which contains one convolution unit; except for the fifth convolutional network structure at the bottom layer, each of the other convolutional network structures is equipped with a convolution compensation unit and an attention mechanism. The specific structure is:

[0089] (1) Convolutional unit

[0090] When constructing a multi-layer convolutional network structure, the kernel function size of the convolution unit of each layer of the convolutional network structure is 3×3, 5×5, 7×7, 11×11 and 15×15 respectively. The downsampling coefficient (i.e., pooling factor) of the convolution units of two adjacent layers is 2.

[0091] (2) Convolution compensation unit

[0092] When extracting lake water bodies, the coherent imaging of the SAR image itself can easily lead to blurred water body boundaries, and the complex scattering mechanism of the water body boundary area can also easily cause misclassification.

[0093] In order to improve the above problems, a convolution compensation unit is set in each convolution network structure of the model except for the bottom layer (fifth layer). When training the deep learning water body extraction model, the adjacent pixels at the boundary of the water body are included in the convolution operation of the layer, where the adjacent pixel sizes of the first to fifth layers are 16, 8, 4, 2 and 1 respectively, and the center offset positions are 10, 8, 6, 4 and 2 respectively.

[0094] In this embodiment, except for the fifth layer unit where no convolution compensation unit is set, the sizes of the compensation units in the other four layers are all 1.

[0095] According to the size of the convolution compensation unit, the center position of the kernel function of the convolution unit is changed, thereby improving the ability of the deep learning water body extraction model to describe the boundary characteristics of the lake, and ultimately improving the accuracy of water body extraction.

[0096] The convolution compensation unit is a zero-filling operation during downsampling. During the encoding process of the convolution unit, the input features of the initial convolution layer (the first convolution layer) are retained, so that the detailed features of the water body SAR image are also retained. These detailed features correspond to the edge of the lake, and after multiple layers of convolution, the main information of the lake (i.e., contour features) is retained. Although the downsampled data lacks the detailed features of the lake edge, the coherent speckle noise in the image is filtered out, which improves the recognition accuracy of the subsequent water body edge.

[0097] (3) Attention Mechanism

[0098] This embodiment introduces an attention mechanism in each layer of the convolutional network structure except the bottom layer (fifth layer). Specifically, different weight coefficients are assigned to each dimension of the convolutional features of each layer of the convolutional network structure. The attention mechanism is used to focus on different features at different scales.

[0099] The first layer of convolutional units has the original image resolution, so a lot of detailed information such as the boundary features of the lake SAR image is retained. Therefore, more attention is paid to the gradient changes in the image, and the areas with large changes are marked as potential water body edges. As the convolutional layers proceed, the attention mechanism of the higher-level convolutional network structure pays more attention to the homogeneity of the image, marking smaller targets with sudden feature changes as interference items and filtering them out. The fifth layer of convolutional units loses some detailed information due to downsampling (i.e., pooling), but suppresses the influence of coherent speckle noise to a certain extent, and the characteristics of the water area are more accurately characterized.

[0100] 2. Training of Deep Learning Water Extraction Model

[0101] The training method of the original deep learning water body extraction model is mainly as follows: using the samples in the training set constructed in step three as data, for each sample, multi-scale convolution features are calculated through the convolution layer; the convolution compensation unit is used to improve the feature description accuracy of the lake boundary area; considering the different contributions of different layers of convolutional network structures to the deep learning water body extraction model, the attention mechanism is used to assign different weight coefficients to the convolution features of each layer of the convolutional network structure, and the features of different dimensions are redistributed; the features extracted from each layer are connected in series to form a multi-dimensional feature vector to describe the water bodies and non-water bodies in the lake area; the test set in step three is used to input the model for testing. If it meets the requirements, the optimal deep learning water body extraction model is obtained, which is used for the extraction of the water body of the lake to be tested; if it does not meet the requirements, the convolution iteration is continued until it meets the requirements.

[0102] The training method of the deep learning water body extraction model provided in this embodiment includes the following steps:

[0103] (1) If Figure 2 As shown in the figure, each sample of the training set obtained in step three (image polarization decomposition matrix, size 512×512) is input into the first layer convolution unit of the original deep learning water body extraction model, and feature extraction is performed in the convolution feature kernel function to obtain the initial convolution output features of the first layer, that is, the initial information of the water body boundary.

[0104] The convolution compensation unit is used to process the initial convolution output features of the first layer (i.e., the initial information of the water body boundary), and the adjacent pixels of the water body boundary are included in the convolution operation of the first layer. The adjacent pixel size of the first layer is 16, and the center offset position is 10; this is used to improve the feature description accuracy of the lake boundary area.

[0105] The sigmoid function is used to set the corresponding weight coefficient for the water body characteristics of each layer of convolutional units as the weight coefficient of the attention mechanism; the attention mechanism of the first layer is used to assign weights to each latitude of the initial convolution output features of this layer to obtain the final output features of the first layer of convolution.

[0106] In the low-level convolutional units, the weight coefficient is small, so more attention is paid to the edge area of ​​the water body. For areas where the boundary characteristics of the water body change less, the boundaries will be blurred in the high-level network based on the similarity of texture features.

[0107] In the high-level convolutional units, the weight coefficient is large, and the high-level convolutional network structure pays more attention to the texture features inside the water body and non-water body. For features with texture mutations, they will be marked as potential uncertainty areas, and a smaller weight coefficient will be given in the low-level network in the next loop iteration.

[0108] The attention mechanism calculates local information through convolution. That is, for the convolution process of different layers in the deep learning water body extraction model, the attention mechanism can extract different local information; through upsampling, the extracted areas of interest are combined as the final result of water body extraction.

[0109] (2) If Figure 2 As shown in the figure, the initial convolution output features of the first layer are downsampled (i.e., pooled), and the downsampled data is processed in the same way as the first layer of convolution network structure to obtain the final output features of the second layer of convolution. That is, the downsampled data is input into the convolution unit of the second layer, and features are extracted in the convolution feature kernel function to obtain the initial convolution output features of the second layer; then the initial convolution output features of the second layer are processed by the convolution compensation unit; finally, the attention mechanism of the second layer is used to process and obtain the final output features of the second layer.

[0110] And so on, the final output features of the third and fourth layers are obtained respectively.

[0111] Since the fifth-layer convolutional network structure does not contain compensation units and attention mechanisms, the initial convolution output features of the fourth layer are downsampled (optionally double downsampling) and directly processed by the fifth-layer convolutional unit as the final output features of this layer.

[0112] After multiple downsampling-convolution operations, the contour information of the water body image is obtained. At this time, the information of the main part of the lake water body is extracted. After the first layer of convolution processing, the edge information of the lake water body is obtained.

[0113] (3) If Figure 2 As shown, the final output features of the fifth-layer convolutional network structure are deconvolved and then fused with the final output features of the fourth layer to obtain the fourth-layer fused data; the final output features of the fifth-layer convolutional network structure are upsampled through the fourth-layer attention mechanism to obtain the fifth-layer upsampled data; the fourth-layer fused data and the fifth-layer upsampled data are deconvolved together to obtain the fourth-layer deconvolution data.

[0114] The deconvolution data of the fourth layer is fused with the final output features of the third layer to obtain the fused data of the third layer; the deconvolution data of the fourth layer is upsampled through the attention mechanism of the third layer to obtain the upsampled data of the fourth layer; the fused data of the third layer and the upsampled data of the fourth layer are deconvolved together to obtain the deconvolution data of the third layer.

[0115] The above-mentioned process of upsampling through the attention mechanism is the process of redistributing the data of the next layer (the second to fourth layers) through the weight coefficients of the attention mechanism of the adjacent previous layer (the first to third layers).

[0116] The convolution kernel function during deconvolution corresponds one to one with the convolution kernel function of the layer during convolution.

[0117] The above method is used and so on to finally obtain the deconvolution data of the first layer.

[0118] (4) The deconvolution data of the first layer is output as a probability value between 0 and 1. The probability value is used to determine whether the deconvolution data of the first layer belongs to a water body or a non-water body, thereby obtaining the final water body extraction result.

[0119] (5) The training set and the test set are input into the deep learning water body extraction model to obtain the output results, that is, the classification results of water bodies and non-water bodies. The cross information entropy loss function is used, and the classification accuracy of water bodies and non-water bodies in the output results is calculated.

[0120] Classification accuracy refers to comparing the water bodies and non-water bodies in the calculated output results with the labeled data in the training set and test set respectively. The percentage of correctly classified water body pixels in the total water body pixels is the water body classification accuracy, and the percentage of correctly classified non-water body pixels in the non-water body pixels is the non-water body classification accuracy.

[0121] When the normalized loss function is less than 0.005 and the classification accuracy of water bodies and non-water bodies is greater than 95, the training is completed. The deep learning water body extraction model at this time is output as the optimal deep learning water body extraction model.

[0122] Otherwise, the parameters of this iterative process are used as the initialization parameters of the model, and the deep learning water extraction model is iterated (trained) again until the above test requirements are met.

[0123] Step 5: Input the SAR image of the lake water body to be tested into the optimal deep learning water body extraction model to obtain the extraction result of the lake water body to be tested

[0124] When using the optimal deep learning water body extraction model obtained through training to extract the water body to be tested, the SAR image of the water body to be tested is first preprocessed and a polarization decomposition matrix is ​​generated. The polarization decomposition matrix is ​​then input into the optimal deep learning water body extraction model obtained in step four, and convolution processing is performed in the same manner as step four. The normalized posterior probability of each pixel, that is, the score of each pixel, is calculated, and pixels with scores greater than 0.5 are marked as water bodies to obtain the final water body extraction result.

[0125] Figure 3This is the monthly average result of the area of ​​the Selin Co Lake in 2020 based on Sentinel-1 polarimetric data. It can be seen from the figure that the method proposed in the present invention can achieve better results in maintaining the lake boundary information and the uniformity of the water body inside the lake.

[0126] Figure 4 This is a time series of the lake area extracted by Sentinel-1 using the deep learning water extraction model and Sentinel-2 during the same period. Figure 6 Extract lake area correlation analysis graphs for Sentinel-1 and Sentinel-2. Figure 4 and 5 It can be seen that the water area estimated based on the two data has good consistency. 2 =0.89. Overall, the lake area calculation results of the two are well consistent from July to October, while the lake area results of Sentinel-1 are larger than those of Sentinel-2 in other months. This is mainly due to the influence of water freezing, which makes the boundary between water and non-water relatively blurred in the Sentinel-2 optical data, so the area estimation results of Sentinel-2 are low. This shows that the use of Sentinel-1 can overcome the influence of water freezing to a certain extent and improve the extraction accuracy of water.

[0127] Figure 6 This is a comparison of the Sentinel-1 Selin Co Lake extraction results and the Sentinel-2 optical image lake extraction results using the deep learning water body extraction model of the present invention. It can be seen that this embodiment can extract obvious water body boundary information; while the optical image has some errors due to the reception of cloud and frozen ice images, which mainly appear in the lake edge area. From the lake boundary images of the two sets of data, it can be seen that for the frozen-thaw state of the southeastern lake, the water body characteristics can be detected in Sentinel-1, while the increased reflectivity of the frozen state of Sentinel-2 leads to missed detection.

[0128] The above specific embodiments describe the implementation of the present invention in detail, but the present invention is not limited to the specific details in the above embodiments. Within the scope of the claims and technical concept of the present invention, the technical solution of the present invention can be modified and changed in many simple ways, and these simple modifications all belong to the protection scope of the present invention.

Claims

1. A lake water extraction method based on deep learning, characterized in that: The following steps are involved: Preprocess the SAR images of the lake water body and the surrounding area to obtain the preprocessed covariance matrix; Perform polarization decomposition on the preprocessed covariance matrix to generate a polarization decomposition matrix; Preliminarily mark the water and non-water areas of the SAR image through visual interpretation, cut the polarization decomposition matrix into a number of samples, divide and form a training set and a test set; Constructing an original deep learning water body extraction model, and training using the training set and the test set to obtain an optimal deep learning water body extraction model; The SAR image of the lake water body to be tested is input into the optimal deep learning water body extraction model, and the extraction result of the lake water body to be tested is output.

2. The lake water extraction method based on deep learning according to claim 1 is characterized in that: The process of preprocessing the SAR images of the lake water body and the surrounding area includes: performing orbit correction, radiation calibration and strip merging on the SAR images of the lake water body and the surrounding area to obtain a scattering matrix of the complete strip; converting the scattering matrix into an original covariance matrix, performing multi-view, geometric correction and terrain correction, and obtaining a preprocessed covariance matrix.

3. The lake water extraction method based on deep learning according to claim 2 is characterized in that: The method of converting the scattering matrix into the original covariance matrix is: The expression of the scattering matrix S is shown in Formula I; Formula I; in, , , and They correspond to different scattering matrix components and satisfy ; The calculation method of the covariance matrix C is shown in Formula II; Formula II; Here, * indicates complex conjugation.

4. The lake water extraction method based on deep learning according to claim 2 is characterized in that: The visual interpretation method is as follows: the observer preliminarily marks the water body and non-water body areas on the SAR image, randomly crops the corresponding polarization decomposition matrix into a number of samples of 512×512 size, and divides the samples into a training set and a test set.

5. The lake water extraction method based on deep learning according to claim 1 is characterized in that: The original deep learning water body extraction model is constructed by several layers of convolutional network structures; the convolutional network structure includes a convolution unit, and the convolutional network structures other than the bottom layer of the convolutional network structure also include a convolutional compensation unit and an attention mechanism; The training method of the original deep learning water body extraction model is: inputting each sample of the training set into the first layer convolution unit of the original deep learning water body extraction model, performing feature extraction in the convolution feature kernel function of the first layer, and obtaining the initial convolution output feature of the first layer; The convolution compensation unit is used to process the initial convolution output features of the first layer, and the adjacent pixels at the water body boundary are included in the convolution processing of the first layer; the adjacent pixel size of the first layer is 16, and the center offset position is 10; A corresponding weight coefficient is set for the water body feature of each layer of convolutional units as the weight coefficient of the attention mechanism; the attention mechanism of the first layer is used to assign a weight to each latitude of the initial convolution output feature of the first layer to obtain the final output feature of the first layer of convolution; The initial convolution output features of the first layer are downsampled and input into the convolution unit of the second layer. The features are extracted in the convolution feature kernel function of the second layer, and then processed by the convolution compensation unit and the attention mechanism in turn to obtain the final output features of the second layer. Similarly, the final output features of the third and fourth layers are obtained respectively. The initial convolution output features of the fourth layer are directly downsampled and directly processed by the fifth layer convolution unit as the final output features of the fifth layer; The final output features of the fifth layer are deconvolved and then fused with the final output features of the fourth layer to obtain the fused data of the fourth layer; the final output features of the fifth layer convolutional network structure are upsampled through the attention mechanism of the fourth layer to obtain the upsampled data of the fifth layer; the fused data of the fourth layer and the upsampled data of the fifth layer are deconvolved together to obtain the deconvolved data of the fourth layer; The deconvolution data of the fourth layer is fused with the final output features of the third layer to obtain the fused data of the third layer; the deconvolution data of the fourth layer is upsampled through the attention mechanism of the third layer to obtain the upsampled data of the fourth layer; the fused data of the third layer and the upsampled data of the fourth layer are deconvolved together to obtain the deconvolution data of the third layer; and so on, the deconvolution data of the first layer is finally obtained; the convolution kernel function of each layer during deconvolution corresponds to the convolution kernel function parameters of the layer during convolution; The deconvolution numbers of the first layer corresponding to the training set and the test set are respectively output as a probability value between 0 and 1, and the probability value is used to determine whether the deconvolution data of the first layer belongs to a water body or a non-water body, that is, to obtain the final water body extraction result; The test set is input into the deep learning water body extraction model to obtain the output results, that is, the classification results of water bodies and non-water bodies; A cross information entropy loss function is used, and the classification accuracy of water bodies and non-water bodies in the output results is calculated; based on the normalized loss function and the classification accuracy of water bodies and non-water bodies, the optimal deep learning water body extraction model is judged and output.

6. The lake water extraction method based on deep learning according to claim 5 is characterized in that: When the normalized loss function is less than 0.005 and the classification accuracy of water bodies and non-water bodies is greater than 95, the training of the deep learning water body extraction model is completed and the output is used as the optimal deep learning water body extraction model; otherwise, the parameters at this time are used as the initialization parameters of the model and re-trained in the deep learning water body extraction model.

7. A lake water extraction system based on deep learning, characterized in that: The lake water extraction system based on deep learning includes: A preprocessing module is used to preprocess the SAR images of the lake water body and the surrounding area to obtain a preprocessed covariance matrix; A model building training module is used to perform polarization decomposition on the preprocessed covariance matrix to obtain a polarization decomposition matrix; to obtain a training set and a test set based on the polarization decomposition matrix; to build an original deep learning water body extraction model; to train and test the original deep learning water body extraction model using the training set and the test set to obtain an optimal deep learning water body extraction model; and to output the extraction results of the SAR image of the lake water body to be tested.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program is used to enable a computer to execute the lake water extraction method based on deep learning as described in any one of claims 1-6.

9. 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 computer program, it implements the lake water extraction method based on deep learning as described in any one of claims 1-6.

Images