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

Abstract

The application discloses a lake water body extraction method and system based on deep learning, a storage medium and an electronic device, and relates to the technical field of remote sensing image processing. The application utilizes the polarization characteristic difference between the lake water body and other ground objects in polarization decomposition to construct a multi-level deep learning water body extraction model. On this basis, the accurate extraction of water boundary information is realized through a feature learning and optimization process. The application provides a new idea for obtaining long-time sequence lake continuous observation data, and provides data support for further carrying out seasonal change analysis of the frozen earth lake water body, flood season rapid early warning, disaster assessment and the like.

Description

Technical Field

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

[0002] Freshwater resources account for only 3% of the Earth's total water storage. Lakes, as vital freshwater reserves, not only influence the regulation and storage of normal water resources through seasonal variations, but also play a crucial role in flood control. These fluctuations further affect various aspects of lake basin ecosystems, including the habitat and reproduction of wetland organisms. These impacts broaden the scope of lake basin ecosystem processes, highlighting the importance of lakes in maintaining the Earth's water and ecological balance. In particular, over the past 50 years, satellite remote sensing has detected the rapid melting and retreat of glaciers on the Tibetan Plateau, triggering changes in the spatiotemporal distribution of lake water resources and water cycle processes in the region. The evolution of the hydrological environment undoubtedly leads to changes in the ecological environment; therefore, monitoring the seasonal changes in lake water bodies is especially important.

[0003] Traditional hydrological field observations are arduous and time-consuming, especially with significant limitations in the observation range of cryosphere lakes, and obtaining large-scale water body information at the watershed scale is even more difficult. Remote sensing technology, which has developed rapidly in recent decades, offers advantages such as wide coverage and short revisit cycles, making it possible to monitor large-scale, continuous lake bodies. With the continuous improvement of satellite remote sensing data acquisition capabilities, lake body monitoring technology based on remote sensing observations has become a current research focus. The cryosphere consists of frozen water bodies, including lakes, rivers, sea ice, glaciers, ice sheets, snow cover, and permafrost. Therefore, any qualitative or quantitative changes in the physical properties and extent of the cryosphere can affect global atmospheric circulation, ocean and temperature, sea level, and ocean current patterns.

[0004] Due to the difficulty in acquiring on-site data for mountain glaciers and glacial lakes, satellite remote sensing provides an effective observation platform for remotely sensing cryosphere lakes at global and regional scales. Through corresponding data analysis algorithms, it can further capture hydrological variables including water boundaries, water depth, water level, water temperature, and runoff estimation. For example, by utilizing the absorption of near-infrared and infrared radiation by water bodies and the rapid attenuation of radar waves in water, accurate water boundary locations can be obtained using optical remote sensing and SAR imagery.

[0005] Optical remote sensing utilizes the difference in spectral reflectance between water bodies and other land features. It extracts lake water information by calculating the normalized difference index of reflectance in the infrared and near-infrared bands, followed by thresholding and feature selection. This is a widely used method. However, optical data is affected by factors such as cloud cover and rain, making image acquisition difficult and posing challenges to subsequent data processing. This is especially true for monitoring lakes during flood season and in the cryosphere, where continuous observation data is often needed to monitor short-term changes in lake water levels. Optical data, affected by cloud cover and rain, is often insufficient to meet these requirements.

[0006] Therefore, how to reduce or avoid the impact of factors such as cloud and rain weather, reduce the difficulty of acquiring water body image data, and improve the accuracy of water body boundary information extraction is an urgent problem to be solved when using optical remote sensing technology to acquire water body information. Summary of the Invention

[0007] This invention, based on synthetic aperture radar data and utilizing a multi-level deep learning model, provides a deep learning-based method, system, storage medium, and electronic device for lake water body extraction. It solves the problem of the inability to continuously monitor lake water optical data due to cloud and rain weather, and achieves accurate extraction of lake water boundary information. This invention can acquire long-term series of lake water data, providing data support for further research on the seasonal changes of cryosphere lake water, rapid flood season early warning, and disaster assessment. Specifically, it is achieved through the following technologies.

[0008] A deep learning-based method for extracting lake water features includes the following steps:

[0009] Preprocessing SAR images of the lake and its surrounding area yields a preprocessed covariance matrix.

[0010] The preprocessed covariance matrix is ​​polarized to generate a polarization decomposition matrix;

[0011] By visually interpreting the SAR image, the water and non-water areas are initially marked. The polarization decomposition matrix is ​​then cropped into several samples, which are divided and formed into training and testing sets.

[0012] Construct an original deep learning water extraction model, and train it using the training set and test set to obtain the optimal deep learning water extraction model.

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

[0014] This invention utilizes imagery data from the Sentinel-1 radar satellite, which possesses a certain degree of penetration, effectively identifying issues caused by lake freezing and thus accurately extracting water areas. Furthermore, the radar satellite is unaffected by atmospheric conditions such as clouds and rain, thereby acquiring accurate water information and improving the precision of lake water extraction.

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

[0016] Furthermore, the method for visually interpreting SAR images is as follows: observers initially mark the areas of water and non-water bodies on the SAR images, then select some of these areas as training samples, and combine them with field data to select a water body sample set.

[0017] Furthermore, the method for converting the scattering matrix into the original covariance matrix is ​​as follows:

[0018] The expression for the scattering matrix S is shown in Equation I;

[0019] Formula I;

[0020] in, , , and Each corresponds to a different scattering matrix component, and satisfies ;

[0021] The method for calculating the covariance matrix C is shown in Equation 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] Where * denotes complex conjugation.

[0024] Furthermore, the visual interpretation method is as follows: the observer initially marks the areas of water and non-water bodies on the SAR image, randomly cuts the corresponding polarization decomposition matrix into several 512×512 samples, and divides the samples into training set and test set.

[0025] Furthermore, the original deep learning water extraction model is constructed from several layers of convolutional network structures; the convolutional network structure includes convolutional units, and the other convolutional network structures besides the bottom layer also include convolutional compensation units and attention mechanisms; furthermore, the deep learning water extraction model is constructed from several layers of convolutional network structures; the convolutional network structure includes convolutional units, and the other convolutional network structures besides the bottom layer also include convolutional compensation units and attention mechanisms;

[0026] The training method of the original deep learning water body extraction model is as follows: input the training set into the convolutional unit of the original deep learning water body extraction model, extract features in the convolutional feature kernel function of the first layer, and obtain the initial convolutional output features of the first layer.

[0027] The initial convolution output features of the first layer are processed using a convolution compensation unit, and the neighboring pixels of the water body boundary are included in the convolution processing of the first layer; the size of the neighboring pixels in the first layer is 16, and the center offset position is 10.

[0028] A corresponding weight coefficient is set for the water features of each convolutional unit, which serves as the weight coefficient of the attention mechanism; the attention mechanism of the first layer is used to assign weights to each dimension of the initial convolutional output features of the first layer, thereby obtaining the final output features of the first convolution.

[0029] The initial convolutional output features of the first layer are downsampled and input into the convolutional unit of the second layer. Feature extraction is performed in the convolutional feature kernel function of the second layer, and then processed by the convolutional compensation unit and attention mechanism in turn to obtain the final output features of the second layer. This process is repeated to obtain the final output features of the third and fourth layers respectively.

[0030] The initial convolutional output features of the fourth layer are directly downsampled and processed by the fifth convolutional unit to become the final output features of the fifth layer.

[0031] The final output features of the fifth layer are deconvolutionally processed 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 deconvolutionally processed together to obtain the deconvolution data of the fourth layer.

[0032] The deconvolutional 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 deconvolutional 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 deconvolutional data of the third layer; and so on, until the deconvolutional data of the first layer is obtained; the convolution kernel function during deconvolution of each layer corresponds one-to-one with the parameters of the convolution kernel function during convolution of that layer;

[0033] Output a probability value between 0 and 1 for the deconvolution numbers of the first layer corresponding to the training set and the test set respectively. Use the probability value to determine whether the deconvolution data of the first layer belongs to water or non-water, and thus obtain the final water extraction result.

[0034] The test set is input into the deep learning water extraction model to obtain the output results, namely 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; based on the normalized loss function and the classification accuracy of water bodies and non-water bodies, the optimal deep learning water extraction model is determined 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 extraction model is completed, and the output is used as the optimal deep learning water extraction model; otherwise, the parameters at this time are used as the initialization parameters of the model, and the deep learning water extraction model is retrained.

[0036] This invention also provides a deep learning-based lake water extraction system, the deep learning-based lake water extraction system comprising:

[0037] The preprocessing module is used to preprocess SAR images of the lake and its surrounding area to obtain a preprocessed covariance matrix.

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

[0039] The present invention also provides a computer-readable storage medium having a computer program stored thereon, the computer program being used to cause a computer to execute the above-described deep learning-based lake water extraction method.

[0040] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the above-described deep learning-based lake water extraction method.

[0041] Compared with existing technologies, the advantages of this invention are as follows: This invention utilizes the differences in polarization characteristics between lake water bodies and other land features in polarization decomposition to construct a multi-level deep learning water extraction model; based on this, it further achieves accurate extraction of water boundary information through feature learning and optimization processes. This invention provides a new approach for obtaining long-term continuous observation data of lakes, and provides data support for further analysis of seasonal changes in cryosphere lake water bodies, rapid early warning during flood season, and disaster assessment, demonstrating significant advantages and innovation. Attached Figure Description

[0042] Figure 1 The overall flowchart of the lake area extraction method provided by the present invention is shown.

[0043] Figure 2 This is a structural diagram of a deep learning water extraction model in lake area extraction methods.

[0044] Figure 3 The following is a detailed implementation of the monthly average area of ​​Siling Co Lake in 2020 based on Sentinel-1 polarization data.

[0045] Figure 4 The time series of lake area extraction results from Sentinel-1 and Sentinel-2 of the deep learning water extraction model of this invention are presented.

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

[0047] Figure 6 Examples of lake area extraction results for Sentinel-1 and Sentinel-2. Detailed Implementation

[0048] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0049] This embodiment uses water extraction from Sentinel-1 polarimetric SAR imagery as an example. The polarimetric SAR lake water extraction method employed is as follows: Figure 1 As shown, the specific steps include the following five steps.

[0050] Step 1: Data Preparation

[0051] The data used in this embodiment comes from Lake Silingco in the Tibet Autonomous Region, which has an area of ​​2395 km². 2 This lake was used as a test subject for cryosphere lakes. Based on the data acquisition time at the sampling sites, two high-quality Sentinel-1 multispectral optical remote sensing images were downloaded: raw polarimetric SAR single-look complex (SLC) data acquired between January and December 2020. The DEM (Digital Elevation Model) data used was 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 to merge multiple strips and then convert them into a covariance matrix form that has been radiometrically calibrated and topographically corrected.

[0054] like Figure 1 As shown, the preprocessing process includes orbit correction, radiometric calibration, and strip merging of SAR images to obtain the scattering matrix of the complete strip; the scattering matrix is ​​converted into the original covariance matrix, and the original covariance matrix is ​​subjected to multi-look, geometric correction, and terrain correction to obtain the preprocessed covariance matrix.

[0055] 1. Track correction.

[0056] To eliminate geometric position changes (phase deviations) caused by real-time orbital parameter errors during Sentinel-1 imaging and improve the positioning accuracy of water body boundaries, precise orbital correction is required for the acquired SAR images. The orbital correction process is implemented using SNAP software.

[0057] 2. Radiation calibration.

[0058] When radar receives reflected signals (echo signals) from lakes and the surrounding area, atmospheric refraction can easily lead to signal attenuation and changes in polarization direction. Changes in the relative positions of satellites and ground targets can cause changes in the center frequency, and changes in ambient temperature can cause changes in system parameters. To eliminate these effects and obtain accurate ground target reflection signals, radiometric calibration of the data is necessary.

[0059] First, ground control point data is extracted from SAR image data and used as reference data for absolute calibration. Based on this, a linear regression equation is established. The amplitude and phase components of SLC data are corrected to eliminate scattering errors caused by atmospheric disturbances and the relative motion of the target and sensor, thus obtaining the results after absolute radiometric calibration.

[0060] 3. Strip merging.

[0061] Sentinel-1 data is typically in multi-strip form; for example, the IW (Interferometric Wide swath) mode contains three strips, and the EW (Extra Wide swath) mode contains five strips. After radiometric calibration, multiple strips need to be merged, and pulse bands (i.e., no-signal portions) in adjacent strips need to be removed to obtain complete SAR image data, i.e., the polarization scattering matrix S of the complete strip, as shown in Equation I below.

[0062] Formula I

[0063] In equation II above, , , and Each corresponds to a different scattering matrix component, and satisfies .

[0064] 4. Generate the covariance matrix.

[0065] The obtained polarization scattering matrix is ​​converted into the original covariance matrix C using Equation 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 Equation II, * denotes 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 ​​represented by Equation III below.

[0069] Formula III

[0070] 5. Multi-view processing.

[0071] SAR imaging improves azimuth resolution using synthetic aperture technology, resulting in significant speckle noise in the images. This embodiment employs multi-look processing, using azimuth averaging to reduce the impact of speckle noise while also significantly reducing the amount of subsequent data during the downsampling process.

[0072] The multi-view processing method involves spatially downsampling the original covariance matrix of the image and then spatially averaging the downsampled data. In multi-view processing, the equivalent number of views is 4.

[0073] 6. Geometric correction

[0074] To eliminate perspective distortion, overlay, and shadows caused by slant-range imaging during radar imaging, geometric correction is required for the multi-view processed data. Common geometric correction methods include range-Doppler, ellipsoidal projection, or collinearity equations.

[0075] 7. Terrain correction.

[0076] Due to geometric and radiation distortions caused by perspective contraction, overlay, and shadows during radar slant range imaging, terrain correction is required for the geometrically corrected data.

[0077] Topographic correction utilizes Digital Elevation Model (DEM) data to geocode the elements in the covariance matrix, assigning them true geographic coordinate information. Simultaneously, it uses the radiation distortion caused by topographic geometry in the DEM data to correct the amplitude and phase of the covariance matrix, eliminating the influence of topography.

[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 the polarization decomposition matrix

[0080] Water bodies primarily exhibit surface scattering, while other land features (vegetation, buildings, etc.) mostly exhibit dihedral and volume scattering. To accurately characterize the scattering differences between lake water bodies and other land features, this embodiment uses the terrain-corrected covariance matrix obtained in step two and the Freeman decomposition method to obtain a polarization decomposition matrix (i.e., the Freeman decomposition matrix) by utilizing the differences in polarization characteristics between the two (water bodies and their surrounding land features).

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

[0082] Visual interpretation refers to the use of manual methods to initially 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 by visual interpretation to construct the training set and the test set. The labeled data corresponding to these samples is obtained by visual interpretation, with 1 representing water bodies and 0 representing non-water bodies.

[0084] The replay period for Sentinel-1 data is 6 days, allowing for approximately 60 images of the same orbit to be acquired per year. This implementation uses samples selected from data spanning 2015-2019, totaling 300 images. Approximately 15 samples are selected from each image, resulting in a total of 4179 samples. Of these, 3643 are randomly selected for model training, and the remainder are used for testing. Therefore, the training set in this implementation contains 3643 samples of size 512×512, and the test set contains 536 samples of size 512×512.

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

[0086] This step constructs and trains a deep learning water extraction model based on multi-layer convolution.

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

[0088] like Figure 2 As shown, the original deep learning water extraction model constructed in this step consists of five downsampling convolutional network layers, each containing one convolutional unit. Except for the bottom fifth layer, each of the other convolutional network layers includes a convolutional compensation unit and an attention mechanism. The specific structure is as follows:

[0089] (1) Convolutional unit

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

[0091] (2) Convolution compensation unit

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

[0093] To address the aforementioned issues, a convolutional compensation unit is incorporated into each layer of the convolutional network structure, except for the lowest layer (the fifth layer). When training the deep learning water extraction model, neighboring pixels at the water body boundary are included in the convolutional operation of that layer. The neighboring pixel sizes for layers 1 through 5 are 16, 8, 4, 2, and 1, respectively, with center offsets of 10, 8, 6, 4, and 2, respectively.

[0094] In this embodiment, except for the fifth layer which does not have a convolution compensation unit, the compensation unit size of the other four layers is 1.

[0095] By adjusting the size of the convolutional compensation unit, the center position of the kernel function of the convolutional unit is changed, thereby improving the ability of the deep learning water extraction model to describe the features of lake boundaries and ultimately improving the accuracy of water extraction.

[0096] The convolutional compensation unit performs zero-padding during downsampling. During the encoding process of the convolutional unit, the input features of the initial convolutional layer (the first convolutional layer) are preserved, thus retaining the detailed features of the water SAR image. These detailed features correspond to the lake edges, and after multiple convolutions, the main information of the lake (i.e., contour features) is preserved. Although the downsampled data lacks the detailed features of the lake edges, it filters out speckle noise in the image, improving the accuracy of subsequent water body edge recognition.

[0097] (3) Attention mechanism

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

[0099] The first convolutional unit uses the original image resolution, thus preserving a large amount of detail information such as boundary features of the lake SAR image. Therefore, it pays more attention to gradient changes in the image, marking areas with large changes as potential water body edges. As multiple convolutional layers progress, the attention mechanism of higher-level convolutional network structures focuses more on the homogeneity of the image, marking small targets with abrupt feature changes as interference terms and filtering them out. The fifth convolutional unit, due to downsampling (i.e., pooling), loses some detail information, but it suppresses the influence of speckle noise to a certain extent, resulting in a more accurate characterization of water body areas.

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

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

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

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

[0104] The initial convolution output features of the first layer (i.e., the initial information of the water body boundary) are processed by the convolution compensation unit. The neighboring pixels of the water body boundary are included in the convolution operation of the first layer. The size of the neighboring pixels in 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 region.

[0105] The sigmoid function is used to set corresponding weight coefficients for the water features of each convolutional unit, which are then used as weight coefficients for the attention mechanism. The attention mechanism of the first layer is used to assign weights to each dimension of the initial convolutional output features of the first layer, thus obtaining the final output features of the first convolution.

[0106] In lower-level convolutional units, the weight coefficients are smaller, thus focusing more on the edge regions of the water body. For regions where the water body boundary features change little, the boundaries are blurred in higher-level networks based on the similarity of texture features.

[0107] In higher-level convolutional units, the weight coefficients are larger, and the higher-level convolutional network structure focuses more on the texture features inside and outside the water body. Features with abrupt texture changes are marked as potential uncertainty regions and given smaller weight coefficients in the lower-level network in the next iteration.

[0108] Attention mechanisms compute local information through convolution. Specifically, for different layers of the convolution process in a deep learning water extraction model, attention mechanisms can extract different local information. By upsampling, the extracted regions of interest are combined as the final result of water extraction.

[0109] (2) For example Figure 2 As shown, the initial convolutional output features of the first layer are downsampled (i.e., pooled). The downsampled data is then processed using the same method as the first layer convolutional network structure to obtain the final output features of the second layer convolution. Specifically, the downsampled data is input into the second layer convolutional unit, where feature extraction is performed in the convolutional feature kernel function to obtain the initial convolutional output features of the second layer. Subsequently, the initial convolutional output features of the second layer are processed using a convolutional compensation unit. Finally, the attention mechanism of the second layer is used for further processing to obtain the final output features of the second layer.

[0110] By analogy, the final output features of the third and fourth layers are obtained in sequence.

[0111] Since the fifth-layer convolutional network structure does not include compensation units and attention mechanisms, the initial convolutional output features of the fourth layer are downsampled (optionally by a factor of two) and then directly processed by the fifth-layer convolutional unit, becoming the final output features of that layer.

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

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

[0114] The deconvolutional 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 deconvolutional 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 deconvolutional data of the third layer.

[0115] The above-described upsampling process using the attention mechanism is a 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 same layer during convolution.

[0117] By repeating the above method, the first layer of deconvolution data is finally obtained.

[0118] (4) Output a probability value between 0 and 1 for the deconvolution data of the first layer. Use the probability value to determine whether the deconvolution data of the first layer belongs to water or non-water, and thus obtain the final water extraction result.

[0119] (5) Input the training set and test set into the deep learning water body extraction model to obtain the output results, namely the classification results of water bodies and non-water bodies. Use the cross-information entropy loss function and calculate the classification accuracy of water bodies and non-water bodies in the output results.

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

[0121] Training is complete when the normalized loss function is less than 0.005 and the classification accuracy for water and non-water bodies is greater than 95%. The output of the deep learning water extraction model at this point is taken as the optimal deep learning water extraction model.

[0122] Otherwise, the parameters from this iteration will be used as the initialization parameters for the model, and the deep learning water extraction model will be iterated (trained) again until the above test requirements are met.

[0123] Step 5: Input the SAR image of the lake to be measured into the optimal deep learning water extraction model to obtain the extraction results of the lake.

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

[0125] Figure 3The figure shows the monthly average area of ​​Siling Co Lake in 2020 based on Sentinel-1 polarization data. As can be seen from the figure, the method proposed in this invention achieves good results in preserving lake boundary information and maintaining the homogeneity of water within the lake.

[0126] Figure 4 The time series of lake area estimates using the Sentinel-1 deep learning water extraction model and the lake area extracted by Sentinel-2 during the same period are presented. Figure 6 Correlation analysis plots of lake area were extracted for Sentinel-1 and Sentinel-2. Figure 4 and 5 It can be seen that the estimated water area based on the two sets of data has good consistency, R 2 =0.89. Overall, the lake area calculations from both Sentinel-1 and Sentinel-2 show good consistency from July to October. However, in other months, the lake area results from Sentinel-1 are larger than those from Sentinel-2. This is mainly due to the effect of water freezing, which blurs the boundary between water and non-water bodies in the Sentinel-2 optical data, thus leading to a lower area estimate from Sentinel-2. This indicates that Sentinel-1 can overcome the effects of water freezing to some extent and improve the accuracy of water body extraction.

[0127] Figure 6 This paper compares the lake extraction results of Sentinel-1 and Sentinel-2 optical imagery using the deep learning water extraction model of this invention. It can be seen that this embodiment can extract clear water boundary information; while the optical imagery, due to the inclusion of cloud and frozen ice, has some errors, mainly appearing in the lake edge region. From the lake boundary images of the two sets of data, it can be seen that for the freeze-thaw state of the southeastern lake, water features can be detected using Sentinel-1, while Sentinel-2, due to increased reflectivity in the frozen state, leads to missed detections.

[0128] The above detailed embodiments describe the implementation of the present invention; however, the present invention is not limited to the specific details described in the above embodiments. Within the scope of the claims and technical concept of the present invention, various simple modifications and changes can be made to the technical solution of the present invention, and these simple modifications all fall within the protection scope of the present invention.

Claims

1. A method for lake water body extraction based on deep learning, characterized in that, Includes the following steps: Orbit correction, radiometric calibration, and strip merging are performed on SAR images of the lake and its surrounding area to obtain the scattering matrix of the complete strip; the scattering matrix is ​​converted into the original covariance matrix, and multi-look, geometric correction, and terrain correction are performed to obtain the preprocessed covariance matrix; The preprocessed covariance matrix is ​​polarized to generate a polarization decomposition matrix; The water and non-water areas of the SAR image are initially marked by visual interpretation. The polarization decomposition matrix is ​​then cropped into several samples, which are divided and formed into training and test sets. Construct an original deep learning water extraction model, and train it using the training set and test set to obtain the optimal deep learning water extraction model. The SAR image of the lake to be tested is input into the optimal deep learning water extraction model, and the extraction result of the lake to be tested is output. The original deep learning water extraction model is constructed from several layers of convolutional network structure; the convolutional network structure includes convolutional units, and the other convolutional network structures besides the bottom layer also include convolutional compensation units and attention mechanisms; The training method of the original deep learning water body extraction model is as follows: each sample of the training set is input into the first layer convolutional unit of the original deep learning water body extraction model, and feature extraction is performed in the convolutional feature kernel function of the first layer to obtain the initial convolutional output features of the first layer. The initial convolution output features of the first layer are processed using a convolution compensation unit, and the neighboring pixels of the water body boundary are included in the convolution processing of the first layer; the size of the neighboring pixels in the first layer is 16, and the center offset position is 10. A corresponding weight coefficient is set for the water features of each convolutional unit, which serves as the weight coefficient of the attention mechanism; the attention mechanism of the first layer is used to assign weights to each dimension of the initial convolutional output features of the first layer, thereby obtaining the final output features of the first convolution. The initial convolutional output features of the first layer are downsampled and input into the convolutional unit of the second layer. Feature extraction is performed in the convolutional feature kernel function of the second layer, and then processed by the convolutional compensation unit and attention mechanism in turn to obtain the final output features of the second layer. This process is repeated to obtain the final output features of the third and fourth layers respectively. The initial convolutional output features of the fourth layer are directly downsampled and processed by the fifth convolutional unit to become the final output features of the fifth layer. The final output features of the fifth layer are deconvolutionally processed 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 deconvolutionally processed together to obtain the deconvolution data of the fourth layer. The deconvolutional 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 deconvolutional 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 deconvolutional data of the third layer; and so on, until the deconvolutional data of the first layer is obtained; the convolution kernel function during deconvolution of each layer corresponds one-to-one with the parameters of the convolution kernel function during convolution of that layer; Output a probability value between 0 and 1 for the deconvolution numbers of the first layer corresponding to the training set and the test set respectively. Use the probability value to determine whether the deconvolution data of the first layer belongs to water or non-water, and thus obtain the final water extraction result. The test set is input into the deep learning water body extraction model to obtain the output results, namely 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. Based on the normalized loss function and the classification accuracy of water bodies and non-water bodies, the optimal deep learning water extraction model is determined and output.

2. The lake water extraction method based on deep learning according to claim 1, characterized in that, The method for converting the scattering matrix into the original covariance matrix is ​​as follows: The expression for the scattering matrix S is shown in Equation I; Formula I; where S HH , S HV , S VH and S VV correspond to different scattering matrix components, respectively, and satisfy S HV =S VH ; The method for calculating the covariance matrix C is shown in Equation II; Formula II; Where * denotes complex conjugation.

3. The lake water extraction method based on deep learning according to claim 1, characterized in that, The visual interpretation method is as follows: the observer initially marks the areas of water and non-water bodies on the SAR image, randomly cuts the corresponding polarization decomposition matrix into several 512×512 samples, and divides the samples into training set and test set.

4. The lake water extraction method based on deep learning according to claim 1, 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 extraction model is completed, and the output is used as the optimal deep learning water extraction model; otherwise, the parameters at this time are used as the initialization parameters of the model, and the deep learning water extraction model is retrained.

5. 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 deep learning-based lake water extraction method according to any one of claims 1-4.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the deep learning-based lake water extraction method according to any one of claims 1-4.

Images