Lake water level and water reserve dynamic monitoring method based on satellite-borne active and passive remote sensing information fusion

Abstract

The invention belongs to the technical field of lake remote sensing monitoring, discloses a lake water level and water reserve dynamic monitoring method based on satellite-borne active and passive remote sensing information fusion, and effectively solves the problem that lake information in remote areas is difficult to efficiently and accurately obtain. Comprising the following steps: S1, preprocessing laser data; s2, preprocessing the multispectral data; s3, taking the ICESat-2 data as a core, combining with a multispectral image covering a corresponding region, and constructing a lake bottom DEM inversion data set based on neighborhood features; s4, constructing and training an elevation inversion convolutional neural network model; s5, predicting and generating a lake bottom DEM by using the trained elevation inversion convolutional neural network model; and S6, calculating the absolute water volume of the lake according to the DEM at the bottom of the lake, and carrying out lake water level-water reserve monitoring. According to the invention, long-time-sequence hydrological observation of remote region lakes can be effectively realized with low cost.

Description

Technical Field

[0001] The present invention belongs to the technical field of lake remote sensing monitoring, and in particular relates to a method for dynamic monitoring of lake water level and water reserves based on the fusion of satellite-borne active and passive remote sensing information, and belongs to a method for estimating lake water reserves without in-situ data based on the fusion of active and passive remote sensing information. Background Art

[0002] The dynamics of lakes are crucial indicators for assessing the regional water cycle. In recent decades, lakes in some plateaus, characterized by vast territories, sparse populations, and harsh natural environments, have experienced significant expansion. This harsh environment, inaccessible to personnel and equipment, makes it extremely difficult to obtain complete and continuous data on lake changes through field measurements. Therefore, efficiently and accurately obtaining lake information in remote areas has become a critical issue that needs to be addressed.

[0003] With the advancement of scientific expeditions to the plateau in recent years, elevation data for some lakes has become available. However, the number of lakes measured represents only a small proportion of the total number of lakes on the plateau, and their spatial distribution is extremely uneven. Based solely on these fragmented data, it is difficult to conduct an in-depth analysis of the regional distribution of lake water reserves. Summary of the Invention

[0004] One purpose of the present invention is to provide a method for dynamic monitoring of lake water level and water reserves based on the fusion of satellite-borne active and passive remote sensing information, so as to effectively solve the problem that lake information in remote areas is difficult to obtain efficiently and accurately.

[0005] To solve the above technical problems, the technical solution adopted by the present invention is: a dynamic monitoring method for lake water level and water storage based on the fusion of satellite-borne active and passive remote sensing information, comprising the following steps: S1, obtaining ICESat-2 data, performing coarse denoising on the ICESat-2 data based on the dual constraints of DEM constraints and sliding window density method, and then manually labeling and saving the data as a .csv table file; S2, obtaining information on the near-infrared band, red band, green band, and blue band of multispectral images to extract lake boundaries; S3, estimating the maximum elevation of the lake based on the relationship between water surface area and water level elevation over multiple periods, The data with elevation greater than the maximum in ICESat-2 data are removed. Then, with ICESat-2 data as the core and multispectral images covering the corresponding area, a high-dimensional training dataset based on neighborhood features is constructed. S4. An elevation inversion convolutional neural network model is constructed and trained. The elevation inversion convolutional neural network model uses ConvNeXt as the backbone network, introduces the SwinTransformer module in the middle part, embeds the CBAM attention module multiple times in the encoding stage, and uses PixelShuffle upsampling in the decoding stage. S5. Extracts 90 pixels around each pixel. 90 4 neighborhood spectral blocks and construct the same input format as the model training stage, where the boundary pixels are filled with boundaries to obtain 90 90 4 neighborhood spectral blocks are input into the trained elevation inversion convolutional neural network model for prediction to obtain the elevation value of the corresponding central pixel; the elevation values ​​of all pixels are obtained with a sliding window with a step size of 1, and the prediction results of all pixels are reconstructed according to their original spatial positions to generate the lake bottom DEM; S6. According to the lake bottom DEM, the absolute water volume of the lake is calculated, and the lake water level-water storage monitoring is carried out.

[0006] Furthermore, in step S1, the ICESat-2 data is ATL03 data provided by the ICESat-2 satellite; first, the ATL03 data corresponding to the RTG number of the ATL08 data is screened and retained; then, the ATL03 data is input, DEM constraint pre-screening is performed, and then sliding window density screening is performed, and combined with dynamic threshold adjustment, the coarse denoised photon data is output; finally, the coarse denoised photon data is imported into the annotation software for manual labeling, and land, noise, water surface and underwater terrain points are respectively marked with different colors.

[0007] Furthermore, in step S1, the method for coarse data denoising is as follows: S11, assuming that the deviation of the photon height distribution of the lake or surface in the study area relative to the DEM is within a limited range, calculate the height difference of each photon point:

[0008] ;

[0009] in, For the The height of a photon, For the The projection position of each photon corresponds to the DEM elevation. For the The height difference of the photons.

[0010] S12. Set upper and lower thresholds , the filter conditions are:

[0011] ;

[0012] in, is the minimum value of the photon point height difference, It is the maximum value of the height difference of the photon points; the photons that meet the screening conditions are retained, otherwise they are deleted.

[0013] S13. For the remaining photons after DEM constraint, use a fixed-length sliding window to perform local photon density screening:

[0014] ;

[0015] in, is the photon density per unit length, is the window length, is the number of photons within each window.

[0016] S14. Set density threshold ,like , all photons in the window are judged as noise and eliminated.

[0017] S15, Dynamically adjust to adapt to terrain changes 、 and , where the local slope according to DEM Calculate the maximum value of the photon point height difference :

[0018] ;

[0019] in, is the upper limit of the base elevation, is the slope sensitivity coefficient.

[0020] Furthermore, in step S2, the date is first defined as a feature set by importing a CSV file containing the date of each ICESat-2 data, providing a basis for the extraction of multispectral images.

[0021] Furthermore, in step S3, assuming that there is a monotonically increasing relationship between the lake water level and the water surface area, then:

[0022] ;

[0023] in, is the lake surface area, is the lake surface elevation.

[0024] The lake surface area was extracted using multi-period multispectral images and the water surface area of ​​each period was calculated. At the same time, the water level elevation at the lake boundary was obtained by combining ICESat-2 data of the same date, and the area-elevation observation pair was constructed. .

[0025] The quadratic polynomial regression is used to fit the area-elevation relationship, and the regression model is expressed as:

[0026] ;

[0027] in, 、 and All are regression coefficients.

[0028] When the lake surface area is zero, the corresponding lake bottom elevation :

[0029] ;

[0030] Combined with the water surface elevation during the highest water level period , the maximum elevation of the lake Expressed as:

[0031] .

[0032] Furthermore, in step S4, the input data of the elevation inversion convolutional neural network model includes the denoised ICESat-2 elevation points containing latitude, longitude and elevation values ​​and the near infrared band, red band, green band and blue band information of the multispectral image. With each elevation point as the center, the 90° angle of the corresponding pixel in the multispectral image data is extracted. 90 4 neighborhood spectral blocks, and use the multidimensional features as the input of the elevation inversion convolutional neural network model, the target elevation value as the label of the elevation inversion convolutional neural network model, and construct the input of the complete elevation inversion convolutional neural network model.

[0033] Furthermore, in step S4, when training the elevation inversion convolutional neural network model: 90 4 neighborhood block F0 is used as input and passes through the Stem module of the ConvNeXt network, using 7 7 convolutions increase the number of channels to 32 and obtain a size of 45 45 Feature map F1 of 32.

[0034] F1 enters the ConvNeXt Stage1 module and performs feature extraction through three ConvNeXt Blocks. The output size is 45 45 The feature map F2 of 64 is generated after the CBAM attention module with a size of 45 45 64 weighted feature map F2_att; each ConvNeXt Block contains depth-separable convolution, layer normalization, GELU activation and 1 1 convolution.

[0035] Next, use 3 with a stride of 2 3 convolutions downsample the feature map F2_att to a size of 23 twenty three 128 feature maps F3.

[0036] On this basis, the Transformer encoder module consisting of two layers of Swin Transformer modules is introduced to obtain a size of 23 twenty three The 128-bit feature map F4 extracts global dependencies through the window multi-head self-attention mechanism.

[0037] F4 through 3 with step size 2 3 convolutions for downsampling, generating a size of 12 2 256 feature map F5, and input it into the ConvNeXt Stage2 module, and the size is calculated by three layers of ConvNeXt Block to obtain a 12 2 256 feature map F6, F6 passes through the CBAM attention module, and the output size is 12 12 256 weighted feature map F6_att.

[0038] F6_att is upsampled by PixelShuffle to obtain a size of 2 twenty four 128 feature map F7, and multi-scale feature fusion with F4, first splicing along the channel dimension to obtain 23 twenty three 256 features, then 3 3 convolutions are compressed to 128 channels, generating a size of 23 twenty three 128 feature maps F8.

[0039] F8 is upsampled to size 46 by PixelShuffle 46 The feature map F9 of 64 is fused with F2_att for multi-scale features and then concatenated through 3 3 convolutions generate a size of 45 45 Feature map F10 of 64.

[0040] Finally, F10 is globally averaged pooled to obtain a size of 1 1 64 vector F11, and predict the elevation value corresponding to the center pixel through full connection layer regression, the output dimension is 1 1 1.

[0041] Furthermore, in step S6, the absolute water volume of the lake is calculated by analyzing the flooding range of the lake at different water levels using the lake bottom DEM, and calculating the water volume of the lake based on the integral. For each water level height, the water volume is calculated. :

[0042] ;

[0043] in, is the maximum water level of the lake, Is the water level The lake area, and is the adjacent water level height, .

[0044] Furthermore, in step S6, the lake water level-water storage monitoring method is: first, a DEM of each period is constructed based on the multispectral data of multiple dates and the corresponding ICESat-2 elevation data through an elevation inversion convolutional neural network model, so as to obtain the absolute water storage capacity of the lake in each period; then, the water level data is obtained by using ICESat-2 laser altimetry, and the absolute water storage capacity of the lake is fitted with the water level data to obtain a water level-water storage regression curve.

[0045] Compared with the existing technology, the beneficial technical effects of the present invention are as follows: (1) The present invention uses a multispectral combined with satellite-borne single-photon laser point cloud method to achieve low-cost estimation of water reserves in plateau lakes where in-situ water depth is difficult to obtain or manual measurement is impossible. The inversion effect of lake water depth, a key parameter for water reserve calculation, has been verified to have an RMSE of <0.5m; therefore, a relatively accurate lake water reserve can be obtained. This is of great significance for the statistics of water resources in lakes that are difficult to access in many plateau areas.

[0046] (2) This invention effectively and cost-effectively implements long-term hydrological observations of lakes in remote areas by combining spaceborne single-photon lidar with satellite remote sensing imagery. By fusing ICESat-2 and WorldView-2 data, this invention can capture detailed surface feature information while ensuring temporal continuity of monitoring, effectively achieving near-real-time tracking and comprehensive analysis of water volume changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a structural diagram of the elevation inversion convolutional neural network model of the present invention.

[0048] Figure 2 It is an underwater topographic map of Lake A generated by inversion using the elevation inversion convolutional neural network model of the present invention.

[0049] Figure 3It is a three-dimensional display view of the underwater terrain of Lake A generated by inversion using the elevation inversion convolutional neural network model of the present invention.

[0050] Figure 4 This is the verification result of the water depth inversion using the elevation inversion convolutional neural network model of the present invention.

[0051] Figure 5 This is the time series change of the water level of Lake A monitored by the present invention.

[0052] Figure 6 This is a regression curve of the water level-water reserve of Lake A obtained by monitoring the present invention. DETAILED DESCRIPTION

[0053] Example 1: This example utilizes a "point-to-surface" integration of spaceborne single-photon lidar and satellite remote sensing imagery to achieve low-cost, long-term hydrological observations of lakes in remote areas. Specifically, ICESat-2's precise land and shallow-water elevation contours can be used to assign elevation information to submerged contours extracted from remote sensing imagery at any given time, thereby determining the relative changes in lake area and water storage at different moments. Determining the absolute water storage capacity at the lake bottom through elevation inversion of optical remote sensing imagery allows the establishment of various reservoir capacity models, such as those based on submerged area versus water storage capacity and water level versus water storage capacity, enabling comprehensive monitoring of lake water level, area, and water storage capacity. By integrating ICESat-2 (Ice, Cloud, and Land Elevation Satellite-2) data with WorldView-2, detailed surface features can be captured while ensuring temporal continuity, enabling near-real-time tracking and comprehensive analysis of water volume changes.

[0054] The method for dynamic monitoring of lake water level and water storage based on the fusion of satellite-borne active and passive remote sensing information provided in this embodiment includes the following steps: S1, laser data preprocessing; S2, multispectral data preprocessing; S3, preparation of a lake bottom DEM inversion dataset; S4, construction and training of an elevation inversion convolutional neural network model; S5, prediction and generation of a lake bottom DEM using the trained elevation inversion convolutional neural network model; S6, calculation of the absolute water volume of the lake based on the lake bottom DEM, and performance of lake water level-water storage monitoring.

[0055] (1) In step S1, it specifically includes: (1) ICESat-2 data acquisition.

[0056] This example uses the ATL03 data product provided by the ICESat-2 satellite, primarily to obtain along-track photon count data for various regions around the globe. ATL03 data contains highly accurate photon echo time and location information, making it suitable for detailed detection of topography and water surfaces. This example focuses on extracting photon points associated with underwater topography from the ATL03 data.

[0057] ATL08 is a surface elevation product based on ATL03 data, specifically designed to extract information such as land, forest, and vegetation height. By processing the raw photon data from ATL03, ATL08 data has already classified features such as ground and vegetation, and excluded obvious noise photons. This means that for terrain near water, the corresponding ATL03 photon points in ATL08 have undergone some preliminary filtering and classification, making them more likely to contain photons with topographic information. Therefore, when filtering data, the RTG number of the ATL08 product is used to search for a corresponding ATL03 product. This improves the efficiency of finding ATL03 products with underwater points.

[0058] The valid original ATL03 product data in .h5 format is further converted into a multi-track original dataset in .mat format. One track of ICESat-2 point cloud data contains a large number of photon points.

[0059] (2) Coarse denoising of ICESat-2 data.

[0060] To effectively remove the significant background noise from ICESat-2 ATL03 photon cloud data, this paper proposes a dual-constraint coarse denoising method based on DEM (digital elevation model) constraints and a sliding window density method. By incorporating prior terrain information and local photon density features, this method achieves rapid and accurate removal of noise photons, reducing the computational complexity of subsequent fine filtering.

[0061] First, assuming that the deviation of the photon height distribution of the lake or surface in the study area relative to the DEM should be within a limited range, the height difference of each photon point is calculated:

[0062] ;

[0063] in, For the The height of a photon, For the The projection position of each photon corresponds to the DEM elevation. For the The height difference of photons. Set the upper and lower thresholds , the filter conditions are:

[0064] ;

[0065] in, is the minimum value of the photon point height difference, is the maximum value of the photon point height difference.

[0066] Photons that meet the screening conditions are retained, otherwise they are deleted. For the remaining photons after DEM constraints, a fixed-length sliding window is used to perform local photon density screening:

[0067] ;

[0068] in, is the photon density per unit length, is the window length, is the number of photons within each window.

[0069] Setting density thresholds ,like , all photons in the window are judged as noise and removed. To adapt to the changes in terrain, dynamic adjustment is made 、 and , where the local slope according to DEM Calculate the maximum value of the photon point height difference :

[0070] ;

[0071] in, is the upper limit of the base elevation, is the slope sensitivity coefficient, It can also be adjusted according to the complexity of the terrain and the expected target point density.

[0072] In general, the traditional DEM constraint and sliding window density method each have their own advantages and disadvantages. This method combines the two to achieve complementarity and introduces DEM slope constraint to adjust the threshold to improve the denoising stability in complex terrain areas. Without significantly increasing the computational burden, it improves the accuracy and fidelity of coarse denoising. After experimental verification, the recommended parameters are 、 、 、 、 .

[0073] (3) ICESat-2 data processing.

[0074] Import the denoised .mat file into the annotation software and manually label it. Yellow represents land, blue represents noise, purple represents water, and red represents underwater terrain points. After manual processing, save the file as a .csv file.

[0075] (2) In step S2, it specifically includes: (1) data acquisition.

[0076] In this example, WorldView-2 remote sensing data with minimal cloud coverage was acquired and processed using the Google Earth (GEE) platform. First, a CSV file containing the date of each ICESat-2 data entry was imported and defined as a feature set, providing the basis for subsequent image extraction. Only the near-infrared, red, green, and blue bands of the multispectral imagery were required. This operation significantly reduced the downloading of unnecessary remote sensing data and significantly reduced the complexity of subsequent calculations.

[0077] (2) Lake boundary extraction.

[0078] This embodiment selects the near-infrared band, red band, green band, and blue band information of the WorldView-2 multispectral image for lake boundary extraction.

[0079] First, calculate the normalized difference water index :

[0080] ;

[0081] in, and Represent the reflectance of the green band and the near-infrared band, respectively.

[0082] Calculated After the index image was extracted, the image was binarized using threshold segmentation to extract the lake's extent. To improve the accuracy of boundary extraction, morphological erosion and dilation were used to remove noise, and connected domain analysis was performed to retain the largest water area, ultimately obtaining the precise lake boundary.

[0083] (3) In step S3, it specifically includes: (1) estimating the maximum elevation of the lake.

[0084] To estimate the maximum elevation of a lake, this embodiment constructs a method based on the relationship between water surface area and water level elevation over multiple periods. Assuming a monotonically increasing relationship between lake water level and water surface area, it can be expressed as:

[0085] ;

[0086] in, is the lake surface area, is the lake surface elevation.

[0087] First, we used WorldView-2 remote sensing images from multiple periods to extract the lake surface area and calculated the surface area of ​​each period. At the same time, we combined ICESat-2 altimetry data from the same date to obtain the water level elevation at the lake boundary and construct an area-elevation observation pair. .

[0088] Subsequently, quadratic polynomial regression was used to fit the area-elevation relationship, and the regression model can be expressed as:

[0089] ;

[0090] in, 、 and are all regression coefficients. When the lake surface area is zero, that is, , the corresponding lake bottom elevation can be calculated :

[0091] ;

[0092] After obtaining the lake bottom elevation, combined with the water surface elevation during the highest water level period , the maximum elevation of the lake It can be calculated as:

[0093] .

[0094] This method is based on the fitting relationship between area and elevation, and can quickly estimate the maximum elevation of a lake in the absence of underwater DEM or elevation measurement data.

[0095] (2) Dataset construction.

[0096] Based on the estimated maximum elevation of the lake, the data in the ICESat-2 elevation data that are greater than the maximum elevation are removed, and the remaining data are matched with the multispectral imagery.

[0097] The sparse elevation point data provided by ICESat-2 is used as the core, combined with WorldView-2 multispectral imagery covering the corresponding area, to construct a high-dimensional training dataset based on neighborhood features. The input data includes denoised ICESat-2 elevation points (including latitude, longitude and elevation values) and the near-infrared, red, green and blue bands of WorldView-2 multispectral imagery. The core goal is to extract the 90% probability of the corresponding pixel in the WorldView-2 data with each elevation point as the center. 90 4 neighborhood spectral values, and use these multidimensional features as the input of the elevation inversion convolutional neural network model, and the target elevation value as the label of the elevation inversion convolutional neural network model to construct the input of the complete elevation inversion convolutional neural network model.

[0098] (IV) In step S4, (1) first, the dataset is divided. The division of the dataset is a key step to ensure the effectiveness of the model training, validation, and testing process. Typically, this process involves dividing the dataset into a training set, a validation set, and a test set.

[0099] In practice, the dataset was divided into a training set (70%), a validation set (15%), and a test set (15%). Furthermore, to more comprehensively evaluate the model's performance, cross-validation was used on the training set to ensure model stability and generalization capabilities.

[0100] (2) Construction and training of elevation inversion convolutional neural network model.

[0101] The elevation inversion convolutional neural network model provided in this embodiment has an overall architecture with ConvNeXt as the backbone network, which fully extracts the local spatial texture and spectral features in the remote sensing image. At the same time, the SwinTransformer module is introduced in the middle part to enhance the modeling ability of the global contextual relationship through the windowed self-attention mechanism. In addition, the model embeds the CBAM attention module multiple times in the encoding stage to further enhance the expression of key features. In the decoding stage, PixelShuffle upsampling is used to replace the traditional transposed convolution to effectively improve the spatial restoration accuracy, and the multi-scale feature fusion structure is used to combine shallow and deep information to enhance the boundary continuity and spatial consistency of the prediction. Overall, the model takes into account the efficiency of local feature extraction and the ability of global relationship modeling, and can provide more accurate and more robust prediction results for elevation inversion tasks.

[0102] The structure of the elevation inversion convolutional neural network model is as follows Figure 1 As shown, the elevation inversion convolutional neural network model is constructed with a size of 90 90 The neighborhood block F0 of 4 is taken as input and passed through the Stem module of the ConvNeXt network, using a step size of 2 and a padding of 3 for 7 7Conv (convolution) increases the number of channels from 4 to 32, and the size is 45 45 Feature map F1 of 32.

[0103] Subsequently, F1 enters the ConvNeXt Stage1 module and performs feature extraction through three ConvNeXt Blocks, with an output size of 45 45 64 feature maps F2. Each ConvNeXt Block contains depth-separable convolution, layer normalization, GELU activation and 1 1 convolution.

[0104] In order to further enhance the attention of the elevation inversion convolutional neural network model to the spatial and channel dimensions, F2 is generated through the CBAM attention module that combines channel and spatial attention to generate a 45-dimensional image. 45 The weighted feature map F2_att of 64.

[0105] Next, use 3 with a stride of 2 3Conv (convolution) downsamples the feature map F2_att to a size of 23 twenty three On this basis, the Transformer encoder module composed of two layers of Swin Transformer modules is introduced to obtain a feature map F3 of size 23 twenty three The 128-bit feature map F4 extracts global dependencies through the window multi-head self-attention mechanism, improving the representation ability of complex spectrum-depth mapping.

[0106] F4 through 3 with step size 2 3 Conv (convolution) downsamples and generates a size of 12 2 256 feature map F5, and input it into the ConvNeXt Stage2 module, and the size is calculated by three layers of ConvNeXt Block to obtain a 12 2 256 feature map F6. To further focus on key features, F6 passes through the CBAM attention module and the output size is 12 12 256 weighted feature map F6_att.

[0107] The decoding stage first performs PixelShuffle upsampling operation to upsample F6_att to 2 twenty four 128 feature map F7, and multi-scale feature fusion with F4 output by Transformer in the encoding stage, first splicing along the channel dimension to obtain 23 twenty three 256 features, then 3 3 Conv (convolution) is compressed to 128 channels, generating a size of 23 twenty three 128 feature maps F8.

[0108] Then, F8 is upsampled to size 46 by PixelShuffle. 46 The feature map F9 of 64 is combined with the F2_att of the encoding stage for multi-scale feature fusion, and then the concatenation is performed through 3 3 Conv (convolution) generates a size of 45 45 Feature map F10 of 64.

[0109] Finally, F10 is globally averaged pooled to obtain a size of 1 1 64 vector F11, and predict the elevation value corresponding to the center pixel through full connection layer regression, the output dimension is 1 1 1.

[0110] The elevation inversion convolutional neural network model of this embodiment achieves the unification of local texture extraction and global relationship modeling through the organic combination of ConvNeXt and Swin Transformer. The CBAM attention module further enhances the feature expression of spatial and channel dimensions. The PixelShuffle upsampling effectively avoids the aliasing effect of traditional upsampling. The overall architecture fully mines the multi-dimensional information of remote sensing images while ensuring the controllable parameters, providing technical support for high-precision elevation inversion.

[0111] (3) In the performance evaluation of the regression model, it is crucial to select appropriate indicators, which can not only verify the effectiveness of the model but also ensure its reliability in practical applications. This embodiment uses the root mean square error ( ), mean absolute error ( ) and the coefficient of determination ( ) as the main evaluation indicator to measure the predictive ability of the model from multiple angles.

[0112] Root mean square error ( ) can be regarded as the mean square error ( ), which is calculated as:

[0113] ;

[0114] in, For the Actual observations, For the The model predicted value, , is the total number of actual observations.

[0115] because The result has the same unit as the actual observation value, so it is more intuitive. At the same time, through the square root operation, It weakens the amplifying effect of abnormal errors on the overall indicators and improves the interpretability.

[0116] The mean absolute error (MAE) measures the average deviation of the model by calculating the average absolute value of the prediction error. The calculation formula is:

[0117] ;

[0118] compared to , MAE does not square the error, so it is more robust to outliers and can truly reflect the average level of the overall prediction error.

[0119] Coefficient of determination ( ) is used to measure the goodness of model fitting, and its formula is as follows:

[0120] ;

[0121] in, is the mean of the observations. It represents the proportion of the total variation of the dependent variable explained by the model, and its value range is between 0 and 1. The larger the value, the better the model can explain the data variation. Taking Lake A as the object, the bottom elevation inversion verification results of the elevation inversion convolutional neural network model provided in this embodiment are as follows: Figure 4 As shown, , , indicating that the deviation between the model inversion elevation of this embodiment and the ICESat-2 elevation is small.

[0122] (V) In step S5, a sliding window method is used to extract the 90° angles around each pixel. 90 4 neighborhood spectral blocks and construct the same input format as the model training stage, where the boundary pixels are filled with boundaries to obtain 90 90 The 4-neighborhood spectral block is input into the trained elevation inversion convolutional neural network model for prediction, and the elevation value of the corresponding central pixel is obtained. During the sliding window extraction process, the sliding step size is usually set to 1 to ensure that every pixel is predicted, thereby generating a spatially continuous elevation inversion map.

[0123] After the prediction is completed, the prediction results of all pixels are reconstructed according to their original spatial positions to generate a two-dimensional elevation distribution map (i.e., lake bottom DEM) that is consistent with the spatial range of the input image, and the predicted value of each pixel represents the elevation information of its corresponding position. The trained elevation inversion convolutional neural network model and method provided in this embodiment are used to invert the underwater topography of Lake A ( Figure 2 ) and underwater terrain 3D display view ( Figure 3 ).

[0124] (VI) In step S6, (1) the absolute water volume of the lake is calculated as follows: after obtaining the complete lake bottom DEM, the water volume can be calculated based on different water levels. The core idea is to use the lake bottom DEM to analyze the flooding range of the lake at different water levels and calculate the water volume of the lake based on the integral. For each water level, the water volume is calculated. , a layer-by-layer accumulation method can be used:

[0125] ;

[0126] in, is the maximum water level of the lake, Is the water level The lake area, and is the adjacent water level height, .

[0127] (2) Lake water level and water storage monitoring.

[0128] First, the DEM of each period was constructed based on the multispectral data of multiple dates and the corresponding ICESat-2 elevation data through the elevation inversion convolutional neural network model, so as to obtain the absolute water storage capacity of the lake in each period.

[0129] Then, water level data was obtained using ICESat-2 laser altimetry, and the absolute water storage capacity of the lake was fitted with the water level data to obtain a water level-water storage regression curve.

[0130] The average annual lake level, lake area and absolute water storage of Lake A since 2016 were observed. The results are as follows: Figure 5 and Figure 6 As shown, Lake A had the lowest annual average water level of 2740.27m and the smallest annual average area of ​​113.54km in 2017. 2 The minimum annual absolute water reserve is 385259153.5m 3 In 2024, the highest annual average water level was 2742.33m and the largest annual average area was 156.03km. 2 The maximum annual absolute water reserve is 696214812.8m 3 .

[0131] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the scope of protection of the present invention.

Claims

1. A method for dynamic monitoring of lake water level and water storage based on the fusion of satellite-borne active and passive remote sensing information, characterized in that: The following steps are involved: S1. Obtain ICESat-2 data, perform coarse denoising on the data based on the dual constraints of DEM and sliding window density method, and then manually label the data and save it as a .csv file. S2, obtain the information of the near-infrared band, red band, green band, and blue band of the multispectral image to extract the lake boundary; S3. Calculate the maximum elevation of the lake based on the relationship between water surface area and water level elevation over multiple time periods, and remove data from the ICESat-2 data that exceeds the maximum elevation. Then, using the ICESat-2 data as the core and combining it with multispectral imagery covering the corresponding area, construct a high-dimensional training dataset based on neighborhood features. S4. Construct and train an elevation inversion convolutional neural network model. The elevation inversion convolutional neural network model uses ConvNeXt as the backbone network, introduces a Swin Transformer module in the middle part, embeds the CBAM attention module multiple times in the encoding stage, and adopts PixelShuffle upsampling in the decoding stage. S5, extract the 90 degrees around each pixel 90 4 neighborhood spectral blocks and construct the same input format as the model training stage, where the boundary pixels are filled with boundaries to obtain 90 90 4 neighborhood spectral blocks are input into the trained elevation inversion convolutional neural network model for prediction to obtain the elevation value of the corresponding central pixel; the elevation values ​​of all pixels are obtained using a sliding window with a step size of 1, and the predicted results of all pixels are reconstructed according to their original spatial positions to generate the lake bottom DEM; S6. Calculate the absolute water volume of the lake based on the lake bottom DEM and monitor the lake water level and water storage.

2. The method for dynamic monitoring of lake water level and water reserves based on satellite-borne active and passive remote sensing information fusion according to claim 1 is characterized in that: In step S1, the ICESat-2 data is ATL03 data provided by the ICESat-2 satellite; First, filter and retain the ATL03 data corresponding to the RTG number of the ATL08 data; Then, the ATL03 data is input, DEM constraint pre-screening is performed, and then sliding window density screening is performed, combined with dynamic threshold adjustment, to output the coarse denoised photon data; Finally, the coarsely denoised photon data were imported into the annotation software for manual labeling, with land, noise, water surface and underwater terrain points marked with different colors.

3. The method for dynamic monitoring of lake water level and water reserves based on satellite-borne active and passive remote sensing information fusion according to claim 2 is characterized in that: In step S1, the method for coarse data denoising is: S11. Assuming that the deviation of the photon height distribution of the lake or surface in the study area relative to the DEM is within a limited range, calculate the height difference of each photon point: ； in, For the The height of a photon, For the The projection position of each photon corresponds to the DEM elevation. For the The height difference of the photons; S12. Set upper and lower thresholds , the filter conditions are: ； in, is the minimum value of the photon point height difference, The maximum value of the photon point height difference; photons that meet the screening conditions are retained, otherwise they are deleted; S13. For the remaining photons after DEM constraint, use a fixed-length sliding window to perform local photon density screening: ； in, is the photon density per unit length, is the window length, is the number of photons in each window; S14. Set density threshold ,like , then all photons in the window are judged as noise and eliminated; S15, Dynamically adjust to adapt to terrain changes 、 and , where the local slope according to DEM Calculate the maximum value of the photon point height difference : ； in, is the upper limit of the base elevation, is the slope sensitivity coefficient.

4. The method for dynamic monitoring of lake water level and water reserves based on satellite-borne active and passive remote sensing information fusion according to claim 3 is characterized in that: In step S2, the date is first defined as a feature set by importing a CSV file containing the date of each ICESat-2 data, providing a basis for the extraction of multispectral images.

5. The method for dynamic monitoring of lake water level and water reserves based on satellite-borne active and passive remote sensing information fusion according to claim 4 is characterized in that: In step S3, assuming that there is a monotonically increasing relationship between the lake water level and the water surface area, then: ； in, is the lake surface area, is the lake elevation; The lake surface area was extracted using multi-period multispectral images and the water surface area of ​​each period was calculated. At the same time, the water level elevation at the lake boundary was obtained by combining ICESat-2 data of the same date, and the area-elevation observation pair was constructed. ; The quadratic polynomial regression is used to fit the area-elevation relationship, and the regression model is expressed as: ； in, 、 and All are regression coefficients; When the lake surface area is zero, the corresponding lake bottom elevation : ； Combined with the water surface elevation during the highest water level period , the maximum elevation of the lake Expressed as: 。 6. The method for dynamic monitoring of lake water level and water reserves based on satellite-borne active and passive remote sensing information fusion according to claim 5 is characterized in that: In step S4, the input data of the elevation inversion convolutional neural network model include the denoised ICESat-2 elevation points containing latitude, longitude and elevation values ​​and the near infrared band, red band, green band and blue band information of the multispectral image. With each elevation point as the center, the 90° angle of the corresponding pixel in the multispectral image data is extracted. 90 4 neighborhood spectral blocks, and use the multidimensional features as the input of the elevation inversion convolutional neural network model, the target elevation value as the label of the elevation inversion convolutional neural network model, and construct the input of the complete elevation inversion convolutional neural network model.

7. The method for dynamic monitoring of lake water level and water reserves based on satellite-borne active and passive remote sensing information fusion according to claim 6 is characterized in that: In step S4, when training the elevation inversion convolutional neural network model: With a size of 90 90 4 neighborhood block F0 is used as input and passes through the Stem module of the ConvNeXt network, using 7 7 convolutions increase the number of channels to 32 and obtain a size of 45 45 Feature map F1 of 32; F1 enters the ConvNeXt Stage1 module and performs feature extraction through three ConvNeXt Blocks. The output size is 45 45 The feature map F2 of 64 is generated after the CBAM attention module with a size of 45 45 64 weighted feature map F2_att; each ConvNeXt Block contains depth-separable convolution, layer normalization, GELU activation and 1 1 convolution; Next, use 3 with a stride of 2 3 convolutions downsample the feature map F2_att to a size of 23 twenty three 128 feature maps F3; On this basis, the Transformer encoder module consisting of two layers of Swin Transformer modules is introduced to obtain a size of 23 twenty three The 128-bit feature map F4 extracts global dependencies through the window multi-head self-attention mechanism; F4 through 3 with step size 2 3 convolutions for downsampling, generating a size of 12 2 256 feature map F5, and input it into the ConvNeXt Stage2 module, and the size is calculated by three layers of ConvNeXt Block to obtain a 12 2 256 feature map F6, F6 passes through the CBAM attention module, and the output size is 12 12 256 weighted feature map F6_att; F6_att is upsampled by PixelShuffle to obtain a size of 2 twenty four 128 feature map F7, and multi-scale feature fusion with F4, first splicing along the channel dimension to obtain 23 twenty three 256 features, then 3 3 convolutions are compressed to 128 channels, generating a size of 23 twenty three 128 feature maps F8; F8 is upsampled to size 46 by PixelShuffle 46 The feature map F9 of 64 is fused with F2_att for multi-scale features and then concatenated through 3 3 convolutions generate a size of 45 45 Feature map F10 of 64; Finally, F10 is globally averaged pooled to obtain a size of 1 1 64 vector F11, and predict the elevation value corresponding to the center pixel through full connection layer regression, the output dimension is 1 1 1.

8. The method for dynamic monitoring of lake water level and water reserves based on satellite-borne active and passive remote sensing information fusion according to claim 7 is characterized in that: In step S6, the absolute water volume of the lake is calculated as follows: The lake bottom DEM is used to analyze the flooding range of the lake at different water levels, and the water volume of the lake is calculated based on the integral. For each water level height, the water volume is calculated. : ； in, is the maximum water level of the lake, Is the water level The lake area, and is the adjacent water level height, .

9. The method for dynamic monitoring of lake water level and water reserves based on satellite-borne active and passive remote sensing information fusion according to claim 8, characterized in that: In step S6, the lake water level-water storage monitoring method is: First, a DEM for each period was constructed using the multispectral data from multiple dates and the corresponding ICESat-2 elevation data using an elevation inversion convolutional neural network model to obtain the absolute water storage capacity of the lake in each period. Then, water level data was obtained using ICESat-2 laser altimetry, and the absolute water storage capacity of the lake was fitted with the water level data to obtain a water level-water storage regression curve.

Images