Method for monitoring lake water level and water storage dynamically based on fusion of spaceborne active and passive remote sensing information

Abstract

The application belongs to the technical field of lake remote sensing monitoring, and discloses a lake water level and water storage dynamic monitoring method based on satellite-borne active and passive remote sensing information fusion, which effectively solves the problem that lake information in remote areas is difficult to be efficiently and accurately acquired. The method comprises the following steps: S1, laser data preprocessing; S2, multispectral data preprocessing; S3, taking ICESat-2 data as the core, combining with multispectral images covering the corresponding area, and constructing a lake bottom DEM inversion dataset based on neighborhood features; S4, construction and training of an elevation inversion convolutional neural network model; S5, generating the 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 lake bottom DEM, and performing lake water level-water storage monitoring. The application can effectively realize low-cost long-time-series hydrological observation of lakes in remote areas.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of lake remote sensing monitoring, and particularly relates to a lake water level and water storage dynamic monitoring method based on satellite-borne active and passive remote sensing information fusion, and belongs to a lake water storage estimation method without in-situ data based on active and passive remote sensing information fusion. BACKGROUND

[0002] The dynamic change of a lake is an important indicator for evaluating the regional water cycle condition. In recent decades, some plateau lakes with vast territory, sparse population and harsh natural environment have experienced significant expansion. Due to the harsh natural environment, it is difficult for personnel and equipment to reach the lakes, so that it is extremely difficult to obtain complete and continuous lake change monitoring data through field measurement. Therefore, how to efficiently and accurately obtain the lake information in remote areas has become a key problem to be solved.

[0003] With the advancement of plateau scientific investigation in recent years, the elevation data of some lakes have been obtained. However, the number of measured lakes accounts for only a small proportion of the total number of plateau lakes, and the spatial distribution is extremely uneven. It is difficult to deeply analyze the distribution characteristics of lake water storage in the region only with these scattered data. SUMMARY

[0004] An object of the application is to provide a lake water level and water storage dynamic monitoring method based on satellite-borne active and passive remote sensing information fusion, which effectively solves the problem that the lake information in remote areas is difficult to be efficiently and accurately obtained.

[0005] To solve the above technical problems, the technical scheme adopted by the application is as follows: the lake water level and water storage dynamic monitoring method based on satellite-borne active and passive remote sensing information fusion comprises the following steps: S1, obtaining ICESat-2 data, performing coarse denoising on the ICESat-2 data based on the double constraints of DEM constraint and sliding window density method, and then performing artificial label annotation and saving as a.csv table file; S2, obtaining the information of the near-infrared band, red band, green band and blue band of the multi-spectral image, and performing lake boundary extraction; S3, based on the relationship between the water surface area and the water level elevation of multiple periods, calculating the maximum elevation of the lake, and removing the data greater than the maximum elevation in the ICESat-2 data; then, taking the ICESat-2 data as the core, combining the multi-spectral image covering the corresponding region, and constructing a high-dimensional training data set based on neighborhood features; S4, constructing and training an elevation inversion convolutional neural network model, wherein the elevation inversion convolutional neural network model takes ConvNeXt as the backbone network, introduces SwinTransformer modules in the middle part, embeds CBAM attention modules multiple times in the encoding stage, and adopts PixelShuffle upsampling in the decoding stage; S5, extracting 90 90 4-neighborhood spectral block, and construct the same input format as the model training stage, where the boundary pixels are filled with boundary padding to obtain 90 90 4-neighborhood spectral block, and input it into the trained height inversion convolutional neural network model for prediction to obtain the height value of the corresponding center pixel; all pixel height values are obtained by 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 a 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 performed.

[0006] Further, 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, combined with dynamic threshold adjustment, to output the rough denoised photon data; finally, the rough denoised photon data is imported into a labeling software for manual label labeling, and land, noise, water surface and underwater terrain points are labeled with different colors respectively.

[0007] Further, in step S1, the data rough denoising method is: S11, assuming that the height distribution of photons of the study area lake or surface relative to the DEM deviation is within a limited range, the height difference of each photon point is calculated:

[0008] ;

[0009] wherein, is the height of the i th photon, is the DEM elevation corresponding to the projection position of the i th photon, is the height difference of the i th photon. S12, set the upper and lower threshold values , and the screening condition is:

[0010]

[0011] ;

[0012] wherein, is the minimum value of the height difference of the photon point, is the maximum value of the height difference of the photon point; the photons meeting the screening condition are retained, otherwise they are deleted.

[0013] S13, for the remaining photons after DEM constraint, a fixed length sliding window is used for local photon density screening:

[0014] ;​​​

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

[0016] S14, set the density threshold , if , all photons within the window are determined to be noise and removed.

[0017] S15, to adapt to the terrain undulation, dynamically adjust , and , wherein the maximum value of the height difference of the photon points is calculated according to the local slope of the DEM :

[0018] ;

[0019] wherein, is the upper limit of the basic elevation, is the slope sensitivity coefficient.

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

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

[0022] ;

[0023] wherein, is the lake water surface area, is the lake surface elevation.

[0024] Using multi-period multispectral images to extract the lake water surface range, calculate the water surface area of each period, at the same time, combined with the ICESat-2 data of the same date to obtain the water level elevation at the lake boundary, construct the area-elevation observation value pair .

[0025] Using a quadratic polynomial regression to fit the area-elevation relationship, the regression model is expressed as:

[0026] ;

[0027] wherein, , and are all regression coefficients.

[0028] ​When the lake's surface area is zero, the corresponding lakebed elevation is... :

[0029] ;

[0030] Combined with the water surface elevation during the period of highest water level Maximum elevation of the lake Represented as:

[0031] .

[0032] Furthermore, in step S4, the input data of the elevation inversion convolutional neural network model includes denoised ICESat-2 elevation points containing latitude, longitude, and elevation values, as well as information from the near-infrared, red, green, and blue bands of multispectral imagery. Centered on each elevation point, 90% of the corresponding pixel's coordinates are extracted from the multispectral image data. 90 The model uses 4 neighborhood spectral blocks and multidimensional features as input to 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.

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

[0034] F1 enters the ConvNeXt Stage1 module, where feature extraction is performed through three ConvNeXt Blocks, with an output size of 45. 45 The 64 feature map F2 is processed by the CBAM attention module to generate a 45-size feature map. 45 The 64-weighted feature map F2_att; each ConvNeXt Block contains depthwise separable convolutions, layer normalization, GELU activation, and 1 1. Convolution.

[0035] Next, use 3 with a step size of 2. 3 convolutions downsample the feature map F2_att, resulting in a size of 23. twenty three Feature map F3 of 128.

[0036] On this basis, the Transformer encoder module composed of two layers of Swin Transformer modules is introduced to obtain a feature map F4 with a size of 23 23 128, and the global dependency relationship is extracted through the window multi-head self-attention mechanism.

[0037] F4 is down-sampled by 3 3 convolution with a step of 2 to generate a feature map F5 with a size of 12 2 256, and is input into the ConvNeXt Stage2 module to obtain a feature map F6 with a size of 12 2 256 through three layers of ConvNeXt Block calculation, and F6 is input into the CBAM attention module to output a weighted feature map F6_att with a size of 12 12 256.

[0038] F6_att is up-sampled by PixelShuffle to obtain a feature map F7 with a size of 2 24 128, and is fused with F4 to obtain a feature map F8 with a size of 23 23 256 through 3 3 convolution to compress to 128 channels to generate a feature map F8 with a size of 23 23 128.

[0039] F8 is up-sampled by PixelShuffle to a feature map F9 with a size of 46 46 64, and is fused with F2_att to obtain a feature map F10 with a size of 45 3 45 64 through 3

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

[0041] Further, in step S6, the calculation method of the absolute water volume of the lake is: the submerged range of the lake at different water levels is analyzed by using the DEM of the bottom of the lake, and the water volume of the lake is calculated according to integration, for each water level height, the volume of the water body is calculated :

[0042] ;

[0043] wherein, is the maximum water level height of the lake, is the area of the lake at the water level height , and is the adjacent water level height, .

[0044] Further, in step S6, the lake water level-water storage monitoring method is: first, according to the multi-date multi-spectral data and the corresponding ICESat-2 elevation data, the DEM of each period is constructed by the elevation inversion convolutional neural network model, so as to obtain the absolute water storage of the lake in each period; then, the water level data is obtained by using the ICESat-2 laser ranging, and the absolute water storage of the lake is fitted with the water level data to obtain the water level-water storage regression curve.

[0045] Compared with the prior art, the beneficial technical effects of the present application are: (1) the present application realizes the low-cost estimation of the water storage of the plateau lake which is difficult to obtain in-situ water depth or cannot be measured manually by combining multi-spectral data with satellite-borne single-photon laser point cloud. The inversion effect of the key parameter of the lake water depth for water storage calculation is verified with RMSE<0.5m; therefore, a relatively accurate lake water storage can be obtained. This has an important role for the lake water resource statistics of many inaccessible lakes in plateau areas.

[0046] (2) the present application effectively realizes the low-cost long-term hydrological observation of remote lakes by combining satellite-borne single-photon lidar and satellite remote sensing images. The present application can capture fine ground feature information and ensure the time continuity of monitoring by fusing ICESat-2 data and WorldView-2 data, so as to effectively realize the near-real-time tracking and comprehensive analysis of water volume changes. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 is the structure diagram of the elevation inversion convolutional neural network model of the present application.

[0048] Figure 2 is the underwater topographic map of A lake generated by using the elevation inversion convolutional neural network model of the present application.

[0049] Figure 3It is a 3D display view of the underwater topography of A lake generated by inversion using the elevation inversion convolutional neural network model of the application.

[0050] Figure 4 It is the water depth verification result of the elevation inversion convolutional neural network model of the application.

[0051] Figure 5 It is the time series change of the water level of A lake monitored by the application.

[0052] Figure 6 It is the water level-water storage regression curve of A lake obtained by monitoring. DETAILED DESCRIPTION

[0053] In this embodiment, the point-plane combination of the spaceborne single-photon lidar and satellite remote sensing images can realize low-cost long-term hydrological observation of lakes in remote areas. Specifically, the submerged contours extracted from the remote sensing images at any time can be given elevation information by relying on the accurate land and shallow water elevation profiles obtained by ICESat-2, so that the relative change of the lake area and water storage at different times can be obtained; by inverting DEM from optical remote sensing images, the absolute water storage of the lake bottom can be obtained, and then various reservoir capacity models such as submerged area-water storage and water level height-water storage can be established to realize full-range monitoring of the lake water level, area and water storage. By fusing ICESat-2 (Ice, Cloud and Land Elevation Satellite 2) data and WorldView-2 (WorldView 2) data, fine ground feature information can be captured, and time continuity of monitoring can be ensured, so as to realize near-real-time tracking and comprehensive analysis of water quantity changes.

[0054] The lake water level and water storage dynamic monitoring method based on the fusion of spaceborne active and passive remote sensing information provided in this embodiment includes the following steps: S1, laser data preprocessing; S2, multispectral data preprocessing; S3, lake bottom DEM inversion dataset making; S4, construction and training of elevation inversion convolutional neural network model; S5, using the trained elevation inversion convolutional neural network model to predict and generate lake bottom DEM; S6, calculating the absolute water quantity of the lake according to the lake bottom DEM, and monitoring the lake water level-water storage.

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

[0056] In this embodiment, ATL03 data product provided by ICESat-2 satellite is used to mainly obtain along-track photon counting data of each region in the world. ATL03 data contains high-precision photon echo time and positioning information, which is suitable for fine detection of terrain and water surface. The focus of this embodiment is to extract photon points related to underwater topography from ATL03 data.

[0057] ATL08 is a ground elevation product based on ATL03 data, which is specifically used for the extraction of information such as land, forest and vegetation height. ATL08 data has classified features such as ground and vegetation by processing the original photon data of ATL03, and has excluded obvious noise photons. This means that for the terrain near the water area, the ATL03 photon points corresponding to ATL08 have been subjected to some preliminary filtering and classification processing, and are more likely to contain photons with terrain information. Therefore, when screening data, check whether there is a corresponding ATL03 product according to the RTG number of the ATL08 product. In this way, the efficiency of finding ATL03 products with underwater points can be improved.

[0058] Further convert the valid ATL03 product data in the.h5 format into a multi-track raw data set in the.mat format. One track of ICESat-2 point cloud data contains a large number of photon points.

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

[0060] In order to effectively remove a large amount of background noise in the ICESat-2 ATL03 photon cloud data, this embodiment proposes a double-constraint rough denoising method based on DEM (Digital Elevation Model) constraint and sliding window density method. This method introduces topographic prior information and local density characteristics of photons to achieve fast and accurate removal of noise photons and reduce the computational load of subsequent fine filtering.

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

[0062] ;

[0063] wherein, is the height of the i-th photon, is the DEM elevation corresponding to the projection position of the i-th photon, is the height difference of the i-th photon. Set the upper and lower threshold values , and the screening condition is:

[0064] ;

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

[0066] ​​​Photons that meet the screening condition are retained, otherwise deleted. For the remaining photons after DEM constraint, local photon density screening is performed using a fixed-length sliding window:

[0067] ;

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

[0069] Set the density threshold , if , all photons in the window are determined to be noise and removed. To adapt to the changes in terrain, dynamically adjust , and , where the maximum value of the height difference of the photon point is calculated according to the local slope of the DEM :

[0070] ;

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

[0072] Overall, the traditional DEM constraint and the sliding window density method have their own advantages and disadvantages. This method combines the two to achieve complementation and introduces DEM slope constraint to adjust the threshold, improving the stability of noise removal in complex terrain areas. Without significantly increasing the computational burden, the accuracy and fidelity of rough denoising are improved. Experimental verification recommends the following parameters: , , , , .

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

[0074] Import the rough denoising.mat file into the labeling software and perform manual label annotation. Yellow is land, blue is noise, purple is water surface, and red is underwater terrain point. After manual processing, save it as a.csv table file.

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

[0076] ​In this embodiment, the GEE (Google Earth) platform is used to obtain WorldView-2 remote sensing data with as few clouds as possible for processing. First, by importing a csv file containing the date of each ICESat-2 data, the dates are defined as a feature set, providing a basis for subsequent image extraction. Only the near-infrared, red, green, and blue band information of the multispectral image needs to be obtained, which can greatly reduce the download of unnecessary remote sensing information and greatly reduce the complexity of subsequent calculations.

[0077] (2) Lake boundary extraction.

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

[0079] First, the normalized difference water index :

[0080] ;

[0081] wherein and represent the reflectivity of the green band and the near-infrared band, respectively.

[0082] After calculating the index image, a threshold segmentation method is used to binarize the image and extract the lake range. To improve the accuracy of boundary extraction, morphological erosion and dilation operations are further used to remove noise, and the largest water area is retained through connected component analysis, and finally the accurate boundary of the lake is obtained.

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

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

[0085] ;

[0086] wherein is the lake water surface area, is the lake surface elevation.

[0087] First, the lake water surface range is extracted using multi-period WorldView-2 remote sensing images, and the water surface area of each period is calculated. At the same time, the water level elevation at the lake boundary is obtained by combining the ICESat-2 height measurement data of the same date, and the area-elevation observation value pair is constructed.

[0088] Subsequently, the area-elevation relationship is fitted using a quadratic polynomial regression, and the regression model can be expressed as:

[0089] ;

[0090] wherein, , and are regression coefficients. When the lake water surface area is zero, i.e. , the corresponding lake bottom elevation can be calculated:

[0091] ;

[0092] After obtaining the lake bottom elevation, the lake maximum elevation can be calculated as:

[0093] .

[0094] This method is based on the fitting relationship between area and elevation, and can quickly calculate the maximum elevation of the lake under the condition of lacking water bottom DEM or elevation measurement data.

[0095] (2) Dataset construction.

[0096] According to the calculated maximum elevation of the lake, the data greater than the maximum elevation in the ICESat-2 elevation data is removed, and the remaining data is matched with the multispectral image.

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

[0098] In step S4, (1) first data set division. The division of the data set is a key step to ensure the effectiveness of the model training, verification and testing process. Usually, this process includes dividing the data set into training set, validation set and test set.

[0099] ​In actual operation, the division ratio of the data set is 70% for the training set, 15% for the validation set, and 15% for the test set. In addition, in order to more comprehensively evaluate the performance of the model, a cross-validation method is used on the training set, which helps to ensure the stability and generalization ability of the model.

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

[0101] The elevation inversion convolutional neural network model provided in the embodiment takes ConvNeXt as the backbone network, fully extracts local spatial texture and spectral features in remote sensing images, and introduces SwinTransformer modules in the middle part to enhance the modeling ability of the model for global context relationships through windowed self-attention mechanisms. In addition, the model embeds CBAM attention modules multiple times in the encoding stage to further strengthen key feature expression. In the decoding stage, PixelShuffle upsampling is used instead of traditional transpose convolution to effectively improve spatial restoration accuracy, and a 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 balances the efficiency of local feature extraction and the ability of global relationship modeling, and can provide more accurate and more robust prediction results for the elevation inversion task.

[0102] The structure of the elevation inversion convolutional neural network model is shown in Figure 1 The elevation inversion convolutional neural network model takes a neighborhood block F0 with a size of 90 90 4 as input, passes through the Stem module of the ConvNeXt network, uses a 7 7Conv (convolution) to increase the channel number from 4 to 32, and obtains a feature map F1 with a size of 45 45 32.

[0103] Subsequently, F1 enters the ConvNeXt Stage1 module, and passes through three ConvNeXt Blocks for feature extraction, and outputs a feature map F2 with a size of 45 45 64. Each ConvNeXt Block includes a depth separable convolution, a layer normalization, a GELU activation, and a 1 1 convolution.

[0104] To further enhance the attention of the elevation inversion convolutional neural network model to the spatial and channel dimensions, F2 passes through the CBAM attention module combining channel and spatial attention, and generates a feature map F3 with a size of 45 45 the weighted feature map F2_att of 64.

[0105] Then, the 3 3 Conv (convolution) downsamples the feature map F2_att to obtain a feature map F3 of size 23 23 128. On this basis, a Transformer encoder module composed of two layers of Swin Transformer modules is introduced to obtain a feature map F4 of size 23 23 128. The window multi-head self-attention mechanism is used to extract global dependency relationships and improve the representation ability of complex spectral-depth mappings.

[0106] F4 is downsampled by 3 3 Conv (convolution) to generate a feature map F5 of size 12 2 256, and is input into a ConvNeXt Stage2 module. After three layers of ConvNeXt Block calculation, a feature map F6 of size 12 2 256 is obtained. To further focus on key features, F6 is input into a CBAM attention module, and a feature map F6_att of size 12 12 256 is output.

[0107] In the decoding stage, F6_att is first upsampled to 2 24 128 by a PixelShuffle upsampling operation, and is fused with F4 output by the Transformer in the encoding stage. First, the channel dimension is concatenated to obtain a feature map F7 of size 23 23 256, which is then compressed to 128 channels by 3 3 Conv (convolution) to generate a feature map F8 of size 23 23 128.

[0108] Subsequently, F8 is upsampled to a feature map F9 of size 46 46 64 by a PixelShuffle upsampling operation, and is fused with F2_att in the encoding stage. After concatenation, a feature map F10 of size 45 3 Conv (convolution) is generated. 45 64.

[0109] Finally, global average pooling is performed on F10 to obtain a vector F11 of size 1 1 64, and a fully connected layer is used to regress the predicted elevation value of the central pixel, with an output dimension of 1 1 1.

[0110] The elevation inversion convolutional neural network model of the embodiment realizes 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 in the spatial and channel dimensions, and the PixelShuffle upsampling effectively avoids the aliasing effect of traditional upsampling. The overall architecture fully excavates the multi-dimensional information of remote sensing images while ensuring 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 not only verify the effectiveness of the model, but also ensure its reliability in practical applications. The root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2) are used as the main evaluation indicators in this embodiment to measure the prediction ability of the model from multiple angles.

[0112] The root mean square error (RMSE) can be regarded as the square root of the mean square error (MSE), and its calculation formula is:

[0113] ;

[0114] wherein, x i is the i th actual observation value, y i is the i th model prediction value, and n is the total number of actual observation values. Since the result of RMSE has the same unit as the actual observation value, it is more intuitive. At the same time, through the square root operation, the amplification effect of abnormal errors on the overall index is weakened, and the interpretability is improved. The mean absolute error (MAE) calculates the average of the absolute values of the prediction errors to measure the average deviation of the model, and the calculation formula is:

[0115]

[0116]

[0117] ;​​​​​​​​​​​​​

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

[0119] Coefficient of determination ( The formula for measuring the goodness of fit of a model is as follows:

[0120] ;

[0121] in, This is the mean of the observed values. This represents the proportion of the total variation in the dependent variable explained by the model, ranging from 0 to 1. A higher value indicates a better explanation of the data variation. Taking Lake A as an example, the verification results of the underwater elevation inversion of the convolutional neural network model provided in this embodiment are as follows: Figure 4 As shown, , This indicates that the deviation between the model-inverted elevation and the ICESat-2 elevation in this embodiment is small.

[0122] (v) In step S5, a sliding window method is used to extract the 90° surrounding area of ​​each pixel. 90 4-neighborhood spectral blocks were constructed, using the same input format as during model training, where boundary cells were padded to obtain 90°. 90 Four neighborhood spectral blocks are input into a trained elevation inversion convolutional neural network model for prediction, yielding the elevation value of the corresponding center pixel. During the sliding window extraction process, the sliding step size is typically set to 1 to ensure that each pixel is predicted, thereby generating a spatially continuous elevation inversion map.

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

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

[0125] ;

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

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

[0128] First, according to the multi-date multi-spectral data and the corresponding ICESat-2 elevation data, the DEM of each period is constructed by the elevation inversion convolutional neural network model, so as to obtain the absolute water storage of the lake in each period.

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

[0130] The average annual lake water level, lake area and absolute water storage of the lake since 2016 are observed, and the observation results are shown in Figure 5 and Figure 6 The average annual water level of A lake in 2017 is the lowest, 2740.27m, the minimum average annual area is 113.54km 2 , and the minimum average absolute water storage is 385259153.5m 3 . In 2024, the average annual water level is the highest, 2742.33m, the maximum average annual area is 156.03km 2 , and the maximum average absolute water storage is 696214812.8m 3 .

[0131] Of course, the above description is not a limitation of the present application, and the present application is not limited to the above examples. Changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present application should also be within the scope of the present application.

Claims

1. A method for dynamic monitoring of lake water level and water storage based on fusion of spaceborne active and passive remote sensing information, characterized in that, The method comprises the following steps: S1, acquiring ICESat-2 data, performing coarse denoising on the ICESat-2 data based on double constraints of DEM constraints and sliding window density method, and then performing manual label annotation and saving as a.csv table file; S2, acquiring information of a near-infrared band, a red band, a green band and a blue band of the multispectral image, and performing lake boundary extraction; S3, calculating the maximum elevation of the lake based on the relationship between the water surface area and the water level elevation in multiple periods, removing data greater than the maximum elevation in the ICESat-2 data, and then taking the ICESat-2 data as the core, combining the multispectral image covering the corresponding region, and constructing a high-dimensional training data set based on neighborhood features; S4, constructing and training an elevation inversion convolutional neural network model, wherein the elevation inversion convolutional neural network model takes ConvNeXt as a backbone network, introduces a Swin Transformer module in the middle part, embeds a CBAM attention module multiple times in the encoding stage, and adopts PixelShuffle up-sampling in the decoding stage; S5, extract 90 90 4 neighborhood spectral blocks, and construct the same input format as the model training stage, where the boundary pixels are filled with boundary padding to obtain 90 90 4 neighborhood spectral blocks, and input them into the trained elevation inversion convolutional neural network model for prediction to obtain the elevation value of the corresponding center pixel; all pixel elevation values are obtained with a sliding window step of 1, and the prediction results of all pixels are reconstructed according to their original spatial positions to generate the lake bottom DEM; S6, calculating the absolute water volume of the lake according to the lake bottom DEM, and performing lake water level-water storage monitoring; In step S4, when the elevation inversion convolutional neural network model is trained: with size 90 90 4 as input, through the Stem module of the ConvNeXt network, using 7 7 convolutions to increase the number of channels to 32 and obtain a feature map F1 with size 45 45 32 F1 enters the ConvNeXt Stage1 module, and feature extraction is performed through three ConvNeXt Blocks, and the output size is 45 45 64; F2 passes through a CBAM attention module to generate a weighted feature map F2_att with a size of 45 45 64; each ConvNeXt Block contains a depth separable convolution, layer normalization, a GELU activation, and a 1 1 convolution; Next, the 3 3convolution down-samples the feature map F2_att to obtain a feature map F3 of size 23 23 128 On this basis, the Transformer encoder module composed of two layers of Swin Transformer modules obtains a feature map F4 with a size of 23 23 128, and the global dependency is extracted through the window multi-head self-attention mechanism. F4 is down-sampled by 3 3 convolution to generate a feature map F5 with size 12 2 256, and input into the ConvNeXt Stage2 module, and calculate through three layers of ConvNeXt Block to obtain a feature map F6 with size 12 2 256, and input into the ConvNeXt Stage2 module, and calculate through three layers of ConvNeXt Block to obtain a feature map F6 with size 12 12 256, and input into the ConvNeXt Stage2 module, and calculate through three layers of ConvNeXt Block to obtain a feature map F6 with size 12 F6_att is up-sampled by a PixelShuffle operation to have a size of 2 24 128 feature maps F7, and multi-scale feature fusion is performed with F4. First, 23 23 256 features are obtained by concatenation along the channel dimension, and then 3 3 convolutional compression is performed to 128 channels to generate a feature map F8 with a size of 23 23 128 F8 is up-sampled by PixelShuffle to size 46 46 64feature map F9, and multi-scale feature fusion is performed with F2_att, and after splicing, a 3 3convolution is performed to generate a 45 45 64feature map F10; Finally, global average pooling is performed on F10 to obtain a vector F11 of size 1 1 64, and the height value corresponding to the center pixel is predicted by a fully connected layer, and the output dimension is 1 1 1.

2. The lake water level and storage dynamic monitoring method based on the fusion of spaceborne active and passive remote sensing information according to claim 1, characterized in that, In step S1, the ICESat-2 data is ATL03 data provided by an ICESat-2 satellite; Firstly, the ATL03 data corresponding to the RTG number of the ATL08 data is selected and reserved; Then, the ATL03 data is inputted, DEM constraint pre-screening is performed, and then sliding window density screening is performed, and a rough denoised photon data is outputted by combining dynamic threshold adjustment; Finally, the rough denoised photon data is imported into a labeling software, and manual label annotation is performed, and land, noise, water surface and underwater terrain points are marked with different colors respectively.

3. The lake water level and storage dynamic monitoring method based on the fusion of spaceborne active and passive remote sensing information according to claim 2, characterized in that, In step S1, the method for data coarse denoising is: S11, assuming that the deviation of the photon height distribution of the lake or the ground in the study area with respect to the DEM is within a limited range, the height difference of each photon point is calculated: ； in, For the first The height of a photon For the first Each photon projection position corresponds to the DEM elevation. For the first The height difference of each photon; S12, setting upper and lower threshold values The screening condition is: ； wherein, is the minimum value of the photon dot height difference, is the maximum value of the photon dot height difference; the photon is retained if the screening condition is met, otherwise it is deleted; S13, for the remaining photons after the DEM constraint, a fixed length sliding window is used for local photon density screening: ； wherein, is the photon density per unit length, is the window length, is the number of photons within each window; S14, set density threshold , if , all photons in the window are determined as noise and rejected; S15, dynamically adjusting to adapt to the changes in terrain , with wherein the maximum value of the height difference of the photon points is calculated according to the local slope of the DEM :​ ； wherein, is the upper limit of the base elevation, is the slope sensitivity factor.

4. The lake water level and storage dynamic monitoring method based on spaceborne active and passive remote sensing information fusion according to claim 3, characterized in that, In step S2, firstly, a.csv file containing the date of each ICESat-2 data is imported, the date is defined as a feature set, and a basis is provided for the extraction of the multispectral image.

5. The lake water level and storage dynamic monitoring method based on the fusion of spaceborne active and passive remote sensing information according to claim 4, 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: ； wherein, is the lake surface area, is the lake surface elevation; The water surface area of each period is calculated by extracting the water surface range of the lake by using multi-period and multi-spectral images, and simultaneously, the water level elevation at the lake boundary is obtained by combining the ICESat-2 data of the same date, and an area-elevation observation value pair is constructed ​ A quadratic polynomial regression is used to fit the area-elevation relationship, and the regression model is represented as: ； wherein , and are regression coefficients; the corresponding lake bottom elevation when the lake surface area is zero : ； the water surface elevation at the time of the highest water level the maximum elevation of the lake is expressed as: 。 6. The lake water level and storage dynamic monitoring method based on spaceborne active and passive remote sensing information fusion according to claim 5, characterized in that, In step S4, the input data of the height inversion convolutional neural network model includes the denoised ICESat-2 height points containing latitude, longitude and height value, and the information of the near-infrared band, red band, green band and blue band of the multispectral image. In the multispectral image data, 90 90 4 neighborhood spectral blocks, and the multi-dimensional features are taken as the input of the height inversion convolutional neural network model, and the target height value is taken as the label of the height inversion convolutional neural network model, to construct the complete input of the height inversion convolutional neural network model.

7. The lake water level and storage dynamic monitoring method based on spaceborne active and passive remote sensing information fusion according to claim 6, characterized in that, In step S6, the calculation method of the absolute water volume of the lake is: The lake bottom DEM is used to analyze the submerged range of the lake at different water levels, and the water volume of the lake is calculated according to integration. For each water level height, the water volume is calculated : ； wherein, is the maximum water level height of the lake, is the water level height is the lake area at the water level height, and is the adjacent water level height, .

8. The lake water level and storage dynamic monitoring method based on spaceborne active and passive remote sensing information fusion according to claim 7, characterized in that, In step S6, the lake water level-water storage monitoring method is: Firstly, the DEM of each period is constructed by using the elevation inversion convolutional neural network model based on the multispectral data of multiple dates and the corresponding ICESat-2 elevation data, so as to obtain the absolute water storage of the lake in each period; Then, the water level data is obtained by using the ICESat-2 laser height measurement, and the absolute water storage of the lake and the water level data are fitted to obtain a water level-water storage regression curve.

Images