Method for constructing lake underwater topography prediction model, lake underwater topography prediction method and related device

Abstract

The application discloses a lake underwater topography prediction model construction method, a lake underwater topography prediction method, a computer program product, an electronic device and a computer readable storage medium. The lake underwater topography prediction model construction method comprises the following steps: extracting real topography data of a plurality of target topographies from an existing DEM topography data set; extracting target data of the target topography from the real topography data; inputting the target data into a preset mathematical model to obtain mathematical simulation topography data of the target topography; and training a lake underwater topography prediction model by taking the mathematical simulation topography data as a training data set. The lake underwater topography prediction model is obtained by taking the error between the output prediction topography data and the real topography data as a loss function. The model prediction accuracy can be improved.

Description

Technical Field

[0001] This application relates to the field of underwater topography prediction, and in particular to a method for constructing an underwater topography prediction model for lakes, a method for predicting underwater topography for lakes, a computer program product, an electronic device, and a computer-readable storage medium. Background Technology

[0002] Lake water storage, as a comprehensive indicator of the hydrological balance of a watershed, can be derived through topographic assessment of the lake. However, measuring lake depth in practice is difficult, and remote sensing satellites have insufficient underwater reflection capabilities. Water storage assessment based on empirical formulas is currently a common method, but it cannot obtain information about the underwater topography of lakes. While current mathematical models based on physical mechanism assumptions can provide theoretical solutions for the underwater topography of any lake globally, their accuracy and applicability are relatively poor. Summary of the Invention

[0003] To address the aforementioned technical problems, this application provides an obstacle detection method, a computer program product, an electronic device, and a computer-readable storage medium, the technical solutions of which are as follows:

[0004] According to a first aspect of this application, a method for constructing a lake underwater topography prediction model is provided, the method comprising:

[0005] From the existing DEM terrain dataset, extract real terrain data for several target terrains, which are terrains with similar topographic structures to lakes and that are not submerged.

[0006] The target data of the target terrain is extracted from the real terrain data. The target data includes boundary data and data of the area enclosed by the boundary. The boundary data includes at least the coordinate information and elevation information of each DEM pixel on the boundary of the target terrain. The data of the area enclosed by the boundary includes at least the coordinate information of each DEM pixel in the area.

[0007] The target data is input into a preset mathematical model to obtain mathematically simulated terrain data of the target terrain. The mathematical model is constructed based on the spatial continuity of the land and underwater terrain of the lake. The mathematically simulated terrain data includes the target data and the simulated elevation information of each DEM pixel in the area surrounded by the boundary.

[0008] The mathematically simulated terrain data is used as a training dataset to train the underwater terrain prediction model of the lake. The underwater terrain prediction model of the lake is trained using the error between the output predicted terrain data and the actual terrain data as a loss function.

[0009] According to a second aspect of this application, a method for predicting underwater topography of lakes is provided, the method comprising:

[0010] In the existing DEM terrain dataset, real target data of real lakes is extracted. The real target data includes at least the coordinate and elevation information of each DEM pixel on the real boundary of the real lake, as well as the coordinate information of each DEM pixel in the area surrounded by the real boundary.

[0011] The real target data is input into a preset mathematical model to obtain mathematically simulated terrain data of the real lake. The mathematical model is constructed based on the spatial continuity of the lake's land and underwater terrain. The mathematically simulated terrain data includes the real target data and the simulated elevation information of each DEM pixel in the area surrounded by the real boundary.

[0012] The mathematically simulated terrain data is input into the underwater terrain prediction model of the lake as described in the first aspect, and the predicted terrain data of the real lake is output through the underwater terrain prediction model of the lake.

[0013] According to a third aspect of this application, a computer program product is provided, the computer program product comprising a computer program that, when executed by a processor, implements the method as described in the first or second aspect.

[0014] According to a fourth aspect of this application, an electronic device is provided, the electronic device comprising:

[0015] processor;

[0016] Memory used to store processor-executable instructions;

[0017] The processor is configured to implement the method as described in the first or second aspect.

[0018] According to a fifth aspect of this application, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method as described in the first or second aspect.

[0019] The technical solution provided in this application constructs a mathematical model based on the physical mechanism assumption of the spatial continuity between the land and underwater topography of lakes. This model can obtain a theoretical solution (i.e., mathematically simulated terrain data) for the target terrain. Based on this theoretical solution, the underwater terrain prediction model further identifies the differences in numerical and spatial distribution between the theoretical solution and the actual terrain, achieving accurate prediction of the actual terrain. Simultaneously, considering the scarcity of actual measured samples of lake terrain, which makes it difficult to meet the quantity and quality requirements of data-driven models (e.g., insufficient training data when building models using deep learning architectures), to construct a large number of training samples suitable for the underwater terrain prediction model of lakes, based on terrain similarity considerations, real terrain data of unsubmerged terrain (e.g., valleys, depressions) with similar terrain structures to lakes are extracted from existing DEM terrain datasets. Based on these real terrain data, a large number of samples are constructed to train the model, thereby greatly improving the model's training effect and thus increasing the model's prediction accuracy.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings.

[0022] Figure 1 This is a schematic diagram of a specific scenario according to an embodiment of this application;

[0023] Figure 2 This is a flowchart illustrating a method for constructing a lake underwater topography prediction model according to an embodiment of this application;

[0024] Figure 3 This is a schematic diagram of a river valley topography according to an embodiment of this application;

[0025] Figure 4 This is a schematic diagram of a river valley splicing according to an embodiment of this application;

[0026] Figure 5 This is a schematic diagram of a lake classification scenario according to an embodiment of this application;

[0027] Figure 6 This is a schematic diagram of the structure of a lake underwater topography prediction model according to an embodiment of this application;

[0028] Figure 7This is a schematic diagram of experimental data from one embodiment of this application;

[0029] Figure 8 This is a schematic diagram of experimental data from another embodiment of this application;

[0030] Figure 9 This is a flowchart illustrating a method for predicting underwater topography of lakes according to an embodiment of this application;

[0031] Figure 10 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0032] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art should fall within the scope of protection of this application.

[0033] Lake water storage is a comprehensive indicator of the hydrological balance of a watershed. Taking the Qinghai-Tibet Plateau as an example, quantitatively assessing the lake water storage of the Qinghai-Tibet Plateau is of great significance for scientifically understanding the value of lake water resources, the response of lakes to climate change, and the water security strategy of river basins. Figure 1 As shown, lake water storage can be determined through topographic assessment. Lake depth (e.g., ...) Figure 1 Determining the depth or elevation of various points below the surface of the lake (as shown in the image) is difficult due to the limited reflectivity of remote sensing satellites beneath the water surface. While empirical formulas are commonly used for water storage assessment, they cannot capture the underwater topography of lakes. Current mathematical models based on physical mechanism assumptions can provide theoretical solutions for the underwater topography of any lake globally, but their accuracy and applicability are limited.

[0034] With the current development of artificial intelligence, AI algorithms driven by big data analysis and multi-source data are accelerating the innovation of water storage assessment methods. However, the scarcity of measured water depth samples (measured samples of lake topographic elevation information) makes it difficult to meet the quantity and quality requirements of data-driven models for training samples. The difficulty and high cost of obtaining measured underwater topographic data for lakes lead to data scarcity, and the use of AI models for underwater topographic prediction also faces the problem of insufficient training data.

[0035] To address the aforementioned problems, this application provides a method for constructing a lake underwater topography prediction model, which can improve the model's training effect, thereby increasing the model's prediction accuracy and enabling precise prediction of the lake's actual topography. Figure 2 As shown, the method includes the following steps:

[0036] S201. Extract the actual terrain data of several target terrains from the existing DEM terrain dataset.

[0037] The target terrain is an unsubmerged terrain with a similar topographical structure to that of a lake.

[0038] S202. Extract the target data of the target terrain from the real terrain data.

[0039] The target data includes boundary data and data of the area enclosed by the boundary. The boundary data includes at least the coordinate information and elevation information of each DEM pixel on the boundary of the target terrain, and the data of the area enclosed by the boundary includes at least the coordinate information of each DEM pixel in the area.

[0040] S203. Input the target data into a preset mathematical model to obtain mathematical simulation terrain data of the target terrain.

[0041] The mathematical model is constructed based on the spatial continuity of the land and underwater topography of the lake. The mathematically simulated terrain data includes the target data and the simulated elevation information of each DEM pixel in the area enclosed by the boundary.

[0042] S204. Use the mathematically simulated terrain data as a training dataset to train the underwater terrain prediction model for the lake.

[0043] The underwater terrain prediction model for the lake is trained using the error between the output predicted terrain data and the actual terrain data as a loss function.

[0044] The technical solution provided in this application, based on the physical mechanism assumption of the spatial continuity between the land and underwater topography of lakes, constructs a mathematical model that can obtain a theoretical solution (i.e., mathematically simulated terrain data) for the target terrain. Building upon this theoretical solution, the underwater terrain prediction model further identifies the differences in numerical and spatial distribution between the theoretical solution and the actual terrain, achieving accurate prediction of the actual terrain. Simultaneously, considering the scarcity of actual measured samples of lake terrain, which makes it difficult to meet the quantity and quality requirements of data-driven models (e.g., insufficient training data when building models using deep learning architectures), to construct a large number of training samples suitable for the underwater terrain prediction model of lakes, based on terrain similarity considerations, real terrain data of unsubmerged terrain (e.g., valleys, depressions) with similar terrain structures to lakes are extracted from existing DEM terrain datasets. Based on these real terrain data, a large number of samples are constructed to train the model, thereby greatly improving the model's training effect and thus increasing the model's prediction accuracy.

[0045] The aforementioned Digital Elevation Model (DEM) terrain dataset is a digital model that records surface elevation information in the form of a regular grid. It is an existing basic data in geographic information systems for describing terrain. Existing DEM terrain datasets are composed of regular grid points (DEM pixels) (plane coordinates X, Y) and their corresponding elevation values ​​(Z), used to accurately represent terrain undulations.

[0046] Existing DEM terrain datasets use the Tag Image File Format (TIFF) to store accurate terrain data of digital elevation models. TIFF is a raster file format that incorporates geospatial information. It stores coordinate systems (such as WGS84, EPSG:4326), projection parameters (such as UTM), resolution, and geographic reference points through built-in tags to ensure that the stored data matches the actual geographical location.

[0047] Existing DEM terrain datasets can be obtained in various ways. As an example, a DEM terrain dataset can be obtained from the Shuttle Radar Topography Mission Digital Elevation Model (SRTM DEM). The SRTM DEM is a global terrain elevation dataset obtained in 2000 by NASA, the National Institute of Geodesy and Mapping (NIMA), and institutions from Germany and Italy through the Shuttle Radar Mapping Mission. It covers the land area from 60° North latitude to 56° South latitude (accounting for 80% of the global land area). As another example, DEM terrain datasets can also be obtained from the Multi-Error-Removed Improved-Terrain Digital Elevation Model (MERIT DEM). MERIT DEM is a high-precision global digital elevation model designed to provide a more accurate representation of terrain by eliminating systematic errors in existing DEM data. Released in 2017 by institutions including the University of Tokyo, it is developed based on multi-source DEM data fusion and error correction technology, including baseline data such as NASA's SRTM3 DEM v2.1, JAXA's AW3D-30m DEM v1, and Viewfinder Panoramas DEM, supplemented by auxiliary data such as ICESat laser altimetry and Landsat forest cover.

[0048] It is worth noting that the above description of the methods for obtaining existing DEM terrain datasets is only an example. In practical applications, other methods may exist, and no specific limitations are made here.

[0049] The target terrain described above can be implemented in various ways. For example, the target terrain could be a depression, or another example, a river valley. In real-world terrain, depressions and river valleys are the most similar to lakes in terms of topographic structure. Their geological formation is also similar to that of lakes. Furthermore, depressions and river valleys are often unsubmerged terrain, making water depth measurement relatively easy. Remote sensing satellites also do not suffer from insufficient underwater reflection capabilities (due to the lack of submersion) for depressions or river valleys. Therefore, there is often a relatively abundant amount of measured data on the overall terrain of depressions and river valleys, making them suitable locations for obtaining training samples. It is worth noting that the above descriptions of specific implementations of the target terrain are merely illustrative. In practical applications, other specific implementations are possible, and no specific limitations are imposed on these.

[0050] There are several ways to extract real terrain data from a target terrain, such as... Figure 3 As shown, considering that the elevation of a river valley typically decreases first and then remains relatively flat, and that the flat areas are often quite large, and that these flat areas have little similarity to lake topography, they are not necessarily the target that the model needs to learn. Learning from them would actually reduce the model's learning efficiency and increase its computational burden. To address this issue, as an example, when the target terrain is a river valley, the original terrain data of the river valley can be extracted. Based on the original terrain data, the lowest position of the river valley can be determined, and the target area of ​​the river valley can be segmented based on this lowest position. This target area is the region segmented by a dividing plane that passes through the lowest position and is perpendicular to the direction of the river valley. This target area is a non-flat area on one side of the river valley. The target area and its mirror image can be stitched together to obtain the stitched result, and this stitched result is determined as the true terrain data of the aforementioned river valley. Figure 3 As shown, a cross-section of a specific river valley terrain is used as an example. The valley's orientation reveals a large, gently sloping area; that is, the valley's terrain slopes down from one side, then flattens out, before rising again on the other side. Based on the valley's lowest point, a dividing plane perpendicular to the valley's orientation can be determined. This lowest point then segments the non-gentle area on one side of the valley terrain, which is the target area. Figure 3 Clearly, the terrain of the target area slopes down as a whole, until it reaches its lowest point. For example... Figure 4 As shown, after obtaining the non-flat area on one side of the valley, i.e., the target area, we can further mirror this target area to obtain the result shown below. Figure 4The "mirror region" shown is stitched together with the target region (non-flat region) to obtain the stitched result (i.e., Figure 4 The valley topography is pieced together to form a bowl-shaped closed region similar to a depression. This pieced-together topography can be identified as the actual topographic data input into the preset mathematical model. By preprocessing the original valley topographic data to resemble a depression (bowl-shaped) (the actual topographic data input into the mathematical model), the valley is cut at its lowest point. The topographic region on the side containing the gentle slope is removed, while the non-gentle slope region on the other side is retained. The retained region is then mirrored and pieced together. The resulting closed region approximates a depression and serves as the final input data for the mathematical model of the valley topography. This improves the efficiency of subsequent learning in the lake underwater topography prediction model and reduces the computational burden on the model.

[0051] There are several ways to perform mirroring at the lowest point of the valley terrain. As an example, one can study and find the theoretical maximum depth of all depressions in the existing measured data (real terrain data), as well as the theoretical maximum depth of lake terrain with measured water depth data. Then, calculate the slope values ​​of each adjacent DEM pixel at these theoretical maximum depth locations along the longitudinal or cross-sectional sections of the terrain, and use the average slope value of each adjacent DEM pixel as the maximum critical threshold for mirroring at the lowest point of the valley terrain.

[0052] As an example, when the target terrain is a depression, a preset contour line can be determined, and data can be extracted from the SRTMDEM 90m data based on the preset contour line to construct depression areas below the preset contour line in batches. For example, depression terrain data at altitudes below 4000 meters on the Qinghai-Tibet Plateau can be extracted.

[0053] Considering that the shape of a lake boundary can reflect the formation and evolution of the lake basin, lakes with similar boundary shapes often exhibit similar topography. Therefore, as an example, the aforementioned underwater topography prediction model for lakes can include multiple sub-models. Based on the boundary data included in the target topography data, the target topography can be classified into boundaries. Then, mathematically simulated topographic data belonging to different boundary types are input into different sub-models. Classifying the boundary shapes of the target topography and placing the mathematically simulated topographic data corresponding to different boundary types into different sub-models for learning allows for a better understanding of the topographic features corresponding to different boundary shapes.

[0054] The target terrain can be classified in various ways, and mathematical simulation terrain data belonging to different boundary types can be input into different sub-models in various ways. As an example, based on the boundary data included in the target terrain data, the ellipticity index corresponding to the target terrain can be determined, and based on the determined ellipticity index, the boundary type to which the target terrain belongs can be determined from the preset boundary types. The preset boundary types can include elliptical, rectangular, and circular shapes. The aforementioned lake underwater terrain prediction model can include a first sub-model, a second sub-model, and a third sub-model corresponding to different preset boundary types. Then, based on the boundary type to which the target terrain belongs, the target sub-model to which the mathematical simulation terrain data of the target terrain should be input is determined. This target sub-model is determined from the first sub-model, the second sub-model, and the third sub-model included in the lake underwater terrain prediction model, which correspond to different preset boundary types. Lake boundary shapes can be roughly divided into three categories based on the ellipticity index: elliptical, rectangular, and circular. Elliptical shapes are usually barrier lakes formed by river damming; rectangular shapes are usually river-formed lakes at the end of rivers; and circular shapes are usually closed lake basins formed by geological structures.

[0055] The following is combined Figure 5 The following is an example illustrating the specific implementation of classifying the aforementioned target terrain:

[0056] like Figure 5 As shown, the ellipticity index (R) can be used. Ei To differentiate between different types of target terrain, the ellipticity index can quantify the geometry and orientation of the target terrain or lake. Its formula is as follows:

[0057]

[0058] R Ei The value range of R is from -1 to 1. Ei When R approaches -1, the boundary shape of the lake or target terrain tends to be elliptical; when R... Ei When REi approaches 0, the boundary shape of the lake or target terrain exhibits a rectangular shape; conversely, when REi approaches 1, the boundary shape of the lake or target terrain tends to be circular.

[0059] In formula (1), E represents the ellipticity of the boundary shape, which can be expressed by the following formula:

[0060]

[0061] Where AcE represents the area formed by the actual boundary of the lake or target terrain (A lake ) and the best elliptical approximation region of lake shape (A) BFE The area formed by the intersection of ).

[0062] R represents the rectangularity of the boundary shape, which can be expressed by the following formula:

[0063]

[0064] Among them, Ac BFR The area formed by the actual boundary of a lake or target terrain (A) lake ) and the best-fit rectangle (A BFR The area of ​​intersection of regions A lake Principal axis and best-fit rectangle (A BFR Alignment. Similarly, Ac MBR The area formed by the actual boundary of a lake or target terrain (A) lake ) and minimum bounding rectangle (A MBR The area of ​​intersection of regions A lake The principal axis and the minimum bounding rectangle (A MBR Alignment.

[0065] Optimal Elliptic Approximation (A) BFE ), minimum bounding rectangle (A MBR ) and best-fit rectangle (A BFR It can be constructed in several ways:

[0066] As an example, for the optimal elliptic approximation (A BFE To construct the polygon (A), we can first calculate the semi-major axis (SMAA) and semi-minor axis (SMIA). Then, we rotate the calculated axes so that they align with the lake polygon (A). lake The orientation is aligned. The calculation formula is as follows:

[0067]

[0068]

[0069] θ=arctan2(λ2,λ1) (6)

[0070]

[0071] The principal axis ratio (AR) represents the aspect ratio of a lake or target terrain and serves as an indicator of its stretching degree, encompassing both its size and orientation. To calculate AR, boundary data (e.g., coordinates) defining the lake or target terrain is first collected. Then, principal component analysis (PCA) is applied to identify the direction of the maximum variance in the coordinates of the lake or target terrain's boundary, thus calculating AR. A rotation angle (denoted as θ) is used to align the principal axis with this direction. The variances of the first and second principal components (denoted as λ1 and λ2, respectively) are crucial for determining the lake polygon (A). lake The directions of the primary and secondary axes are crucial.

[0072] As an example, for the minimum bounding rectangle (A) MBR The construction of ) can be based on the lake polygon (A lake The primary and secondary axes generate an outer rectangle aligned with them, as shown in the following formula:

[0073]

[0074]

[0075] A MBR =H MBR ·W MBR (10)

[0076] Among them, W MBR and H MBR These represent the minimum bounding rectangle (A) MBR ) width and height.

[0077] As an example, for the best-fit rectangle (A) BFR The construction of ) can be achieved by using the minimum bounding rectangle (A MBR This was derived by applying an area scaling factor to adjust A. MBR The dimensions can be used to obtain A. BFR Its formula can be as follows:

[0078]

[0079] W BFR =k·W MBR (12)

[0080] H BFR =k·H MBR (13)

[0081] A BFR =H BFR ·W BFR(14)

[0082] Where k is the area scaling factor, W BFR and H BFR These represent the best-fit rectangles (A and B) BFR ) width and height.

[0083] It is worth noting that the above description of the classification methods for target terrain is only an example. In practical applications, other classification methods may exist, and no specific limitations are made.

[0084] As an example, before inputting the target data extracted from the real terrain data of the target terrain into the preset mathematical model, the initial elevation values ​​of each DEM pixel in the region enclosed by the boundary of the target terrain can be filled with zeros to obtain preprocessed target data. This preprocessed target data is then input into the aforementioned preset mathematical model. The target data input into the mathematical model consists of the boundary data of the target terrain and the data of the region enclosed by that boundary. The data on the boundary is known and real, while the elevation information contained in the data of the region enclosed by the boundary is initially filled with zeros; that is, the elevation information contained in the data of the region enclosed by the boundary needs to be simulated by the mathematical model to obtain a theoretical solution.

[0085] As another example, the boundary data contained in the target data input to the above mathematical model can exist in the form of a binary mask describing the boundary contour of the target terrain, while the data of the area enclosed by the boundary contained in the target data input to the above mathematical model can be understood as a DEM image with the internal elevation filled with zero for the target terrain.

[0086] Understandably, the aforementioned mathematical model is constructed based on the spatial continuity of the land and underwater topography of lakes. This spatial continuity stems from the existing lake-underwater topography similarity assumption, a theoretical model based on the spatial continuity or similarity between land and underwater topography, primarily used to infer underwater topography in lakes, rivers, and other bodies where measured data is lacking. The core idea of ​​this assumption is to extrapolate or interpolate known topographic features (such as slope and elevation distribution) of the lakeshore or adjacent areas to construct a model of the underwater topography. By exploring the spatial correlation between land and underwater topography, the lake-underwater topography similarity assumption provides an efficient and low-cost solution for modeling water bodies lacking measured data.

[0087] As an example, the mathematical model described above can simulate underwater terrain based on Hermite interpolation. Specifically, it identifies and processes DEM image regions based on Hermite interpolation until the elevation drops to the same height or slope threshold. As another example, the Hermite interpolation process can start from the highest point on the boundary of the target terrain or lake, using this point as a reference. From this reference point, the elevation values ​​of surrounding pixels are gradually reduced, generating a series of gradually decreasing contour lines representing the underwater terrain. As yet another example, the mathematical model can reduce the elevation in 1-meter increments in each iteration. Since the target data input to the mathematical model can contain a binary mask describing the boundary contour of the lake or target terrain, and a DEM image with zeros filled inside the (virtual) lake or target terrain—meaning the information inside the lake is unknown while the land surface information is known—the Hermite interpolation operation also proceeds from the boundary of the lake or target terrain inwards. This means that the DEM value calculated under normal circumstances will always be lower than the DEM value at the highest point of the boundary.

[0088] Based on this, using the highest point on the boundary of the target terrain or lake as a reference point ensures that the subsequent computational domain is either positive or negative, avoiding the possibility of crossing the zero point. However, if a value near the zero point is chosen as the reference point, unpredictable changes will occur regardless of whether scaling or division is used in the calculation, increasing computational complexity.

[0089] Considering that any DEM pixel on the boundary of a lake or target terrain has multiple directions (e.g., north, south, east, west), and that different boundary points require different slope directions, and that the slope decrease at the boundary calculation point does not occur in only one direction (e.g., a lower boundary point may need to retreat upwards and to the left (or right) simultaneously as it retreats into the lake or target terrain), it is necessary to calculate the slope values ​​in multiple directions for the same DEM pixel. As an example, in each iteration, the slope of each boundary DEM pixel in the four main directions (north, south, east, west) can be calculated. Directions towards water can be excluded from slope calculation to maintain the integrity of the lake or target terrain boundary. Furthermore, morphological operations such as dilation and erosion can be applied to further refine the terrain delineation. The formula for slope calculation is as follows:

[0090] P(i) = (DEM(land(i)) - DEM( lake (i))) / cellsize (15)

[0091]

[0092] Among them, DEM(land(i)) and DEM( lake(i) represents the elevation of the land and lake surfaces at a given point i, respectively; P(i) represents the slope receding inward from the lake (or target terrain) boundary, calculated in the specified directions (such as north, south, east, and west as mentioned above); cellsize refers to the resolution of the digital elevation model (DEM). As an example, cellsize can be 3 arcseconds; k1 represents the average slope calculated for each point on the boundary in all directions; n is the total number of directions considered in the analysis.

[0093] Considering that the sample data used to train the aforementioned underwater topography prediction model for lakes are all constructed based on data of non-submerged terrain (such as valleys and depressions) with similar topographic structures to lakes, rather than based on measured topographic data of real lakes, the prediction results of the trained underwater topography prediction model may still have some errors compared to the existing and relatively limited measured topographic data of real lakes. To reduce this error, as an example, after training the aforementioned underwater topography prediction model, several existing measured topographic data of real lakes can be obtained. This measured topographic data is obtained by actually measuring the topography of real lakes. Measured target data of real lakes is extracted from this measured topographic data. This measured target data includes measured boundary data and data of the area enclosed by the measured boundary. The measured boundary data includes at least the measured coordinates and measured elevation information of each DEM pixel on the boundary of the real lake, while the data of the area enclosed by the measured boundary includes at least the measured coordinates and measured elevation information of each DEM pixel in that area. The coordinate information is measured; the measured target data is input into the above mathematical model to obtain the mathematically simulated terrain data of the real lake. The mathematically simulated terrain data includes the measured target data and the simulated elevation information of each DEM pixel in the area surrounded by the measured boundary; the mathematically simulated terrain data of the real lake is input into the trained underwater terrain prediction model of the lake to obtain the predicted terrain data of the real lake; the parameters of the trained underwater terrain prediction model of the lake are optimized based on the error between the predicted terrain data of the real lake and the measured terrain data of the real lake to obtain the parameter-optimized underwater terrain prediction model of the lake.

[0094] As can be seen, after the underwater topography prediction model for lakes is trained on samples of lake-like terrain such as valleys and depressions, it can be further optimized using a small amount of existing measured data from real lakes. This involves extracting the boundaries of the real lakes, inputting these boundaries into the mathematical model to obtain the overall topography simulation results of the real lakes, and then inputting the simulation results into the prediction model to obtain the prediction results. The model can then be optimized by analyzing the deviation between the prediction results and the measured results, which can further reduce the prediction error of the optimized underwater topography prediction model for lakes.

[0095] The aforementioned underwater terrain prediction model for lakes can be implemented in various ways. As an example, the underwater terrain prediction model for lakes can be an Enhanced Deep Residual Networks for Single Image Super-Resolution (EDSR) model. EDSR is a deep residual network architecture widely used in computer vision tasks, capable of identifying and compensating for differences in numerical and spatial distribution between mathematically simulated terrain data and real terrain data.

[0096] As an example, such as Figure 6 As shown, the EDSR model architecture first extracts features through a convolutional layer, followed by several improved ResNet modules. Each module contains two convolutional layers interspersed with ReLU activation functions. The output of each ResNet module is passed through an element-wise summation layer (e.g., ...). Figure 6 The "+" sign in the model indicates further optimization to refine the feature maps. The final layer of the EDSR model is a convolutional layer used to map the coarse mathematical solution to the final terrain output. As another example, given computational resource constraints, the performance of the EDSR model can be improved by increasing the number of ResNet modules.

[0097] It should be noted that, Figure 6 The "virtual lake (depression, valley) mathematical simulation terrain sample library" in the "model training" stage can correspond to the mathematical simulation terrain data of the above-mentioned target terrain, and the "virtual lake (depression, valley) real terrain sample library" can correspond to the predicted terrain data output by the above-mentioned underwater terrain prediction model of lakes. Figure 6 The “model prediction” stage is the parameter tuning stage mentioned above. The “sample library of typical measured lakes and other real terrain” in the “model prediction” stage can correspond to the measured terrain data of the real lakes mentioned above. The “sample library of typical measured lakes and other mathematically simulated terrain” can correspond to the mathematically simulated terrain data of the real lakes mentioned above. The “prediction results of typical measured lakes” can correspond to the predicted terrain data of the real lakes mentioned above.

[0098] The performance of the aforementioned underwater topography prediction model for lakes can be evaluated in several ways. As an example, datasets used to validate the model's performance include: ① a test set (20%) and a validation set (10%) of land depressions and valleys; ② measured water depth data of typical lakes; ③ topography of the lake perimeter inundation area caused by changes in lake water volume recorded by the DEM. As another example, correlation R², root mean square error RMSE, absolute error AE, and relative error RE can be used to compare the elevation of the validation dataset pixel-by-pixel, and the number of residual blocks and learning rate of the EDSR model can be tuned to improve model performance.

[0099] As another example, for the aforementioned EDSR model, the dataset can be divided into 80% for training, 10% for validation, and 10% for testing. Training uses 50×50 pixel image patches. Data augmentation is performed through sliding cropping, and DEM grayscale values ​​are normalized to the range of 0 to 1 to accelerate model convergence. The network structure consists of an initial 3×3 convolutional layer and 32 residual blocks, with a total of 38,932,992 parameters. The training set contains 85.1 billion pixels of elevation data. Experimental results show that training and convergence are completed within 300 epochs. As another example, the EDSR model is trained using the Adam optimizer with an initial learning rate of 0.00001 on an A6000 processor, taking approximately 12 hours per epoch.

[0100] As an example, when evaluating the accuracy and generalization ability of the above EDSR model, the evaluation dataset can include two parts: ① water depth data from six lakes on the Tibetan Plateau, representing different geographical and climatic regions, and diverse lake geometries (size, shape, depth; as shown in Table 1 below); ② water depth data from more than 1400 lakes larger than 1 km. 2 Among the lakes on the Qinghai-Tibet Plateau, 597 lakes experienced an expansion in lake area and a change in inundation zone exceeding two pixels (180 meters) between 2000 and 2022.

[0101] Table 1

[0102]

[0103] Table 1 shows the underwater topography reconstruction performance of the coupled model and the mathematical model for three types of virtual lakes (three lake boundary types: elliptical, rectangular, and circular) during the testing and validation periods. All values ​​are presented as mean. The improvement ratio (%) indicates the improvement in reconstruction performance of the coupled model (the model formed by coupling the aforementioned preset mathematical model with the aforementioned lake underwater topography prediction model) compared to the existing mathematical model.

[0104] The coupled model proposed in this application (the model formed by coupling the aforementioned preset mathematical model with the aforementioned lake underwater topography prediction model) significantly improves the numerical accuracy and spatial distribution of underwater topography reconstruction. In tests and verifications of three lake boundary types (ellipse, rectangle, and circle), it achieved an average R0. 2 =0.48, NRMSE=16.15%, RMSE=0.13 and RE=118% (see Table 1 above). Compared with existing mathematical models (R 2 =-1.72, NRMSE=30.79%, RMSE=0.22, RE=131%), accuracy improved by approximately 50%, terrain similarity (R 2 This has improved by over 100%. Visualization of the reconstructed terrain (see...) Figure 7), which further verifies the capability of the coupled model in improving spatial distribution, especially in capturing subtle geomorphic features.

[0105] See Figure 7 , the coupled model( Figure 7 "Reconstruction" in) and the original DEM( Figure 7 "MERIT DEM" in) show examples of underwater terrain for three lake boundary types (ellipse, rectangle, circle) during the test period and the validation period. Figure 7 where (a)-(h) show lakes with -1<REi<-0.01, among which (a-d) are valley-type and (e-h) are depression-type; by analogy, (i)-(p) and (q)-(x) are lakes with -0.01<REi<0.01 and 0.01<REi<1, respectively.

[0106] Figure 8 shows the reconstruction results of typical lake underwater terrain on the Qinghai-Tibet Plateau, where (a) is Buruo Co, (b) is Qinghai Lake, (c) is Longmu Co, (d) is Mapam Yumco, (e) is Taruo Co, and (f) is Yanhu. (a)-(b) represent lakes with REi in the range of -1<REi<-0.01, (c) represents a lake with REi in the range of -0.01<REi<0.01, and (d)-(f) represent lakes with REi in the range of 0.01<REi<1. From Figure 8 , it can be seen that the coupled model of the embodiment of the present application successfully reconstructs the underwater terrain of 6 measured lakes on the Qinghai-Tibet Plateau, with an average RMSE of maximum water depth of 3.62 meters and an average relative error (RE) of 6.57%. Compared with the existing mathematical models, the coupled model of the embodiment of the present application is significantly improved in maximum water depth reconstruction, with the RMSE reduced by 80.75% and the RE reduced by 73.41%, demonstrating its capability to capture complex geomorphic features. The analysis of the volume-height curves of the 6 lakes shows that there is strong consistency between the observed data and the reconstructed data, R 2 ranges from 0.75 to 0.96, and NRMSE ranges from 6.24% to 16.17% (see Figure 8 ). The coupled model effectively represents the underwater terrain, including the unimodal and bimodal terrain of Qinghai Lake and Yanhu, as well as the central island in the lake.

[0107] Based on the lake underwater terrain prediction model constructed by the method for constructing a lake underwater terrain prediction model described in any one of the above embodiments, that is, the preset lake underwater terrain prediction model, the embodiment of the present application further provides a lake underwater terrain prediction method, as shown in Figure 9 , the method includes the following steps:

[0108] S901. Extract real target data of real lakes from the existing DEM terrain dataset.

[0109] The real target data includes at least the coordinate and elevation information of each DEM pixel on the real boundary of the real lake, and the coordinate information of each DEM pixel in the area enclosed by the real boundary.

[0110] S902. Input the real target data into a preset mathematical model to obtain mathematical simulation terrain data of the real lake.

[0111] The mathematical model is constructed based on the spatial continuity of the land and underwater topography of the lake. The mathematically simulated terrain data includes the real target data and the simulated elevation information of each DEM pixel in the area surrounded by the real boundary.

[0112] S903. Input the mathematically simulated terrain data into a preset underwater terrain prediction model for lakes, and output the predicted terrain data of the real lake through the underwater terrain prediction model for lakes.

[0113] The construction method of the above-mentioned preset lake underwater topography prediction model can be found in any of the above embodiments describing the construction method of the lake underwater topography prediction model, and will not be repeated here.

[0114] Understandably, when a well-constructed underwater topography prediction model for lakes is actually applied to the prediction stage of real lakes, the measured boundary data of real lakes can often be obtained from existing DEM datasets, meaning they are known. The unknown part is the area enclosed by the boundary, i.e., the underwater topography. This needs to be simulated by a mathematical model, and then the deviation needs to be corrected by the well-constructed underwater topography prediction model for lakes.

[0115] As an example, the water storage capacity of a real lake can be determined based on the predicted topographic data of the real lake output by the underwater topographic prediction model.

[0116] As an example, a well-constructed underwater topography prediction model for lakes can contain multiple sub-models. Real lakes can be classified according to the ellipticity index mentioned above, and mathematical simulation topography data of real lakes can be input into different sub-models for prediction based on different boundary shapes of real lakes. The specific classification and input sub-model determination process can refer to the construction method of the underwater topography prediction model for lakes described in any of the above embodiments, and will not be repeated here.

[0117] Corresponding to the above method embodiments, this application also provides a computer program product, including a computer program, which, when executed by a processor, implements the method for constructing a lake underwater topography prediction model or the lake underwater topography prediction method described in any of the above embodiments.

[0118] This application also provides an electronic device, such as Figure 10 As shown, the electronic device includes:

[0119] Processor 1001;

[0120] Memory 1002 is used to store processor-executable instructions;

[0121] The processor 1001 is configured to implement the method for constructing a lake underwater topography prediction model or the lake underwater topography prediction method described in any of the embodiments above.

[0122] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for constructing a lake underwater topography prediction model or the method for predicting lake underwater topography as described in any of the embodiments above.

[0123] The above description is only a specific embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for constructing a lake underwater topography prediction model, characterized in that, The method includes: From the existing DEM terrain dataset, extract real terrain data for several target terrains, which are terrains with similar topographic structures to lakes and that are not submerged. The target data of the target terrain is extracted from the real terrain data. The target data includes boundary data and data of the area enclosed by the boundary. The boundary data includes at least the coordinate information and elevation information of each DEM pixel on the boundary of the target terrain. The data of the area enclosed by the boundary includes at least the coordinate information of each DEM pixel in the area. The target data is input into a preset mathematical model to obtain mathematically simulated terrain data of the target terrain. The mathematical model is constructed based on the spatial continuity of the land and underwater terrain of the lake. The mathematically simulated terrain data includes the target data and the simulated elevation information of each DEM pixel in the area surrounded by the boundary. The mathematically simulated terrain data is used as a training dataset to train the underwater terrain prediction model of the lake. The underwater terrain prediction model of the lake is trained using the error between the output predicted terrain data and the actual terrain data as a loss function.

2. The method according to claim 1, characterized in that, The lake underwater topography prediction model includes multiple sub-models, and the method further includes: Based on the boundary data, the target terrain is classified into boundaries; The step of inputting the mathematically simulated terrain data into a preset lake underwater terrain prediction model includes: Mathematical simulation terrain data of target terrain belonging to different boundary types are input into different sub-models.

3. The method according to claim 2, characterized in that, The boundary classification of the target terrain includes: Based on the boundary data, the ellipticity index corresponding to the target terrain is determined, and based on the determined ellipticity index, the boundary type to which the target terrain belongs is determined from the preset boundary types; wherein, the preset boundary types include: ellipse, rectangle and circle; the lake underwater terrain prediction model includes a first sub-model, a second sub-model and a third sub-model corresponding to different preset boundary types respectively; The step of inputting mathematically simulated terrain data belonging to different boundary types into different sub-models includes: Based on the boundary type to which the target terrain belongs, the target sub-model into which the mathematical simulation terrain data of the target terrain should be input is determined. The target sub-model is determined among the first sub-model, the second sub-model, and the third sub-model.

4. The method according to claim 1, characterized in that, The target terrain includes: depression terrain, and / or, valley terrain; The extraction of real terrain data for several target terrain features includes: When the target terrain is a river valley, the original terrain data of the river valley is extracted; The lowest position of the valley topography is determined based on the original topographic data, and the target area of ​​the valley topography is segmented based on the lowest position. The target area is the area segmented by a dividing plane that passes through the lowest position and is perpendicular to the direction of the valley topography. The target area is a non-flat area on one side of the valley topography. The target area and its mirror image are stitched together to obtain a stitched result, which is then determined as the true terrain data of the valley.

5. The method according to claim 1, characterized in that, The method further includes: Acquire measured topographic data of several existing real lakes, wherein the measured topographic data is obtained by actually measuring the topography of the real lakes; The measured target data of the real lake is extracted from the measured terrain data; the measured target data includes the measured boundary data and the data of the area enclosed by the measured boundary. The measured target data is input into the mathematical model to obtain the mathematically simulated terrain data of the real lake; The mathematical simulation topographic data of the real lake is input into the trained underwater topographic prediction model of the lake to obtain the predicted topographic data of the real lake. Based on the error between the predicted terrain data and the measured terrain data of the actual lake, the parameters of the trained underwater terrain prediction model of the lake are optimized to obtain the optimized underwater terrain prediction model of the lake.

6. The method according to claim 1, characterized in that... Before inputting the target data into the preset mathematical model, the method further includes: In the target data, the initial elevation values ​​of each DEM pixel in the area enclosed by the boundary of the target terrain are filled with zeros to obtain the preprocessed target data. The step of inputting the target data into a preset mathematical model includes: The preprocessed target data is input into the mathematical model.

7. The method according to claim 1, characterized in that, The underwater topography prediction model for the lake is the Enhanced Depth Super-Resolution Network (EDSR) model.

8. A method for predicting underwater topography in lakes, characterized in that, The method includes: In the existing DEM terrain dataset, real target data of real lakes is extracted. The real target data includes at least the coordinate and elevation information of each DEM pixel on the real boundary of the real lake, as well as the coordinate information of each DEM pixel in the area surrounded by the real boundary. The real target data is input into a preset mathematical model to obtain mathematically simulated terrain data of the real lake. The mathematical model is constructed based on the spatial continuity of the lake's land and underwater terrain. The mathematically simulated terrain data includes the real target data and the simulated elevation information of each DEM pixel in the area surrounded by the real boundary. The mathematically simulated terrain data is input into the lake underwater terrain prediction model in any one of claims 1 to 7, and the predicted terrain data of the real lake is output through the lake underwater terrain prediction model.

9. The method according to claim 8, characterized in that, The method further includes: Based on the predicted terrain data, the actual water storage of the lake is determined.

10. A computer program product comprising a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 9.

11. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to implement the method of any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 9.

Images