A method for constructing virtual karst landform scenes based on DEM and deep learning

Abstract

The invention discloses a method for constructing a virtual karst landform scene based on DEM and deep learning. The method comprises the following steps: firstly extracting saddle points, mountain top points and slope foot lines of a karst landform DEM; then converting terrain feature elements into three-dimensional terrain feature elements in a grayscale value range of 0 to 255, merging the above feature elements into an RGB three-band image, cropping them into terrain feature samples of a size of 256×256, and splicing them with a normalized DEM of a corresponding position to generate a training sample set; then using the training sample set to train a CGANs network; finally, hand-drawing three-dimensional saddle points, mountain top points and slope foot lines according to development needs, synthesizing them into three-dimensional terrain feature elements, and inputting them into a trained karst landform construction model to construct a virtual karst landform scene. The method has the advantages of low cost and less time consumption, and provides a virtual karst landform scene with high authenticity and meeting geographical laws for virtual scene application fields such as games and movies.

Description

Technical Field

[0001] The present invention relates to the field of three-dimensional modeling application technology, and in particular to a method for constructing a virtual karst landform scene based on DEM and deep learning. Background Art

[0002] Virtual terrain is a fictitious surface form depicted based on limited information under the guidance of existing spatial cognitive knowledge. Virtual terrain is widely used in virtual scenes of games and movies, and is an important factor affecting the landscape and experience of virtual scenes. Karst landform is a commonly used landform type in virtual scenes of games and movies due to its unique landform and high landscape value.

[0003] At present, in the construction of virtual geographical scenes such as games and movies, the generation of virtual terrain is mainly controlled by fractal index. Although the terrain constructed using fractal parameters has infinite details, the spatial relationship is difficult to guarantee and does not conform to the visual spatial structure of geology. In addition, karst landforms are relatively complex, and collecting terrain data through traditional field measurement methods will consume a lot of manpower and material resources. Therefore, it is necessary to propose a method that can quickly construct a virtual karst landform scene with high realism and meet geographical laws. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a method for constructing a virtual karst landform scene based on DEM and deep learning, comprising the following steps:

[0005] S1. Obtain DEM data of karst landforms, perform depression filling, flow direction calculation and flow calculation on DEM, and then select flow threshold to extract valley lines;

[0006] S2, using the maximum elevation value to subtract from the original DEM to obtain the inverse terrain data, and then perform depression filling, flow direction calculation and flow calculation, and then select the flow threshold to extract the ridge line;

[0007] S3, extracting saddle points by determining the intersection of the ridge line and the valley line;

[0008] S4, use the maximum neighborhood window analysis method to extract the mountain top points in the study area;

[0009] S5. Calculate the positive and negative openings of the DEM terrain in the study area, subtract the two, segment the difference results based on the optimal segmentation threshold to obtain a binary image, and extract its boundary, which is the slope foot line;

[0010] S6, reclassifying the saddle point, mountain top point and slope foot line data obtained in steps S1 to S5 into two values ​​of 0-1, that is, the background value is 0 and the foreground value is 1;

[0011] S7, normalize the DEM and convert the DEM grayscale value to the range of 0 to 255;

[0012] S8, multiplying the binary saddle point, mountain top point and slope foot line obtained in step S6 with the DEM obtained in the previous step to obtain three-dimensional terrain feature elements endowed with relative elevation features, namely, three-dimensional saddle points, mountain top points and slope foot lines with grayscale values ​​in the range of 0 to 255;

[0013] S9, synthesize the three-dimensional saddle points, mountain top points and slope foot lines obtained in the previous step according to the three RGB bands to obtain the three-dimensional terrain feature elements of the study area, forming a three-dimensional terrain feature training set;

[0014] S10, using a sliding window to crop the three-dimensional terrain feature training set obtained in the previous step and the DEM data obtained in step S7 into an image of size 256×256;

[0015] S11, horizontally splicing the three-dimensional terrain feature elements obtained in the previous step and the DEM image of the corresponding position to obtain a three-dimensional terrain feature training sample set;

[0016] S12, using the three-dimensional terrain element training sample set obtained in the previous step to train the CGANs network, so that the CGANs network learns the karst landform characteristics and their mapping relationship with the saddle point, the mountain top point and the slope foot line;

[0017] S13. According to actual development needs, hand-draw the 3D saddle point, mountain top point and slope foot line, and synthesize them according to the three RGB bands;

[0018] S14, input the hand-drawn saddle points, mountain top points and slope foot lines into the CGANs network trained in step S12 to generate the required virtual karst landform scene.

[0019] The technical solution further defined in the present invention is:

[0020] Furthermore, in steps S1 and S2, the D8 single flow direction model is used to perform flow direction calculation.

[0021] As described above, in the method for constructing a virtual karst landform scene based on DEM and deep learning, in steps S1 and S2, the result of the flow direction calculation is the cumulative flow, and the flow threshold is set to 2000.

[0022] As described above, a method for constructing a virtual karst landform scene based on DEM and deep learning is described. In step S4, a maximum neighborhood window analysis is performed on the DEM, the window size is set to 11×11, the step size is set to 1, the statistical type is set to maximum value, and raster data containing the maximum value point in each analysis window is generated; the generated raster data is differenced with the original DEM, and the point with a difference result of 0 is extracted as the mountain top point.

[0023] As described above, in a method for constructing a virtual karst landform scene based on DEM and deep learning, in step S5, the calculation method of the positive and negative openings of the DEM terrain in the study area is specifically as follows: eight rays are drawn from the central pixel to eight directions of 0°, 45°, 90°, 135°, 180°, 225°, 270° and 315°, and pixels on the eight rays whose distance from the central pixel is less than or equal to the analysis radius are selected to participate in the calculation;

[0024] By calculating the maximum positive altitude angle and the maximum negative altitude angle within the analysis radius in each direction, the minimum zenith angle and the minimum nadir angle within the analysis radius in each direction are obtained. The average of the minimum zenith angles in the eight directions is the positive opening of the position, and the average of the minimum nadir angles in the eight directions is the negative opening. The specific formula is as follows:

[0025]

[0026] in, represents the positive opening of the point, Indicates the negative opening of the point, , Respectively represent the minimum zenith angle and minimum nadir angle within the analysis radius in each direction, where n=0, 45, 90, 135, 180, 225, 270, 315, and ; It represents the angle between the center point in the n direction and the line connecting the pixel with the highest gray value within the analysis radius and the horizontal direction. It represents the angle between the horizontal direction and the line connecting the center point and the pixel with the lowest gray value within the analysis radius in the n direction.

[0027] As described above, a method for constructing a virtual karst landform scene based on DEM and deep learning is described. In step S5, the positive and negative openings of the DEM terrain in the study area are subjected to difference calculation, and the optimal segmentation threshold is set to 9°. The difference result greater than the optimal segmentation threshold is positive terrain, and vice versa. The positive and negative opening difference results of the terrain are reclassified into a binary image according to the optimal segmentation threshold, and the positive terrain area is assigned a value of 1 and the negative terrain area is assigned a value of 0. Subsequently, the positive terrain area is converted into vector data, and the slope foot line is obtained by extracting the boundary of the positive terrain area.

[0028] As described above, in the method for constructing a virtual karst landform scene based on DEM and deep learning, in step S7, the DEM is normalized to convert the DEM grayscale value into the range of 0 to 255. The normalization formula is:

[0029]

[0030] Among them, O(i,j) represents the output grayscale value, I(i,j) represents the grayscale value of the i-th row and j-th column of the DEM image, and I min Indicates the minimum gray value of the DEM image, I max Indicates the maximum grayscale value of the DEM image.

[0031] As described above, in a method for constructing a virtual karst landform scene based on DEM and deep learning, in step S10, the sliding window size is set to 256×256, the step size is set to 64 pixels, and sliding starts from the upper left corner of the DEM image in the horizontal direction. After traversing a whole row horizontally, the window moves a step distance in the vertical direction and returns to the leftmost side of the image to continue traversing until the entire image is traversed.

[0032] As described above, in a method for constructing a virtual karst landform scene based on DEM and deep learning, in step S12, the generator of the CGANs network adopts a U-Net structure, the convolution kernel size of the generator and discriminator of the CGANs network is 4×4, and the step size is set to 2.

[0033] As described above, in a method for constructing a virtual karst landform scene based on DEM and deep learning, in step S12, the CGANs network uses the Adam optimizer, the learning rate is set to 0.0002, and the momentum is 0.5.

[0034] The beneficial effects of the present invention are:

[0035] In the present invention, firstly, the existing 5-meter high-resolution DEM is used to extract three-dimensional terrain feature elements, deep learning sample image pairs are produced, and a conditional generative adversarial network model is trained; then, the hand-drawn three-dimensional terrain feature elements are input into the trained high-precision model to generate a virtual karst landform scene, so that a virtual karst landform scene can be quickly constructed in a low-cost and low-time manner, providing a virtual karst landform scene with high authenticity and that meets geographical laws for virtual scene application fields such as games, movies and television. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0037] Figure 2 It is a regional map of the research sample area and a schematic diagram of 5-meter DEM cut samples in an embodiment of the present invention;

[0038] Figure 3 Schematic diagram of the calculation principle of the positive and negative opening of terrain in the embodiment of the present invention;

[0039] Figure 4 A flow chart for preparing a sample of three-dimensional terrain feature elements in an embodiment of the present invention;

[0040] Figure 5 Schematic diagram of the working principle of the CGANs model in an embodiment of the present invention;

[0041] Figure 6 A schematic diagram of the hill shadow results constructed by outputting images of the model with different training times in an embodiment of the present invention;

[0042] Figure 7 This is a schematic diagram of the construction result of the virtual karst landform scene in an embodiment of the present invention. DETAILED DESCRIPTION

[0043] This embodiment provides a method for constructing a virtual karst landform scene based on DEM and deep learning. Figure 1 As shown, the following steps are included:

[0044] S1. Obtain the digital elevation model (DEM) data of karst landforms, perform depression filling, flow direction calculation and flow volume calculation on the DEM, and then select the flow threshold to extract the valley line.

[0045] S2. Subtract the maximum elevation from the original DEM to obtain the inverse terrain data, then perform depression filling, flow direction calculation, and flow calculation, and then select the flow threshold to extract the ridge line.

[0046] In steps S1 and S2, the D8 single flow direction model is used to perform flow direction calculation. The result of the flow direction calculation is the cumulative flow, and the flow threshold is set to 2000.

[0047] S3. Extract the saddle point by determining the intersection of the ridge line and the valley line.

[0048] S4. Use the maximum neighborhood window analysis method to extract the mountain top points in the study area; perform maximum neighborhood window analysis on the DEM, set the window size to 11×11, the step size to 1, and the statistical type to maximum value, and generate raster data containing the maximum value points in each analysis window; perform a difference operation on the generated raster data and the original DEM, and extract the points with a difference result of 0 as the mountain top points.

[0049] S5. Calculate the positive and negative openings of the DEM terrain in the study area, subtract the two, segment the difference results based on the optimal segmentation threshold to obtain a binary image, and extract its boundary, which is the slope foot line.

[0050] The calculation method of the positive and negative openings of the DEM terrain in the study area is as follows: eight rays are drawn from the central pixel to eight directions, namely 0°, 45°, 90°, 135°, 180°, 225°, 270° and 315°, and the pixels on these eight rays whose distance from the central pixel is less than or equal to the analysis radius are selected to participate in the calculation.

[0051] By calculating the maximum positive altitude angle and the maximum negative altitude angle within the analysis radius in each direction, the minimum zenith angle and the minimum nadir angle within the analysis radius in each direction are obtained. The average of the minimum zenith angles in the eight directions is the positive opening of the position, and the average of the minimum nadir angles in the eight directions is the negative opening. The specific formula is as follows:

[0052]

[0053] in, represents the positive opening of the point, Indicates the negative opening of the point, , Respectively represent the minimum zenith angle and minimum nadir angle within the analysis radius in each direction, where n=0, 45, 90, 135, 180, 225, 270, 315, and ; It represents the angle between the center point in the n direction and the line connecting the pixel with the highest gray value within the analysis radius and the horizontal direction. It represents the angle between the horizontal direction and the line connecting the center point and the pixel with the lowest gray value within the analysis radius in the n direction.

[0054] The difference operation of the positive and negative openings of the DEM terrain in the study area was performed, and the optimal segmentation threshold was set to 9°. The difference result greater than the optimal segmentation threshold was positive terrain, otherwise it was negative terrain. The positive and negative terrain opening difference results were reclassified into a binary image according to the optimal segmentation threshold, and the positive terrain area was assigned a value of 1 and the negative terrain area was assigned a value of 0. The positive terrain area was then converted into vector data, and the slope foot line was obtained by extracting the boundary of the positive terrain area.

[0055] S6. Reclassify the saddle point, mountain top point and slope foot line data obtained in steps S1 to S5 into two values ​​of 0-1, that is, the background value is 0 and the foreground value is 1.

[0056] S7. Normalize the DEM and convert the DEM grayscale value to the range of 0 to 255. The normalization formula is:

[0057]

[0058] Among them, O(i,j) represents the output grayscale value, I(i,j) represents the grayscale value of the i-th row and j-th column of the DEM image, and Imin Indicates the minimum gray value of the DEM image, I max Indicates the maximum grayscale value of the DEM image.

[0059] S8. Multiply the binary saddle points, mountain top points and slope foot lines obtained in step S6 with the DEM obtained in the previous step to obtain three-dimensional terrain feature elements endowed with relative elevation features, i.e., three-dimensional saddle points, mountain top points and slope foot lines with grayscale values ​​in the range of 0 to 255.

[0060] S9. The three-dimensional saddle points, mountain top points and slope foot lines obtained in the previous step are synthesized according to the three RGB bands to obtain the three-dimensional terrain feature elements of the study area to form a three-dimensional terrain feature training set.

[0061] S10. The three-dimensional terrain feature training set obtained in the previous step and the DEM data obtained in step S7 are cropped using a sliding window to a 256×256 image; the sliding window size is set to 256×256, the step size is set to 64 pixels, and the sliding starts from the upper left corner of the DEM image in the horizontal direction. After traversing a whole row horizontally, the window moves a step distance in the vertical direction and returns to the leftmost side of the image to continue traversing until the entire image is traversed.

[0062] S11, horizontally splicing the three-dimensional terrain feature elements obtained in the previous step and the DEM image of the corresponding position to obtain a three-dimensional terrain feature training sample set.

[0063] S12. Use the three-dimensional terrain element training sample set obtained in the previous step to train the Conditional Generative Adversarial Networks (CGANs) to let the CGANs network learn the karst landform characteristics and their mapping relationship with the saddle point, mountain top point and slope foot line; the generator of the CGANs network adopts the U-Net structure, and the convolution kernel size of the generator and discriminator of the CGANs network is 4×4, and the step size is set to 2; the CGANs network uses the Adam optimizer, the learning rate is set to 0.0002, and the momentum is 0.5.

[0064] S13. According to actual development needs, hand-draw the three-dimensional saddle point, mountain top point and slope foot line, and synthesize them according to the three RGB bands.

[0065] S14, input the hand-drawn saddle points, mountain top points and slope foot lines into the CGANs network trained in step S12 to generate the required virtual karst landform scene.

[0066] In this embodiment, for example Figure 2The embodiment shown studies the sample area and 5-meter high-resolution DEM data samples of the karst landform area in Guilin, Guangxi. First, the 5-meter high-resolution DEM of the karst landform is aligned with the geographic coordinate system and the projection coordinate system, and the DEM data is clipped to the scope of the study area.

[0067] Then, the DEM is sequentially subjected to depression filling, flow direction calculation, and flow confluence calculation. The cumulative flow threshold (set to 2000) is selected to extract the valley line, and the maximum elevation value is subtracted from the original DEM to obtain the inverse terrain data. The inverse terrain data is sequentially subjected to depression filling, flow direction calculation, and flow confluence calculation. The cumulative flow threshold is selected as 2000 to extract the ridge line, and then the saddle point is extracted based on the intersection of the ridge line and the valley line.

[0068] Then, the maximum neighborhood window analysis method was used to extract the mountain top points in the study area. The window size was set to 11×11, the step size was set to 1, and the statistical type was set to maximum value. The raster data containing the maximum value points in each analysis window were generated. The generated raster data was subtracted from the original DEM, and the mountain top points were extracted by judging that the difference result was 0.

[0069] Then, if Figure 3 As shown, the positive and negative openings of the terrain are calculated, and the difference operation is performed on the positive and negative openings of the terrain. The optimal segmentation threshold is set to 9°. The difference result greater than the optimal segmentation threshold is positive terrain, otherwise it is negative terrain. According to the optimal segmentation threshold, the positive and negative opening difference results of the terrain are reclassified into a binary image, the positive terrain area is assigned a value of 1, and the negative terrain area is assigned a value of 0. Then the positive terrain area is converted into vector data, and the slope foot line is obtained by extracting the boundary of the positive terrain area.

[0070] The extracted saddle points, mountain top points, and slope foot lines are then reclassified into 0-1 binary data and multiplied with the DEM data normalized to the range of 0-255 to obtain three-dimensional terrain feature elements with relative elevation data.

[0071] Then, the three-dimensional terrain feature elements of the saddle point, the mountain top point, and the slope foot line are merged according to the RGB three bands, and the merged feature element image and the normalized DEM are cropped using a sliding window. The sliding window size is set to 256×256 pixels, the step size is 64 pixels, and the feature element image and DEM image are segmented into images of 256×256 pixels (1.28×1.28 km). The cropped three-dimensional terrain feature samples are horizontally spliced ​​with the normalized DEM at the corresponding position to obtain the training sample set, as shown in Figure 4 shown.

[0072] Then the obtained training sample set is input into the CGANs model. The working principle of CGANs is as follows Figure 5 shown; according to Figure 6 As shown in the figure, as the number of training times increases, the performance of the model-generated images gradually improves.

[0073] Then, according to the development needs of the game scene, hand-draw the three-dimensional saddle point, mountain top point, and slope foot line, and synthesize them according to the three RGB bands.

[0074] Finally, the hand-drawn 3D terrain feature elements are input into the trained model to construct a virtual karst landform scene. Figure 7 As shown, the method of this embodiment quickly constructs a virtual karst landform scene in a low-cost and short-time manner, providing a virtual karst landform scene with high realism and in line with geographical laws for virtual scene application fields such as games, movies and television.

[0075] In addition to the above embodiments, the present invention may also have other implementation modes. Any technical solution formed by equivalent replacement or equivalent transformation falls within the protection scope required by the present invention.

Claims

1. A method for constructing a virtual karst landform scene based on DEM and deep learning, characterized by: The following steps are involved: S1. Obtain DEM data of karst landforms, perform depression filling, flow direction calculation and flow calculation on DEM, and then select flow threshold to extract valley lines; S2, using the maximum elevation value to subtract from the original DEM to obtain the inverse terrain data, and then perform depression filling, flow direction calculation and flow calculation, and then select the flow threshold to extract the ridge line; S3, extracting saddle points by determining the intersection of the ridge line and the valley line; S4, use the maximum neighborhood window analysis method to extract the mountain top points in the study area; S5. Calculate the positive and negative openings of the DEM terrain in the study area, subtract the two, segment the difference results based on the optimal segmentation threshold to obtain a binary image, and extract its boundary, which is the slope foot line; S6, reclassifying the saddle point, mountain top point and slope foot line data obtained in steps S1 to S5 into two values ​​of 0-1, that is, the background value is 0 and the foreground value is 1; S7, normalize the DEM and convert the gray value of the DEM into the range of 0 to 255; S8, multiplying the binary saddle point, mountain top point and slope foot line obtained in step S6 with the DEM obtained in the previous step to obtain three-dimensional terrain feature elements endowed with relative elevation features, namely, three-dimensional saddle points, mountain top points and slope foot lines with grayscale values ​​in the range of 0 to 255; S9, synthesize the three-dimensional saddle points, mountain top points and slope foot lines obtained in the previous step according to the three RGB bands to obtain the three-dimensional terrain feature elements of the study area, forming a three-dimensional terrain feature training set; S10, using a sliding window to crop the three-dimensional terrain feature training set obtained in the previous step and the DEM data obtained in step S7 into an image of size 256×256; S11, horizontally splicing the three-dimensional terrain feature elements obtained in the previous step and the DEM image of the corresponding position to obtain a three-dimensional terrain feature training sample set; S12, using the three-dimensional terrain element training sample set obtained in the previous step to train the CGANs network, so that the CGANs network learns the karst landform characteristics and their mapping relationship with the saddle point, the mountain top point and the slope foot line; S13. According to actual development needs, hand-draw the 3D saddle point, mountain top point and slope foot line, and synthesize them according to the three RGB bands; S14, inputting the hand-drawn saddle points, mountain top points and slope foot lines into the CGANs network trained in step S12 to generate the required virtual karst landform scene; In step S5, the calculation method of the positive and negative opening of the DEM terrain in the study area is specifically as follows: eight rays are drawn from the central pixel to eight directions of 0°, 45°, 90°, 135°, 180°, 225°, 270° and 315°, and the pixels on these eight rays whose distance from the central pixel is less than or equal to the analysis radius are selected to participate in the calculation; By calculating the maximum positive altitude angle and the maximum negative altitude angle within the analysis radius in each direction, the minimum zenith angle and the minimum nadir angle within the analysis radius in each direction are obtained. The average of the minimum zenith angles in the eight directions is the positive opening of the position, and the average of the minimum nadir angles in the eight directions is the negative opening. The specific formula is as follows: in, represents the positive opening of the point, Indicates the negative opening of the point, , Respectively represent the minimum zenith angle and minimum nadir angle within the analysis radius in each direction, where n=0, 45, 90, 135, 180, 225, 270, 315, and ; It represents the angle between the center point in the n direction and the line connecting the pixel with the highest gray value within the analysis radius and the horizontal direction. It represents the angle between the center point in the n direction and the line connecting the pixel with the lowest gray value within the analysis radius and the horizontal direction; In step S5, the difference operation is performed on the positive and negative openings of the DEM terrain in the study area, and the optimal segmentation threshold is set to 9°. The difference result greater than the optimal segmentation threshold is positive terrain, otherwise it is negative terrain; according to the optimal segmentation threshold, the positive and negative opening difference results of the terrain are reclassified into a binary image, the positive terrain area is assigned a value of 1, and the negative terrain area is assigned a value of 0, and then the positive terrain area is converted into vector data, and the slope foot line is obtained by extracting the boundary of the positive terrain area.

2. The method for constructing a virtual karst landform scene based on DEM and deep learning according to claim 1, characterized in that: In the steps S1 and S2, the D8 single flow direction model is used to perform flow direction calculation.

3. The method for constructing a virtual karst landform scene based on DEM and deep learning according to claim 2, characterized in that: In the steps S1 and S2, the result of the flow direction calculation is the cumulative flow, and the flow threshold is set to 2000.

4. The method for constructing a virtual karst landform scene based on DEM and deep learning according to claim 1, characterized in that: In step S4, a maximum neighborhood window analysis is performed on the DEM, the window size is set to 11×11, the step size is set to 1, the statistical type is set to maximum value, and raster data containing the maximum value point in each analysis window is generated; the generated raster data is differenced with the original DEM, and the point with a difference result of 0 is extracted as the mountain top point.

5. The method for constructing a virtual karst landform scene based on DEM and deep learning according to claim 1, characterized in that: In step S7, the DEM is normalized to convert the grayscale value of the DEM into a range of 0 to 255. The normalization formula is: Among them, O(i,j) represents the output grayscale value, I(i,j) represents the grayscale value of the i-th row and j-th column of the DEM image, and I min Indicates the minimum gray value of the DEM image, I max Indicates the maximum grayscale value of the DEM image.

6. The method for constructing a virtual karst landform scene based on DEM and deep learning according to claim 1, characterized in that: In step S10, the sliding window size is set to 256×256, the step size is set to 64 pixels, and the sliding starts from the upper left corner of the DEM image in the horizontal direction. After traversing a whole row horizontally, the window moves a step distance in the vertical direction and returns to the leftmost side of the image to continue traversing until the entire image is traversed.

7. The method for constructing a virtual karst landform scene based on DEM and deep learning according to claim 1, characterized in that: In step S12, the generator of the CGANs network adopts a U-Net structure, the convolution kernel size of the generator and the discriminator of the CGANs network adopts 4×4, and the step size is set to 2.

8. The method for constructing a virtual karst landform scene based on DEM and deep learning according to claim 7, characterized in that: In step S12, the CGANs network uses the Adam optimizer, the learning rate is set to 0.0002, and the momentum is 0.5.

Images