A Remote Sensing Map Analysis Method Based on Elevation Information

Abstract

This invention relates to a remote sensing map analysis method based on elevation information. First, the terrain is segmented into pixels, and the height of each pixel is determined. When two adjacent pixels have the same height, they are determined to be the same terrain. Then, the terrain of each pixel and the prediction loss of each pixel are determined. Finally, the loss is fed back to update the parameters. This invention uses elevation information as an auxiliary input and utilizes the continuity of terrain as an image segmentation method, which can significantly improve the accuracy of terrain judgment.

Description

TECHNICAL FIELD

[0001] The present application relates to a remote sensing map analysis method, in particular to a remote sensing map analysis method based on elevation information. BACKGROUND

[0002] In recent years, with the development of unmanned aerial vehicle technology, aerial photography data taken by unmanned aerial vehicles are used to analyze rural and urban topography, which has gradually become an important tool and means for rural and urban construction. In this process, the analysis of urban roads has quite a number of technical means and research experiences, but the analysis technology of aerial photography data related to rural areas is still in the vacancy stage. Common semantic segmentation methods are usually directly based on RGB data, which is prone to visual ambiguity due to various reasons when facing different lightings, occlusions and other situations, resulting in that the segmentation model cannot achieve the optimal experimental effect. How to reduce the terrain misjudgment problem caused by occlusion and other reasons and automatically analyze and identify key rivers, houses and roads in rural areas has become a problem to be solved. SUMMARY

[0003] In view of the above problems, the present application provides a remote sensing map analysis method based on elevation information, which comprises the following steps:

[0004] S1: performing pixel segmentation on the terrain and judging the height of each pixel;

[0005] S2: judging whether the heights of two adjacent pixels are the same, if yes, judging as the same terrain, if not, judging as different terrains;

[0006] S3: determining the terrain of each pixel;

[0007] S4: determining the prediction loss of each pixel;

[0008] S5: loss back propagation and parameter updating.

[0009] Further, step S2 comprises the following steps:

[0010] S21: selecting any two pixels, one of which is set as (x1, y1, h1) and the second pixel is set as (x2, y2, h2), wherein (x, y) is the spatial coordinate and h represents the height of the pixel;

[0011] S22: judging whether the two pixels are adjacent points, if yes, proceeding to step S23, if not, returning to step S21;

[0012] S23: Determine whether h1-h2≤λ is satisfied. If yes, the two pixels belong to the same terrain. If no, the two pixels do not belong to the same terrain. λ is the set threshold.

[0013] Further, step S22 includes the following steps: when y1 = y2, if x1 - x2 = 1, then the two pixels are determined to be neighboring points; when x1 = x2, if y1 - y2 = 1, then the two pixels are determined to be neighboring points.

[0014] Furthermore, step S3 also includes the following steps:

[0015] S31: Merge the RGB image X and the elevation image H into the input image Z;

[0016] S32: The input image Z is input to the encoder, passed through a dilated convolutional network, then fed into a 1×1 convolutional network, and then input to the decoder for 4x upsampling to obtain the first feature output. The input image Z is then passed through a dilated convolutional network and input to the decoder. In the decoder, it passes through a 1×1 convolutional network to obtain the second feature output. The first feature output and the second feature output are merged and then passed through a 3×3 convolutional network for 4x upsampling to finally output the segmented image P.

[0017] Further, step S31 is linear fusion: Z = αX + (1-α)H, where Z is the image fused with RGB information and elevation information, X is the image with RGB information, H is the image with elevation information, and α is a constant coefficient.

[0018] Further, step S31 is convolutional fusion: Z = W1X + W2H, where Z is an image that fuses RGB information and elevation information, and W1 and W2 are 1×1 convolutional networks.

[0019] Further, step S31 is a linear-to-convolutional fusion: Z = W1(αX + (1-α)H), where Z is an image that fuses RGB information and elevation information, X is an image with RGB information, H is an image with elevation information, α is a constant coefficient, and W1 is a 1×1 convolutional network.

[0020] Furthermore, the dilated convolutional network in step S32 includes a 1×1 convolutional network, a 3×3 convolutional network with a spatial stride of 6, a 3×3 convolutional network with a spatial stride of 12, a 3×3 convolutional network with a spatial stride of 18, and a pooling network.

[0021] Furthermore, in step S4, the predicted loss is Loss = (LP). 2 Where Loss is the prediction loss, L is the label, and P is the final output segmented image P, P = D(E(Z,W)). E ),WD ), where D is the decoder, E is the encoder, and W D For all parameters of the decoder, W E The encoder provides all the parameters, and the final output segmented image P contains a prediction of the terrain to which each pixel location belongs.

[0022] Further, step S5 calculates the gradient for each parameter of the encoder and decoder based on the currently calculated loss, and updates the gradient.

[0023] This invention provides a remote sensing map analysis method based on elevation information. First, the terrain is segmented into pixels, and the height of each pixel is determined. When two adjacent pixels have the same height, they are determined to be the same terrain. Then, the terrain of each pixel and the prediction loss of each pixel are determined. Finally, the loss is fed back to update the parameters. This invention uses elevation information as an auxiliary input and utilizes the continuity of terrain as an image segmentation method, which can significantly improve the accuracy of terrain judgment. Attached Figure Description

[0024] Figure 1 This is a schematic diagram illustrating the method for determining the terrain of each pixel according to the present invention.

[0025] Explanation of reference numerals in the attached figures

[0026] 1 encoder, 2 decoders, 11 dilated convolutional networks. Detailed Implementation

[0027] To provide a further understanding of the purpose, structure, features, and functions of the present invention, detailed descriptions are provided below with reference to specific embodiments.

[0028] To address the above problems, this invention provides a remote sensing map analysis method based on elevation information, which includes the following steps:

[0029] S1: Segment the terrain into pixels and determine the height of each pixel;

[0030] S2: Determine whether two adjacent pixels have the same height. If they do, they are considered to be the same terrain; otherwise, they are considered to be different terrains.

[0031] S3: Determine the terrain for each pixel;

[0032] S4: Determine the prediction loss for each pixel;

[0033] S5: Loss feedback, parameter update;

[0034] First, the terrain is segmented into pixels, and the height of each pixel is determined. When two adjacent pixels have the same height, they are considered to be the same terrain. Then, the terrain of each pixel and the prediction loss of each pixel are determined. Finally, the loss is fed back to update the parameters.

[0035] Since the collected data contains elevation information, and the terrain has a certain degree of continuity, meaning that if two locations have almost identical elevation information and are very close in location, it can be assumed that these two locations belong to the same terrain type. The above step S2 includes the following steps:

[0036] S21: Select any two pixels, set one pixel as (x1, y1, h1) and the second pixel as (x2, y2, h2), where (x, y) are spatial coordinates and h represents the height of the pixel;

[0037] S22: Determine whether the two pixels are adjacent points. If yes, proceed to step S23; otherwise, return to step S21. When y1 = y2, if x1 - x2 = 1, then the two pixels are adjacent points. When x1 = x2, if y1 - y2 = 1, then the two pixels are adjacent points.

[0038] S23: Determine whether h1-h2≤λ is satisfied. If yes, the two pixels belong to the same terrain. If no, the two pixels do not belong to the same terrain. λ is the set threshold, usually set to λ=0.5m. When multiple pixels belong to the same terrain, we can use the existing terrain prior information to determine the terrain of other multiple pixels with the same terrain.

[0039] Combining elevation information with terrain continuity allows for a better determination of whether the current pixel and its neighboring pixels belong to the same terrain. For pixels whose terrain is unknown, it is also necessary to determine the terrain of the current pixel. (See [link to documentation]). Figure 1 , Figure 1 This is a schematic diagram of the method for determining the terrain of each pixel according to the present invention. Step S3 above further includes the following steps:

[0040] S31: Merge the RGB image X and the elevation image H into the input image Z; where the RGB image X is a color image taken by the drone, and the elevation image H is an image containing altitude information, and the RGB image X and the elevation image H have the same resolution.

[0041] S32: The input image Z is input to encoder 1, passes through dilated convolutional network 11, is then fed into a 1×1 convolutional network, and is then input to decoder for 4x upsampling to obtain the first feature output. The input image Z passes through dilated convolutional network and is then input to decoder 2. In decoder 2, it passes through a 1×1 convolutional network to obtain the second feature output. The first feature output and the second feature output are merged and then passed through a 3×3 convolutional network for 4x upsampling to finally output segmented image P.

[0042] In step S32, the dilated convolutional network includes a 1×1 convolutional network, a 3×3 convolutional network with a spatial stride of 6, a 3×3 convolutional network with a spatial stride of 12, a 3×3 convolutional network with a spatial stride of 18, and a pooling network. The input image Z passes through the 1×1 convolutional network to output the first feature map Z1, through the 3×3 convolutional network with a spatial stride of 6 to output the second feature map Z2, through the 3×3 convolutional network with a spatial stride of 12 to output the third feature map Z3, through the 3×3 convolutional network with a spatial stride of 18 to output the fourth feature map Z4, and through the pooling network to output the fifth feature map Z5. Finally, the first feature map Z1, the second feature map Z2, the third feature map Z3, the fourth feature map Z4, and the fifth feature map Z5 are merged.

[0043] There are several ways to merge an RGB image X with an elevation image H into an input image Z, as follows:

[0044] The first type is linear fusion: Z = αX + (1-α)H, where Z is the input image, which is an image that integrates RGB information and elevation information, X is an image with RGB information, H is an image with elevation information, and α is a constant coefficient.

[0045] The second type is convolutional fusion: Z = W1X + W2H, where Z is the input image, which is an image that fuses RGB information and elevation information, and W1 and W2 are 1×1 convolutional networks.

[0046] The third type is a fusion of linear and convolutional methods: Z = W1(αX + (1-α)H), where Z is the input image, which is an image that combines RGB information and elevation information, X is an image with RGB information, H is an image with elevation information, α is a constant coefficient, and W1 is a 1×1 convolutional network.

[0047] Furthermore, in step S4, the predicted loss is Loss = (LP). 2 Where Loss is the prediction loss, L is the label, and P is the final output segmented image P, P = D(E(Z,W)). E ),W D ), where D is the decoder, E is the encoder, and W D For all parameters of the decoder, WE The encoder provides all the parameters, and the final output segmented image P contains a prediction of the terrain to which each pixel location belongs.

[0048] Further, step S5 calculates the gradient for each parameter of the encoder and decoder based on the currently calculated loss, and updates the gradient.

[0049] This invention provides a remote sensing map analysis method based on elevation information. First, the terrain is segmented into pixels, and the height of each pixel is determined. When two adjacent pixels have the same height, they are determined to be the same terrain. Then, the terrain of each pixel and the prediction loss of each pixel are determined. Finally, the loss is fed back to update the parameters. This invention uses elevation information as an auxiliary input and utilizes the continuity of terrain as an image segmentation method, which can significantly improve the accuracy of terrain judgment.

[0050] The present invention has been described in the above-described embodiments; however, these embodiments are merely examples for implementing the present invention. It must be noted that the disclosed embodiments do not limit the scope of the present invention. Conversely, any modifications and refinements made without departing from the spirit and scope of the present invention are within the scope of patent protection of the present invention.

Claims

1. A remote sensing map analysis method based on elevation information, characterized by, The method comprises the following steps: S1: pixel segmentation is performed on the terrain, and the height of each pixel is determined; S2: it is determined whether the heights of two adjacent pixels are the same, if yes, it is determined that they belong to the same terrain, and if not, it is determined that they belong to different terrains; S3: if it is determined that they belong to the same terrain, the terrain of each pixel is determined; S4: the prediction loss of each pixel is determined; S5: loss back propagation and parameter updating are performed; The step S3 comprises: S31: the RGB picture X and the elevation picture H are fused into an input picture Z; S32: the input picture Z is input into an encoder, and after a hollow convolution network, it is input into a 1*1 convolution network, and then input into a decoder for 4 times of upsampling to obtain a first feature output, the input picture Z is input into the decoder after the hollow convolution network, and after a 1*1 convolution network in the decoder, a second feature output is obtained, the first feature output and the second feature output are combined, and then input into a 3*3 convolution network, and then 4 times of upsampling are performed, and finally a segmentation picture P is output.

2. The remote sensing map analysis method based on elevation information according to claim 1, characterized in that, The step S2 comprises the following steps: S21: select any two pixels, one of which is set as , and the other is set as , where is the spatial coordinate, and h represents the height of the pixel; S22: it is determined whether the two pixels are adjacent points, if yes, the step S23 is performed, and if not, the step S21 is returned; S23: judge whether or not if yes, the two pixels belong to the same terrain, if no, the two pixels do not belong to the same terrain, and λ is a set threshold value.

3. The method of claim 2, wherein the elevation information is obtained from a digital elevation model (DEM) or a digital terrain model (DTM). Step S22 comprises the following steps: when , if , the two pixel points are determined as adjacent points; when , if , the two pixel points are determined as adjacent points.

4. The remote sensing map analysis method based on elevation information according to claim 1, characterized in that, Step S31 is linear fusion: wherein Z is a picture fused with RGB information and height information, X is a picture with RGB information, H is a picture with height information, and a is a constant coefficient.

5. The method of claim 1, wherein the method further comprises: Step S31 is convolution fusion: wherein Z is a picture fused with RGB information and height information, and is a 1x1 convolution network.

6. The remote sensing map analysis method based on elevation information according to claim 1, wherein, Step S31 is a fusion of linear addition and convolution: wherein Z is a picture fused with RGB information and height information, X is a picture with RGB information, H is a picture with height information, and a is a constant coefficient, is a 1x1 convolution network.

7. The method of claim 1, wherein the method is based on elevation information. The hollow convolution network in the step S32 comprises a 1*1 convolution network, a 3*3 convolution network with a spatial step of 6, a 3*3 convolution network with a spatial step of 12, a 3*3 convolution network with a spatial step of 18 and a pooling network.

8. The remote sensing map analysis method based on elevation information according to claim 1, wherein, Prediction loss in step S4 wherein Loss is the prediction loss, L is the label, P is the final output segmentation picture P, wherein D is the decoder, E is the encoder, all parameters of the decoder, WE are all parameters of the encoder, the final output segmentation picture P contains a prediction of the terrain to which each pixel position belongs. 9.The method of claim 1, wherein, In the step S5, the gradient of each parameter of the encoder and the decoder is calculated according to the loss obtained by the current calculation, and the gradient is updated.

Images