Underlying surface engineering and ground feature intelligent identification method based on small sample learning

Abstract

The invention provides an underlying surface engineering and surface feature intelligent identification method based on small sample learning, which comprises the steps of collecting surface feature sample data, and further comprises the following steps: constructing a learning sample set; performing feature extraction; and carrying out sample identification. According to the underlying surface engineering and surface feature intelligent identification method based on small sample learning, a machine learning technology and a manifold learning algorithm are introduced into underlying surface engineering and surface feature intelligent identification, and a small sample feature learning technology is utilized to analyze and mine historical image samples, so that the underlying surface engineering and surface feature intelligent identification method is realized. And a newly collected image sample can be quickly identified and intelligently judged.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing target recognition, and in particular to an underlying surface engineering and ground object intelligent recognition method based on small sample learning. Background Art

[0002] With the advancement of digital twin work in the operation and management of reservoirs and sluice gates, satellite remote sensing and drone images have played a huge role in practical work. Through the mining and analysis of remote sensing or drone image data and the identification of information, the scientific, accurate and effective application of digital twin results is supported.

[0003] At present, the recognition technology of remote sensing or UAV images is mainly based on deep learning methods, but this requires a large number of high-quality training samples. This method is not ideal for the recognition of water conservancy projects or land objects with fewer training samples. In addition, this method requires manual labeling of a large number of training samples and the determination of the categories and names of the training samples in advance, which is difficult to use in actual work and cannot be widely promoted and applied.

[0004] The invention patent application with application number CN117789036A discloses a method, system, device and storage medium for identifying the type of underlying land surface. The identification method first removes fog from the optical remote sensing image that can reflect the pixel features of the ground object, and then extracts the features of the channel dimension and the spatial dimension to obtain an optical remote sensing feature map with rich features and high resolution; then, the radar remote sensing image with elevation information and the optical remote sensing feature map with pixel features are subjected to feature-level fusion processing to obtain a fused remote sensing feature map that is easy to classify and identify; finally, the fused remote sensing feature map is processed by using the frame regression processing and probability mapping, and the identification frame with the most accurate classification result is screened out based on the intersection-over-union ratio and the confidence level, and the classification result corresponding to the maximum classification probability of the identification frame is used as the underlying surface type identification result. The disadvantages of this method are: 1) the number of samples required is high, and it may be difficult to learn effective features when the data is insufficient; 2) the convolutional neural network has a complex structure, a large amount of calculation, and requires high computer computing power and a slow calculation speed; 3) it is not suitable for practical work, especially for the identification of engineering ground objects with a small number of samples.

[0005] The invention patent application with application number CN110135364A discloses a method and device for identifying a land object, which includes: obtaining feature information of each sub-region in the image to be annotated; using a clustering algorithm to obtain cluster regions of the same land object type according to the feature information of each sub-region in the image to be annotated; calculating the similarity information between the feature information of the cluster region and the features of the annotated samples; and determining the land object type of the cluster region according to the similarity information. The disadvantage of this method is that it is difficult to accurately cluster high-dimensional data into a specified category by clustering alone, the error rate is high, and the recognition speed is slow; and the diversity, quantity and quality of samples are required to be high. Summary of the invention

[0006] In order to solve the above-mentioned technical problems, the present invention proposes a method for intelligent identification of underlying surface engineering and land objects based on small sample learning, which introduces machine learning technology and manifold learning algorithm into the intelligent identification of underlying surface engineering and land objects, and utilizes small sample feature learning technology. On the basis of analyzing and mining historical image samples, the newly collected image samples can be quickly identified and intelligently judged.

[0007] The purpose of the present invention is to provide a method for intelligently identifying underlying surface engineering and ground objects based on small sample learning, comprising collecting ground object sample data, and further comprising the following steps: Step 1: Construct a learning sample set; Step 2: Perform feature extraction; Step 3: Perform sample identification.

[0008] Preferably, the ground feature sample data includes at least one type of image sample of a reservoir, a river and a lake.

[0009] In any of the above schemes, preferably, step 1 includes the following sub-steps: Step 11: performing filtering and denoising, image erosion, expansion and opening and closing operations on the image samples, and determining the area where the ground object is located; Step 12: performing front and back background segmentation, image data cropping and target edge extraction on the processed image samples to extract the range of the ground object to the maximum extent; Step 13: Label the names of the features and build a learning sample set Q .

[0010] In any of the above schemes, preferably, step 2 includes analyzing the learning sample data set using the ISOMAP algorithm, projecting it into a low-dimensional space and extracting features while retaining its features in the high-dimensional space, and constructing a low-dimensional feature subspace Φ.

[0011] In any of the above solutions, preferably, the ISOMAP algorithm comprises the following sub-steps: Step 21: Determine the neighborhood points on the manifold M and construct a neighbor graph; Step 22: Estimating the Manifold M Geodesic distance on d M ( i , j ); Step 23: Apply MDS Construction d Dimensional embedding; in, i and j are any two points on M.

[0012] In any of the above solutions, preferably, step 21 includes setting the input space X Any two sample vectors x i With x j The Euclidean distance is d E ( i , j ), and then use all sample vectors x i (1≤ i ≤N) construct a weighted graph G, including: For x i The nearest k points as neighborhood; In graph G, if x j is x i Neighborhood points, connect them, and let the length of the connecting line be their Euclidean distance d E ( i , j ); Set Input Space X Any two sample vectors x p and x q The Euclidean distance is d E ( p , q ), and then use all sample vectors x k (1≤ k ≤ N ) Construct a weighted graph G; Where N is the number of all points.

[0013] In any of the above solutions, preferably, step 22 includes calculating the shortest pathd G ( i , j ) method to approximate the manifold M Geodesic distance on d M ( i , j ), including the following sub-steps: Step 221: In the figure G In the example, let any two sample vectors x i and x j The shortest distance between d G ( i , j ),if x i and x j If there is a connection between d G ( i , j )= d E ( i , j ), otherwise d G ( i , j )=∞; Step 222: Update d G ( i , j ) value, for k =1,2,3,… N ,make d G ( i , j )=min{ d G ( i , j ), d G ( i , k )+ d G ( i , k+1 )}; Step 223: After multiple iterations, the shortest path matrix between sample vectors D G ={ d G ( i, j )} gradually converge.

[0014] In any of the above solutions, preferably, step 23 includes the following sub-steps: Step 231: Use classic MDS to compress the sample vector into d Dimension, and make the Euclidean distance between the compressed sample vectors as close as possible to the shortest path that has been found; Step 232: Assume that the matrix τ( D G ) d Eigenvalues λ 1≥ λ 2≥ λ 3≥…≥ λ d The corresponding eigenvector is v 1 , v 2 , v 3 ,…, v d ,but d dimensional embedding vector y i No. p The components are equal to, so the embedding coordinates of each point in the high-dimensional space in the low-dimensional space Y , expressed as:

[0015] in, λp It is p eigenvalues, v p i yes λ p The corresponding eigenvector i A quantity.

[0016] In any of the above schemes, it is preferred that step 3 includes projecting a new ground object sample to be identified into a low-dimensional feature subspace Φ when it appears, and finding the point closest to the new ground object sample to be identified in the feature subspace, and judging the category of the ground object sample to be identified based on the state of the most similar point in the learning sample library.

[0017] In any of the above schemes, preferably, step 3 includes the following sub-steps: Step 31: Project the ground feature sample to be identified into the low-dimensional feature subspace Φ of the sample library through the optimal projection matrices U and V to obtain the point of the sample in the low-dimensional feature subspace; Step 32: According to the principle of minimum distance, find the learning sample that is most similar to the new sample in the low-dimensional feature subspace; Step 33: Extract the state of the learning sample as the recognition result. min( d 2D ( Y t ; Y i ))=min(|| Y t - Y i ||2) in, d 2D yes Y t and Y i The distance between Y t is the feature matrix of the ground object samples to be identified, Y i is the feature matrix of the small sample library; Step 34: Determine the category of the ground object of the new sample to be identified based on the appearance of the ground object in the learning sample library that is most similar to the sample to be identified.

[0018] The present invention proposes a method for intelligent recognition of underlying surface engineering and ground objects based on small sample learning, which can automatically realize ground object labeling, does not require manual labeling, does not require a large number of learning samples, and does not require advance learning and training. It can automatically realize the recognition and judgment of ground objects, has low requirements on the number and quality of samples and the computing power of the computer, and has high recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The present invention is a flowchart of a preferred embodiment of the method for intelligently identifying underlying surface engineering and ground objects based on small sample learning.

[0020] Figure 2 The present invention is a technical flow chart of a preferred embodiment of the underlying surface engineering and ground object intelligent identification method based on small sample learning according to the present invention.

[0021] Figure 3 It is a schematic diagram of an embodiment of the location identification of ground objects according to the underlying surface engineering and ground object intelligent identification method based on small sample learning of the present invention.

[0022] Figure 4 It is a schematic diagram of an embodiment of the underlying surface engineering and intelligent ground object recognition method based on small sample learning of the present invention, which uses the front and back background segmentation to extract the ground objects in the maximum range.

[0023] Figure 5 It is a schematic diagram of an embodiment of the recognition results of ground objects and engineering according to the method for intelligent recognition of underlying surface engineering and ground objects based on small sample learning of the present invention. DETAILED DESCRIPTION

[0024] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0025] Example 1 like Figure 1 As shown, a method for intelligent recognition of underlying surface engineering and ground objects based on small sample learning is performed, step 1000 is executed to collect ground object sample data, and the ground object sample data at least includes an image sample of a reservoir, a river and a lake.

[0026] Execute step 1100 to construct a learning sample set, including the following sub-steps: Execute step 1110 to perform filtering and denoising, image erosion, dilation, and opening and closing operations on the image sample, and determine the area where the object is located; Execute step 1120 to perform front and back background segmentation, image data cropping, and target edge extraction on the processed image sample, and extract the range of the ground object to the maximum extent; Execute step 1130 to label the names of the features and construct a learning sample set Q .

[0027] Execute step 1200 to perform feature extraction, use the ISOMAP algorithm to analyze the learning sample data set, retain its features in the high-dimensional space, project it into the low-dimensional space and perform feature extraction, and construct a low-dimensional feature subspace Φ.

[0028] The ISOMAP algorithm includes the following sub-steps: Execute step 1210 to determine the neighborhood points on the manifold M and construct a neighbor graph, including setting the input space X Any two sample vectors x i With x j The Euclidean distance is d E ( i , j ), and then use all sample vectors x i (1≤ i ≤N) construct a weighted graph G, where N is the number of all points, including: For x i The nearest k points as neighborhood; In graph G, if x j is x iNeighborhood points, connect them, and let the length of the connecting line be their Euclidean distance d E ( i , j ); Set Input Space X Any two sample vectors x p and x q The Euclidean distance is d E ( p , q ), and then use all sample vectors x k (1≤ k ≤ N )Construct a weighted graph G.

[0029] Execute step 1220 to estimate the manifold M Geodesic distance on d M ( i , j ),in, i and j For any two points on M, including using the shortest path calculation d G ( i , j ) method to approximate the manifold M Geodesic distance on d M ( i , j ), including the following sub-steps: Execute step 1221, in FIG. G In the example, let any two sample vectors x i and x j The shortest distance between d G ( i , j ),if x i and x j If there is a connection between d G ( i , j )= d E ( i , j ), otherwise dG ( i , j )=∞; Execute step 1222 to update d G ( i , j ) value, for k =1,2,3,… N ,make d G ( i , j )=min{ d G ( i , j ), d G ( i , k )+ d G ( i , k+1 )}; Execute step 1223, after multiple iterations, the shortest path matrix between sample vectors D G ={ d G ( i , j )} gradually converge.

[0030] Execute step 1230 and apply MDS construction d dimensional embedding, the step 23 comprises the following sub-steps: Execute step 1231 and use classic MDS to compress the sample vector into d Dimension, and make the Euclidean distance between the compressed sample vectors as close as possible to the shortest path that has been found; Execute step 1232, set the matrix τ( D G ) d Eigenvalues λ 1≥ λ 2≥ λ 3≥…≥ λ d The corresponding eigenvector is v 1 , v 2 , v 3 ,…, v d ,but d dimensional embedding vector y i No.p The components are equal to, so the embedding coordinates of each point in the high-dimensional space in the low-dimensional space Y , expressed as:

[0031] in, λp It is p eigenvalues, v p i yes λ p The corresponding eigenvector i A quantity.

[0032] Execute step 1300 to perform sample recognition, including projecting a new ground object sample to be recognized into a low-dimensional feature subspace Φ when it appears, and finding the point closest to the new ground object sample to be recognized in the feature subspace, and judging the category of the ground object sample to be recognized according to the state of the most similar point in the learning sample library, including the following sub-steps: Execute step 1310 to project the ground feature sample to be identified into the low-dimensional feature subspace Φ of the sample library through the optimal projection matrices U and V to obtain the point of the sample in the low-dimensional feature subspace; Execute step 1320 to find the learning sample that is most similar to the new sample in the low-dimensional feature subspace according to the principle of minimum distance; Execute step 1330 to extract the state of the learning sample as a recognition result. min( d 2D ( Y t ; Y i ))=min(|| Y t - Y i ||2) in, d 2D yes Y t and Y i The distance between Y t is the feature matrix of the ground object samples to be identified, Y i is the feature matrix of the small sample library; Step 1340 is executed to determine the category of the ground object of the new sample to be identified based on the appearance of the ground object in the learning sample library that is most similar to the sample to be identified.

[0033] Example 2 The present invention introduces computer vision and machine learning technology into image analysis, and uses small sample feature learning technology to perform feature recognition on newly collected image samples and historical samples based on the analysis and mining of historical image samples. The present invention can automatically realize the annotation of objects, does not require manual annotation, does not require a large number of learning samples, and does not require advance learning and training, and can automatically realize the recognition and judgment of objects. The requirements for the number and quality of samples and the computing power of computers are low, the recognition accuracy is high, and it can be widely promoted and applied in practical work.

[0034] The present invention is mainly aimed at the recognition of underlying surface reservoirs, rivers, lakes and other landforms. These landforms are mainly black. Therefore, before the construction of the small sample training sample library, the black block area in the image is first located and represented by a red frame. And further through opening and closing operations, front and back background segmentation, target edge extraction, etc., the area where the landform is located is finally determined and marked with a red frame. This pre-processing method is suitable for the recognition requirements of landforms and meets practical needs. By constructing a learning sample library by processing samples in advance, features that are easier to identify, although the samples are small, the recognition accuracy is greatly improved.

[0035] The present invention uses this method to identify reservoirs, rivers and lakes in images. The results show that this method can effectively identify the appearance status of the project. The accuracy of river identification is above 98%, and the accuracy of identification of large reservoirs and lakes is above 87%. In addition, the method has fast calculation speed and high recognition accuracy.

[0036] like Figure 2 As shown, the main steps of this application include: 1. Construction of learning sample set. This includes the collection, cleaning, and identification of sample data of reservoirs, rivers, lakes, etc., segmentation of the background and front and back of the cleaned image samples, cropping of image data, etc., and labeling of the names of the features to build a learning sample set. Q .

[0037] Select the target, use the drone to take pictures of reservoirs, lakes, rivers, etc. in the area, and collect samples. Due to the influence of the drone's cruising route, shooting angle, and light during collection, the collected image samples cannot be analyzed directly, and the image needs to be processed in advance, mainly including filtering and denoising the sample image, image corrosion, expansion, opening and closing operations, and finally determining the area where the object is located, such as Figure 3 As shown in the figure, the red frame indicates the location of the object. Figure 4 As shown, the algorithm is used to segment the foreground and background, extract the target edge, etc., and extract the range of the ground object to the maximum extent.

[0038] After the image samples are processed by the above algorithm, the shape of the ground object is automatically located and recognized to the maximum extent and marked in the image. Then, with the marked ground object as the center, the image sample is automatically cropped into a single ground object sample with a resolution of 100×100, and its status is annotated one by one as a training positive sample.

[0039] In order to ensure the diversity of the sample library, this paper collected ground object samples through multiple channels and obtained sample images of various states of the ground objects. After the same data processing method, they were cropped into single bolt samples with a resolution of 100×100, and their states were labeled one by one as training negative samples to expand and improve the professional sample library.

[0040] 2. Feature extraction: Use the ISOMAP algorithm to analyze the learning sample data set, retain its features in the high-dimensional space, project it into the low-dimensional space and extract features to construct a low-dimensional feature subspace Φ.

[0041] The learning sample library of the features is a nonlinear high-dimensional data space, so the present invention first uses the Isomap algorithm, an algorithm for nonlinear data dimensionality reduction, to perform nonlinear dimensionality reduction analysis on the high-dimensional data. The Isomap algorithm is based on CMDS, and attempts to maintain the intrinsic geometric properties of the data and obtain the geodesic distance between data points on the manifold. The Isomap algorithm replaces the Euclidean distance with the geodesic distance, and uses CMDS to perform low-dimensional embedding of the geodesic distance, thus overcoming the limitations of CMDS. The geodesic distance on a manifold can be expressed as the sum of the distances between a series of neighborhood points. The key to the algorithm is to use the Euclidean distance d between sample vectors. E (i,j) calculates the geodesic distance d between samples G (i, j), and then use the classic CMDS algorithm to construct a new d-dimensional (d is the dimension of the reduced dimensionality space) Euclidean space Y , maximize the Euclidean distance d between samples E (i,j) and d G (i, j) error is the smallest, thus playing a role in dimensionality reduction. For neighborhood points, Isomap directly obtains their geodesic distance from the input space; for non-neighborhood points, their geodesic distance can be approximated as the sum of the geodesic distances of a series of neighborhood points.

[0042] The Isomap algorithm has three steps: the first step is to determine which points on the manifold M are neighbors of each other. The second step is to calculate the shortest path d G ( i , j ) method to estimate the geodesic distance on the manifold M d M (i , j ). The third step is to apply CMDS to construct d Dimensional embedding. The specific algorithm is as follows: The first step is to determine the neighborhood points on the manifold M and construct a neighbor graph. Suppose the input space X Any two sample vectors x i With x j The Euclidean distance is d E ( i , j ), and then use all sample vectors x i (1≤i≤N) constructs a weighted graph G. Use the following method to determine x i Neighborhood of x i The nearest k points as neighborhood. In graph G, if x j is x i Neighborhood points, connect them, and let the length of the connecting line be their Euclidean distance d E ( i , j ). Perform the same operation as above for all points in the input sample set, and the weighted graph G can be obtained.

[0043] The second step is to estimate the manifold M Geodesic distance on d M ( i , j ). Use the shortest path calculation d G ( i , j ) method to approximate the manifold M Geodesic distance on d M ( i , j ). G In the example, let any two sample vectors x i and x j The shortest distance between d G ( i , j ),if x i and x j If there is a connection between d G ( i ,j )= d E ( i , j ), otherwise d G ( i , j )=∞. Then update d G ( i , j ) value, for k =1,2,3,… N ,make d G ( i , j )=min{ d G ( i , j ), d G ( i , k )+ d G ( i , k+ 1 )}, after multiple iterations, the shortest path matrix between sample vectors D G ={ d G ( i , j )} will converge. The shortest path matrix can be calculated using the Dijkstra algorithm.

[0044] The third step is to apply the MDS construct d dimensional embedding. Use classic MDS to compress the sample vector into d dimension, and make the Euclidean distance between the compressed sample vectors as close as possible to the shortest path that has been found. Let the matrix τ( D G ) d Eigenvalues λ 1≥ λ 2≥ λ 3≥…≥ λ d The corresponding eigenvector is v 1 , v 2 , v 3 ,…, v d ,λp It is p eigenvalues,v p i yes λp The th component of the corresponding eigenvector is d dimensional embedding vector y i No. p The components are equal to, so the embedding coordinates of each point in the high-dimensional space in the low-dimensional space Y , expressed as:

[0045] This achieves dimensionality reduction of high-dimensional data.

[0046] 3. Sample recognition. When a new ground object sample to be recognized appears, it is first projected into the low-dimensional feature subspace Φ, and the point closest to the new sample is found in the feature subspace, that is, the most similar learning sample. According to the state of the most similar point in the learning sample library, the category of the ground object sample to be recognized is determined, and the recognition process is finally completed.

[0047] The identification of new samples is to project the ground object samples to be identified into the low-dimensional feature subspace Φ of the sample library through the optimal projection matrices U and V, and obtain the points of the samples in the low-dimensional feature subspace. Then, according to the principle of minimum distance, find the learning sample that is most similar to the new sample in the low-dimensional feature subspace, that is, the one with the smallest distance. The state of the labeled known learning sample corresponding to the new sample is the recognition result.

[0048] min( d 2D ( Y t ; Y i ))=min(|| Y t - Y i ||2) In the formula, Y t is the feature matrix of the ground object samples to be identified, Y i is the feature matrix of the small sample library. Finally, the object category of the new sample to be identified is determined based on the appearance of the objects in the learning sample library that are most similar to the sample to be identified.

[0049] Example 3 The algorithm of the present invention is used to automatically identify and classify reservoirs, lakes and rivers. The algorithm does not require a large number of samples, has low requirements on computer computing power, has a fast recognition speed, and can complete the intelligent recognition and labeling of engineering or land objects in an instant.

[0050] like Figure 5 As shown, (1) is a known reservoir sample, (2) is a known lake sample, and (3) is a known river sample. Using the algorithm of the present invention, the images to be identified are classified, (1a), (1b) and (1c) are the identification results of the images being reservoirs, (2a), (2b) and (2c) are the identification results of the images being lakes, and (3a), (3b) and (3c) are the identification results of the images being rivers.

[0051] from Figure 5 It can be seen that the algorithm can effectively identify reservoirs, rivers and lakes with high accuracy and fast recognition speed, and can be widely promoted and applied.

[0052] This algorithm is a deep mining of drone or remote sensing images based on data mining technology, with different types of projects and landforms as samples. It not only does not require a high number of samples, but also has low requirements for computing power. In actual work, it can quickly identify projects and landforms in the "Water Conservancy Three-Dimensional Digital Twin Platform", providing new ideas for the application of artificial intelligence technology in digital twin construction.

[0053] In order to better understand the present invention, the above is described in detail in conjunction with the specific embodiments of the present invention, but it is not intended to limit the present invention. Any simple modifications made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention. Each embodiment in this specification focuses on the differences from other embodiments, and the same or similar parts between the embodiments can be referenced to each other. For the system embodiment, since it basically corresponds to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

Claims

1. A method for intelligent identification of underlying surface engineering and ground objects based on small sample learning, comprising collecting ground object sample data, characterized in that: The following steps are also included: Step 1: Construct a learning sample set; Step 2: Perform feature extraction; Step 3: Perform sample identification, including the following sub-steps: Step 31: Project the ground feature sample to be identified into the low-dimensional feature subspace Φ of the sample library through the optimal projection matrices U and V to obtain the point of the sample in the low-dimensional feature subspace; Step 32: According to the principle of minimum distance, find the learning sample that is most similar to the new sample in the low-dimensional feature subspace; Step 33: Extract the state of the learning sample as the recognition result. my( d 2D ( Y t ; Y i ))=min(|| Y t - Y i ||2) in, d 2D yes Y t and Y i The distance between Y t is the feature matrix of the ground object samples to be identified, Y i is the feature matrix of the small sample library; Step 34: Determine the category of the ground object of the new sample to be identified based on the appearance of the ground object in the learning sample library that is most similar to the sample to be identified.

2. The method for intelligent recognition of underlying surface engineering and ground objects based on small sample learning according to claim 1, characterized in that: The ground feature sample data includes at least one type of image sample of a reservoir, a river and a lake.

3. The method for intelligent recognition of underlying surface engineering and ground objects based on small sample learning according to claim 2, characterized in that: The step 1 includes the following sub-steps: Step 11: performing filtering and denoising, image erosion, expansion and opening and closing operations on the image samples, and determining the area where the ground object is located; Step 12: performing front and back background segmentation, image data cropping and target edge extraction on the processed image samples to extract the range of the ground object to the maximum extent; Step 13: Label the names of the features and build a learning sample set Q .

4. The method for intelligent identification of underlying surface engineering and ground objects based on small sample learning according to claim 3, characterized in that: The step 2 includes analyzing the learning sample data set using the ISOMAP algorithm, projecting it into a low-dimensional space and extracting features while retaining its features in the high-dimensional space, and constructing a low-dimensional feature subspace Φ.

5. The method for intelligent identification of underlying surface engineering and ground objects based on small sample learning according to claim 4, characterized in that: The ISOMAP algorithm includes the following sub-steps: Step 21: Determine the neighborhood points on the manifold M and construct a neighbor graph; Step 22: Estimating the Manifold M Geodesic distance on d M ( i , j ); Step 23: Apply MDS Construction d Dimensional embedding; in, i and j are any two points on M.

6. The method for intelligent identification of underlying surface engineering and ground objects based on small sample learning according to claim 5, characterized in that: The step 21 includes setting the input space X Any two sample vectors x i With x j The Euclidean distance is d E ( i , j ), and then use all sample vectors x i (1≤ i ≤N) construct a weighted graph G, include: For x i The nearest k points as neighborhood; In graph G, if x j is x i Neighborhood points, connect them, and let the length of the connecting line be their Euclidean distance d E ( i , j ); Set Input Space X Any two sample vectors x p and x q The Euclidean distance is d E ( p , q ), and then use all sample vectors x k (1≤ k ≤ N ) Construct a weighted graph G; Where N is the number of all points.

7. The method for intelligent identification of underlying surface engineering and ground objects based on small sample learning according to claim 6, characterized in that: The step 22 includes calculating the shortest path d G ( i , j ) method to approximate the manifold M Geodesic distance on d M ( i , j ), including the following sub-steps: Step 221: In the figure G In the example, let any two sample vectors x i and x j The shortest distance between d G ( i , j ),if x i and x j If there is a connection between d G ( i , j )= d E ( i , j ), otherwise d G ( i , j )=∞; Step 222: Update d G ( i , j ) value, for k =1,2,3,… N ,make d G ( i , j )=min{ d G ( i , j ), d G ( i , k )+ d G ( i , k+1 )}; Step 223: After multiple iterations, the shortest path matrix between sample vectors D G ={ d G ( i , j )} gradually converge.

8. The method for intelligent identification of underlying surface engineering and ground objects based on small sample learning according to claim 7, characterized in that: The step 23 includes the following sub-steps: Step 231: Use classic MDS to compress the sample vector into d Dimension, and make the Euclidean distance between the compressed sample vectors as close as possible to the shortest path that has been found; Step 232: Assume that the matrix τ( D G ) d Eigenvalues λ 1≥ λ 2≥ λ 3≥…≥ λ d The corresponding eigenvector is v 1 , v 2 , v 3 ,…, v d ,but d dimensional embedding vector y i No. p The components are equal to, so the embedding coordinates of each point in the high-dimensional space in the low-dimensional space Y , expressed as: in, λp It is p eigenvalues, v p i yes λ p The corresponding eigenvector i A quantity.

9. The method for intelligent identification of underlying surface engineering and ground objects based on small sample learning according to claim 8, characterized in that: The step 3 includes projecting a new ground object sample to be identified into a low-dimensional feature subspace Φ when it appears, and finding the point closest to the new ground object sample to be identified in the feature subspace, and judging the category of the ground object sample to be identified according to the state of the most similar point in the learning sample library.

Images