Erosion gully extraction method based on noise label

Abstract

According to the erosion gully extraction method based on the noise label, firstly, an erosion gully remote sensing image is obtained, multi-scale transformation and extraction are carried out on the erosion gully remote sensing image to obtain a result image, and a multi-scale similarity loss item is calculated according to the obtained result image to improve the multi-scale aggregation capability and the anti-noise capability of a model; secondly, according to the erosion gully remote sensing image and the result image, calculating to obtain a refined label image; and finally, according to a cross entropy loss function of the original scale extraction result graph and the noise label graph, a cross entropy loss function of the original scale extraction result graph and the refined label graph, and a multi-scale similarity loss item, summation is carried out to obtain an erosion gully extraction model loss function. Optimizing the erosion gully extraction model according to the obtained erosion gully extraction model loss function; the method has excellent effects in erosion gully segmentation tasks and erosion gully image information extraction.

Description

Technical Field

[0001] The present application relates to the field of erosion gully extraction, and in particular to an erosion gully extraction method based on noise labels. Background Art

[0002] As an important representation of land degradation and soil erosion, erosion gullies not only destroy the balance of regional ecosystems, but also pose a serious threat to agricultural production and sustainable land use. However, traditional erosion gully survey methods usually rely on field surveys or manual interpretation of aerial and satellite remote sensing images. This method is not only time-consuming and labor-intensive, but also inefficient. In recent years, with the rapid development of convolutional neural networks (CNNs), researchers have begun to use this advanced technology to automatically extract erosion gully information from remote sensing images. However, convolutional neural network models usually rely on high-precision pixel-level labels, and the labeling process of erosion gully pixel-level labels is costly and requires a lot of human resources. Therefore, the automatic extraction of erosion gully images still faces many challenges. Summary of the invention

[0003] In order to solve the problem of difficulty in erosion gully extraction in the existing technology, this paper proposes an erosion gully extraction method using noise labels. This method can accurately extract erosion gully image information from erosion gully remote sensing images using noise labels with low annotation cost;

[0004] A method for extracting erosion gullies based on noise labels, comprising:

[0005] Step 1: Construct a training set for the erosion gully extraction model. The training set includes the erosion gully remote sensing image and the erosion gully noise label map Y corresponding to the erosion gully remote sensing image.

[0006] Step 2: Perform multiple scale transformations on the erosion gully remote sensing images in the training set to obtain multiple scale-transformed erosion gully remote sensing images, and extract the multiple scale-transformed erosion gully remote sensing images respectively through the erosion gully extraction model to obtain extraction result images corresponding to each scale-transformed erosion gully remote sensing image; the extraction result images include down-sampling extraction result images, original scale extraction result images F and up-sampling extraction result images;

[0007] Step 3: The multi-scale similarity loss term L is obtained by averaging the similarity divergence between the down-sampled extraction result graph, the original scale extraction result graph F, and the up-sampled extraction result graph. KL ;

[0008] Step 4: Convert the gully remote sensing image into an undirected graph, constrain the distance between any two vertices in the undirected graph, and refine the original scale extraction result map according to the constrained distance to obtain the gully refinement label map corresponding to the gully remote sensing image.

[0009] Step 5: The erosion gully extraction model is optimized as the loss function to obtain the optimized erosion gully extraction model;

[0010] L CE (F,Y) is the cross entropy loss function between the original scale extraction result image and the noise label image. is the cross entropy loss function between the original scale extraction result map and the refined label map, L KL is the multi-scale similarity loss term;

[0011] Step 6: Input the remote sensing image of the erosion gully to be measured into the optimized erosion gully extraction model, and output the erosion gully extraction result map corresponding to the remote sensing image of the erosion gully to be measured.

[0012] Beneficial Effects

[0013] The present invention discloses a method for extracting gullies based on noise labels. First, a remote sensing image of an erosion gully, a noise label map corresponding to the remote sensing image of the erosion gully, and a refined label map corresponding to the remote sensing image of the erosion gully are obtained to construct a training set. The remote sensing image of the erosion gully is subjected to multi-scale transformation. The scale-transformed image is extracted by an erosion gully extraction model to obtain a result map. The multi-scale similarity loss term is calculated based on the result map to improve the multi-scale aggregation ability and noise resistance ability of the model. Secondly, the refined label map is calculated based on the remote sensing image of the erosion gully and the result map. Finally, the cross entropy loss function of the original scale extraction result map and the noise label map, and the cross entropy loss function of the original scale extraction result map and the refined label map are calculated. The loss function of the gully extraction model is obtained by summing the cross entropy loss function of the original scale extraction result map and the noise label map, the cross entropy loss function of the original scale extraction result map and the refined label map, and the multi-scale similarity loss term. The gully extraction model is optimized based on the obtained gully extraction model loss function to obtain an optimized gully extraction model. The method of the present invention achieves excellent results in gully segmentation tasks and gully image information extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is a flow chart of a method for extracting erosion gullies based on noise labels according to a specific implementation of the present application. DETAILED DESCRIPTION

[0015] Specific implementation method 1: The following will be combined with the attached embodiment of the present invention Figure 1 , illustrate this implementation mode, and clearly and completely describe the technical solutions in the embodiments of the present invention:

[0016] A method for extracting erosion gullies based on noise labels, comprising:

[0017] Step 1: Construct a training set for the erosion gully extraction model. The training set includes the erosion gully remote sensing image and the erosion gully noise label map Y corresponding to the erosion gully remote sensing image.

[0018] Step 2: Perform multiple scale transformations on the erosion gully remote sensing images in the training set to obtain multiple scale-transformed erosion gully remote sensing images, and extract the multiple scale-transformed erosion gully remote sensing images respectively through the erosion gully extraction model to obtain extraction result images corresponding to each scale-transformed erosion gully remote sensing image; the extraction result images include down-sampling extraction result images, original scale extraction result images F and up-sampling extraction result images;

[0019] Step 3: The multi-scale similarity loss term L is obtained by averaging the similarity divergence between the down-sampled extraction result graph, the original scale extraction result graph F, and the up-sampled extraction result graph. KL ;

[0020] Step 4: Convert the gully remote sensing image into an undirected graph, constrain the distance between any two vertices in the undirected graph, and refine the original scale extraction result map according to the constrained distance to obtain the gully refinement label map corresponding to the gully remote sensing image.

[0021] Step 5: The erosion gully extraction model is optimized as the loss function to obtain the optimized erosion gully extraction model;

[0022] L CE (F,Y) is the cross entropy loss function between the original scale extraction result image and the noise label image. is the cross entropy loss function between the original scale extraction result map and the refined label map, L KL is the multi-scale similarity loss term;

[0023] Step 6: Input the remote sensing image of the erosion gully to be measured into the optimized erosion gully extraction model, and output the erosion gully extraction result map corresponding to the remote sensing image of the erosion gully to be measured.

[0024] Specifically, facing the task of erosion gully extraction with noise labels, this paper designs an erosion gully extraction method based on noise labels to improve the multi-scale aggregation ability and anti-noise performance of the model; the color and spatial information of the erosion gully remote sensing image is used to constrain the prediction of the model, so as to refine the label and obtain the refined label. Finally, by introducing a multi-scale similarity loss term, a noise label map, and a refined label map, a more reliable learning target is provided for the model.

[0025] Further, the erosion gully remote sensing images in the training set are transformed into multiple scales to obtain multiple scale-transformed erosion gully remote sensing images, and the erosion gully remote sensing images after multiple scale transformations are respectively extracted by the erosion gully extraction model to obtain extraction result images corresponding to each scale-transformed erosion gully remote sensing image, including:

[0026] The gully remote sensing images were transformed into ×0.75 scale, ×1 scale and ×1.25 scale respectively to obtain the original scale ×0.75 image, the original scale ×1 image and the original scale ×1.25 image;

[0027] The erosion gully extraction model is used to extract the original scale × 0.75 image, the original scale × 1 image, and the original scale × 1.25 image, respectively, to obtain the down-sampling extraction result Figure F down , original scale extraction result image F and up-sampled extraction result image F up .

[0028] Furthermore, the formula for the multi-scale similarity loss term is obtained by taking the average of the similarity divergences between the down-sampled extraction result graph, the original scale extraction result graph and the up-sampled extraction result graph:

[0029] L KL =-[KL(F,F down )+KL(F,F up )+KL(F down ,F up )] / 3

[0030] Among them, L KL is the multi-scale similarity loss term, KL(F,F down ) is the original scale extraction result image F and the downsampled extraction result image F down The KL divergence of the similarity between up ) is the original scale extraction result image F and the up-sampled extraction result image F up The KL divergence of the similarity between down ,F up ) is the downsampling extraction result figure F down And the upsampling extraction result figure F up The KL divergence of the similarity between .

[0031] Specifically, the semantic features of the image should be consistent under any semantic space perturbation. Therefore, a multi-scale similarity loss is designed to effectively prevent the model from overfitting to noise.

[0032] Furthermore, step 4: convert the erosion gully remote sensing image into an undirected graph, constrain the distance between any two vertices in the undirected graph, and refine the original scale extraction result map according to the constrained distance to obtain the erosion gully refined label map corresponding to the erosion gully remote sensing image include:

[0033] The erosion gully remote sensing image is represented as an undirected graph G = (V, E);

[0034] Among them, V is the vertex set consisting of all pixels in the undirected graph, and E is the boundary set consisting of the boundaries between any two adjacent vertices;

[0035] Calculate the color feature similarity weight σ of any two adjacent vertices i and j on the boundary set E i,j ;

[0036]

[0037] Where C(i) and C(j) are the spectral values ​​of vertex i and vertex j respectively, ||C(i)|| and ||C(j)|| represent the modulus of the spectral values ​​of vertex i and vertex j respectively, and C(i)·C(j) is the dot product operation of the spectral value of vertex i and the spectral value of vertex j;

[0038] Calculate the color feature similarity weights of all two adjacent vertices in the boundary set E, and construct a minimum spanning tree based on the color feature similarity weights of all two adjacent vertices;

[0039] According to the minimum spanning tree, we can get the path set S of the shortest path between any two vertices m and n on the edge set E. m,n ;

[0040] Calculate the distance θ between any two vertices m and n on the boundary set E m,n ;

[0041]

[0042] Among them, x and y are the shortest path set S m,n Two adjacent vertices in , σ x,y is the color feature similarity weight between vertex x and vertex y.

[0043] The distance θ between any two vertices m and n on the boundary set E m,n Apply constraints to obtain the distance D after constraints m,n ;

[0044] D m,n =exp(-θ m,n / B)

[0045] Among them, B is the constant value of color feature adjustment, and exp is the natural exponential function;

[0046] According to the distance D after constraint m,n Refine the original scale extraction result graph F to obtain the vertex refinement label of vertex m in the boundary set E

[0047]

[0048] in, is the normalization term, F m Extract the predicted value of vertex m in the result graph F for the original scale;

[0049] Calculate the vertex refinement label of each vertex in the boundary set E, and arrange the vertex refinement label of each vertex in the boundary set according to the vertex position to form a refinement label graph

[0050] Specific implementation method 2: A computer-readable storage medium stores a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the method described in specific implementation method 1.

[0051] Specific implementation method three: A erosion gully extraction device based on noise labels, comprising a storage device, a processor, and a computer program stored in the storage device and executable on the processor, characterized in that the processor executes the computer program to implement a erosion gully extraction method based on noise labels as described in specific implementation method one.

[0052] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that features described in conjunction with individual embodiments may be used in other described embodiments.

Claims

1. A method for extracting erosion gullies based on noise labels, characterized in that: include: Step 1: Construct a training set for the erosion gully extraction model. The training set includes the erosion gully remote sensing image and the erosion gully noise label map Y corresponding to the erosion gully remote sensing image. Step 2: Perform multiple scale transformations on the erosion gully remote sensing images in the training set to obtain multiple scale-transformed erosion gully remote sensing images, and extract the multiple scale-transformed erosion gully remote sensing images respectively through the erosion gully extraction model to obtain extraction result images corresponding to each scale-transformed erosion gully remote sensing image; the extraction result images include down-sampling extraction result images, original scale extraction result images F and up-sampling extraction result images; Step 3: The multi-scale similarity loss term L is obtained by averaging the similarity divergence between the down-sampled extraction result graph, the original scale extraction result graph F, and the up-sampled extraction result graph. KL ; Step 4: Convert the gully remote sensing image into an undirected graph, constrain the distance between any two vertices in the undirected graph, and refine the original scale extraction result map according to the constrained distance to obtain the gully refinement label map corresponding to the gully remote sensing image. Step 5: The loss function is used to optimize the erosion gully extraction model, and the optimized erosion gully extraction model is obtained; L CE (F,Y) is the cross entropy loss function between the original scale extraction result image and the noise label image. is the cross entropy loss function between the original scale extraction result map and the refined label map, L KL is the multi-scale similarity loss term; Step 6: Input the remote sensing image of the erosion gully to be measured into the optimized erosion gully extraction model, and output the erosion gully extraction result map corresponding to the remote sensing image of the erosion gully to be measured.

2. The method for extracting erosion gullies based on noise labels according to claim 1, characterized in that: Performing multiple scale transformations on the erosion gully remote sensing images in the training set to obtain multiple scale-transformed erosion gully remote sensing images, respectively extracting the multiple scale-transformed erosion gully remote sensing images through the erosion gully extraction model to obtain extraction result images corresponding to each scale-transformed erosion gully remote sensing image, including: The gully remote sensing images were transformed into ×0.75 scale, ×1 scale and ×1.25 scale respectively to obtain the original scale ×0.75 image, the original scale ×1 image and the original scale ×1.25 image; The erosion gully extraction model is used to extract the original scale × 0.75 image, the original scale × 1 image, and the original scale × 1.25 image, respectively, to obtain the down-sampling extraction result Figure F down , original scale extraction result image F and up-sampled extraction result image F up .

3. The method for extracting gullies based on noise labels according to claim 2, characterized in that: The formula for the multi-scale similarity loss term is obtained by taking the average of the similarity divergence between the down-sampled extraction result graph, the original scale extraction result graph and the up-sampled extraction result graph: L KL =-[KL(F,F down )+KL(F,F up )+KL(F down ,F up )] / 3 Among them, L KL is the multi-scale similarity loss term, KL(F,F down ) is the original scale extraction result image F and the downsampled extraction result image F down The KL divergence of the similarity between up ) is the original scale extraction result image F and the up-sampled extraction result image F up The KL divergence of the similarity between down ,F up ) is the downsampling extraction result figure F down And the upsampling extraction result figure F up The KL divergence of the similarity between .

4. The method for extracting gullies based on noise labels according to claim 3 is characterized in that: The erosion gully remote sensing image is converted into an undirected graph, the distance between any two vertices in the undirected graph is constrained, and the original scale extraction result map is refined according to the constrained distance to obtain the erosion gully refined label map corresponding to the erosion gully remote sensing image include: The erosion gully remote sensing image is represented as an undirected graph G = (V, E); Among them, V is the vertex set consisting of all pixels in the undirected graph, and E is the boundary set consisting of the boundaries between any two adjacent vertices in the undirected graph; Calculate the color feature similarity weight σ of any two adjacent vertices i and j on the boundary set E i,j ; Where C(i) and C(j) are the spectral values ​​of vertex i and vertex j respectively, ||C(i)|| and ||C(j)|| represent the modulus of the spectral values ​​of vertex i and vertex j respectively, and C(i)·C(j) is the dot product operation of the spectral value of vertex i and the spectral value of vertex j; Calculate the color feature similarity weights of all two adjacent vertices in the boundary set E, and construct a minimum spanning tree based on the color feature similarity weights of all two adjacent vertices; According to the minimum spanning tree, we can get the path set S of the shortest path between any two vertices m and n on the edge set E. m,n ; Calculate the distance θ between any two vertices m and n on the boundary set E m,n ; Among them, x and y are the shortest path set S m,n Two adjacent vertices in , σ x,y is the color feature similarity weight between vertex x and vertex y. The distance θ between any two vertices m and n on the boundary set E m,n Apply constraints to obtain the distance D after constraints m,n ; D m,n =exp(-θ m,n / B) Among them, B is the constant value of color feature adjustment, and exp is the natural exponential function; According to the distance D after constraint m,n Refine the original scale extraction result graph F to obtain the vertex refinement label of vertex m in the boundary set E in, is the normalization term, F m Extract the predicted value of vertex m in the result graph F for the original scale; Calculate the vertex refinement label of each vertex E in the boundary set, and arrange the vertex refinement label of each vertex in the boundary set according to the vertex position to form an erosion gully refinement label map 5. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 4 are implemented.

6. A device for extracting erosion gullies based on noise labels, comprising a storage device, a processor, and a computer program stored in the storage device and executable on the processor, characterized in that: The processor executes the computer program to implement the method for extracting gullies based on noise labels as described in any one of claims 1 to 4.

Images