An automatic extraction method of unsupervised remote sensing fields based on adaptive method
Through adaptive deep learning networks and unsupervised learning methods, the problem of relying on supervised learning and manual intervention in remote sensing technology is solved, efficient and accurate field extraction is achieved, and the scientificity and accuracy of agricultural monitoring is improved.
Patent Information
- Application Number
- CN202410863792.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-29
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-06-29
AI Technical Summary
Existing remote sensing technologies rely on supervised learning and manual intervention, resulting in complex data acquisition, slow processing speed, high cost, and difficult to ensure the accuracy and consistency of results, making it difficult to effectively extract field information in massive remote sensing images.
Adopting an unsupervised remote sensing field automatic extraction method based on adaptive methods, by collecting and preprocessing remote sensing image data, an adaptive deep learning network is constructed, unsupervised learning and loss function optimization, field distribution map is generated, and regularly updated to Adapt to seasonal changes.
It realizes efficient and accurate field extraction, reduces costs, improves the scientificity and accuracy of agricultural monitoring and management, and ensures data timeliness.
Smart Images

Figure CN118521915B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of remote sensing image processing, and in particular to an unsupervised remote sensing field automatic extraction method based on an adaptive method. Background Art
[0002] With the rapid development of modern remote sensing technology, how to quickly extract detailed and accurate field information from massive remote sensing images in an efficient and accurate manner has gradually become the core key to comprehensive monitoring and scientific management in the agricultural field.
[0003] Most of the existing related technical means are heavily dependent on the supervised learning model, which means that a large amount of labeled sample data is required. Not only is the data acquisition and labeling process cumbersome and complicated, but it also consumes a lot of time and manpower costs. In addition, some technologies rely on direct manual intervention, which not only greatly limits the processing speed and scale, but also makes it difficult to cope with large-scale remote sensing image data. Moreover, due to the subjectivity and uncertainty of manual operation, it is easy to affect the accuracy and consistency of the extraction results, resulting in extremely low efficiency or extremely high cost. Therefore, it is necessary to provide an unsupervised remote sensing field automatic extraction method based on an adaptive method to solve the above technical problems. Summary of the invention
[0004] The present invention provides an unsupervised remote sensing field automatic extraction method based on an adaptive method, which solves the problems mentioned in the above background technology.
[0005] In order to solve the above technical problems, the present invention provides an unsupervised remote sensing field automatic extraction method based on an adaptive method, comprising the following steps:
[0006] Step 1: collect remote sensing image data of the set field area and perform preprocessing;
[0007] Step 2: Use the preprocessed remote sensing image data to build an adaptive deep learning network to extract image features; organize the extracted image features into a feature map; each pixel in the feature map is represented by a feature vector; the construction of the adaptive deep learning network includes adaptive convolutional layers, multi-scale feature fusion and residual connections;
[0008] Step 3: Use feature maps in remote sensing image data for unsupervised learning;
[0009] Step 4: Calculate and optimize the loss function of the field area;
[0010] Step 5: Evaluate the clustering results, extract the boundaries of the fields, and generate a field distribution map;
[0011] Step six: Regularly acquire new remote sensing image data and update the adaptive deep learning network to adapt to seasonal changes and crop growth cycles.
[0012] Preferably, based on step one, the specific steps of preprocessing include denoising, atmospheric correction, radiation correction, geometric correction, and converting the remote sensing image data into a unified format and scale for subsequent processing.
[0013] As a preferred embodiment, the specific logic of constructing an adaptive deep learning network is:
[0014] Establish an adaptive convolution layer: Set the input feature map to I, where I represents a three-dimensional tensor with a size of H×W×C. H, W, and C represent the height, width, and number of channels of the remote sensing image, respectively. Each channel contains information of a specific spectral band. The adaptive convolution operation is expressed as: , and get the output feature map ; Among them, h' and w' represent the height and width of the output features after the convolution operation, and the corresponding calculation formula is expressed as: , , i and j represent the row and column positions of the convolution kernel when it slides on the input feature map, k represents the number corresponding to the kth channel of the input feature map, Pi and Pj represent the total number of padding added on both sides of the i-th row and j-th column of the input feature map respectively;
[0015] Multi-scale feature fusion: Multi-scale feature fusion technology combines input feature maps of different resolutions. is the input feature map of scale s, and the input feature maps of different scales are fused. The formula is: , and obtain the fused output feature map O; where Fuse represents the fusion operation, Indices representing different scales;
[0016] Residual connection: Let D be the input of a layer in the neural network, D′ be the output of the layer, and use batch normalization and activation function for processing. The formula is: Q=ReLU(BN(D′)+D), and the output Q of the residual connection is obtained; where ReLU represents the activation function and BN represents the batch normalization operation.
[0017] Preferably, based on step 5, the clustering results are evaluated, the boundaries of the plots are extracted, and the specific method of generating the plot distribution map is as follows:
[0018] Set the group processing iteration number threshold JU, repeat the group processing to obtain the corresponding intra-cluster error sum of squares Je, e represents the time sequence of the group processing to obtain the number of the intra-cluster error sum of squares;
[0019] Record the number of repeated grouping processes, calculate the difference between the grouping process iteration threshold and the number of repeated grouping processes, and obtain the remaining number of iterations P1; when the remaining number of iterations is zero, stop the grouping process; calculate the difference between the adjacent intra-cluster error sums of squares in chronological order to obtain the error neighbor value P2; calculate the variance of the intra-cluster error sums of squares obtained by the current repeated grouping process, and use the formula to obtain the significant change value P3; where represents the mean of the intra-cluster error sums of squares of repeated grouping processes;
[0020] The remaining number of iterations, the error neighbor value and the significant change value are weighted, and the cluster convergence value P is obtained using the formula P=P1×p1+P2×p2+P3×p3; where p1, p2, and p3 represent the weights corresponding to the remaining number of iterations, the error neighbor value and the significant change value, respectively;
[0021] Set the convergence threshold and compare the cluster convergence value with its convergence threshold. If the cluster convergence value is greater than or equal to its convergence threshold, it means that the cluster center has not reached the expected result; conversely, if the cluster convergence value is less than its convergence threshold, it means that the cluster center no longer changes significantly or reaches the expected effect.
[0022] The pixels that are clustered and belong to the same cluster are considered as part of the same field, thus identifying the boundaries of the field;
[0023] Use morphological operations or edge detection algorithms to refine field boundaries;
[0024] The boundaries of the identified fields are marked on the map to generate a field distribution map.
[0025] As a preferred embodiment, step 4 is to calculate and optimize the loss function of the field area, specifically:
[0026] Measure the difference between the clustering results and the actual field distribution, and optimize it, using the formula , we get the loss function M of the field; where v x represents the data point at position x in the field area, A represents the total number of data points in the field area, represents the reconstructed field characteristics;
[0027] The network weights are updated using an optimization algorithm based on the gradient of the loss function. The formula is: ;in, Respectively represent the network weights at the tth and t+1th time steps, represents the network weight, η represents the learning rate, represents the gradient, represents the gradient of the loss function.
[0028] Preferably, based on step three, the specific method of unsupervised learning is:
[0029] Use the clustering algorithm to group any pixel in the feature map using the formula: The sum of squared errors within the cluster J is calculated; where v x represents the data point in the field area at pixel point x, A represents the total number of data points in the field area, is the center of the y-th cluster and its value represents the mean of all points in the cluster, min y Represents each data point v x Find the distance Smallest cluster center , represents the Euclidean norm.
[0030] Preferably, the method of the present invention also provides a user interface, allowing the user to select a specific field area, input parameters, or view and export field distribution maps.
[0031] Compared with the related art, the unsupervised remote sensing field automatic extraction method based on the adaptive method provided by the present invention has the following beneficial effects:
[0032] 1. The present invention adopts an adaptive method to realize automatic extraction of remote sensing fields. By preprocessing the unified data format, an adaptive deep learning network is constructed to achieve efficient feature extraction, unsupervised learning and precise loss function calculation optimization, which greatly improves the accuracy and efficiency of extraction, reduces costs, and provides strong support for agricultural monitoring and management.
[0033] 2. The present invention can accurately extract the boundaries of fields and generate distribution maps. The way of evaluating clustering results is scientific and reasonable. The network is regularly updated to adapt to seasonal and crop cycle changes, ensuring the timeliness and practicality of the data. This solution improves the scientificity and accuracy of agricultural management. BRIEF DESCRIPTION OF THE DRAWINGS
[0034] Figure 1 A flowchart of an unsupervised remote sensing field automatic extraction method based on an adaptive method is provided in the present invention. DETAILED DESCRIPTION
[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0036] The terms used in this disclosure are for the purpose of describing specific embodiments only and are not intended to limit the disclosure. The singular forms of "group", "class" and "the" used in this disclosure and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.
[0037] It should be understood that although the terms first, second, third, etc. may be used in the present disclosure to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present disclosure, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".
[0038] Please refer to Figure 1 An unsupervised remote sensing field automatic extraction method based on an adaptive method comprises the following steps:
[0039] Step 1: collect remote sensing image data of the set field area and perform preprocessing;
[0040] Step 2: Use the preprocessed remote sensing image data to build an adaptive deep learning network to extract image features; organize the extracted image features into a feature map; each pixel in the feature map is represented by a feature vector; the construction of the adaptive deep learning network includes adaptive convolutional layers, multi-scale feature fusion and residual connections;
[0041] Step 3: Use feature maps in remote sensing image data for unsupervised learning;
[0042] Step 4: Calculate and optimize the loss function of the field area;
[0043] Step 5: Evaluate the clustering results, extract the boundaries of the fields, and generate a field distribution map;
[0044] Step six: Regularly acquire new remote sensing image data and update the adaptive deep learning network to adapt to seasonal changes and crop growth cycles.
[0045] In the present invention, based on step one, the specific steps of preprocessing include denoising, atmospheric correction, radiation correction, geometric correction, and converting remote sensing image data into a unified format and scale for subsequent processing.
[0046] In the present invention, the specific logic of constructing an adaptive deep learning network is:
[0047] Establish an adaptive convolution layer: Set the input feature map to I, where I represents a three-dimensional tensor with a size of H×W×C. H, W, and C represent the height, width, and number of channels of the remote sensing image, respectively. Each channel contains information of a specific spectral band. The adaptive convolution operation is expressed as: , and get the output feature map ; Among them, h' and w' represent the height and width of the output features after the convolution operation, and the corresponding calculation formula is expressed as: , , i and j represent the row and column positions of the convolution kernel when it slides on the input feature map, k represents the number corresponding to the kth channel of the input feature map, Pi and Pj represent the total number of padding added on both sides of the i-th row and j-th column of the input feature map respectively;
[0048] Multi-scale feature fusion: Multi-scale feature fusion technology combines input feature maps of different resolutions. is the input feature map of scale s, and the input feature maps of different scales are fused. The formula is: , and obtain the fused output feature map O; where Fuse represents the fusion operation, Indices representing different scales;
[0049] Residual connection: Let D be the input of a layer in the neural network, D′ be the output of the layer, and use batch normalization and activation function for processing. The formula is: Q=ReLU(BN(D′)+D), and the output Q of the residual connection is obtained; where ReLU represents the activation function and BN represents the batch normalization operation.
[0050] In the present invention, based on step 5, the clustering results are evaluated, the boundaries of the plots are extracted, and the specific method of generating the plot distribution map is as follows:
[0051] Set the group processing iteration number threshold JU, repeat the group processing to obtain the corresponding intra-cluster error sum of squares Je, e represents the time sequence of the group processing to obtain the number of the intra-cluster error sum of squares;
[0052] Record the number of repeated grouping processes, calculate the difference between the grouping process iteration threshold and the number of repeated grouping processes, and obtain the remaining number of iterations P1; when the remaining number of iterations is zero, stop the grouping process; calculate the difference between the adjacent intra-cluster error sums of squares in chronological order to obtain the error neighbor value P2; calculate the variance of the intra-cluster error sums of squares obtained by the current repeated grouping process, and use the formula to obtain the significant change value P3; where represents the mean of the intra-cluster error sums of squares of repeated grouping processes;
[0053] The remaining number of iterations, the error neighbor value and the significant change value are weighted, and the cluster convergence value P is obtained using the formula P=P1×p1+P2×p2+P3×p3; where p1, p2, and p3 represent the weights corresponding to the remaining number of iterations, the error neighbor value and the significant change value, respectively;
[0054] Set the convergence threshold and compare the cluster convergence value with its convergence threshold. If the cluster convergence value is greater than or equal to its convergence threshold, it means that the cluster center has not reached the expected result; conversely, if the cluster convergence value is less than its convergence threshold, it means that the cluster center no longer changes significantly or reaches the expected effect.
[0055] The pixels that are clustered and belong to the same cluster are considered as part of the same field, thus identifying the boundaries of the field;
[0056] Use morphological operations or edge detection algorithms to refine field boundaries;
[0057] The boundaries of the identified fields are marked on the map to generate a field distribution map.
[0058] In the present invention, step 4 is to calculate and optimize the loss function of the field area, specifically:
[0059] Measure the difference between the clustering results and the actual field distribution, and optimize it, using the formula , we get the loss function M of the field; where v x represents the data point at position x in the field area, A represents the total number of data points in the field area, represents the reconstructed field characteristics;
[0060] The network weights are updated using an optimization algorithm based on the gradient of the loss function. The formula is: ;in, Respectively represent the network weights at the tth and t+1th time steps, represents the network weight, η represents the learning rate, represents the gradient, represents the gradient of the loss function.
[0061] In the present invention, based on step three, the specific method of unsupervised learning is:
[0062] Use the clustering algorithm to group any pixel in the feature map using the formula: The sum of squared errors within the cluster J is calculated; where v x represents the data point in the field area at pixel point x, A represents the total number of data points in the field area, is the center of the y-th cluster and its value represents the mean of all points in the cluster, min y Represents each data point vx Find the distance Smallest cluster center , represents the Euclidean norm.
[0063] In the present invention, the method of the present invention also provides a user interface, allowing the user to select a specific field area, input parameters, or view and export field distribution maps.
[0064] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary techniques in the art that are not disclosed in this disclosure. The specification and examples are to be considered exemplary only, and the true scope and spirit of the present invention are indicated by the following claims.
[0065] It should be understood that the present invention is not limited to the exact construction that has been described above and shown in the drawings and that various modifications and changes may be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.
Claims
1. An unsupervised remote sensing field automatic extraction method based on an adaptive method, characterized in that: The following steps are involved: Step 1: collect remote sensing image data of the set field area and perform preprocessing; Step 2: Use the preprocessed remote sensing image data to build an adaptive deep learning network to extract image features. The specific logic of building an adaptive deep learning network is as follows: Establish an adaptive convolution layer: Set the input feature map to I, where I represents a three-dimensional tensor with a size of H×W×C. H, W, and C represent the height, width, and number of channels of the remote sensing image, respectively. Each channel contains information of a specific spectral band: The adaptive convolution operation is expressed as: O h,w,c =∑ i,j,k I h′,w′,k ·W i,j,k , get the output feature map O h,w,c ; Wherein, h' and w' respectively represent the height and width of the output feature after the convolution operation, and the corresponding calculation formula is: h'=h-i+1+2Pi, w'=w-j+1+2Pj, i and j represent the row and column positions when the convolution kernel slides on the input feature map, k represents the number corresponding to the kth channel of the input feature map, Pi and Pj represent the total number of padding added on both sides of the i-th row and j-th column of the input feature map respectively; Multi-scale feature fusion: Multi-scale feature fusion technology combines input feature maps of different resolutions. sn is the input feature map of scale s, and the input feature maps of different scales are fused. The formula is: O = Fuse (O s1 , O s2 , ..., O sn ), and obtain the fused output feature map O; where Fuse represents the fusion operation, O s1 , O s2 , ..., O sn Indices representing different scales; Residual connection: Let D be the input of a layer in the neural network, D′ be the output of the layer, and use batch normalization and activation function for processing. The formula is: Q = ReLU (BN (D′) + D), and the output Q of the residual connection is obtained; where ReLU represents the activation function and BN represents the batch normalization operation; The extracted image features are organized into a feature map, where each pixel in the feature map is represented by a feature vector; the construction of the adaptive deep learning network includes adaptive convolutional layers, multi-scale feature fusion, and residual connections; Step 3: Use the feature map in the remote sensing image data for unsupervised learning; the specific method of unsupervised learning is: Use the clustering algorithm to group any pixel in the feature map using the formula: The sum of squared errors within the cluster J is calculated; where v x represents the data point in the field area at pixel x, A represents the total number of data points in the field area, μ y is the center of the y-th cluster and its value represents the mean of all points in the cluster, min y Represents each data point v x Find the distance ||v x -μ y || 2 The smallest cluster center μ y , || || represents the Euclidean norm; Step 4: Calculate and optimize the loss function of the field area, specifically: Measure the difference between the clustering results and the actual field distribution, and optimize it, using the formula Get the loss function M of the field; where v x represents the data point at position x in the field area, A represents the total number of data points in the field area, represents the reconstructed field characteristics; The network weights are updated using an optimization algorithm based on the gradient of the loss function. The formula is: Among them, δ t+1 , δ t They represent the network weights at the tth and t+1th time steps, δ represents the network weight, η represents the learning rate, represents the gradient, represents the gradient of the loss function; Step 5: Evaluate the clustering results, extract the boundaries of the fields, and generate the field distribution map in the following way: Set the group processing iteration number threshold JU, repeat the group processing to obtain the corresponding intra-cluster error sum of squares Je, e represents the time sequence of the group processing to obtain the number of the intra-cluster error sum of squares; Record the number of repeated grouping processes, calculate the difference between the grouping process iteration threshold and the number of repeated grouping processes, and obtain the remaining number of iterations P1; when the remaining number of iterations is zero, stop the grouping process; calculate the difference between the adjacent intra-cluster error sums of squares in chronological order to obtain the error neighbor value P2; calculate the variance of the intra-cluster error sums of squares obtained by the current repeated grouping process, and use the formula to obtain the significant change value P3; where P3 represents the mean of the intra-cluster error sums of squares of repeated grouping processes; The remaining number of iterations, the error neighbor value and the significant change value are weighted, and the cluster convergence value P is obtained using the formula P = P1×p1+P2×p2+P3×p3; where p1, p2, and p3 represent the weights corresponding to the remaining number of iterations, the error neighbor value and the significant change value, respectively; Set the convergence threshold and compare the cluster convergence value with its convergence threshold. If the cluster convergence value is greater than or equal to its convergence threshold, it means that the cluster center has not reached the expected result; conversely, if the cluster convergence value is less than its convergence threshold, it means that the cluster center no longer changes significantly or reaches the expected effect. The pixels that are clustered and belong to the same cluster are considered as part of the same field, thus identifying the boundaries of the field; Use morphological operations or edge detection algorithms to refine field boundaries; Mark the boundaries of the identified fields on the map to generate a field distribution map; Step six: Regularly acquire new remote sensing image data and update the adaptive deep learning network to adapt to seasonal changes and crop growth cycles.
2. The unsupervised remote sensing field automatic extraction method based on an adaptive method according to claim 1 is characterized in that: Based on step one, the specific steps of preprocessing include denoising, atmospheric correction, radiation correction, geometric correction, and converting remote sensing image data into a unified format and scale for subsequent processing.
3. The unsupervised remote sensing field automatic extraction method based on an adaptive method according to claim 1 is characterized in that: The method also provides a user interface that allows the user to select specific field areas, input parameters, or view and export field distribution maps.
