Land change detection method based on matrix adaptive cross decomposition

Abstract

The invention discloses a land change detection method based on matrix adaptive cross decomposition, and relates to the technical field of land change detection, and the method specifically comprises the steps: obtaining a dual-time-phase land change image data set, and carrying out the preprocessing of the dual-time-phase land change image data set; for the preprocessed dual-temporal land change images, obtaining background dictionaries corresponding to two sub-images in each group of dual-temporal land change images by using an improved singular value decomposition algorithm, and further constructing a sample set; constructing a to-be-trained network comprising a feature extraction module and a global information module, and training the to-be-trained network based on each sample in the sample set to obtain a land change detection model; and based on the background dictionary of the target image, utilizing the land change detection model to obtain a land change detection result of the target image. According to the method, the calculation and memory overhead of global information modeling can be effectively reduced, and efficient and accurate land change detection is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of land change detection, and particularly to a land change detection method based on matrix adaptive cross decomposition. Background Art

[0002] With the development of society and the continuous improvement of the ability to utilize nature, a series of increasingly serious global problems such as soil erosion, river flooding, water body depletion, and environmental pollution have emerged, resulting in a continuous reduction in the quantity of land resources and a deterioration in quality.

[0003] To reveal the surface changes, domestic and foreign scholars use remote sensing technology and change detection algorithms for land change detection, and conduct quantitative analysis on remote sensing images of different time phases in the same area to achieve accurate detection of land changes. However, traditional land change detection methods have large computational and memory overheads. Therefore, how to obtain a change detection method that can not only maintain high detection performance but also reduce complexity remains a technical problem to be solved. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a land change detection method based on matrix adaptive cross decomposition.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A land change detection method based on matrix adaptive cross decomposition constructs a land change detection model according to steps S1 to S3, and obtains the land change detection result of the target image according to step A:

[0007] Step S1, obtain a dataset of dual-temporal land change images with a preset number of groups and land change regions marked, and preprocess each group of dual-temporal land change images in the dataset;

[0008] Step S2, for the preprocessed dual-temporal land change images, use an improved singular value decomposition algorithm to obtain the background dictionaries corresponding to the two sub-images in each group of dual-temporal land change images, and construct samples with the dual-temporal land change images, the land change regions corresponding to the dual-temporal land change images, and the background dictionaries corresponding to the two sub-images in the dual-temporal land change images respectively, and then obtain each sample to form a sample set;

[0009] Step S3, construct a network to be trained including a feature extraction module and a global information module, and based on each sample in the sample set, use the background dictionaries corresponding to the dual-temporal land change image and its two sub-images in the sample as inputs, and the land change region corresponding to the dual-temporal land change image as the output, and train the network to be trained to obtain a land change detection model;

[0010] Step A: Based on the background dictionary of the target image obtained in step S2, use the land change detection model to test the target image and obtain the land change detection result.

[0011] Furthermore, the data preprocessing in step S1 includes: geometric correction, radiometric correction, registration, denoising, cloud and shadow removal, band selection and fusion.

[0012] Furthermore, the step S2 uses an improved singular value decomposition algorithm to obtain the background dictionaries corresponding to the two sub-images in each group of double-temporal land change images, including the following steps:

[0013] Step S21: Obtain the original data matrix and the initial background dictionary of each group of double-temporal land change images after preprocessing, and use the orthogonal matching pursuit algorithm to calculate the sparse coefficient matrix γ, and then execute step S22;

[0014] Step S22: Based on the sparse coefficient matrix γ, use the singular value decomposition algorithm to update the initial background dictionary and the sparse coefficient matrix γ, and then execute step S23;

[0015] Step S23: If the termination condition is satisfied, output the updated background dictionary; otherwise, iterate and repeat step S22 until the termination condition is satisfied.

[0016] Furthermore, the step S22 includes:

[0017] Step S221: Based on the sparse coefficient matrix γ, define the non-zero index set of the sparse coefficient matrix as ω m , and calculate the error matrix E according to the following formula m :

[0018]

[0019] where y is the original data matrix, i is the index variable, and α i is the coefficient value in the sparse coefficient matrix, is the i-th atom in the dictionary;

[0020] Step S222: Based on the non-zero index set ω of the sparse coefficient matrix m , define the columns corresponding to ω in the error matrix that are not zero as E', and perform singular value decomposition on E' using the following formula to update the background dictionary and the sparse coefficient matrix: m

[0021] E' = U∑V T

[0022] where U is the left singular vector corresponding to each group of double-temporal land change images, and V T is the right singular vector corresponding to each group of double-temporal land change images.​

[0023] Further, the input end of the feature extraction module constitutes the input end of the land change detection model; the input end of the global information module is connected to the output end of the feature extraction module; the output end of the global information module constitutes the output end of the land change detection model;

[0024] The feature extraction module receives training samples, performs feature difference and feature fusion using a U-Net network, and outputs a tensor in matrix form; the global information module uses a matrix adaptive cross-approximate decomposition algorithm to decompose the tensor in matrix form and outputs a land change detection image.

[0025] Further, the matrix adaptive cross-approximate decomposition algorithm performs approximate estimation according to the following formula:

[0026]

[0027] where, Z m×n is an m×n-dimensional matrix, representing the feature matrix corresponding to each group of double-temporal land change images, is an approximate matrix, representing the error magnitude generated during the approximation process, r is the effective rank of matrix Z m×n , U m×r is a full-rank matrix of rank r, used to extract the main features related to land change in the feature matrix, V r×n is a full-rank matrix of rank r, used to combine and represent the main features.

[0028] Further, the matrix adaptive cross-approximate decomposition algorithm performs error control according to the following formula:

[0029]

[0030] where, Z m×n is an m×n-dimensional matrix, representing the feature matrix corresponding to each group of double-temporal land change images, is an approximate matrix, used to reflect the error magnitude generated during the approximation process, τ is the error iteration threshold, R m×n is the error matrix, ||·|| is the Frobenius norm of the matrix, r is the effective rank of matrix Z m×n .

[0031] Further, when training the land change detection model, it is jointly optimized through a total loss function including a cross-entropy loss function and a generalized dice loss function. The cross-entropy loss function, the generalized dice loss function, and the total loss function are as follows:

[0032]

[0033] L = C + G

[0034] Among them, C is the cross-entropy loss function, G is the generalized dice loss function, L is the total loss function, N is the total number of pixels in each group of dual-temporal land change images, L is the number of categories, is the one-hot encoded true value of category l at the nth position, is the predicted probability value of category l at the nth position, is the true pixel probability value of category l at the nth position, ω l is the weight of category l.

[0035] Furthermore, the accuracy rate, recall rate, and F1 score are used as evaluation indicators to test and evaluate the land change detection model:

[0036]

[0037] Among them, P is the accuracy rate, R is the recall rate, F is the F1 score, T p is the correct observation value, F p is the wrong observation value, F N is the missed observation value.

[0038] Beneficial effects brought by adopting the above technical solutions:

[0039] (1) The present invention uses the background dictionary to provide the land background feature reference for the model, facilitating the model to compare the features of dual-temporal land change images and locate the land change areas; and the background dictionary can be used to optimize the problem of background interference during model training, improve the detection accuracy and efficiency, reduce the misjudgment rate, and ensure the stable and efficient operation of the model in complex scenarios;

[0040] (2) The present invention combines matrix adaptive cross-decomposition and deep learning technology to achieve precise detection of land changes in dual-temporal land change images;

[0041] (3) The present invention uses the global attention mechanism of matrix adaptive cross-decomposition to reduce the computational and memory overhead of global information modeling and achieve efficient land change detection. Brief Description of the Drawings

[0042] Figure 1 is the flowchart of the present invention;

[0043] Figure 2 is the flowchart of matrix recombination calculation of the present invention;

[0044] Figure 3 is the flowchart of the matrix adaptive cross-approximation algorithm of the present invention;

[0045] Figure 4This is a comparison chart of the test results between the present invention and the comparative model;

[0046] Figure 5 This is a visualization chart of the training progress of the land change detection model in the application example of the present invention. Specific implementation manners

[0047] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings.

[0048] Reference Figure 1 , a land change detection method based on matrix adaptive cross decomposition, constructs a land change detection model according to steps S1 to S3, and obtains the land change detection result of the target image according to step A:

[0049] Step S1, based on the Arc Gsi map, obtain 9000 groups of double-temporal land change image datasets with labeled land change regions. One group of double-temporal land change images consists of single land change images of the same region at two different time points, and preprocess each group of double-temporal land change images in the dataset by using geometric correction, radiometric correction, registration, denoising, cloud and shadow removal, and band selection and fusion;

[0050] Step S2, for the preprocessed double-temporal land change images, use the improved singular value decomposition algorithm to obtain the background dictionaries corresponding to the two sub-images in each group of double-temporal land change images, and construct samples with the double-temporal land change images, the land change regions corresponding to the double-temporal land change images, and the background dictionaries corresponding to the two sub-images in the double-temporal land change images respectively, and then obtain each sample to form a sample set; in this embodiment, the background dictionary can accurately depict the land background features of the two sub-images in each group of double-temporal land change images and represent the background information of the land change images;

[0051] Step S3, construct a network to be trained including a feature extraction module and a global information module. Based on each sample in the sample set, use the background dictionaries corresponding to the double-temporal land change image and its two sub-images in the sample as the input, and the land change region corresponding to the double-temporal land change image as the output, and train the network to be trained to obtain a land change detection model; in this embodiment, using the background dictionary as the input can provide a reference for distinguishing the land change region, assist in separating the change region from the background region, and effectively enhance the resistance of the model to noise and complex environments, such as misjudgment caused by different seasonal lighting differences;

[0052] Step A1, based on the background dictionary of the target image obtained in step S2, use the land change detection model to test the target image to obtain the land change detection result.

[0053] Further, the steps of obtaining the background dictionaries corresponding to the two sub-images in each group of double-temporal land change images by using the improved singular value decomposition algorithm in step S2 are as follows:

[0054] Step S21: Obtain the original data matrix and the initial background dictionary of each group of double-temporal land change images after preprocessing, and use the orthogonal matching pursuit algorithm to calculate the sparse coefficient matrix γ, and then execute step S22;

[0055] Step S22: Based on the sparse coefficient matrix γ, use the singular value decomposition algorithm to update the initial background dictionary and the sparse coefficient matrix γ, and then execute step S23;

[0056] Step S23: If the termination condition is satisfied, output the updated background dictionary; otherwise, iterate and repeat step S22 until the termination condition is satisfied.

[0057] Further, the step S22 includes:

[0058] Step S221: Based on the sparse coefficient matrix γ, define the non-zero index set of the sparse coefficient matrix as ω m , and calculate the error matrix E according to the following formula m :

[0059]

[0060] where y is the original data matrix, i is the index variable, and α i is the coefficient value in the sparse coefficient matrix, is the i-th atom in the dictionary;

[0061] Step S222: Based on the non-zero index set ω of the sparse coefficient matrix m , define the columns corresponding to ω in the error matrix that are not zero as E', and perform singular value decomposition on E' using the following formula to update the background dictionary and the sparse coefficient matrix: m E' = U∑V

[0062] E' = U∑V T

[0063] where U is the left singular vector corresponding to each group of double-temporal land change images, used to capture the main change directions of land types, and V T is the right singular vector corresponding to each group of double-temporal land change images, used to further refine the analysis of land changes, such as distinguishing the detailed features of different land change types.

[0064] Further, the input end of the feature extraction module constitutes the input end of the land change detection model; the input end of the global information module is connected to the output end of the feature extraction module; the output end of the global information module constitutes the output end of the land change detection model;

[0065] The feature extraction module receives training samples, performs feature difference and feature fusion using a U-Net network, and outputs a tensor in matrix form; the global information module uses the matrix adaptive cross-approximate decomposition algorithm to decompose the tensor in matrix form and outputs a land change detection image; in this embodiment, the background dictionary is combined with matrix decomposition and adaptive propagation. The background dictionary can provide an initial feature reference for matrix decomposition, enabling the model to focus on land change-related features; in adaptive propagation, the background dictionary can enable effective transmission of feature information, effectively improving detection accuracy and stability.

[0066] Further, referring to Figure 2 and Figure 3 , the matrix adaptive cross-approximate decomposition algorithm performs approximate estimation according to the following formula:

[0067]

[0068] where Z m×n is an m×n-dimensional matrix representing the feature matrix corresponding to each group of double-temporal land change images, is an approximate matrix used to reflect the error magnitude generated during the approximation process, r is the effective rank of matrix Z m×n , U m×r is a full-rank matrix of rank r used to map the original high-dimensional feature space to a low-dimensional r-dimensional space and extract the main features related to land change in the feature matrix, and V r×n is a full-rank matrix of rank r used to combine and represent the main features.

[0069] Further, the matrix adaptive cross-approximate decomposition algorithm performs error control according to the following formula:

[0070]

[0071] where Z m×n is an m×n-dimensional matrix representing the feature matrix corresponding to each group of double-temporal land change images, is an approximate matrix used to reflect the error magnitude generated during the approximation process, τ is the error iteration threshold, R m×n is the error matrix, ||·|| is the Frobenius norm of the matrix. By calculating the norm of the error matrix and the norm of the original matrix and comparing them with the error iteration threshold, the error degree between the approximate matrix and the original matrix can be quantified, and r is the rank of matrix Zm×n Effective rank.

[0072] Furthermore, when training the land change detection model, it is jointly optimized through a total loss function including a cross-entropy loss function and a generalized dice loss function. The cross-entropy loss function, the generalized dice loss function, and the total loss function are as follows:

[0073]

[0074] L = C + G

[0075] where C is the cross-entropy loss function, G is the generalized dice loss function, L is the total loss function, N is the total number of pixels in each group of double-temporal land change images, L is the number of categories, is the one-hot encoded true value of category l at the nth position, is the predicted probability value of category l at the nth position, is the true pixel probability value of category l at the nth position, ω l is the weight of category l.

[0076] Furthermore, the accuracy rate, recall rate, and F1 score are used as evaluation indicators to test and evaluate the land change detection model:

[0077]

[0078] where P is the accuracy rate, R is the recall rate, F is the F1 score, T p is the correct observation value, F p is the wrong observation value, F N is the missed observation value.

[0079] The technical effects of the present invention will be further described in detail below in combination with experimental results.

[0080] As Figure 4 shown, when comparing the present invention with the comparative model, compared with the comparative model, the method proposed by the present invention can accurately identify the change area of the target double-temporal land change image and achieve accurate detection of land changes.

[0081] As Figure 5As shown, remote sensing images of the target city in 2010 and 2020 were obtained, and the remote sensing images were preprocessed using geometric correction, radiometric correction, registration, denoising, cloud and shadow removal, and band selection and fusion. The background dictionary corresponding to the remote sensing images was obtained using an improved singular value algorithm, and the land change detection model was further trained. The changes in accuracy and loss value during the training process can be observed from the figure. The accuracy of the trained land change detection model fluctuates between 50% and 60%, and the loss value drops from nearly 0.9 to about 0.7 and gradually stabilizes.

[0082] The above are only the preferred embodiments of the present invention, and do not impose any limitations on the present invention. Any simple modifications, changes, and equivalent structural changes made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A land change detection method based on matrix self-adaptive cross decomposition, characterized in that: Construct a land change detection model according to steps S1 to S3, and obtain the land change detection result of the target image according to step A: Step S1: Obtain a dataset of dual-temporal land change images with a preset number of groups and marked land change regions, and preprocess each group of dual-temporal land change images in the dataset; Step S2: For the preprocessed dual-temporal land change images, use an improved singular value decomposition algorithm to obtain the background dictionaries corresponding to the two sub-images in each group of dual-temporal land change images, and construct samples with the dual-temporal land change images, the land change regions corresponding to the dual-temporal land change images, and the background dictionaries corresponding to the two sub-images in the dual-temporal land change images respectively, and then obtain each sample to form a sample set; Step S3: Construct a network to be trained including a feature extraction module and a global information module. Based on each sample in the sample set, use the background dictionaries corresponding to the dual-temporal land change image and its two sub-images in the sample as the input, and the land change region corresponding to the dual-temporal land change image as the output, and train the network to be trained to obtain a land change detection model; Step A: Based on the background dictionary of the target image obtained in step S2, use the land change detection model to test the target image to obtain the land change detection result.

2. The land change detection method based on matrix adaptive cross decomposition according to claim 1, characterized in that, The data preprocessing described in step S1 includes: geometric correction, radiometric correction, registration, denoising, cloud and shadow removal, band selection and fusion.

3. A land change detection method based on matrix adaptive cross decomposition according to claim 1, characterized in that, The step S2 uses an improved singular value decomposition algorithm to obtain the background dictionaries corresponding to the two sub-images in each group of dual-temporal land change images, including the following steps: Step S21: Obtain the original data matrix and the initial background dictionary of each group of preprocessed dual-temporal land change images, use the orthogonal matching pursuit algorithm to calculate the sparse coefficient matrix γ, and then execute step S22; Step S22: Based on the sparse coefficient matrix γ, use the singular value decomposition algorithm to update the initial background dictionary and the sparse coefficient matrix γ, and then execute step S23; Step S23: If the termination condition is satisfied, output the updated background dictionary; otherwise, iterate and repeat step S22 until the termination condition is satisfied.

4. A method for land change detection based on matrix adaptive cross decomposition according to claim 3, characterized in that The step S22 includes: Step S221: Based on the sparse coefficient matrix γ, define the non-zero index set of the sparse coefficient matrix as ω m , and calculate the error matrix E according to the following formula m : where y is the original data matrix, i is the index variable, and α i is the coefficient value in the sparse coefficient matrix, and is the i-th atom in the background dictionary; Step S222, based on the non-zero index set ω of the sparse coefficient matrix m , define the columns corresponding to ω in the error matrix as E', and perform singular value decomposition on E' using the following formula to update the background dictionary and the sparse coefficient matrix: m where the columns for which ω is not zero are E', and update the background dictionary and the sparse coefficient matrix by performing singular value decomposition on E' using the following formula: E' = UΣV T Among them, U is the left singular vector corresponding to each group of double-temporal land change images, and V T is the right singular vector corresponding to each group of double-temporal land change images.

5. A method for land change detection based on matrix adaptive cross decomposition according to claim 1, characterized in that, The input end of the feature extraction module constitutes the input end of the land change detection model; the input end of the global information module is connected to the output end of the feature extraction module; the output end of the global information module constitutes the output end of the land change detection model; The feature extraction module receives training samples, performs feature difference and feature fusion using a U-Net network, and outputs a tensor in matrix form; The global information module uses the matrix adaptive cross-approximate decomposition algorithm to decompose the tensor in matrix form and outputs the land change detection image.

6. The land change detection method based on matrix adaptive cross decomposition according to claim 5, characterized in that The matrix adaptive cross-approximate decomposition algorithm performs approximate estimation according to the following formula: Among them, Z m×n is an m×n dimensional matrix, representing the feature matrix corresponding to each group of double-temporal land change images. is an approximation matrix, representing the error magnitude generated during the approximation process. r is the effective rank of matrix Z m×n and U m×r is a full-rank matrix of rank r, used to extract the main features related to land change in the feature matrix. V r×n is a full-rank matrix of rank r, used to combine and represent the main features.

7. A method for land change detection based on matrix adaptive cross decomposition according to claim 6, characterized in that, The matrix adaptive cross-approximate decomposition algorithm performs error control according to the following formula: Among them, Z m×n is an m×n dimensional matrix, representing the feature matrix corresponding to each group of dual-temporal land change images, is the approximation matrix, representing the error magnitude generated during the approximation process, τ is the error iteration threshold, and R m×n is the error matrix, ||·|| is the Frobenius norm of the matrix, and r is the effective rank of the matrix Z m×n of.

8. A method for land change detection based on matrix adaptive cross decomposition according to claim 1, characterized in that When training the land change detection model, it is jointly optimized through a total loss function including a cross-entropy loss function and a generalized dice loss function. The cross-entropy loss function, the generalized dice loss function, and the total loss function are as follows: L = C + G Among them, C is the cross-entropy loss function, G is the generalized dice loss function, L is the total loss function, N is the total number of pixels in each group of dual-temporal land change images, L is the number of categories, is the one-hot encoded true value of category l at the nth position, is the predicted probability value of category l at the nth position, is the true pixel probability value of category l at the nth position, ω l is the weight of category l.

9. A land change detection method based on matrix adaptive cross decomposition according to claim 1, characterized in that Using accuracy, recall, and F1-score as evaluation metrics, the land change detection model is tested and evaluated: Among them, P is the accuracy rate, R is the recall rate, F is the F1 score, T p is the correct observation value, F p is the incorrect observation value, F N is the missing observation value.

Images