Unsupervised learning forest fire change detection method and device integrating prior knowledge

Abstract

The invention discloses an unsupervised learning forest fire change detection method and device incorporating prior knowledge, the method comprising: obtaining dual-phase remote sensing images of a fire area before and after a forest fire and performing preprocessing; respectively calculating NBRSWIR indexes to obtain NBRSWIR index graphs; performing uncertainty analysis to obtain training samples; using the training samples to train two symmetrical deep network branches; using the trained deep network branches to respectively extract initial features of the NBRSWIR index graphs; performing slow feature analysis on the initial features to obtain feature differences of the initial features; calculating a chi-square distance to obtain a change intensity graph; performing K-means threshold segmentation to obtain forest fire areas. The invention provides a novel forest fire change detection framework, which not only incorporates prior geological knowledge and improves the spectral separation capability of a fire area and background information, but also suppresses background information changes to extract more complex features of the fire area, and still ensures classification accuracy without manually annotating training samples.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and remote sensing image processing, and mainly solves the problem of forest fire change detection based on remote sensing images, and specifically relates to an unsupervised learning forest fire change detection method and device that incorporates prior knowledge. Background Art

[0002] Forest fire is one of the most common disasters. It not only seriously affects the stability of forest ecosystems, but also brings environmental pollution and related secondary geological disasters, threatening human life and property safety. Therefore, the acquisition of information such as the location of the fire area and the burned area is crucial for the assessment and restoration of economic and ecological losses. However, due to the limitations of complex terrain and bad weather conditions, traditional ground surveys are usually difficult and costly. Satellite remote sensing can cover a large area at a high frequency and provide invisible spectral information, playing an important role in fire monitoring and burn area mapping. However, since the characteristics of the fire area are often complex and there are spectral similarities between different land cover categories, although domestic and foreign scholars have conducted a lot of research in this area, there is currently no universal method with strong universality. Therefore, forest fire remote sensing image change detection has always been a hot spot and difficulty in the international remote sensing field.

[0003] Among the traditional forest fire change detection algorithms, the most widely used methods are those based on spectral index and those based on image classification. The spectral index-based method mathematically combines the bands according to the differences in spectral information of different types of objects in different wavelength ranges to achieve the classification of fire areas and other types of objects. There are two main indexes that can be used for forest fire change detection: vegetation index and fire index. Due to its simple implementation and high accuracy, this method has been widely used in fire area detection. However, since it is usually based on threshold segmentation or clustering methods for binarization to achieve fire area detection, and the background information changes are often complex, this method often leads to unnecessary false detection or missed detection. Image classification-based methods separate fire areas and regions by minimizing intra-class differences and maximizing inter-class differences using a set of attributes, such as support vector machines (SVM), random forests (RF), principal component analysis (PCA), etc. However, the key step of these algorithms is the selection of attribute features, which takes a lot of time.

[0004] In recent years, the rapid development of deep learning technology has promoted the progress of forest fire change detection. At present, it is mainly based on the convolutional neural network (CNN) method, and more features can be automatically extracted. However, due to the complexity of the causes of forest fires and the uncertainty of wind direction, the boundary features of the fire area are often more complex. At the same time, due to the spectral similarity between the fire area and water bodies, smoke, bare land and other objects, the background information has a greater impact. However, existing methods rarely consider the above issues. In addition, there are few publicly available forest fire datasets, and supervised deep learning methods require a large number of datasets for training. Summary of the invention

[0005] The main technical problem to be solved by the present invention is to provide a forest fire change detection method and device based on unsupervised learning that incorporates prior knowledge, so as to ensure the accuracy of forest fire change detection without manually labeled training samples.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is as follows:

[0007] According to one aspect of the present invention, there is provided an unsupervised learning forest fire change detection method incorporating prior knowledge, comprising the following steps:

[0008] S1: Acquire dual-temporal remote sensing images of the fire area before and after the forest fire, and preprocess the dual-temporal remote sensing images;

[0009] S2: Calculate the NBRSWIR index of the pre-processed dual-temporal remote sensing images respectively to incorporate prior knowledge, and obtain NBRSWIR index graphs X and Y of the dual-temporal remote sensing images;

[0010] S3: Perform uncertainty analysis on the NBRSWIR index graphs X and Y to obtain training samples X train , Y train ;

[0011] S4: training two symmetrical deep network branches using the training samples, and obtaining trained deep network branches after the training is completed;

[0012] S5: Use the trained deep network branch to extract the initial features X of the NBRSWIR index graphs X and Y respectively φ , Y φ ;

[0013] S6: For the initial feature X φ , Y φ Perform slow feature analysis to obtain the initial feature X φ , Y φ The characteristic difference of

[0014] S7: Calculate the chi-square distance of each pixel point according to the feature difference to obtain a change intensity map;

[0015] S8: Perform K-means threshold segmentation on the change intensity map to obtain the final forest fire area.

[0016] Furthermore, in step S1, the preprocessing step includes:

[0017] S11: The ESA Copernicus Data Center downloads the L1C-level multispectral data of the Sentinel-2 in the forest fire area before and after the disaster;

[0018] S12: Use the sen2cor tool to perform radiometric calibration and atmospheric correction on the L1C multispectral data, thereby obtaining L2A products;

[0019] S13: Using SNAP software to perform super-resolution synthesis on the L2A-level products, synthesizing all bands into bands with a spatial resolution of 10m, and then obtaining pre-disaster and post-disaster remote sensing images with a resolution of 10m;

[0020] S14: The pre-disaster and post-disaster remote sensing images with a resolution of 10 m are respectively clipped according to the scope of the study area, so as to obtain the pre-processed dual-temporal remote sensing images.

[0021] Furthermore, in step S2, the two shortwave infrared bands of the pre-processed dual-temporal remote sensing image are subjected to band operation to obtain the NBRSWIR index map before and after the fire. The specific calculation formula is:

[0022]

[0023] Among them, the NBRSWIR index is a new fire index, and SWIR1 and SWIR2 are the 11th and 12th band data of the preprocessed dual-phase remote sensing image.

[0024] Furthermore, step S3 specifically includes:

[0025] S31: Difference the NBRSWIR index graphs X and Y before and after the fire to obtain a NBRSWIR index difference graph reflecting the fire area information. The specific calculation formula is:

[0026] d NBRSWIR =NBRSWIR post -NBRSWIR pre

[0027] Among them, NBRSWIR post NBRSWIR is the NBRSWIR index map of remote sensing images after the fire. pre is the NBRSWIR index map of remote sensing images before the fire, dNBRSWIR This is the difference map of NBRSWIR index before and after the fire;

[0028] S32: performing fuzzy C-means clustering on the NBRSWIR index difference map to achieve threshold segmentation, and dividing the study area into a confirmed burned area, an uncertain area, and a confirmed unburned area;

[0029] S33: Randomly select pixels in the NBRSWIR index map before and after the fire in the unburned area as training samples X train , Y train .

[0030] Furthermore, step S32 specifically includes:

[0031] S321: Set the precision e of the objective function, the fuzzy index m, the number of clusters c and the maximum number of iterations t of the algorithm. The objective function J is:

[0032]

[0033] The constraints of the objective function are:

[0034]

[0035]

[0036] Where c represents the number of clusters, n represents the total number of pixels in the NBRSWIR index difference map, m represents the fuzzy index, and u represents the ij Represents sample x j The membership degree of class i, j represents the jth pixel, x represents the sample represented by the NBRSWIR index difference map, i represents the i-th cluster, v i represents the center of class i, and d() represents the distance measure;

[0037] S322: Randomly initialize the membership matrix u ij and cluster center v i ;

[0038] S323: Update the membership matrix and cluster center, specifically:

[0039]

[0040]

[0041] Among them, k represents the kth cluster, v k represents the center of k class;

[0042] S324: If the objective function satisfies |J(t)-J(t+1)|<e, the iteration ends and goes to step S325, otherwise repeat step S323;

[0043] S325: According to the obtained membership matrix, the class corresponding to the maximum sample membership is taken as the result of sample clustering, and the clustering is completed, and the samples are further divided into a certain burned area, an uncertain area, and a certain unburned area.

[0044] Furthermore, step S4 specifically includes:

[0045] S41: Construct two symmetrical deep network branches, both of which are fully connected layers, including input layer, hidden layer and output layer, and each hidden layer has the same number of nodes;

[0046] S42: Initialize the parameters {θ1,θ2} of two symmetric deep network branches;

[0047] S43: Calculate the training samples X before and after the fire respectively train , Y train Projected features after transformation by deep network branches

[0048] S44: Calculate the loss function. According to the slow feature analysis theory, the unchanged quantity has the smallest eigenvalue. Therefore, the variance of the unchanged pixels is suppressed by minimizing the total square of all eigenvalues. The loss function Specifically:

[0049]

[0050] Where:

[0051]

[0052]

[0053] in, After centralization n represents the number of pixels, is a matrix whose elements are all 1, I represents the identity matrix, r represents the regularization constant, (.) T Represents the transposed matrix of a matrix, (.) -1 is the inverse matrix of the matrix, tr(.) is the trace of the matrix;

[0054] S45: Calculate the gradient: and

[0055] S46: Update parameters {θ1,θ2} using gradient descent algorithm;

[0056] S47: Repeat steps S43-S46 until the maximum number of iterations is reached to obtain a trained deep network branch.

[0057] Furthermore, step S43 specifically includes:

[0058] S431: For the training sample X before the fire train :

[0059] The output of the first hidden layer is expressed as: in m represents the number of bands, and n represents the number of pixels; represents the weight matrix, represents the bias vector, h1 is the number of nodes in the first hidden layer; s(·) represents the activation function;

[0060] The output of each hidden layer l is then expressed as: in

[0061] Finally, the projection features after the network conversion are: Where o represents the number of nodes in the output layer, is the weight matrix, h l is the number of hidden layer nodes, is the deviation vector;

[0062] S432: For training sample Y after fire train , the projected features after transformation by the deep network branch The process is the same as step S431.

[0063] Furthermore, step S6 specifically includes:

[0064] S61: Obtaining initial features X before and after the fire φ , Y φ The mean and standard deviation

[0065] S62: Standardize feature X according to the mean and standard deviation φ , Y φ , specifically:

[0066] For the initial characteristics X before the fire φ , using the formula The initial features obtained after the normalized deep network conversion, where Yes X φ The normalized value of the i-th pixel; X φ i Yes X φ The value of the i-th pixel; Yes X φ The mean of each pixel value; Yes X φStandard deviation of each pixel value; initial feature Y after fire φ The standardization process is the same as;

[0067] S63: Based on standardized features Get feature X φ , Y φ The covariance matrix A of the difference and the features X φ and Y φ The covariance matrix and matrix B of are specifically:

[0068] The matrix A is:

[0069] The matrix B is:

[0070] in, and are the normalized values ​​of the i-th pixel of the initial feature of the remote sensing image NBRSWIR index map before and after the fire after conversion by the deep network, and P is the total number of pixels;

[0071] S64: solving the generalized eigenvalues ​​of matrix A relative to matrix B and the eigenvectors corresponding to the generalized eigenvalues, sorting the corresponding eigenvectors from small to large according to the size of the generalized eigenvalues, and obtaining an ordered eigenvector matrix;

[0072] S65: Use an ordered feature vector matrix to transform the standardized features Project it into the feature space and obtain X φ , Y φ The characteristic difference is:

[0073]

[0074] Among them, SFA represents the characteristic difference, w represents the ordered eigenvector matrix, Represents the initial feature X after the NBRSWIR index map of the remote sensing image before and after the fire is converted by the deep network φ , Y φ A matrix of normalized values ​​for each pixel.

[0075] Furthermore, in step S7, the step of calculating the chi-square distance of each pixel point includes:

[0076] Divide the square of each pixel feature difference by the square root of the corresponding generalized eigenvalue. The calculation formula is as follows:

[0077]

[0078] Among them, ch i is the chi-square distance of the i-th pixel, SFA iis the characteristic difference of the i-th pixel, and λ is the generalized eigenvalue of the band.

[0079] According to another aspect of the present invention, the present invention provides an unsupervised learning forest fire change detection device incorporating prior knowledge, comprising the following modules:

[0080] An image preprocessing module, used to obtain dual-temporal remote sensing images of the fire area before and after the forest fire, and to preprocess the dual-temporal remote sensing images;

[0081] The NBRSWIR index calculation module is used to calculate the NBRSWIR index of the pre-processed dual-temporal remote sensing images respectively to incorporate prior knowledge and obtain NBRSWIR index graphs X and Y of the dual-temporal remote sensing images;

[0082] The training sample acquisition module is used to perform uncertainty analysis on the NBRSWIR index graphs X and Y to obtain training samples X train , Y train ;

[0083] A network training module, used to train two symmetrical deep network branches through the training samples, and after the training is completed, obtain the trained deep network branches;

[0084] The initial feature extraction module is used to use the trained deep network branch to extract the initial features X of the NBRSWIR index graphs X and Y respectively. φ , Y φ ;

[0085] A slow feature analysis module, used for performing slow feature analysis on the initial feature to obtain a feature difference value of the initial feature;

[0086] A chi-square distance calculation module, used to calculate the chi-square distance of each pixel point according to the feature difference to obtain a change intensity map;

[0087] The threshold segmentation module is used to perform K-means threshold segmentation on the change intensity map to obtain the final forest fire area.

[0088] The technical solution provided by the present invention has the following beneficial effects:

[0089] A forest fire change detection framework is proposed. By incorporating prior knowledge into the NBRSWIR index, the spectral separation ability of the fire area and the complex background is improved. At the same time, the powerful feature extraction ability of the deep network is combined to extract a more complete fire area. An unsupervised symmetric deep network with strong generalization ability is provided. More reliable pseudo samples are generated through uncertainty analysis based on the fire index map. The changes in complex background information are suppressed by combining the deep network and slow feature analysis (SFA) to extract the fire area. The accuracy of forest fire change detection is guaranteed without manually labeled training samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0091] Figure 1 It is a flow chart of a forest fire change detection method in a specific implementation manner of the present invention;

[0092] Figure 2 It is an overall framework diagram of the forest fire change detection method in a specific implementation manner of the present invention;

[0093] Figure 3 It is a flowchart of the slow feature analysis process in a specific implementation manner of the present invention;

[0094] Figure 4 This is a comparison result diagram of fire burnt area detection on the Xichang fire dataset in a specific implementation manner of the present invention;

[0095] Figure 5 It is a structural diagram of a forest fire change detection device in a specific implementation manner of the present invention. DETAILED DESCRIPTION

[0096] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0097] DETAILED DESCRIPTION OF THE EMBODIMENTS A method and apparatus for detecting forest fire changes by unsupervised learning incorporating prior knowledge is disclosed.

[0098] Please refer to Figure 1-2 , Figure 1 This is a flow chart of a forest fire change detection method in a specific implementation manner of the present invention. Figure 2 It is an overall framework diagram of the forest fire change detection method in a specific implementation manner of the present invention;

[0099] An unsupervised learning forest fire change detection method that incorporates prior knowledge. The specific steps are as follows:

[0100] S1. Obtain dual-temporal remote sensing images of the fire area before and after the forest fire, and preprocess the dual-temporal remote sensing images.

[0101] First, the Copernicus Data Center of the European Space Agency (https: / / scihub.copernicus.eu / dhus / # / home) downloads the L1C multispectral data of the Sentinel-2 in the forest fire area before and after the disaster according to the location of the fire area; the sen2cor tool is used to perform radiation calibration and atmospheric correction on the L1C multispectral data to obtain the L2A product; then the SNAP software is used to perform super-resolution synthesis on the L2A product, and all bands are synthesized into bands with a spatial resolution of 10m; finally, the pre-disaster and post-disaster remote sensing images are respectively cut into images of the same size according to the scope of the study area, and then the pre-processed dual-phase remote sensing images are obtained. This embodiment uses a total of three forest fire data sets, but only the Xichang fire data set is analyzed and discussed in this example. On March 30, 2020, the Xichang fire occurred in Xichang City, Liangshan Prefecture, Sichuan Province. The image imaging time before the fire was March 25, 2020, and the image imaging time after the fire was April 4, 2020. The surface coverage of the image is mainly forest area, with a spatial resolution of 10m, a total of 12 bands, and a size of 953 × 1501. The spectral image contains fire-burned areas, vegetation, buildings, bare land, smoke and other types of land features. The fire-burned area accounts for 17% of the entire study area.

[0102] S2: Calculate the NBRSWIR index of the preprocessed dual-temporal remote sensing images to incorporate prior knowledge and obtain the NBRSWIR index graphs X and Y of the dual-temporal remote sensing images;

[0103] The NBRSWIR index maps before and after the fire were obtained by performing band calculation on the two shortwave infrared bands of the preprocessed dual-temporal remote sensing image. The NBRSWIR index maps before and after the fire have only one band. The specific calculation method is as follows:

[0104]

[0105] Among them, the NBRSWIR index (Normalized Burned Ratio SWIR) is a new fire index, and SWIR1 and SWIR2 are the 11th and 12th band data of the preprocessed dual-temporal remote sensing image, respectively.

[0106] S3: Perform uncertainty analysis on the NBRSWIR index graphs X and Y to obtain training samples X train , Y train ;

[0107] Step S3 specifically includes:

[0108] S31: Difference the NBRSWIR index graphs X and Y before and after the fire to obtain a NBRSWIR index difference graph reflecting the fire area information. The specific calculation formula is:

[0109] d NBRSWIR =NBRSWIR post -NBRSWIR pre

[0110] Among them, NBRSWIR post NBRSWIR is the NBRSWIR index map of remote sensing images after the fire. pre is the NBRSWIR index map of remote sensing images before the fire, d NBRSWIR This is the difference map of NBRSWIR index before and after the fire;

[0111] S32: performing fuzzy C-means clustering on the NBRSWIR index difference map to achieve threshold segmentation, and dividing the study area into a confirmed burned area, an uncertain area, and a confirmed unburned area;

[0112] Furthermore, S32 specifically includes:

[0113] S321, set the precision e of the objective function to 0.0001, the fuzzy index m to 2, the number of clusters c to 3 and the maximum number of iterations t of the algorithm to 1000, and the objective function J is:

[0114]

[0115] The constraints of the objective function are:

[0116]

[0117]

[0118] Where c represents the number of categories, n represents the total number of pixels in the NBRSWIR index difference map described in S31, m represents the fuzzy index, and u represents ij Represents sample x j The membership degree of class i, j represents the jth pixel, x represents the sample represented by the NBRSWIR index difference map described in S31, i represents the i-th cluster, v i represents the center of class i, and d() represents the distance measure;

[0119] S322, randomly initialize the membership matrix u ij and cluster center v i ;

[0120] S323, updating the membership matrix and cluster center, specifically:

[0121]

[0122]

[0123] Among them, k represents the kth cluster, v k represents the center of k class;

[0124] S324: If the objective function satisfies |J(t)-J(t+1)|<e, the iteration ends and goes to step S325, otherwise repeat step S323;

[0125] S325: According to the obtained membership matrix, the class corresponding to the maximum sample membership is taken as the result of sample clustering, and the clustering is completed, and the samples are further divided into a certain burned area, an uncertain area, and a certain unburned area.

[0126] S33: Since the purpose of the present invention is to suppress the change of background information and then extract the fire area, using unchanged pixels (determining the unburned area) as training samples is helpful to improve the final accuracy. 2.5% of the pixels in the NBRSWIR index map before and after the fire in the unburned area are randomly selected as training samples X train , Y train .

[0127] S4: The training sample X described in S3 train , Y train The two symmetrical deep network branches are trained, and after the training is completed, two trained symmetrical deep network branches are obtained;

[0128] S4 specifically includes:

[0129] S41: Construct two symmetrical deep network branches, both of which are fully connected layers, including an input layer, three hidden layers and an output layer, and each hidden layer has the same number of nodes;

[0130] S42: Initialize the parameters of the dual-branch deep network {θ1,θ2}, where θ1 includes and θ2 includes and

[0131] S43: Calculate the training samples X before and after the fire respectively train , Y train Projected features after deep network transformation Specifically:

[0132] Take the pre-fire training sample as an example:

[0133] The output of the first hidden layer can be expressed as: in m represents the number of bands, and n represents the number of pixels; represents the weight matrix, represents the bias vector, h1 is the number of nodes in the first hidden layer, and s(·) represents the activation function.

[0134] Then the output of each hidden layer l can be expressed as: in

[0135] Finally, the projection features after the network conversion are: Where o represents the number of nodes in the output layer, is the weight matrix, h l is the number of hidden layer nodes, is the deviation vector;

[0136] Post-fire training sample Y train The process is the same as the pre-fire training samples.

[0137] S44: Calculate the loss function. According to the slow feature analysis theory, the unchanged quantity has the smallest eigenvalue. Therefore, the variance of the unchanged pixels can be suppressed by minimizing the total square of all eigenvalues. The loss function Specifically:

[0138]

[0139] Where:

[0140]

[0141]

[0142] in, After centralization n represents the number of pixels, is a matrix whose elements are all 1, I represents the identity matrix, r represents the regularization constant, (.) T Represents the transposed matrix of a matrix, (.) -1 is the inverse matrix of the matrix, tr(.) is the trace of the matrix;

[0143] S45: Calculate the gradient: and

[0144] S46: Update parameters using gradient descent algorithm;

[0145] S47: To ensure the convergence of the model, repeat steps S43-S46 until the maximum number of iterations is reached.

[0146] S5: Use the two symmetrical deep network branches trained in S4 to extract the initial features X of the spectral index maps X and Y of the remote sensing images before and after the fire obtained in S2. φ , Y φ ;

[0147] S6: For the initial feature X φ , Y φ Perform slow feature analysis (SFA), project it into feature space, and obtain the initial feature X φ , Y φ The feature difference of the non-fire area pixels will be suppressed in this feature difference space, so the separability between the fire area pixels and the non-fire area pixels is enhanced. The specific steps of slow feature analysis are as follows: Figure 3 As shown, including:

[0148] S61: Obtaining initial features X before and after the fire φ , Y φ The mean and standard deviation

[0149] S62: Standardize feature X by mean and standard deviation φ , Y φ , specifically:

[0150] The initial characteristics before the fire are X φ For example, using the formula The initial features obtained after the normalized deep network conversion, where Yes X φ The normalized value of the i-th pixel; X φ i Yes X φ The value of the i-th pixel; Yes X φ The mean of each pixel value; Yes X φ Standard deviation of each pixel value; initial feature Y after fire φ The solution process is similar;

[0151] S63: Based on standardized features Get feature X φ , Y φ The covariance matrix A of the difference and the features X φ and Y φ The covariance matrix and matrix B of are specifically:

[0152] The matrix A is:

[0153] The matrix B is:

[0154] in, and are the normalized values ​​of the i-th pixel of the initial feature of the remote sensing image NBRSWIR index map before and after the fire after conversion by the deep network, and P is the total number of pixels;

[0155] S64: solving the generalized eigenvalues ​​of matrix A relative to matrix B and the eigenvectors corresponding to the generalized eigenvalues, sorting the corresponding eigenvectors from small to large according to the size of the generalized eigenvalues, and obtaining an ordered eigenvector matrix;

[0156] S65: Use an ordered feature vector matrix to transform the standardized features Project it into the feature space and obtain X φ , Y φ The characteristic difference is:

[0157]

[0158] Where SFA represents the characteristic difference, w represents the ordered eigenvector matrix, Represents the initial feature X after the NBRSWIR index map of the remote sensing image before and after the fire is converted by the deep network φ , Y φ A matrix of normalized values ​​for each pixel.

[0159] S7: Calculate the chi-square distance of each pixel point according to the characteristic difference value obtained in S6 to obtain a change intensity map. Specifically, divide the square of the characteristic difference value of each pixel point by the square root of the corresponding generalized eigenvalue. The calculation formula is as follows:

[0160]

[0161] Among them ch i is the chi-square distance of the i-th pixel, SFA i is the characteristic difference of the i-th pixel, and λ is the generalized eigenvalue of the band.

[0162] S8: Perform K-means threshold segmentation on the change intensity map described in S7 to obtain the final forest fire area.

[0163] In order to verify the unsupervised learning forest fire change detection method incorporating prior knowledge of the present invention, in this embodiment, the classification results are also analyzed and evaluated.

[0164] Figure 4 The fire area detection results of different methods on the Xichang fire dataset are shown. Figure 4 (a) is a composite image of the red band (R), near infrared band (NIR), and short-wave infrared band (SWIR1) after the fire. Figure 4 (b) is the real surface change label used for reference (black areas are non-fire burned areas, white areas are fire burned areas), Figure 4 (c)- Figure 4 (h) Classification diagrams of algorithms BAI (Burn Area Index), NBR (Normalized Burn Ratio), CVA (Change Vector Analysis), PCA (Principal Component Analysis), SFA (Slow Feature Analysis), and DSFA (Deep Slow Feature Analysis). Figure 4 (i) is the classification diagram of the method proposed in this experiment. By comparison, it can be seen that the results of the method proposed in this invention are closest to the real landmark change labels, effectively reducing the misclassification of fire areas and landforms such as smoke, water, and vegetation, and can detect a relatively complete fire burn area. Table 1 shows the comparison of the accuracy evaluation results of different methods. Overall, it can be seen that the method proposed in this experiment has the best detection accuracy, which illustrates the effectiveness of the method of the present invention.

[0165] Table 1 is a comparison of the accuracy evaluation results of each classification method

[0166]

[0167] Please refer to Figure 5 , Figure 5 It is a structural diagram of a forest fire change detection device in a specific implementation manner of the present invention.

[0168] An unsupervised learning forest fire change detection device incorporating prior knowledge is used to implement the steps of the unsupervised learning forest fire change detection method described above, specifically comprising the following modules:

[0169] An image preprocessing module 1 is used to obtain dual-temporal remote sensing images of the fire area before and after the forest fire, and to preprocess the dual-temporal remote sensing images;

[0170] NBRSWIR index calculation module 2, used for respectively calculating NBRSWIR index of the pre-processed dual-temporal remote sensing images to incorporate prior knowledge and obtain NBRSWIR index graphs X and Y of the dual-temporal remote sensing images;

[0171] The training sample acquisition module 3 is used to perform uncertainty analysis on the NBRSWIR index graphs X and Y to obtain training samples X. train , Y train ;

[0172] A network training module 4 is used to train two symmetrical deep network branches through the training samples, and after the training is completed, a trained deep network branch is obtained;

[0173] The initial feature extraction module 5 is used to use the trained deep network branch to extract the initial features X of the NBRSWIR index graphs X and Y respectively. φ , Y φ ;

[0174] Slow feature analysis module 6, used to perform slow feature analysis on the initial feature to obtain the initial feature X φ , Y φ The characteristic difference of

[0175] A chi-square distance calculation module 7, used to calculate the chi-square distance of each pixel point according to the feature difference to obtain a change intensity map;

[0176] The threshold segmentation module 8 is used to perform K-means threshold segmentation on the change intensity map to obtain the final forest fire area.

[0177] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0178] The serial numbers of the embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments. In a unit claim that lists several means, several of these means may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order and these words may be interpreted as identifiers.

[0179] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An unsupervised learning forest fire change detection method incorporating prior knowledge, characterized in that: It includes the following steps: S1: Obtain the two-temporal remote sensing images of the fire area before and after the forest fire, and preprocess the two-temporal remote sensing images; S2: Calculate the NBRSWIR index for the preprocessed two-temporal remote sensing images respectively to incorporate prior knowledge, and obtain the NBRSWIR index maps X and Y of the two-temporal remote sensing images; S3: Perform uncertainty analysis on the NBRSWIR index graphs X and Y to obtain training samples X train , Y train ; S4: Train two symmetric deep network branches with the training samples. After the training is completed, obtain the trained deep network branches; S5: Use the trained deep network branch to extract the initial features X of the NBRSWIR index graphs X and Y respectively φ , Y φ ; S6: For the initial feature X φ , Y φ Perform slow feature analysis to obtain the initial feature X φ , Y φ The characteristic difference of S7: Calculate the chi-square distance of each pixel point according to the feature difference to obtain the change intensity map; S8: Perform K-means threshold segmentation on the change intensity map to obtain the final forest fire area.

2. The unsupervised learning forest fire change detection method incorporating prior knowledge as claimed in claim 1, characterized in that: In step S1, the steps of the preprocessing include: S11: Download the L1C-level multispectral data of Sentinel-2 before and after the forest fire in the fire area from the Copernicus Data Center of the European Space Agency; S12: Use the sen2cor tool to perform radiometric calibration and atmospheric correction on the L1C-level multispectral data, and then obtain the L2A-level product; S13: Use the SNAP software to perform super-resolution synthesis on the L2A-level product, synthesize all bands into a band with a spatial resolution of 10m, and then obtain the pre-disaster and post-disaster remote sensing images with a resolution of 10m; S14: Crop the pre-disaster and post-disaster remote sensing images with a resolution of 10m according to the study area range respectively, and then obtain the preprocessed two-temporal remote sensing images.

3. The unsupervised learning forest fire change detection method incorporating prior knowledge as claimed in claim 1, characterized in that: In step S2, perform band operation on the two shortwave infrared bands of the preprocessed two-temporal remote sensing images to obtain the NBRSWIR index maps of the pre-fire and post-fire phases. The specific calculation formula is: Among them, the NBRSWIR index is a new fire index, and SWIR1 and SWIR2 are the data of the 11th and 12th bands of the preprocessed two-temporal remote sensing images respectively.

4. The unsupervised learning forest fire change detection method incorporating prior knowledge as claimed in claim 1, characterized in that: Step S3 specifically includes: S31: Subtract the NBRSWIR index maps X and Y before and after the fire to obtain the NBRSWIR index difference map reflecting the fire area information. The specific calculation formula is: d NBRSWIR =NBRSWIR post -NBRSWIR pre Among them, NBRSWIR post NBRSWIR is the NBRSWIR index map of remote sensing images after the fire. pre is the NBRSWIR index map of remote sensing images before the fire, d NBRSWIR This is the difference map of NBRSWIR index before and after the fire; S32: Perform fuzzy C-means clustering on the NBRSWIR index difference map to achieve threshold segmentation, and divide the study area into a determined burned area, an uncertain area, and a determined unburned area; S33: Randomly select pixels in the NBRSWIR index map before and after the fire in the unburned area as training samples X train , Y train .

5. The unsupervised learning forest fire change detection method incorporating prior knowledge as claimed in claim 4, characterized in that: Step S32 specifically includes: S321: Set the accuracy e of the objective function, the fuzzy index m, the number of clusters c, and the maximum number of iterations t of the algorithm. The objective function J is: The constraint conditions of the objective function are: Where c represents the number of clusters, n represents the total number of pixels in the NBRSWIR index difference map, m represents the fuzzy index, and u represents the ij Represents sample x j The membership degree of class i, j represents the jth pixel, x represents the sample represented by the NBRSWIR index difference map, i represents the i-th cluster, v i represents the center of class i, and d() represents the distance measure; S322: Randomly initialize the membership matrix u ij and cluster center v i ; S323: Update the membership matrix and the cluster centers, specifically: Among them, k represents the kth cluster, v k represents the center of k class; S324: If the objective function satisfies |J(t)-J(t + 1)|<e, the iteration ends and proceeds to step S325; otherwise, repeat step S323; S325: According to the obtained membership matrix, take the class corresponding to the maximum sample membership as the result of sample clustering. After the clustering ends, the samples are divided into a determined burned area, an uncertain area, and a determined unburned area.

6. The unsupervised learning forest fire change detection method incorporating prior knowledge as claimed in claim 1, characterized in that: Step S4 specifically includes: S41: Construct two symmetric deep network branches, both of which are fully connected layers, including an input layer, a hidden layer, and an output layer. Each hidden layer has the same number of nodes; S42: Initialize the parameters {θ1,θ2} of two symmetric deep network branches; S43: Calculate the training samples X before and after the fire respectively train , Y train Projected features after transformation by deep network branches S44: Calculate the loss function. According to the slow feature analysis theory, the unchanged quantity has the smallest eigenvalue. Therefore, the variance of the unchanged pixels is suppressed by minimizing the total square of all eigenvalues. The loss function Specifically: Where: in, After centralization n represents the number of pixels, is a matrix whose elements are all 1, I represents the identity matrix, r represents the regularization constant, (.) T Represents the transposed matrix of a matrix, (.) -1 is the inverse matrix of the matrix, tr(.) is the trace of the matrix; S45: Calculate the gradient: and S46: Update parameters {θ1,θ2} using gradient descent algorithm; S47: Repeat steps S43-S46 until the maximum number of iterations is reached to obtain a trained deep network branch.

7. The unsupervised learning forest fire change detection method incorporating prior knowledge as claimed in claim 6, characterized in that: Step S43 specifically includes: S431: For the training sample X before the fire train : The output of the first hidden layer is expressed as: in m represents the number of bands, and n represents the number of pixels; represents the weight matrix, represents the bias vector, h1 is the number of nodes in the first hidden layer; s(·) represents the activation function; The output of each hidden layer l is then expressed as: in Finally, the projection features after the network conversion are: Where o represents the number of nodes in the output layer, is the weight matrix, h l is the number of hidden layer nodes, is the deviation vector; S432: For training sample Y after fire train , the projected features after transformation by the deep network branch The process is the same as step S431.

8. The unsupervised learning forest fire change detection method incorporating prior knowledge according to claim 1 is characterized in that: Step S6 specifically includes: S61: Obtaining initial features X before and after the fire φ , Y φ The mean and standard deviation S62: Standardize feature X according to the mean and standard deviation φ , Y φ , specifically: For the initial characteristics X before the fire φ , using the formula The initial features obtained after the normalized deep network conversion, where Yes X φ The normalized value of the i-th pixel; Yes X φ The value of the i-th pixel; Yes X φ The mean of each pixel value; Yes X φ Standard deviation of each pixel value; initial feature Y after fire φ The standardization process is the same as; S63: Based on standardized features Get feature X φ , Y φ The covariance matrix A of the difference and the features X φ and Y φ The covariance matrix and matrix B of are specifically: The matrix A is: The matrix B is: in, and are the normalized values ​​of the i-th pixel of the initial feature of the remote sensing image NBRSWIR index map before and after the fire after conversion by the deep network, and P is the total number of pixels; S64: solving the generalized eigenvalues ​​of matrix A relative to matrix B and the eigenvectors corresponding to the generalized eigenvalues, sorting the corresponding eigenvectors from small to large according to the size of the generalized eigenvalues, and obtaining an ordered eigenvector matrix; S65: Use an ordered feature vector matrix to transform the standardized features Project it into the feature space and obtain X φ , Y φ The characteristic difference is: Among them, SFA represents the characteristic difference, w represents the ordered eigenvector matrix, Represents the initial feature X after the NBRSWIR index map of the remote sensing image before and after the fire is converted by the deep network φ , Y φ A matrix of normalized values ​​for each pixel.

9. The unsupervised learning forest fire change detection method incorporating prior knowledge according to claim 1 is characterized in that: In step S7, the step of calculating the chi-square distance of each pixel point includes: Divide the square of each pixel feature difference by the square root of the corresponding generalized eigenvalue. The calculation formula is as follows: Among them, ch i is the chi-square distance of the ith pixel, SFA i is the characteristic difference of the i-th pixel, and λ is the corresponding generalized eigenvalue.

10. An unsupervised learning forest fire change detection device incorporating prior knowledge, characterized in that: Includes the following modules: An image preprocessing module, used to obtain dual-temporal remote sensing images of the fire area before and after the forest fire, and to preprocess the dual-temporal remote sensing images; The NBRSWIR index calculation module is used to calculate the NBRSWIR index of the pre-processed dual-temporal remote sensing images respectively to incorporate prior knowledge and obtain NBRSWIR index graphs X and Y of the dual-temporal remote sensing images; The training sample acquisition module is used to perform uncertainty analysis on the NBRSWIR index graphs X and Y to obtain training samples X train , Y train ; A network training module, used to train two symmetrical deep network branches through the training samples, and after the training is completed, obtain the trained deep network branches; The initial feature extraction module is used to use the trained deep network branch to extract the initial features X of the NBRSWIR index graphs X and Y respectively. φ , Y φ ; A slow feature analysis module is used to perform slow feature analysis on the initial feature to obtain the initial feature X φ , Y φ The characteristic difference of A chi-square distance calculation module, used to calculate the chi-square distance of each pixel point according to the feature difference to obtain a change intensity map; The threshold segmentation module is used to perform K-means threshold segmentation on the change intensity map to obtain the final forest fire area.

Images