Semi-supervised change detection method for high-resolution remote sensing images based on transfer learning

Abstract

The present invention belongs to the field of image processing technology, and discloses a semi-supervised high-resolution remote sensing image change detection method based on transfer learning, wherein the model parameters obtained by training in the full supervision process are used as the initial model parameters in the unsupervised learning process; in the unsupervised learning, the output feature map of the deep feature extraction module is subjected to random perturbation differentiation to obtain multiple perturbation feature maps, and the output feature map of the deep feature extraction module and the multiple perturbation feature maps are respectively obtained by a decoder to obtain multiple predicted change maps containing change information, thereby obtaining the model parameters after unsupervised learning; transfer learning is performed based on the model parameters after unsupervised learning to obtain new model parameters; and the new model parameters are used to perform change detection on the target domain data. The present invention applies the deep learning model obtained by training in the source domain to the target domain through the transfer learning technology, thereby helping the change detection model in the target domain to have better detection performance.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a semi-supervised high-resolution remote sensing image change detection method based on transfer learning. Background Art

[0002] Remote sensing image change detection is widely used in land use monitoring, urban planning, natural disaster monitoring, etc. With the deepening of deep learning technology in the field of image processing and analysis, the existing remote sensing image change detection is mainly focused on the fully supervised deep neural network model, which requires a large number of labeled samples to support the training process of the network model. However, there are certain differences in the changes detected in the remote sensing image change detection requirements in different fields, resulting in the need for a large number of different labeled samples in different fields.

[0003] With the rapid development of remote sensing detection technology, it is easier to obtain a large number of unlabeled remote sensing images in the application of remote sensing change detection. Since existing deep learning algorithms mostly rely on a large amount of labeled data for training, this requires a lot of time and computing power. In some tasks, labeled data may be very scarce, making it difficult to train the model from scratch. However, in other similar tasks, there is a large amount of labeled data, and a better model can be obtained after training. Transfer learning The knowledge learned in the source domain can be transferred to the target domain. Using transfer learning of deep learning, the weight parameters of the model do not need to be trained from scratch, which can not only significantly reduce the training time, but also reduce the demand for labeled data.

[0004] Existing transfer learning mainly uses the fully supervised transfer learning process. However, due to the large differences between the source and target domains, inter-domain offsets, or interference from multiple factors, the effect transfer is not effective. Summary of the invention

[0005] In view of the domain deviation between the source domain and the target domain in the remote sensing image change detection task in transfer learning, the present invention proposes a semi-supervised high-resolution remote sensing image change detection method based on transfer learning. Compared with the full supervision-transfer learning process, the present invention adopts a fully supervised-unsupervised semi-supervised model in the source domain and introduces an unlabeled training process in the source domain, which can better adapt to the model optimization of a small number of samples in the target domain; and for the feature difference analysis in change detection, a feature value rotation-permutation method is proposed to adapt to the difference in the understanding of the change target between the source domain and the target domain.

[0006] The present invention is implemented by the following technical solutions: A semi-supervised high-resolution remote sensing image change detection method based on transfer learning, including a fully supervised learning process, an unsupervised learning process and a transfer learning process;

[0007] Source domain data includes labeled data and unlabeled data;

[0008] The fully supervised learning process uses the label data of the source domain, uses an encoder to obtain the hidden features of the labeled images before and after the change, and calculates the feature difference map. The feature difference map is extracted by the deep feature extraction module to obtain the output feature map of the deep feature extraction module. The output feature map is decoded by the decoder to obtain the predicted change map containing the change information.

[0009] The model parameters trained in the fully supervised process are used as the initial model parameters in the unsupervised learning process;

[0010] The unsupervised learning process uses the unlabeled data of the source domain, uses an encoder to obtain the hidden features of the unlabeled images before and after the change, and calculates the feature difference map. After feature extraction of the feature difference map, the output feature map of the deep feature extraction module is obtained. The output feature map of the deep feature extraction module is differentiated by random perturbation to obtain multiple perturbation feature maps. The output feature map of the deep feature extraction module and the multiple perturbation feature maps are respectively obtained by a decoder to obtain multiple predicted change maps containing change information; the model parameters after unsupervised learning are obtained;

[0011] Transfer learning is performed based on the model parameters after unsupervised learning. The transfer learning process uses a small amount of labeled data and some unlabeled data in the target field to perform fully supervised learning and unsupervised learning again to obtain new model parameters; the new model parameters are used to detect changes in the target field data.

[0012] Further preferably, the deep feature extraction module includes n residual groups, the input of residual group 1 is the feature difference map, the output of residual group j-1 is used as the input of residual group j, j=2,…,n-1,n; the outputs of residual group 1 to residual group n are input into the global feature fusion module, and the result output by the global feature fusion module is added to the feature difference map to obtain the output feature map of the deep feature extraction module.

[0013] Further preferably, the residual group consists of L residual attention modules and 1 3×3 convolution; first, the two 3×3 convolutions in the residual attention module are connected by a ReLU activation function, and secondly, a coordinate attention module is embedded in the residual group; the input feature is , represents the output feature map of the L-1th residual attention module in the Kth residual group, and the output feature map of the Lth residual attention module in the Kth residual group for:

[0014] ;

[0015] in, represents the Lth residual attention module in the Kth residual group, represents the L-1th residual attention module in the Kth residual group, represents the first residual attention module in the Kth residual group.

[0016] Further preferably, in the global feature fusion module, the input feature is first processed by a 1×1 convolution operation and then sent to the channel shuffling module; the channel shuffling module subdivides the feature map into 4 sub-feature maps and re-divides each sub-feature map to obtain Represents the generated sub-feature map, i , Represents the convolution processing corresponding to each sub-feature map, represent Output of With the previous convolutional layer Output Add them together and then send the result to the next convolutional layer to be processed.

[0017] Further preferably, in the unsupervised learning process, the feature difference map is divided into two branches after passing through the deep feature extraction module, and the first branch directly outputs the output feature map of the deep feature extraction module. , After the decoder, a predicted change map containing change information is obtained ; The second branch extracts the output feature map of the deep feature extraction module After random perturbation processing, NP perturbation feature maps are obtained ,..., , the perturbation feature map is decoded to obtain NP predicted change maps containing change information , ..., ,in Represents the output feature map of the deep feature extraction module After random perturbation processing, the first perturbation feature map is obtained. Represents the output feature map of the deep feature extraction module After random perturbation processing, the NPth perturbation feature graph is obtained. represent The predicted change graph obtained by decoding, represent The predicted change graph obtained by decoding;

[0018] Further preferably, in the unsupervised process, the random perturbation processing includes the eigenvalue interval rotation processing:

[0019] ;

[0020] ;

[0021] in, represents the eigenvalue of the uth row and vth column of the feature map before rotation, represents the intermediate eigenvalue of the uth row and vth column of the feature map during rotation processing, represents the eigenvalue of the uth row and vth column of the feature map after rotation, represents an eigenvalue interval, Indicates the number of interval spans for rotation, Indicates the range of characteristic values; Indicates the range of eigenvalues.

[0022] Further preferably, in the unsupervised process, the random perturbation processing includes the eigenvalue interval replacement processing, which divides the eigenvalue range into 8 intervals on average, expressed as: , represents the first eigenvalue interval, represents the second eigenvalue interval, represents the third eigenvalue interval, represents the fourth eigenvalue interval, represents the fifth eigenvalue interval, represents the sixth eigenvalue interval, represents the 7th eigenvalue interval, Indicates the 8th eigenvalue interval; eigenvalue interval , then , , ..., ; Indicates the maximum eigenvalue; Indicates the minimum eigenvalue; perform two eigenvalue interval permutations, and the order after the first permutation is ; The order after the second permutation is .

[0023] Predicted change graph output by the first branch As the baseline, each predicted change map output by the second branch is compared with the baseline and the loss is calculated.

[0024] Further optimization, using cross entropy loss as the full supervision loss .

[0025] Further preferably, in the unsupervised stage, the unlabeled dataset is defined as ,in is the i-th pair of unlabeled variation images, is the i-th unlabeled image before the change, is the i-th unlabeled changed image, For the size of the unlabeled dataset, an unsupervised loss based on unlabeled data is adopted , using unlabeled variation image pairs The obtained feature difference map ,and As a set of feature maps after random perturbation, is the unlabeled variation image pair The corresponding i-th perturbation feature map after random perturbation, is the unlabeled variation image pair The corresponding p-th perturbation feature map after random perturbation, is the unlabeled variation image pair The corresponding NP-th perturbation feature map after random perturbation;

[0026] Through the decoder Get the predicted change graph , through NP auxiliary decoders Process each disturbance feature map to obtain the corresponding predicted change map :

[0027] ;

[0028] ;

[0029] The unsupervised loss is:

[0030] ;

[0031] in is a distance metric.

[0032] Further optimization, the overall loss function The definition is as follows:

[0033] ;

[0034] in is a regularization constant.

[0035] The present invention targets the deep learning model of changes in high-resolution remote sensing images. The deep learning model trained in the source domain is applied to the target domain through transfer learning technology, so as to help the change detection model in the target domain have better detection performance.

[0036] The present invention considers a large amount of unlabeled data in the target field, establishes a semi-supervised model in the source field, enables the model to better adapt to the unlabeled data in the source field, and then enables the model to better adapt to the unlabeled data in the target field. At the same time, in the unsupervised random perturbation process, the remote sensing image change detection mainly analyzes the difference in features to detect the target. In order to adapt to the difference in understanding of the changed target between the source field and the target field, a eigenvalue rotation-permutation method is proposed. On the basis of retaining a certain relationship between eigenvalues, the strength of the eigenvalue is changed to adapt to the difference in understanding of the changed target.

[0037] The present invention uses labeled data on the data set of the source domain for fully supervised learning, uses unlabeled data for unsupervised learning, and adopts a semi-supervised learning method based on consistency regularization to extract the semantic information of unlabeled data. In the process of feature extraction and fusion, the present invention proposes a deep feature extraction module and a global feature fusion module to extract richer semantic information from remote sensing images and enhance the learning and generalization capabilities of the model; in the random perturbation processing of unsupervised learning, the present invention proposes a feature value rotation and permutation method to improve the model's ability to adapt to the deviation between the source domain and the target domain. The overall network uses a joint training method of fully supervised learning and unsupervised learning, which can not only reduce the number of labeled data used in the source domain, but also use unsupervised learning to improve the model's ability to adapt to unlabeled data in the target domain; at the same time, the overall network uses a small amount of labeled data sets in the target domain to optimize the model, and applies the model to the target domain to solve the problem of few samples in the target domain. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is the overall framework diagram of the present invention.

[0039] Figure 2 Schematic diagram of the deep feature extraction module.

[0040] Figure 3 It is the residual group structure diagram.

[0041] Figure 4 Schematic diagram of the global feature fusion module.

[0042] Figure 5 Schematic diagram of eigenvalue interval permutation processing. DETAILED DESCRIPTION

[0043] 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 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 those skilled in the art without creative work are within the scope of protection of the present invention.

[0044] In view of the different needs for remote sensing image change detection in different industries and fields, the present invention uses a partially annotated dataset in the source field to establish a semi-supervised change detection deep learning model, uses a small amount of annotated dataset in the target field to optimize the model, and applies the model to the target field to solve the problem of few samples in the target field.

[0045] Reference Figure 1 , a semi-supervised high-resolution remote sensing image change detection method based on transfer learning, including a fully supervised learning process, an unsupervised learning process and a transfer learning process;

[0046] Source domain data includes labeled data and unlabeled data;

[0047] The fully supervised learning process uses the label data of the source domain, uses an encoder to obtain the hidden features of the labeled images before and after the change, and calculates the feature difference map. The feature difference map is extracted by the deep feature extraction module to obtain the output feature map of the deep feature extraction module. The output feature map is decoded by the decoder to obtain the predicted change map containing the change information.

[0048] The model parameters (including the parameters of the encoder, deep feature extraction module, and decoder) obtained through full supervision are used as the initial model parameters in the unsupervised learning process;

[0049] The unsupervised learning process uses the unlabeled data in the source domain, uses an encoder to obtain the hidden features of the unlabeled images before and after the change, and calculates the feature difference map. After feature extraction on the feature difference map, the output feature map of the deep feature extraction module is obtained. The output feature map of the deep feature extraction module is differentiated by random perturbation to obtain multiple perturbation feature maps. The output feature map of the deep feature extraction module and the multiple perturbation feature maps are respectively obtained by a decoder to obtain multiple predicted change maps containing change information; the adaptability of the method to unlabeled data is enhanced through the unsupervised learning process, and the model parameters after unsupervised learning (including new parameters of the encoder, deep feature extraction module, and decoder) are obtained;

[0050] Transfer learning is performed based on the model parameters after unsupervised learning. The transfer learning process uses a small amount of labeled data and some unlabeled data in the target field to perform fully supervised learning and unsupervised learning again to obtain new model parameters. The new model parameters are used to detect changes in the target field data.

[0051] The loss calculation of the learning process includes full supervision loss and unsupervised loss. The loss calculation is used to select the predicted change graph that best meets expectations.

[0052] The present invention uses a consistency regularization technique in the semi-supervised learning process formed by fully supervised learning and unsupervised learning, so that the model produces similar prediction distributions on disturbed data points, and continuously optimizes and iterates in similar prediction distributions.

[0053] In this embodiment, the fully supervised learning process is as follows:

[0054] The input image pair is the labeled image before the change And the changed image , respectively, using the encoder extract and The features of the image before the change are obtained And the image feature map after the change , two encoders Share weights and calculate and The difference between them is the feature difference map .

[0055] ;

[0056] ;

[0057] Construct deep feature extraction modules (such as Figure 2 The feature difference diagram For further feature extraction, the deep feature extraction module includes n residual groups, and the input of residual group 1 is the feature difference map , the output of residual group j-1 is used as the input of residual group j, j=2,…,n-1,n; the output of residual group 1 to residual group n is input to the global feature fusion module, and the result of the global feature fusion module output is compared with the feature difference map Add together to get the output feature map of the deep feature extraction module .

[0058] The residual group in the deep feature extraction module (such as Figure 3 shown) by The residual attention module and one 3×3 convolution are composed. First, the two 3×3 convolutions in the residual attention module are connected through the ReLU activation function to improve the model's expressiveness and alleviate the gradient vanishing problem. Secondly, in order to further explore the semantic information of the feature map in the spatial and channel dimensions, a coordinate attention module is embedded in the residual group. , is the output feature map of the L-1th residual attention module in the Kth residual group, and the output feature map of the Lth residual attention module in the Kth residual group for:

[0059] ;

[0060] in, represents the Lth residual attention module in the Kth residual group, and the other subscripts are deduced accordingly. represents the L-1th residual attention module in the Kth residual group, Represents the first residual attention module in the Kth residual group. The residual connection in the residual group can reduce the loss of feature information to a certain extent to output a more accurate feature map, thereby improving the performance of the network model.

[0061] The global feature fusion module in the deep feature extraction module (such as Figure 4 As shown in Figure 1, the input feature is first processed by a 1×1 convolution operation and then sent to the channel shuffle module. The channel shuffle module subdivides the feature map into 4 sub-feature maps and re-divides each sub-feature map to promote information interaction between different parts. Represents the generated sub-feature map, i , Represents the convolution processing corresponding to each sub-feature map, represent The output of the sub-feature map With the previous convolutional layer Output Add them together and then send the result to the next convolutional layer to be processed.

[0062] Output of the deep feature extraction module Using a decoder Get the predicted change graph containing change information .Decoder It is a series of sub-pixel convolution upsampling modules to achieve a predicted change map with the same spatial resolution as the input image. The process inside the decoder can be:

[0063] ;

[0064] The unsupervised learning process is as follows:

[0065] The input image pair is the unlabeled before-change image And the changed image , respectively, using the encoder extract and The features of the image before the change are obtained And the image feature map after the change , and calculate and The difference between them is the feature difference map .

[0066] Feature Difference Map After passing through the deep feature extraction module, it is divided into two branches. The first branch directly outputs the output feature map of the deep feature extraction module. , After the decoder, a predicted change map containing change information is obtained ; The second branch extracts the output feature map of the deep feature extraction module After random perturbation processing, NP perturbation feature maps are obtained ,..., These perturbation feature maps are decoded to obtain NP predicted change maps containing change information ,..., .in Represents the output feature map of the deep feature extraction module After random perturbation processing, the first perturbation feature map is obtained. Represents the output feature map of the deep feature extraction module After random perturbation processing, the NPth perturbation feature graph is obtained. represent The predicted change graph obtained by decoding, represent The predicted change graph obtained by decoding;

[0067] In addition to the usual geometric transformations such as scaling, rotation, and cropping, random perturbation processing also includes eigenvalue interval rotation processing and eigenvalue interval replacement processing, specifically:

[0068] (1) Divide the eigenvalue range into 8 intervals, expressed as: , represents the first eigenvalue interval, represents the second eigenvalue interval, represents the third eigenvalue interval, represents the fourth eigenvalue interval, represents the fifth eigenvalue interval, represents the sixth eigenvalue interval, represents the 7th eigenvalue interval, Indicates the 8th eigenvalue interval; eigenvalue interval , then , , ..., ; Indicates the maximum eigenvalue; Indicates the minimum eigenvalue;

[0069] (2) Random perturbation processing includes eigenvalue interval rotation processing:

[0070] ;

[0071] ;

[0072] in, represents the eigenvalue of the uth row and vth column of the feature map before rotation, represents the intermediate eigenvalue of the uth row and vth column of the feature map during rotation processing, represents the eigenvalue of the uth row and vth column of the feature map after rotation, represents an eigenvalue interval, Indicates the number of interval spans for rotation, Indicates the maximum eigenvalue; Indicates the minimum eigenvalue.

[0073] (3) Eigenvalue interval replacement processing: perform eigenvalue interval replacement processing twice (such as Figure 5 ), where the order after the first permutation is ; The order after the second permutation is .when The interval is replaced by When the interval position is, the eigenvalue calculation formula is:

[0074] ;

[0075] in, Indicates being in The eigenvalue of the interval row u and column v, Indicates being in The eigenvalue of the interval row u and column v, and express Interval and Minimum value of the interval.

[0076] Predicted change graph output by the first branch As the baseline, each predicted change map output by the second branch is compared with the baseline and the loss is calculated.

[0077] The overall loss function is designed as follows:

[0078] In the fully supervised learning stage, the labeled dataset is defined as , using mathematical notation as ,in{ } represents the i-th pair of images with label changes, is the i-th labeled image before the change, is the i-th labeled changed image, is the true change mask of the i-th image, is the size of the label dataset. The present invention uses cross entropy loss (CE) as the full supervision loss , calculate the predicted change map With the real change mask The loss can be expressed in mathematical notation as:

[0079] ;

[0080] In the unsupervised stage, the unlabeled dataset is defined as ,in is the i-th pair of unlabeled variation images, is the i-th unlabeled image before the change, is the i-th unlabeled changed image, is the size of the unlabeled dataset, which is usually assumed to be much larger than the size of the labeled dataset The present invention adopts an unsupervised loss based on unlabeled data. , which provides an additional training signal to optimize the parameters of the overall model using unlabeled variation image pairs The obtained feature difference map ,and As a set of feature maps after random perturbation, is the unlabeled variation image pair The corresponding i-th perturbation feature map after random perturbation, is the unlabeled variation image pair The corresponding p-th perturbation feature map after random perturbation, is the unlabeled variation image pair The corresponding NP-th perturbation feature map after random perturbation.

[0081] Through the decoder Get the predicted change graph , similarly, by designing a set of Similar NP auxiliary decoders Process each disturbance feature map to obtain the corresponding predicted change map .

[0082] ;

[0083] ;

[0084] The unsupervised loss is:

[0085] ;

[0086] in is a distance metric that measures the predicted change graph and The difference between Choose to use mean square error (MSE).

[0087] Overall loss function The definition is as follows:

[0088] ;

[0089] in is a regularization constant, which is the number of training iterations In particular, From 0 to 1, just like the first iteration The same Gaussian function is used for the remaining iterations. Keep it at 1.

[0090] The transfer learning process includes a small amount of labeled data and some unlabeled data in the target domain. The model parameters obtained by fully supervised learning and unsupervised learning of the source domain data set are trained again using a small amount of labeled data and some unlabeled data to obtain new model parameters. The new model parameters are used to detect changes in the data in the target domain.

[0091] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A semi-supervised high-resolution remote sensing image change detection method based on transfer learning, including a fully supervised learning process, an unsupervised learning process, and a transfer learning process; Features: Source domain data includes labeled data and unlabeled data; The fully supervised learning process uses the label data of the source domain, uses an encoder to obtain the hidden features of the labeled images before and after the change, and calculates the feature difference map. The feature difference map is extracted by the deep feature extraction module to obtain the output feature map of the deep feature extraction module. The output feature map is decoded by the decoder to obtain the predicted change map containing the change information. The model parameters trained in the fully supervised process are used as the initial model parameters in the unsupervised learning process; The unsupervised learning process uses the unlabeled data of the source domain, uses an encoder to obtain the hidden features of the unlabeled images before and after the change, and calculates the feature difference map. After feature extraction of the feature difference map, the output feature map of the deep feature extraction module is obtained. The output feature map of the deep feature extraction module is differentiated by random perturbation to obtain multiple perturbation feature maps. The output feature map of the deep feature extraction module and the multiple perturbation feature maps are respectively obtained by a decoder to obtain multiple predicted change maps containing change information; the model parameters after unsupervised learning are obtained; Transfer learning is performed based on the model parameters after unsupervised learning. The transfer learning process uses the labeled data and some unlabeled data in the target field to perform fully supervised learning and unsupervised learning again to obtain new model parameters; the new model parameters are used to detect changes in the target field data; In the unsupervised process, random perturbation processing includes eigenvalue interval rotation processing: ...,7； ； in, represents the eigenvalue of the uth row and vth column of the feature map before rotation, represents the intermediate eigenvalue of the uth row and vth column of the feature map during rotation processing, represents the eigenvalue of the uth row and vth column of the feature map after rotation, represents an eigenvalue interval, Indicates the number of interval spans for rotation, Indicates the maximum eigenvalue; Indicates the minimum eigenvalue; In the unsupervised process, the random perturbation processing includes the eigenvalue interval replacement processing, which divides the eigenvalue range into 8 intervals on average, expressed as: , represents the first eigenvalue interval, represents the second eigenvalue interval, represents the third eigenvalue interval, represents the fourth eigenvalue interval, represents the fifth eigenvalue interval, represents the sixth eigenvalue interval, represents the 7th eigenvalue interval, Indicates the 8th eigenvalue interval; eigenvalue interval , then , , ..., ; Indicates the maximum eigenvalue; Indicates the minimum eigenvalue; perform two eigenvalue interval permutations, and the order after the first permutation is ; The order after the second permutation is .

2. According to claim 1, the semi-supervised high-resolution remote sensing image change detection method based on transfer learning, Features: The deep feature extraction module includes n residual groups, the input of residual group 1 is the feature difference map, the output of residual group j-1 is used as the input of residual group j, j=2,…,n-1,n; the outputs of residual group 1 to residual group n are input to the global feature fusion module, and the result output by the global feature fusion module is added to the feature difference map to obtain the output feature map of the deep feature extraction module.

3. According to claim 2, the semi-supervised high-resolution remote sensing image change detection method based on transfer learning, Features: The residual group consists of L residual attention modules and 1 3×3 convolution. First, the two 3×3 convolutions in the residual attention module are connected by the ReLU activation function. Second, the coordinate attention module is embedded in the residual group. The input feature is , represents the output feature map of the L-1th residual attention module in the Kth residual group, and the output feature map of the Lth residual attention module in the Kth residual group for: ； in, represents the Lth residual attention module in the Kth residual group, represents the L-1th residual attention module in the Kth residual group, represents the first residual attention module in the Kth residual group.

4. The semi-supervised high-resolution remote sensing image change detection method based on transfer learning according to claim 2, Features: In the global feature fusion module, the input feature is first processed by a 1×1 convolution operation and then sent to the channel shuffle module; the channel shuffle module subdivides the feature map into 4 sub-feature maps and re-divides each sub-feature map to obtain represents the generated sub-feature map, , Represents the convolution processing corresponding to each sub-feature map, represent Output of With the previous convolutional layer Output Add them together and then send the result to the next convolutional layer to be processed.

5. The semi-supervised high-resolution remote sensing image change detection method based on transfer learning according to claim 1, Features: In the unsupervised learning process, the feature difference map is divided into two branches after passing through the deep feature extraction module. The first branch directly outputs the output feature map of the deep feature extraction module. , After the decoder, a predicted change map containing change information is obtained ; The second branch extracts the output feature map of the deep feature extraction module After random perturbation processing, NP perturbation feature maps are obtained ,..., , the perturbation feature map is decoded to obtain NP predicted change maps containing change information , ..., ,in Represents the output feature map of the deep feature extraction module After random perturbation processing, the first perturbation feature map is obtained. Represents the output feature map of the deep feature extraction module After random perturbation processing, the NPth perturbation feature graph is obtained. represent The predicted change graph obtained by decoding, represent The predicted change graph obtained by decoding; Predicted change graph output by the first branch As the baseline, each predicted change map output by the second branch is compared with the baseline and the loss is calculated.

6. The semi-supervised high-resolution remote sensing image change detection method based on transfer learning according to claim 1, Features: Use cross entropy loss as the fully supervised loss .

7. The semi-supervised high-resolution remote sensing image change detection method based on transfer learning according to claim 6, Features: In the unsupervised stage, the unlabeled dataset is defined as ,in is the i-th pair of unlabeled variation images, is the i-th unlabeled image before the change, is the i-th unlabeled changed image, For the size of the unlabeled dataset, an unsupervised loss based on unlabeled data is adopted , using unlabeled variation image pairs The obtained feature difference map ,and As a set of feature maps after random perturbation, is the unlabeled variation image pair The corresponding i-th perturbation feature map after random perturbation, is the unlabeled variation image pair The corresponding p-th perturbation feature map after random perturbation, is the unlabeled variation image pair The corresponding NP-th perturbation feature map after random perturbation; Through the decoder Get the predicted change graph , through NP auxiliary decoders Process each disturbance feature map to obtain the corresponding predicted change map : ； ； The unsupervised loss is: ； in is a distance metric.

8. The semi-supervised high-resolution remote sensing image change detection method based on transfer learning according to claim 6, Features: Overall loss function The definition is as follows: ； in, is a regularization constant.

Images