Semi-supervised multispectral change detection method based on edge lifting
By introducing feature pyramid network and channel prior convolution attention module in remote sensing image change detection, combined with multi-scale feature fusion and perturbation processing, the problems of lack of marking samples and poor edge extraction in the prior art are solved, and the problem of improved Accuracy and robustness of change detection.
Patent Information
- Application Number
- CN202411203689.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-30
- Publication Date
- 2025-01-03
AI Technical Summary
The existing remote sensing image change detection method based on deep learning faces the problems of lack of labeled samples and poor edge extraction results, especially when processing high-resolution image data, the false alarm rate is high and the detection accuracy is low.
Using a semi-supervised multispectral change detection method based on edge lift, connectivity edge extraction is performed by introducing feature pyramid networks and channel prior convolution attention modules, combined with multi-scale feature fusion modules, the composite loss of supervised losses and unsupervised losses are utilized. Functions, perform adaptive fusion of semantic features and texture features, and enhance model robustness through perturbation at image levels and feature levels.
The edge accuracy of change detection and the model's sensitivity to change regions are improved, the recognition ability and robustness in complex scenarios are enhanced, and the accurate detection of change regions is achieved.
Smart Images
Figure CN119251149A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of remote sensing image change detection, and in particular to a semi-supervised multispectral change detection method based on edge lifting. Background Art
[0002] Cities are ecosystems that are closely related to humans and have a profound and extensive impact on the geographical environment. Changes in land use and land cover reflect the process of urbanization and are crucial to the study of urban expansion, economic development, sociology and related fields. Remote sensing image change detection is a technology that obtains change information through multi-temporal satellite images acquired at different times. With the continuous improvement of computer performance and the continuous development of deep learning technology, change detection based on high-resolution remote sensing images can be divided into two stages, namely the traditional change detection stage and the change detection stage based on deep learning algorithms. Traditional change detection algorithms can be summarized into the following three categories: direct comparison method, transformation-based method and classification-based method. These methods perform well in lower-resolution image data, but face certain limitations when processing high-resolution image data. Deep learning requires a large number of labeled samples to train the model to extract change features, which are relatively difficult to obtain in change detection tasks. To address this problem, Sudhanshu Mittal proposed a semi-supervised learning method to enhance the model using abundant unlabeled samples. The essence of change detection is binary semantic segmentation, so semantic segmentation and semi-supervised learning are applied to change detection. Daifeng Peng introduced a semi-supervised algorithm for change detection, learning from a small number of labeled samples and abundant unlabeled samples to enhance the performance of change detection.
[0003] Existing deep learning-based change detection methods face many challenges. First, the lack of labeled samples is an important issue. Existing methods often rely on a large number of labeled samples for training. However, in practical applications, it is very difficult to obtain sufficient and accurate labeled samples, which limits the training effect of the model. Secondly, in terms of edge extraction effect of change detection, existing deep learning algorithms have difficulty in handling complex changes in background and change areas, resulting in poor edge extraction effect, high false alarm rate and low detection accuracy. Summary of the invention
[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a semi-supervised multispectral change detection method based on edge lifting that effectively improves the edge accuracy of change detection.
[0005] The purpose of the present invention can be achieved by the following technical solutions:
[0006] A semi-supervised multispectral change detection method based on edge lifting includes the following steps:
[0007] The change detection image to be detected is input into a pre-trained semi-supervised network to obtain a change detection result. The semi-supervised network includes:
[0008] The encoder is used to extract features from the input change detection image to obtain semantic features C;
[0009] A connectivity edge extraction module is used to perform connectivity edge extraction based on feature maps of multiple scales obtained during feature extraction by the encoder to obtain texture features M;
[0010] Multi-scale feature fusion module, used to add semantic features C and texture features M to obtain feature map F Fuse ; For the feature map F Fuse Perform layer-by-layer pooling and downsampling fusion to obtain multi-level feature fusion RA, thereby obtaining weights of different positions in the feature map, realizing the fusion of semantic features C and texture features M, and obtaining the fusion result F;
[0011] The decoder is used to decode the fusion result F and perform change detection based on the decoding result.
[0012] Furthermore, the processing process of the connectivity edge extraction module specifically includes the following steps:
[0013] The feature maps of multiple scales obtained by feature extraction by the encoder are fused to obtain a multi-scale feature fusion map Es, which is upsampled to the input size to obtain a preliminary output X; the preliminary output X is globally average pooled and mapped to the same dimension as the multi-scale feature fusion map Es to obtain a vector V; the vector V is convolved and activated to obtain a directional prior weight a, and the multi-scale feature fusion map Es and the directional prior weight a are divided into multiple groups of slices in the channel dimension, each group of slices corresponds to different semantic information or target features, and a feature slice e is obtained; each feature slice e is subjected to feature extraction through a dual path of ordinary convolution and dilated convolution, and then the local feature information and context information are integrated through a channel prior convolution attention module to obtain an optimized feature representation e'; the directional prior weight a and the feature representation e' are channel-level multiplied to obtain f; the convolution kernel is used to re-encode each group of results f, and the residual of the encoded result and the corresponding original feature slice e is added to obtain an optimized feature representation M'; the output M' of each sub-path is stacked and re-encoded to generate a new feature map M as a texture feature.
[0014] Furthermore, the acquisition process of the multi-scale feature fusion map Es is specifically as follows:
[0015] The last multi-layer feature map obtained by the encoder through feature extraction is fused using the feature pyramid network attention mechanism to obtain a multi-scale feature fusion map Es.
[0016] Furthermore, the process of the multi-scale feature fusion module fusing and obtaining the multi-level feature fusion RA is specifically as follows:
[0017] For the feature map F Fuse Pooling is performed in sequence to obtain multiple pooled feature maps; for the feature map F Fuse Downsampling is performed, and the result is added to the pooled feature map of the corresponding size, and then channel compression and convolution operations are performed. The result is downsampled again in a cycle, and the result is added to the pooled feature map of the corresponding size, and then channel compression and convolution operations are performed, until all the pooled feature maps are added, and the multi-level feature fusion RA is obtained.
[0018] Furthermore, the calculation expression for obtaining the fusion result F by the multi-scale feature fusion module is:
[0019] F=C×a+M×(1-a)
[0020] a=Sigmoid(RA)
[0021] Where a is the weight of different positions in the feature map.
[0022] Furthermore, the process of acquiring the training data of the semi-supervised network includes:
[0023] Obtain training samples, add image-level perturbations and feature-level perturbations to the training samples, thereby obtaining training data for a semi-supervised network;
[0024] The disturbance adding process includes:
[0025] Divide the training samples into three groups of data streams, add weak perturbations to one group of data streams, and obtain data x w ; Add random strong disturbances to the other two groups of data streams to obtain data x s1 、x s2 ; Carry out feature-level strong perturbation on a set of weakly perturbed data streams and a set of strongly perturbed data streams to obtain Fs(x w )、Fs(x s2 ), select data Fs(x w )、Fs(x s2 ) and x s1 , as the training data.
[0026] Further, the weak perturbation includes resizing, cropping and flipping;
[0027] The strong disturbance includes Gaussian blur, contrast adjustment, sharpening adjustment, brightness adjustment, hue adjustment and exposure adjustment.
[0028] Furthermore, the training samples include labeled training samples and unlabeled training samples.
[0029] Furthermore, the loss function L in the semi-supervised network training process includes the supervision loss L S and the unsupervised loss L U , the supervision loss L S is the cross entropy loss between the model prediction and the true value label, and the unsupervised loss L U In order to regularize the sample prediction under strong perturbation to the same loss as the prediction under weak perturbation, the unsupervised loss L U The corresponding calculation expression is:
[0030]
[0031] In the formula, B U is the batch size of unlabeled data, γ is the predefined confidence threshold, and p w is the weak perturbation prediction, c is the cross entropy loss function, p is the sample prediction, and m is the sample label.
[0032] Furthermore, the calculation expression of the loss function L in the semi-supervised network training process is:
[0033]
[0034] Compared with the prior art, the present invention has the following advantages:
[0035] (1) The present invention constructs a connectivity edge extraction module by introducing a feature pyramid network and a channel prior convolution attention module. This module achieves high-precision edge detection and feature extraction by fusing multi-scale feature maps and global average pooling; by integrating the global and local information of the feature map, the directional prior weight is used to guide the model to focus on specific feature channels, thereby improving the accuracy of edge detection and the sensitivity of the model to changing areas; through multi-path processing and attention mechanism, the model can effectively learn and express different scales and semantic information, and improve the detection accuracy of changing areas; feature slicing processing and adaptive weight learning further improve the stability and recognition ability of the model in complex scenarios.
[0036] (2) The present invention designs a multi-scale feature fusion module, which realizes the adaptive fusion of semantic features and texture features through pooling operation and multi-level feature fusion. A multi-scale attention mechanism (RA) is constructed to adaptively adjust the weights of different positions in the feature map, and fuse semantic features and texture features through element-by-element multiplication.
[0037] The composite loss function contains supervised loss and unsupervised loss, which ensures the prediction consistency of the model under different degrees of disturbance, thereby improving the accuracy and robustness of change detection. This module can capture contextual information of different scales, realize the adaptive fusion of semantic features and texture features, enhance the model's understanding of multi-scale semantic information and texture details of the image, and realize accurate detection of changed areas.
[0038] (3) Aiming at the complex distribution characteristics of the changing area and the low labeling rate of unlabeled samples, the present invention introduces random perturbation methods at the image level and feature level, and enhances the robustness of the model through weak perturbations (resizing, cropping and flipping) and strong perturbations (Gaussian blur, contrast, sharpening, brightness, hue, exposure, etc.), so that the background distribution of multispectral data can be robustly learned through a variety of perturbation methods, thereby effectively avoiding the interference of unlabeled samples on network training. In addition, a feature-level perturbation method for weakly and strongly perturbed data streams is designed, and perturbation processing is performed at the input image and feature levels, which effectively prevents the model from over-relying on specific features, captures the intrinsic structure and semantic information of the data, and improves the adaptability to different input changes.
[0039] (4) This method has the advantages of semi-supervised learning and change detection based on multi-scale feature fusion, and can be applied to multi-spectral change detection. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 A schematic diagram of a process framework of a semi-supervised multi-spectral change detection method based on edge lifting provided in an embodiment of the present invention;
[0041] Figure 2 A schematic diagram of disturbance addition of a semi-supervised multi-spectral change detection method based on edge lifting provided in an embodiment of the present invention;
[0042] Figure 3 A schematic diagram of a connectivity edge extraction module provided in an embodiment of the present invention;
[0043] Figure 4 A schematic diagram of the structure of a multi-scale feature fusion module provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0044] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings 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. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.
[0045] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0046] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.
[0047] Example 1
[0048] like Figure 1 As shown, this embodiment provides a semi-supervised multi-spectral change detection method based on edge lifting, comprising the following steps:
[0049] Input the change detection image to be detected into the pre-trained semi-supervised network to obtain the change detection result. The semi-supervised network includes:
[0050] The encoder is used to extract features from the input change detection image to obtain semantic features C;
[0051] A connectivity edge extraction module is used to perform connectivity edge extraction based on feature maps of multiple scales obtained during feature extraction by the encoder to obtain texture features M;
[0052] Multi-scale feature fusion module, used to add semantic features C and texture features M to obtain feature map F Fuse ; For the feature map F Fuse Perform layer-by-layer pooling and downsampling fusion to obtain multi-level feature fusion RA, thereby obtaining weights of different positions in the feature map, realizing the fusion of semantic features C and texture features M, and obtaining the fusion result F;
[0053] The decoder is used to decode the fusion result F and perform change detection based on the decoding result.
[0054] Preferably, the process of acquiring training data of the semi-supervised network includes:
[0055] Obtain training samples, add image-level perturbations and feature-level perturbations to the training samples, thereby obtaining training data for a semi-supervised network;
[0056] The process of adding disturbances includes:
[0057] Divide the training samples into three groups of data streams, add weak perturbations to one group of data streams, and obtain data x w ; Add random strong disturbances to the other two groups of data streams to obtain data xs1 、x s2 ; Carry out feature-level strong perturbation on a set of weakly perturbed data streams and a set of strongly perturbed data streams to obtain Fs(x w )、Fs(x s2 ), select data Fs(x w )、Fs(x s2 ) and x s1 , as training data.
[0058] Further preferably, the loss function L in the semi-supervised network training process includes the supervision loss L S and the unsupervised loss L U , supervision loss L S is the cross entropy loss between the model prediction and the true value label, and the unsupervised loss L U In order to regularize the sample prediction under strong perturbation to the same loss as the prediction under weak perturbation, the unsupervised loss L U The corresponding calculation expression is:
[0059]
[0060] In the formula, B U is the batch size of unlabeled data, γ is the predefined confidence threshold, and p w is the weak perturbation prediction, c is the cross entropy loss function, p is the sample prediction, and m is the sample label.
[0061] The specific implementation process of the above solution in this embodiment includes:
[0062] Step 1: Normalize the original multispectral data; construct training samples for the semi-supervised change detection model, including labeled samples and unlabeled samples.
[0063] Step 2: Construct a semi-supervised framework based on encoding and decoding, including a semi-supervised framework, a connectivity edge extraction module, and a multi-scale feature fusion module. The corresponding processing process is as follows:
[0064] Before the training samples are input into the network, three groups of data streams are generated from unlabeled samples, and random perturbations are uniformly performed at the image level. One group undergoes weak perturbation x w , including resizing, cropping, and flipping, these perturbations are designed to simulate common image transformations and maintain the basic features of the image. The other two groups carried out random strong perturbations x s1 、x s2, including Gaussian blur, contrast, sharpening, brightness, hue, exposure, etc. These strong perturbations increase the diversity of the image and enhance the robustness of the model. After the perturbation processing at the image level, the three groups of feedforward streams are immediately input into the encoding layer of the network, and the feature level perturbation Fs is further introduced. A group of weakly perturbed data streams and a group of strongly perturbed data streams are subjected to feature level strong perturbations, namely channel dropout, to obtain Fs(x w )、Fs(x s2 ). This operation can effectively prevent the model from over-relying on certain specific features, thereby enhancing its adaptability to different input changes. Figure 2 As shown in the figure, the data flow of the training process is to introduce different degrees of perturbations at the image level and feature level. The framework can better capture the intrinsic structure and semantic information of the data, thereby achieving better performance in semi-supervised learning tasks.
[0065] Construct a connectivity edge extraction module. The feature pyramid network attention mechanism (FPN) is applied to the last four layers of the encoder, which contain rich high-level semantic information, to obtain a feature map Es that integrates multi-scale information. The output Es is upsampled to the input size to obtain a preliminary output X. In order to further extract global context information, global average pooling (GAP) is performed on X to compress the spatial dimension to 1×1 to obtain global context information. Next, a 1×1 convolution kernel W1 is used to map the output of GAP to the same dimension as Es to obtain a vector V. Then, a 1×1 convolution and a sigmoid activation function are applied to V to obtain directional prior weights. These weights are used to guide the model to pay more attention to specific feature channels and enhance sensitivity to changing areas. The multi-scale feature fusion map Es and the directional prior weight a are divided into 4 slices in the channel dimension, and each group of slices corresponds to different semantic information or target features. The feature slice e is passed through the dual paths of ordinary convolution and void convolution, and the channel prior convolution attention module (CPCA) is applied to integrate local feature information and contextual information to obtain the optimized feature representation e', which helps to capture the subtle differences in the changing area and improve the accuracy of edge detection. The prior information a is multiplied by e' at the channel level to obtain f, and the importance of features in different directions is flexibly adjusted to highlight or suppress features with specific directional information. The 4 sets of results f are re-encoded using a 1×1 convolution kernel W to achieve adaptive weight learning, which helps to improve the stability of the model in different changing scenarios. The residual of the encoded result is added to the original slice e to obtain the optimized feature representation M'. Finally, all sub-path outputs M' are stacked and re-encoded, integrating information from different processing paths to generate a new feature map M. Connectivity edge extraction module such as Figure 3As shown in the figure, this process not only retains multi-scale features, but also enhances the feature expression capability through multi-path processing and attention mechanism, thereby improving the model's ability to understand and recognize complex scenes.
[0066] Construct a multi-scale feature fusion module. Feature map F Fuse ∈R C×H×W The feature map F is obtained by adding the feature layer C containing semantic information and the connectivity edge feature M containing texture information. Fuse ∈R C×H×W Pooling is performed to obtain f c1 ∈R C×1×1 、f c4 ∈R C×4×4 、f c8 ∈R C×8×8 、f c16 ∈R C×16×16 and f c32 ∈R C×32×32 , the pooled feature map f c16 With downsampling to 16×16 f c32 After adding, channel compression and convolution operations are performed to obtain f'c 16 ∈R C×16×16 . Then the pooled feature map f c8 With f'c downsampled to 8×8 16 After adding, channel compression and convolution operations are performed to obtain f c8 ', and so on to form a multi-level feature fusion RA, which can capture contextual information of different scales and help the model focus on important areas. Finally, the weights of different positions in the feature map are obtained through multi-scale attention RA. The derivation formula of the weights is as follows:
[0067] a=Sigmoid(RA)
[0068] The semantic feature C and texture feature M are fused by element-by-element multiplication to achieve adaptive fusion of semantic features and texture features. The fusion formula is as follows:
[0069] F=C×a+M×(1-a)
[0070] a=Sigmoid(RA)
[0071] Multi-scale feature fusion module such as Figure 4 As shown in the figure, through this multi-scale feature fusion, the model can better understand the multi-scale semantic information and texture details of the image, thereby achieving better performance in the change detection task.
[0072] Step 3: Set the loss function L including the supervision loss L S and the unsupervised loss L U The supervision loss LS is the cross entropy loss between the model prediction and the true value label. The unsupervised loss L U Regularizing the sample prediction under strong perturbation to be the same as the prediction under weak perturbation can be expressed as:
[0073]
[0074] In the formula, B U is the batch size of unlabeled data, γ is the predefined confidence threshold, and p w is the weak perturbation prediction, c is the cross entropy loss function, p is the sample prediction, and m is the sample label. The unsupervised loss term encourages the model to maintain consistent predictions for the same sample under different degrees of perturbation, thereby improving the robustness of the model.
[0075] The overall loss L can be expressed as:
[0076]
[0077] Step 4: Train the above semi-supervised framework, which includes the following sub-steps:
[0078] Step 401: After completing the first three steps, start training the reconstructed edge-lifting-based semi-supervised change detection method. According to the characteristics and requirements of the multispectral change detection dataset, the ratio of labeled and unlabeled training samples is divided into 5%, 10%, 20% and 40%. The number of training times is set to 100, and the number of training times can be appropriately adjusted according to the actual training situation. For example, when the loss value does not decrease for a long time, the training can be terminated in advance.
[0079] Step 402: After starting training, all training images are input into the network at the beginning of each round of training. The semi-supervised change detection network will continuously adjust the parameters of the network according to the principle of minimizing the loss function until the loss value reaches the minimum, thereby achieving the best training effect.
[0080] Step 403: Repeat steps 401 and 402 to perform multiple trainings. In each training, various parameters of the network can be modified, and the network model parameters with the best performance are selected and saved as the best model to obtain the final change detection result.
[0081] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.
Claims
1. A semi-supervised multispectral change detection method based on edge lifting, characterized in that: The following steps are involved: The change detection image to be detected is input into a pre-trained semi-supervised network to obtain a change detection result. The semi-supervised network includes: The encoder is used to extract features from the input change detection image to obtain semantic features C; A connectivity edge extraction module is used to perform connectivity edge extraction based on feature maps of multiple scales obtained during feature extraction by the encoder to obtain texture features M; Multi-scale feature fusion module, used to add semantic features C and texture features M to obtain feature map F Fuse ; For the feature map F Fuse Perform layer-by-layer pooling and downsampling fusion to obtain multi-level feature fusion RA, thereby obtaining weights of different positions in the feature map, realizing the fusion of semantic features C and texture features M, and obtaining the fusion result F; The decoder is used to decode the fusion result F and perform change detection based on the decoding result.
2. The method for semi-supervised multi-spectral change detection based on edge lifting according to claim 1, characterized in that: The processing process of the connectivity edge extraction module specifically includes the following steps: The feature maps of multiple scales obtained by feature extraction by the encoder are fused to obtain a multi-scale feature fusion map Es, which is upsampled to the input size to obtain a preliminary output X; the preliminary output X is globally average pooled and mapped to the same dimension as the multi-scale feature fusion map Es to obtain a vector V; the vector V is convolved and activated to obtain a directional prior weight a, and the multi-scale feature fusion map Es and the directional prior weight a are divided into multiple groups of slices in the channel dimension, each group of slices corresponds to different semantic information or target features, and a feature slice e is obtained; each feature slice e is subjected to feature extraction through a dual path of ordinary convolution and dilated convolution, and then the local feature information and context information are integrated through a channel prior convolution attention module to obtain an optimized feature representation e'; the directional prior weight a and the feature representation e' are channel-level multiplied to obtain f; the convolution kernel is used to re-encode each group of results f, and the residual of the encoded result and the corresponding original feature slice e is added to obtain an optimized feature representation M'; the output M' of each sub-path is stacked and re-encoded to generate a new feature map M as a texture feature.
3. The method for semi-supervised multi-spectral change detection based on edge lifting according to claim 1, characterized in that: The acquisition process of the multi-scale feature fusion map Es is specifically as follows: The last multi-layer feature map obtained by the encoder through feature extraction is fused using the feature pyramid network attention mechanism to obtain a multi-scale feature fusion map Es.
4. The method for semi-supervised multi-spectral change detection based on edge lifting according to claim 1, characterized in that: The process of the multi-scale feature fusion module fusing and obtaining the multi-level feature fusion RA is specifically as follows: For the feature map F Fuse Pooling is performed in sequence to obtain multiple pooled feature maps; for the feature map F Fuse Downsampling is performed, and the result is added to the pooled feature map of the corresponding size, and then channel compression and convolution operations are performed. The result is downsampled again in a cycle, and the result is added to the pooled feature map of the corresponding size, and then channel compression and convolution operations are performed, until all the pooled feature maps are added, and the multi-level feature fusion RA is obtained.
5. The method for semi-supervised multi-spectral change detection based on edge lifting according to claim 1, characterized in that: The calculation expression for obtaining the fusion result F by the multi-scale feature fusion module is: F=C×a+M×(1-a) a=Sigmoid(RA) Where a is the weight of different positions in the feature map.
6. The method for semi-supervised multi-spectral change detection based on edge lifting according to claim 1, characterized in that: The process of acquiring the training data of the semi-supervised network includes: Obtain training samples, add image-level perturbations and feature-level perturbations to the training samples, thereby obtaining training data for a semi-supervised network; The disturbance adding process includes: The training samples are divided into three groups of data streams. We add weak perturbations to one group of data streams to obtain data xw. We add three kinds of random strong perturbations to the other two groups of data streams to obtain data xs. 1 、xs 2 ; Carry out strong perturbation on a set of weakly perturbed data streams and a set of strongly perturbed data streams at the feature level to obtain Fs(xw), Fs(xs 2 ), select data Fs(xw), Fs(xs 2 ) and xs 1 , as the training data.
7. The method for semi-supervised multi-spectral change detection based on edge lifting according to claim 6, characterized in that: The weak perturbations include resizing, cropping, and flipping; The strong disturbance includes Gaussian blur, contrast adjustment, sharpening adjustment, brightness adjustment, hue adjustment and exposure adjustment.
8. The method for semi-supervised multi-spectral change detection based on edge lifting according to claim 6, characterized in that: The training samples include labeled training samples and unlabeled training samples.
9. The method for semi-supervised multi-spectral change detection based on edge lifting according to claim 6, characterized in that: The loss function L in the semi-supervised network training process includes the supervision loss L S and the unsupervised loss L U , the supervision loss L S is the cross entropy loss between the model prediction and the true value label, and the unsupervised loss L U In order to regularize the sample prediction under strong perturbation to the same loss as the prediction under weak perturbation, the unsupervised loss L U The corresponding calculation expression is: In the formula, B U is the batch size of unlabeled data, γ is the predefined confidence threshold, and p w is the weak perturbation prediction, c is the cross entropy loss function, p is the sample prediction, and m is the sample label.
10. The method for semi-supervised multi-spectral change detection based on edge lifting according to claim 8, characterized in that: The calculation expression of the loss function L in the semi-supervised network training process is: