Remote sensing image change detection method and system based on self-supervised feature difference learning

Abstract

The invention discloses a remote sensing image change detection method and system based on self-supervised feature difference learning, and the method comprises the steps: S1, collecting and making a remote sensing image data set, enabling each image to have a pixel-level label, dividing the images into a training set, a verification set and a test set, and cutting the images into uniform sizes; s2, constructing a Barlow Twins self-supervised comparison pre-training Self-supervised control learning, SSCL and fine tuning CD architecture based on feature difference, including a self-supervised comparison pre-training stage and a fine tuning stage, and using a feature extractor, a projection network and an improved UNet + + model; s3, pre-training the unlabeled image by using a self-supervision algorithm; s4, selecting an optimal feature encoder component for a subsequent encoder; s5, inputting an image with a label to the improved UNet + + model, and performing deep supervision; and S6, storing the model weight, testing by using the test set, and obtaining a change detection result. The method can effectively detect the change in the remote sensing image when limited marked data samples are used.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing change detection, and in particular to a remote sensing image change detection method and system based on self-supervised feature difference learning. Background Art

[0002] Land use and land cover (LULC) change detection (CD) uses dual-temporal remote sensing images to identify surface changes. Similar to other remote sensing image interpretation tasks, LULC CD involves processing multi-scene remote sensing images covering the same area at different times. Deep learning uses an end-to-end learning model to directly extract change area information from dual-temporal remote sensing images, which completely changes the traditional methods. A representative method is the fully convolutional twin network. The key to its effectiveness lies in a large number of manually labeled change samples, which are very time-consuming to collect and annotate from dual-temporal remote sensing images and require professional domain knowledge. To overcome this challenge, many researchers have tried transfer learning strategies to use knowledge learned from the large natural image dataset ImageNet. In order to solve the significant limitation of the domain gap between natural images and remote sensing images, the idea of ​​using a small number of manually labeled samples or even completely unlabeled samples has become increasingly popular in LULCCD applications. There are two main research directions: building massive remote sensing image datasets and using unsupervised contrastive learning to pre-train unlabeled massive remote sensing image datasets. As a unique form of unsupervised learning, self-supervised contrastive learning (SSCL) trains the model through self-designed proxy tasks, which can learn meaningful feature representations by utilizing a large amount of unlabeled remote sensing image data;

[0003] Despite the great potential of self-supervised contrastive pre-training, many advanced SSCL methods are currently difficult to implement in practice. They require a large number of data samples and computing resources to be effectively applied, limiting their practicality. For example, although SimCLR and its improved versions achieved state-of-the-art (SOTA) performance on the ImageNet classification task, they require a batch size of 4096; similarly, BYOL achieved cutting-edge performance through positive contrastive learning, symmetric networks, and stopped gradient methods, but required 64 graphics processing units (GPUs); MoCo solved the negative sample problem by introducing a momentum encoder and a queue sampling strategy, but its computational speed was limited by the need to sample negative samples in the entire queue; SimSiam does not require negative samples and adopts an asymmetric network structure and cross-gradient updates to combat trivial solutions. However, it requires more computing resources and is sensitive to hyperparameters; SwAV utilizes online clustering and avoids the need for negative samples through multi-view predictive coding, which has advantages when label costs are high. However, since it utilizes an online clustering algorithm, it has high requirements on computing resources and the training process is relatively complicated. Summary of the invention

[0004] The purpose of the present invention is to solve the problems existing in the prior art and provide a remote sensing image change detection method and system based on self-supervised feature difference learning, which can effectively detect changes in remote sensing images even when using limited labeled data samples, which is particularly valuable for CD applications where it is difficult and expensive to obtain labeled data for changed areas. Therefore, in order to solve the above technical problems, the remote sensing image change detection method based on self-supervised feature difference learning provided by the present invention includes:

[0005] S1. Collect dual-temporal remote sensing images, create a remote sensing image data set, divide the remote sensing image data set into a training set, a validation set, and a test set, and cut the dual-temporal remote sensing images in the remote sensing image data set into the same size, wherein the dual-temporal remote sensing images all have pixel-level annotations, and the same size is 256×256;

[0006] S2. Construct a BTCD algorithm architecture of Barlow Twins self-supervised contrastive pre-training and fine-tuning CD based on feature differences, including a self-supervised contrastive pre-training Self-supervised contrastive learning, SSCL stage and a fine-tuning CD stage. The self-supervised contrastive pre-training Self-supervised contrastive learning, SSCL stage includes a feature extractor and a projection network, and the fine-tuning CD stage includes an improved UNet++ model. The improved UNet++ model uses an encoder-decoder architecture including a spatial and channel compression and excitation module, and the feature extractor uses ResNet50;

[0007] Preferably, the operation process of the BTCD architecture of BT self-supervision versus pre-training and fine-tuning CD based on feature differences is:

[0008] S201: Definition and is at time t 1 and t 2 Bi-temporal images of the same geographic location captured separately are pre-trained on an unlabeled training set using a self-supervised algorithm so that the pre-trained encoder is fine-tuned to the labeled training set superior;

[0009] S202: In the self-supervised contrast pre-training SSCL stage, only bi-temporal images are used and For training, in the fine-tuning CD phase, bi-temporal images are used and And the labeled images Y in the training set train Conduct training;

[0010] S203: After the self-supervised contrast pre-training SSCL phase is completed, a pre-trained encoder with initialization parameters is obtained;

[0011] S204: In the fine-tuning CD stage, the bands of the dual-phase image are further superimposed to create a new image with twice the number of bands, which is then input into the improved UNet++ model, and the improved UNet++ model is end-to-end trained using a pre-trained encoder. The improved UNet++ model extracts features from the given dual-phase image and classifies it as “changed” or “unchanged” to generate a prediction map;

[0012] According to the above technical solution, the self-supervised contrast pre-training SSCL stage in S202 further includes: initializing the self-supervised contrast pre-training model with supervised ImageNet weights, training an encoder E θ , where θ is a learnable parameter used to solve a pre-training task;

[0013] The BTCD architecture uses three objective functions during training: BT loss function, IP loss function and CCR loss function, and performs self-supervision through hybrid loss as the loss function of the pre-training task "contrastive learning". The encoder E θ The parameter θ at the i-th iteration i The update looks like this:

[0014]

[0015] Among them, θ i is the iteration parameter, κ is the learning rate, represents the partial derivative, represents the loss function;

[0016] Preferably, the three objective functions: BT loss function, IP loss function and CCR loss function further include:

[0017] The BT loss function is defined as follows:

[0018]

[0019]

[0020] Among them, λ is a weight constant, b is the index batch sample, i and j are the vector dimensions of the index network output, C is the cross-correlation matrix between the difference representations D1 and D2, by maximizing the information of the difference representation (D1, D2) between the corresponding bi-temporal images, making them similar in the feature space, and then the model uses the modified BT loss function Loss BTTrain in a self-supervised manner;

[0021] Preferably, the IP loss function is defined as follows:

[0022]

[0023]

[0024]

[0025]

[0026]

[0027] Where σ is the softmax function, DKL is the Kullback–Leibler divergence, AP1 and AP2 are the average probabilities of belonging to the corresponding features. When using bi-temporal images for change detection, it is necessary to ensure that samples with semantic similarity produce similar predictions. In order to achieve IP loss, the features between the enhanced views (X′1, X′2) and (X″1, X″2) of the bi-temporal images are enhanced by using the Jensen-Shannon divergence D JS As a goal, the loss function Loss is used during network training. IP , to force the model to learn features that are robust to the radiometric differences between the two-phase images;

[0028] Preferably, the CCR loss function is defined as follows:

[0029]

[0030] where ||.|| F Represents the Frobenius norm. The network encoder in the BTCD architecture generates activation maps from the enhanced transformed input image X1, which are represented by E'1∈R b×c×h×w and E″1∈R b×c×h×w , where b represents the batch size, c represents the number of output channels, h and w represent the spatial dimensions. The network encoder generates activation maps for the transformed X2 image, represented by E'2 and E″2 respectively. In order to maintain the temporal consistency of the input images X1 and X2 in the feature space, a pairwise similarity matrix G'= between the feature maps E' and E'2 is constructed.<E'1,E'2> , a pairwise similarity matrix G″ between feature graphs E″1 and E″2 is constructed.<E″1,E″2> , where <, > represents the dot product.

[0031] Preferably, in the total loss, the three loss functions (BT, IP and CCR) are weighted combined to derive the final network loss function:

[0032] Loss Total =Loss BT +αLoss IP +βLoss CCR

[0033] Where α and β are loss balance weights;

[0034] According to the above technical solution, the S204 further includes: the improved UNet++ model captures multi-scale features through skip connections and integrates them together through feature superposition, using an encoder-decoder architecture, which includes a spatial and channel compression and excitation scSE module, and in the fine-tuning CD stage, using the pre-trained encoder E Φ The improved UNet++ model is trained end-to-end to improve the UNet++ model from the given dual-phase image and Extract features from the dataset and classify them as “changed” or “unchanged” to predict the change graph. The calculation formula is as follows:

[0035]

[0036] in Indicates the decoder and CD head, is a learnable parameter, using the parameters Φ and The improved UNet++ model is updated as follows:

[0037]

[0038]

[0039] Among them, Loss CD represents the loss function, Y train represents the true label.

[0040] S3, using the self-supervised contrast pre-training BTCD algorithm to perform self-supervised pre-training processing based on the BT algorithm on the unlabeled dual-temporal remote sensing images;

[0041] Preferably, the operation process of the self-supervised comparison pre-training BTCD algorithm is:

[0042] S301: The BTCD algorithm takes the bi-temporal image (X1, X2) as input, applies color distortion and Gaussian blur transformation to obtain an enhanced image (X1→X′1, X″1″, X2→X′2, X″2″);

[0043] S302: adopting a twin network architecture, wherein the twin network architecture uses ResNet50 excluding a final classification layer as a feature extractor, followed by a projection network;

[0044] S303: The input image is subjected to multi-layer convolution and pooling operations, and the receptive field is further enhanced by using dilated convolution. A convolution kernel with a step size of 16 is used to reduce the dimension of the input image, and a feature vector with a dimension of 16×16×2048 is generated to capture the basic features.

[0045] S304: Apply two-dimensional adaptive average pooling to the feature vector and pass it to the projection head; the projection network contains two linear layers, each linear layer has a hidden layer size of 512 output units, and the first layer of projection is followed by a batch normalization layer and a rectified linear unit;

[0046] S305: applying the absolute difference to the output embedding to obtain a difference embedding; applying the BT loss function to the difference embedding of size 1×256 to generate a cross-correlation matrix of size 256×256;

[0047] S306: Using a twin encoder (E θ ), extract feature vectors E′1 and E′2 from the bi-temporal image, and use the nonlinear projection Applied to the encoded feature vectors E′1 and E′2 to generate representations Z′1 and Z′2, the absolute feature difference is then applied to the projected output to learn representations of the changing features between the input images, as described below:

[0049] D1=|Z′1-Z′2|=|G(E(X′1))–G(E(X′2))|

[0051] D2=|Z″1-Z″2|=|G(E(X″1))–G(E(X″2))|

[0052] When the semantic changes between the bi-temporal image features (X'1, X'2) and (X"1, X"2) are consistent, the model is constrained and ensured by applying the BT objective function on the difference feature representations D1 and D2. Achieve maximum cross-correlation of changing features;

[0053] S307: The model learns specific information about the changes between the two-phase images. After the self-supervised contrast pre-training phase is completed, the encoder (E Φ ) parameters are passed to the downstream fine-tuning CD stage tasks

[0054] S4, selecting the encoder component that provides the best features for change detection and using it in the encoder of the next stage;

[0055] S5. Input the labeled dual-temporal remote sensing image into the improved UNet++ model, and use the prediction map and loss function output by the model for deep supervision;

[0056] S6. Save the trained model weights and use the test set to test to obtain the change detection result of the test image.

[0057] Preferably, the testing using the test set includes:

[0058] S601: Select a change detection data set to verify the performance of the BTCD algorithm, the change detection data set includes: a LEVIR-CD data set and a SYSU-CD data set;

[0059] S602: the change detection datasets are uniformly cropped to 256×256, and for each dataset, samples are randomly divided into three parts: a training set, a validation set, and a test set. The training set is used for model fitting, the validation set is used for adjusting the model during training, and the test set is used for evaluating the generalization performance of the model.

[0060] S603: In the LEVIR-CD dataset, the number of training sets, validation sets, and test sets are 7120, 1024, and 2048, respectively; in SYSU-CD, the number of training sets, validation sets, and test sets are 12000, 4000, and 4000, respectively;

[0061] S604: In the self-supervised comparative pre-training SSCL stage, the ResNet50 model is trained using image pairs from the training set and the validation set, the batch size is set to 40, and the LARS optimizer is used to train 400 epochs, the initial learning rate is 0.003, which is adjusted by multiplying the batch size and dividing by 256, and the trade-off parameter λ of the loss function is set to λ = 5×10 -3 , using 1.5×10 -6 The weight decay parameter value of , the hyperparameters α and β are set to 100 and 2000 respectively;

[0062] S605: Evaluate the performance of the BTCD algorithm in subsequent CD tasks by fine-tuning, using ResNet50 as the feature extractor, fine-tuning with the SOTAUNet++ model, and applying data augmentation techniques including random flipping, cropping, and Gaussian blurring; that is, using the AdamW optimizer, with an initial learning rate of 0.003, a batch size of 30, and training for 150 epochs;

[0063] S606: Use four indicators, accuracy, recall, F1 score and intersection over union (IoU), to compare the performance of the algorithm with the SOTA method. Compare the proposed CD method with several SOTA methods. Fine-tune the improved UNet++ model on the LEVIR-CD and SYSU-CD datasets using three strategies: random initialization, supervised ImageNet pre-training, and pre-training using the BTCD algorithm. In order to evaluate the effectiveness of the proposed pre-training method, a comparative analysis is performed with several other self-supervised comparison pre-training methods. Output the change detection results of the test image.

[0064] According to the above technical solution, a remote sensing image change detection system based on self-supervised feature difference learning is further provided, the system comprising:

[0065] Remote sensing image dataset module, used to collect dual-temporal remote sensing images and produce remote sensing image datasets;

[0066] BTCD algorithm architecture module, used to build BTCD architecture for BT self-supervision contrast pre-training and fine-tuning CD based on feature differences;

[0067] A pre-training processing module for pre-training an unlabeled training set using a self-supervised algorithm;

[0068] UNet++ model module, which is used to perform deep supervision on labeled bi-temporal remote sensing images and loss functions based on the improved UNet++ model;

[0069] The test module is used to perform testing using the test set to obtain the change detection result of the test image.

[0070] Preferably, the remote sensing image dataset module comprises:

[0071] Training set module, used for model fitting;

[0072] A validation set module, used to adjust the model during training;

[0073] The test set module is used to evaluate the generalization performance of the model based on the test set.

[0074] Preferably, the BTCD algorithm architecture module includes:

[0075] The self-supervised contrastive pre-training SSCL stage module is used to initialize the self-supervised contrastive pre-training model using supervised ImageNet weights;

[0076] The fine-tuning CD stage module is used to perform end-to-end training of the improved UNet++ model using the pre-trained encoder.

[0077] Preferably, the UNet++ model module includes:

[0078] Loss function module, including: BT loss function, IP loss function, CCR loss function;

[0079] The deep supervision module is used to perform deep supervision using the prediction graph and loss function output by the improved UNet++ model.

[0080] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0081] 1) The present invention provides a new framework named BTCD, which consists of two cascaded stages: self-supervised contrastive pre-training and fine-tuning. In the first stage, the algorithm uses the absolute feature difference self-supervised contrastive pre-training method to learn task-specific variation representations for CD; in the second stage, the pre-trained encoder is fine-tuned, which effectively enhances the downstream CD network by leveraging favorable parameter initialization.

[0082] 2) The present invention adds IP loss and CCR loss to the self-supervised contrast pre-training of the original BT loss function in order to reduce the impact of radiation differences, noise changes and imaging perspective differences caused by dual-phase images. This method improves the performance of the pre-training process.

[0083] 3) We evaluate the proposed method on two public LULCCD datasets and demonstrate that fine-tuning using only the pre-trained encoder outperforms the ImageNet supervised pre-training method and several recently proposed SSCL pre-training methods without the need for additional data.

[0084] 4) This paper verifies the effectiveness of self-supervised contrast pre-training when sample data is insufficient. When labeled data is insufficient, the proposed pre-training method significantly improves the performance of the CD model. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0086] In the attached picture:

[0087] Figure 1 is a flowchart of the steps of a remote sensing image change detection method based on self-supervised feature difference learning provided by an embodiment of the present invention;

[0088] Figure 2 It is an overall framework diagram provided by an embodiment of the present invention;

[0089] Figure 3 is a flow chart of a self-supervised comparative pre-training algorithm provided by an embodiment of the present invention;

[0090] Figure 4 is an example diagram of the LEVIR-CD data set provided by an embodiment of the present invention;

[0091] Figure 5 is an example diagram of a SYSU-CD data set provided by an embodiment of the present invention;

[0092] Figure 6 is a visual comparison example diagram of applying the UNet++ model to the LEVIR-CD dataset using different pre-training methods provided by an embodiment of the present invention;

[0093] Figure 7 is a visual comparison example diagram of applying the UNet++ model to the SYSU-CD dataset using different pre-training methods provided by an embodiment of the present invention;

[0094] Figure 8 is an example diagram of the performance (F1 score) of different pre-training methods evaluated using the improved UNet++ model under different label availability provided by an embodiment of the present invention;

[0095] Fig. 9 This is an example diagram of the WHU-BCD dataset (including a dual-phase image, a true value, and a predicted change diagram) provided by an embodiment of the present invention;

[0096] Fig.10 It is a schematic diagram of the module composition of a remote sensing image change detection system based on self-supervised feature difference learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0097] 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 only part of the embodiments of the present invention, rather than all of the embodiments. These embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. In addition, it should be understood that after reading the content of the present invention, based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0098] The present invention provides a technical solution: a remote sensing image change detection method based on self-supervised feature difference learning, comprising the following steps:

[0099] S1. Collect dual-temporal remote sensing images and create a remote sensing image dataset. Ensure that each dual-temporal remote sensing image has pixel-level annotations. Divide the remote sensing image dataset into a training set, a validation set, and a test set. At the same time, cut the dual-temporal remote sensing images in the remote sensing image dataset into the same size, such as 256×256.

[0100] S2. Construct a BTCD architecture based on Barlow Twins self-supervised contrast pre-training and fine-tuning CD with feature difference, including a self-supervised contrast pre-training stage and a fine-tuning stage. The first stage includes a feature extractor and a projection network, and the second stage includes an improved UNet++ model, which utilizes an encoder-decoder architecture, including a spatial and channel compression and excitation (scSE) module, wherein the feature extractor can use ResNet50;

[0101] In this embodiment, combined with Figure 1 , Figure 1 An overview of the BTCD architecture is shown. The architecture of BT Self-Supervised Contrastive Pre-training and Fine-tuning CD (BTCD) based on Feature Difference is:

[0102] definition and is at time t 1 and t 2 Bi-temporal images of the same geographic location captured separately, pre-trained on an unlabeled training set using a self-supervised algorithm so that the pre-trained encoder can be easily fine-tuned to the labeled training set The proposed framework consists of two different stages: In the self-supervised contrastive pre-training (SSCL) stage, only the bi-temporal images are used. and In the subsequent CD phase, the framework uses bi-temporal images and And the labeled images Y in the training set train For training, the entire framework is enhanced in two key steps; the first CNN is trained on a pre-training task to select the encoder component that provides the best features for change detection, and then the selected encoder is used by the second CNN for change detection.

[0103] Exemplarily, the self-supervised contrastive learning (SSCL) stage is combined with Figure 2 , Figure 2 The execution process of the self-supervised contrastive pre-training algorithm (BTCD) is shown. The process of the self-supervised contrastive pre-training stage is as follows:

[0104] In the self-supervised contrastive learning (SSCL) stage, the self-supervised contrastive pre-training model is initialized with supervised ImageNet weights instead of random initialization, and an encoder E is trained. θ, where θ is a learnable parameter used to solve a pre-training task. The BTCD algorithm uses three objective functions during training: BT loss, IP loss, and CCR loss. It uses a hybrid loss for self-supervision as the loss function of the pre-training task “contrastive learning”. The encoder E θ The parameter θ at the i-th iteration i Updated as follows:

[0105]

[0106] Among them, θ i is the iteration parameter, κ is the learning rate, represents the partial derivative, represents the loss function;

[0107] Exemplarily, the improved UNet++ model included in the fine-tuning stage is:

[0108] The improved UNet++ model was chosen because it was able to achieve SOTA results on the CD dataset. The improved UNet++ model can capture multi-scale features through skip connections and integrate them together through feature stacking. It utilizes an encoder-decoder architecture that includes a spatial and channel compression and excitation (scSE) module. To train the model, the bands of the dual-phase image are stacked together to create a new image with twice the number of bands, which is then input into the improved UNet++ model. The advantage of this approach is that it allows the model to learn both low-level and high-level features of the input image, which the Siamese network cannot do. It also simplifies the CD problem into a semantic segmentation task. In the fine-tuning CD stage, the pre-trained encoder E is used to train the model. Φ The improved UNet++ model is trained end-to-end to improve the UNet++ model from the given dual-phase image and Extract features from the dataset and classify them as “changed” or “unchanged” to predict the change graph. The calculation formula is as follows:

[0109]

[0110] in Indicates the decoder and CD head, is a learnable parameter, using the parameters Φ and The improved UNet++ model is updated as follows:

[0111]

[0112]

[0113] Among them, LossCD represents the loss function, Y train represents the true label.

[0114] S3, using the self-supervised contrast pre-training BTCD algorithm to pre-train the unlabeled dual-temporal remote sensing images;

[0115] In this embodiment, the specific implementation process of the self-supervised comparison pre-training BTCD algorithm is as follows:

[0116] The BTCD algorithm takes a bi-temporal image (X1, X2) as input and applies color distortion and Gaussian blur transformation to obtain the enhanced image (X1→X′1, X″1, X2→X′2, X″2); it adopts a Siamese network architecture that uses ResNet50 (excluding the final classification layer) as a feature extractor followed by a projection network;

[0117] Exemplarily, the input image is initially subjected to multiple layers of convolution and pooling operations, further using dilated convolution to enhance the receptive field without significantly increasing the number of parameters, and then a convolution kernel with a stride of 16 is used to reduce the dimensionality of the input image, resulting in a more compact representation; this series of operations ultimately produces a feature vector of dimension 16×16×2048, which effectively captures the essential features while maintaining computational efficiency and a wide range of context;

[0118] Specifically, 2D adaptive average pooling is applied to the feature vector and passed to the projection head; the projection network contains two linear layers, each with a hidden layer size of 512 output units; due to the large amount of computation, the output of the projection network is modified to achieve an embedding size of 256, while the embedding size of the BT network is 8192; the first layer of projection is followed by a batch normalization layer and rectified linear units;

[0119] Furthermore, the absolute difference is applied to the output embedding to obtain the difference embedding; the BT loss function is used on the difference embedding of size 1×256 to generate a cross-correlation matrix of size 256×256;

[0120] For example, the twin encoder (E θ ) has a shared parameter θ, which is used to extract feature vectors E'1 and E'2 from the bi-temporal images by nonlinearly projecting Applied to the encoded feature vectors E'1 and E'2 to generate representations Z'1 and Z'2, the absolute feature difference is then applied to the projected output to learn representations of the changing features between the input images, as described below:

[0122] D1=|Z′1-Z′2|=|G(E(X′1))–G(E(X′2))|

[0124] D2=|Z″1-Z″2|=|G(E(X″1))–G(E(X″2))|

[0125] Assume that the semantic changes between the bi-temporal image features (X′1, X′2) and (X″1″, X″2″) should be consistent; therefore, the model needs to be constrained to ensure Regardless of the image enhancement method applied, to achieve this, a BT objective function is applied on the difference feature representations D1 and D2 to maximize the cross-correlation of the change features; this encourages the model to learn specific information about the changes that occur between the two-phase images; after the self-supervised contrastive pre-training phase is completed, the encoder (E Φ ) parameters are passed to downstream CD tasks.

[0126] In this embodiment, the loss function is represented by a BT loss function, an IP loss function, a CCR loss function, and a total loss function;

[0127] For example, in the BT loss function: by maximizing the information of the difference representation (D1, D2) between the corresponding bi-temporal images, making them similar in the feature space, and then the model uses the modified BT loss function Loss BT Training is performed in a self-supervised manner, and the function is defined as follows:

[0128]

[0129]

[0130] Where λ is a weight constant, b is the index batch sample, i and j are the vector dimensions of the index network output, and C is the cross-correlation matrix between the differential representations D1 and D2;

[0131] Exemplarily, in the IP loss function: When using bi-temporal images for change detection, it is necessary to ensure that samples with semantic similarity produce similar predictions regardless of the radiation conditions. In order to achieve IP, the features between the enhanced views (X′1, X′2) and (X″1″, X″2″) of the bi-temporal images are enhanced. To achieve this, the Jensen-Shannon divergence D is used. JS As a goal to implement IP:

[0132]

[0133]

[0134]

[0135]

[0136]

[0137] Where σ is the softmax function, DKL is the Kullback–Leibler divergence, AP1 and AP2 are the average probabilities belonging to the corresponding features, and the loss function Loss is used during network training. IP , to force the model to learn features that are robust to the radiometric differences between the two-phase images;

[0138] For example, in the CCR loss function, the challenge of change detection is that dual-phase images are severely affected by noise, seasonal changes, and imaging angle differences. Therefore, it is crucial to strengthen the similarity between images in the feature space, because the probability of changes between images is usually lower than the probability of no changes. In order to strengthen the similarity between images, a CCR is developed that retains the semantic similarity between temporal image pairs in the feature space. Specifically, the network encoder in the BTCD algorithm is forced to generate similar activations of dual-phase images, thereby incorporating temporal invariance into the network model, which in turn implicitly enhances the model's robustness to noise changes in the pre-training stage. The BTCD network encoder generates activation maps from the enhanced transformed input image X1, which are represented by E′1∈R b×c×h×w and E″1″∈R b×c×h×w , where b represents the batch size, c represents the number of output channels, h and w represent the spatial dimensions. Similarly, the network encoder generates activation maps for the transformed X2 image, represented by E′2 and E″2″ respectively. In order to maintain the temporal consistency of the input images X1 and X2 in the feature space, a pairwise similarity matrix G′ between the feature maps E′ and E′2 is constructed.<E′1,E′2> , Similarly, the pairwise similarity matrix G″ between the feature graphs E″1″ and E″2″ is constructed.<E″1″,E″2″> , <, > represents the dot product, and the CCR loss function is defined as follows:

[0139]

[0140] where ||.|| F represents the Frobenius norm;

[0141] Exemplarily, in the total loss function, the three loss functions (BT, IP and CCR) are weighted and combined to derive the final network loss function:

[0142] Loss Total =Loss BT +αLoss IP +βLoss CCR

[0143] where α and β are loss balancing weights.

[0144] S4, selecting the encoder component that provides the best features for change detection and using it in the next stage encoder;

[0145] S5, input the labeled dual-temporal remote sensing image into the improved UNet++ model, and then use the prediction map and loss function output by the model for deep supervision;

[0146] S6. Save the trained model weights and use the test set to test to obtain the change detection result of the test image;

[0147] In this embodiment, combined with Figure 4 , Figure 5 ,Change detection datasets are selected to verify the performance of the BTCD method.,These datasets include LEVIRbuilding Change Detection dataset (LEVIR-CD) and Sun Yat-SenUniversity Change Detection dataset (SYSU-CD).,Due to the differences in the sizes of these datasets, they are uniformly cropped to 256×256.,For each dataset, the samples are randomly divided into three parts: training set,,validation set and test set;

[0148] For example, the training set is used for model fitting, the validation set is used to adjust the model during training, and the test set is used to evaluate the generalization performance of the model. In LEVIR-CD, the number of training sets, validation sets, and test sets is 7120, 1024, and 2048, respectively. Figure 3 An example of the LEVIR-CD dataset is shown. In SYSU-CD, the number of training sets, validation sets, and test sets are 12000, 4000, and 4000, respectively. Figure 4 The example of SYSU-CD dataset is shown. In the SSCL stage, a ResNet50 model is trained by using image pairs from the training set and the validation set (without labels). The batch size is set to 40 and trained for 400 epochs using the LARS optimizer. The initial learning rate is 0.003, which is adjusted by multiplying the batch size and dividing by 256. The trade-off parameter λ of the loss function is set to λ = 5×10 -3 , using 1.5×10 -6 The weight decay parameter value of , the hyperparameters α and β are set to 100 and 2000 respectively;

[0149] Furthermore, the performance of the BTCD algorithm in subsequent CD tasks was evaluated by fine-tuning. For consistency, ResNet50 was used as the feature extractor and the SOTAUNet++ model was used for fine-tuning. Data augmentation techniques including random flipping, cropping, and Gaussian blur were applied; that is, the AdamW optimizer was used, the initial learning rate was 0.003, the batch size was 30, and training was performed for 150 epochs.

[0150] In this embodiment, four indicators, accuracy, recall, F1 score and intersection over union (IoU), are used to compare the performance of the algorithm with the SOTA method. The proposed CD method is compared with several SOTA methods, including FC-EF, FC-Siam-Diff, FC-Siam-Conc, DASNet, STANet, SNUNet, DSIFN, DSAMNet, SRCDNet, BiDateNet, and the latest transformer-based methods BIT, ChangeFormer, MSTDSNet-CD and Hybrid-TransCD. We fine-tune the improved UNet++ model on the LEVIR-CD and SYSU-CD datasets using three strategies: random initialization (Rand-init), supervised ImageNet pre-training (ImageNet-sup) and pre-training using BTCD. Therefore, in order to evaluate the effectiveness of the pre-training method proposed in the technical solution of the present invention, we applied the above technical solution to the actual remote sensing image data change detection, input the actual remote sensing image data into the above model for detection, and output the change detection results of the remote sensing image based on the above model. We also conducted a detailed comparative analysis with several other self-supervised comparative pre-training methods, including CMC, MoCov2, SimCLR and BT.

[0151] Table 1 compares the performance of other SOTA models on the LEVIR-CD dataset. The comparison results are shown in Table 1:

[0152]

[0153]

[0154] As shown in Table 1, our method achieves the highest F1 score (91.30%) and IoU (83.81%) among all tested methods. Hybrid-TransCD and ChangeFormer achieve the second and third highest F1 scores (90.06% and 90.04%, respectively), and the third and second highest IoU (81.92% and 82.48%, respectively). Both methods use the self-attention mechanism, which can effectively capture global information and map it to multiple spaces using MultiHead, thereby enhancing the expressive power of the model. Therefore, Hybrid-TransCD and ChangeFormer outperform CNN-based methods in terms of CD results.

[0155] Furthermore, to evaluate the effectiveness of different pre-training methods, we fine-tuned the downstream CD network (improved UNet++) using the pre-trained models CMC, MoCov2, SimCLR, and BT. To ensure a fair comparison, we only used the pre-trained model parameters to initialize the backbone of the improved UNet++ model, i.e., the parameters of the ResNet50 encoder. In addition, we compared our BTCD method with several mainstream self-supervised comparative pre-training methods, including random initialization and ImageNet pre-training. These mainstream self-supervised comparative pre-training methods use the ResNet-50 model to pre-train on the ImageNet dataset and achieve good parameter initialization effects. We fine-tuned the encoder part of the ResNet-50 model obtained by these methods and used it for downstream CD tasks. The performance results of the improved UNet++ model on the LEVIR-CD dataset using different pre-training methods are shown in Table 2 below:

[0156]

[0157] As shown in Table 2, our pre-trained method obtains the highest F1 score (91.30%) and IoU (83.81%). MoCov2 and BT obtain the second and third highest F1 scores (89.19% and 89.06%, respectively) and IoU (80.49% and 80.28%, respectively).

[0158] For example, the visual comparison results of the LEVIR-CD dataset are shown in Figure 6As shown in Figure 2, these scenes reflect the growth of buildings, mainly including the change from soil, grass and paved ground to newly built areas. Since CMC, MoCov2, SimCLR and BT are all self-supervised contrastive pre-training methods based on the ImageNet dataset, they lack prior knowledge about the CD task, which limits the performance of the CD model to a certain extent. Considering the domain gap between ImageNet and remote sensing images, our BTCD method directly performs self-supervised contrastive pre-training on the LEVIR-CD training dataset to learn representations within the domain. The results show that this pre-training strategy is more suitable for downstream CD tasks.

[0159] The performance comparison results with other SOTA models on the SYSU-CD dataset are shown in Table 3 below:

[0160]

[0161]

[0162] As shown in Table 3, our method achieved the highest F1 score (81.69%) and IoU value (69.04%). MSTDSNetCD ranked second with an F1 score of 80.33% and an IoU value of 67.13%, while Hybrid-TransCD ranked third with an F1 score of 80.13% and an IoU value of 66.84%. MSTDSNet-CD combines multi-scale SwinTransformer and deep supervision network for CD tasks, while the Hybrid-TransCD method adopts a hybrid multi-scale Transformer module and simulates the multi-scale hybrid attention representation of dual-phase images using a fine-grained self-attention mechanism. Both methods can enhance the expressive power of the model, thereby outperforming CNN-based methods in CD results. For the SYSU-CD dataset, the F1 score of the ImageNet pre-training method is slightly higher than that of the random initialization method, but the difference is not significant.

[0163] Due to the significant distribution difference between the ImageNet dataset and the SYSU-CD dataset, domain shift occurs. Therefore, training a general model in an unsupervised manner is more effective and robust for downstream CD tasks. Obtaining the ImageNet pre-trained model requires millions of labeled samples for supervised learning. The results in Table 3 further show that our BTCD method achieves the best results in detection performance and has higher accuracy than other methods.

[0164] Exemplarily, the performance of the improved UNet++ model on the SYSU-CD dataset using different pre-training methods is shown in Table 4 below:

[0165]

[0166] As shown in Table 4, for the SYSU-CD dataset, we fine-tuned the downstream CD network (improved UNet++) using the pre-trained models CMC, MoCov2, SimCLR, and BT. Our BTCD pre-training method achieved the highest F1 score (81.69%) and IoU (69.04%), the MoCov2 method ranked second with an F1 score of 81.31% and an IoU of 68.50%, and BT ranked third with an F1 score of 80.35% and an IoU of 67.19%;

[0167] Specifically, the visual comparison results of the SYSU-CD dataset are as follows: Figure 7 As shown in Figure 2, these scenes depict various types of changes, including new urban construction, preliminary foundation work before construction, and offshore construction activities. The CD results obtained using the random initialization, ImageNet-sup, and CMC-based methods have more omissions ( Figure 7 In comparison, MoCov2 and BTCD perform better and provide more complete detection results. Figure 7 Rows 1-3 in show that for complex CD scenes, our BTCD method can detect contour information better than other tested methods. Overall, our BTCD method has fewer errors and higher robustness than all other tested methods, providing superior performance for the CD task.

[0168] Preferably, we fine-tune the UNet++ model on the LEVIR-CD and SYSU-CD datasets through three strategies: Rand-init, ImageNet-sup, and pre-training with BTCD. To evaluate the effectiveness of our proposed pre-training method, we conduct a comprehensive comparative analysis of several other self-supervised pre-training methods, including CMC, MoCov2, SimCLR, and BT.

[0169] The results of the ablation experiments on the LEVIR-CD and SYSU-CD datasets are shown in Table 5 below:

[0170]

[0171] As shown in Table 6, our proposed pre-training method (BTCD) produces the best performance. Compared with Rand-init, applying our BTCD pre-training process improves the F1 score of LEVIR-CD by nearly 3.50% and the IoU by about 5.56%. For SYSU-CD, the F1 score rises by about 3.52% and the IoU by about 4.88%. On the ImageNet dataset, self-supervised comparative pre-training using CMC, MoCov2, SimCLR, and BT outperforms ImageNet-sup in terms of F1 score and IoU when transferred to the LEVIR-CD and SYSU-CD datasets.

[0172] Specifically, on the LEVIRCD dataset, our proposed pre-training method outperforms the best self-supervised contrastive pre-training method, MoCov2, by 2.11% and nearly 3.32% in IoU. For the SYSU-CD dataset, our proposed pre-training method is roughly comparable to MoCov2 in terms of F1 score and IoU, with slight improvements. In summary, ablation experiments show that our proposed self-supervised contrastive pre-training method for unlabeled images can reach or even exceed the widely used pre-training method based on the ImageNet dataset. In addition, our method alleviates the domain shift problem caused by transferring knowledge from ImageNet weights (obtained by pre-training with a dataset that is very different from the CD dataset). Qualitative and quantitative comparisons show that our proposed BTCD shows significant effectiveness and superiority for the CD task.

[0173] For example, the sensitivity analysis table of different loss balance hyperparameter settings for the YSU-CD dataset is shown in Table 6 below:

[0174]

[0175] Specifically, inspired by the observation that semantically similar remote sensing images should have consistent prediction results under different lighting conditions, in order to achieve invariant prediction, we use the IP loss function to force the BTCD model to learn features that are not affected by imaging lighting conditions. In addition, in order to mitigate the impact of illumination, noise, seasonal changes, and remote sensing imaging angle differences on CD results, we adopt the CCR loss function. This dual approach not only enhances the matching of dual-phase images in decision and feature spaces, but also reduces the impact of these confounding factors on CD results. To further analyze the impact of different loss balance hyperparameter settings on the final results, similar to the method of Tung and Mori

[30] , we determine the CD accuracy results of different hyperparameter settings for the SYSU-CD dataset as shown in Table 7. To evaluate the performance of IP loss, we keep the coefficient β of CCR loss unchanged and change the value of α. Increasing the value of α has limited impact on the downstream CD task performance of the BTCD method (Table 6). Therefore, we fixed the value of α to 100 and further analyzed the final impact on the accuracy of the CD task by changing the coefficient of the CCR loss, that is, modifying β. Since the scale of CCR loss is smaller than that of BT loss and IP loss, we increased β in increments of 1000 to balance the losses on the same scale; as shown in Table 6, increasing β initially slightly improved the performance; when α = 100 and β = 2000, the BTCD method achieved the highest F1 score and IoU on the SYSU-CD dataset. After further increasing β, the effect reached saturation. Therefore, the optimal settings of α and β for the SYSU-CD dataset are 100 and 2000, respectively. Subsequently, we also applied the same loss function hyperparameter values ​​to the LEVIR-CD dataset and obtained good CD results, which shows that the performance of the BTCD algorithm proposed in this paper is not sensitive to the hyperparameters of the final loss function. In addition, the parameters can be further optimized by cross-validating the validation sets of the two CD datasets.

[0176] For example, manually marking the changed areas is an expensive process that requires a lot of manpower and resources. Therefore, the lack of access to large-scale annotated data remains a major challenge in change detection using dual-temporal remote sensing images, such as Figure 8As shown in the figure, we tested various ratios of training data (20%, 40%, 60% and 100%) and tried various pre-training methods on the LEVIR-CD and SYSU-CD datasets. Compared with pre-training methods such as Rand-init, ImageNet-sup, CMC, MoCOv2, SimCLR and BT, our BTCD pre-training method performed well on both datasets. Even in the case of limited labeled data, the F1 score gradually improved with the increase of the ratio of training data. When using 60% of the training data, the BTCD method was applied to LEVIR-CD and SYSU-CD. When applied to the SU-CD dataset, BTCD obtained F1 scores of 0.8973 and 0.788, respectively, both of which are higher than the F1 scores of CNN-based methods using all training data. When the training data ratio is further reduced to 40%, BTCD obtains an F1 score of 0.8943 when applied to the LEVIR-CD dataset, which is only slightly lower than the scores achieved by ChangeFormer and Hybrid-TransCD; and an F1 score of 0.7787 when applied to the SYSU-CD dataset, which is comparable to the performance indicators of DSAMNet, SRCDNet, and BIT. When the training data ratio is reduced to 20%, BTCD obtains F1 scores of 0.8486 and 0.7504 when applied to the LEVIR-CD and SYSU-CD datasets, respectively, which is slightly higher than the F1 scores of FC-EF, FC-Siam-Diff, FC-Siam-Conc, and DASNet. We demonstrate that the BTCD pre-training method we developed can significantly improve the generalization ability and data efficiency of downstream change detection models. It is worth noting that we achieve competitive results using only 20% of the training samples for model training. In summary, our BTCD pre-training method can significantly improve the robustness and performance of change detection models when labeled data is limited.

[0177] Preferably, LULCCD requires the model to exhibit true generalization capabilities. To validate the generalization performance of the proposed BTCD framework, we conduct experiments on the extensive WHU Building CD (WHU-BCD) dataset. This dataset covers the area affected by the 6.3 magnitude earthquake in February 2011, which was rebuilt in the following years. Our focus is on Christchurch, New Zealand, a city that has seen a large number of new buildings since the earthquake. The dataset includes aerial images taken in April 2012, covering an area of ​​20.5 km. 2, contains 12,796 buildings (the 2016 dataset contains 16,077 buildings in the same area), and the ground sampling distance of these images is 0.2 meters / pixel; in addition, the dataset also provides a reference change mask and a pair of registered aerial images with a combined size of 15,354×32,507×3 pixels. The remote sensing images of 2012 and 2016 are shown in Figure 2. Fig. 9 a and 9b, while the ground truth of building changes is shown in Fig. 9 c, in which the black area represents the unchanged area and the white area represents the changed building area. We crop the bi-temporal image into smaller blocks of 256×256×3 pixels and resize the edge blocks to 256×256×3 pixels, finally obtaining 7620 pairs of image blocks.

[0178] Exemplarily, in the self-supervised contrastive pre-training stage, we use these 7620 pairs of unlabeled image patches to perform SSCL. SSCL involves designing an instance discrimination task to identify the image in the dataset that is most similar to the input features. It maximizes the similarity between two positive samples while increasing the distance between positive and negative samples. In the fine-tuning CD stage, the trained encoder is equipped with a robust feature representation and is transferred to the downstream CD task. We conducted experimental analysis on the WHU-BCD dataset using the improved UNet++ model. The results are spliced ​​together according to the relative positions of the cropped image patches, and the edge patches are resized back to their original size. The final experimental results on the WHU-BCD dataset are shown in Figure 2. Fig. 9 As shown in Figure 2, we evaluated the accuracy of the results and obtained a precision of 0.7674, a recall of 0.7771, an F1 score of 0.7722, and an IoU of 0.6289. Both quantitative and qualitative experimental results show that the proposed BTCD framework has strong generalization performance in practical applications.

[0179] In addition, it is worth noting that our method is pre-trained on RGB images due to the nature of the two public datasets used, which only contain RGB images. Therefore, we have to exclude information outside the visible spectrum. However, in practical applications, most remote sensing images contain not only RGB bands but also information outside the visible spectrum. This limitation affects the application of our method on multispectral satellite image data such as those from Sentinel-2 and Landsat. In future work, we plan to extend the proposed BTCD algorithm and study the adaptation strategy from RGB images to multispectral images, thereby improving the wide applicability of the method.

[0180] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the embodiments of the present invention are described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations shall fall within the scope defined by the appended claims.

[0181] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device 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 device.

[0182] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A remote sensing image change detection method based on self-supervised feature difference learning, characterized by: The following steps are involved: S1. Collect dual-temporal remote sensing images, create a remote sensing image dataset, divide the remote sensing image dataset into a training set, a validation set, and a test set, and cut the dual-temporal remote sensing images in the remote sensing image dataset into the same size, wherein the dual-temporal remote sensing images all have pixel-level annotations; S2. Construct a BTCD algorithm architecture of Barlow Twins self-supervised contrastive pre-training and fine-tuning CD based on feature differences, including a self-supervised contrastive pre-training Self-supervised contrastive learning, SSCL stage and a fine-tuning CD stage. The self-supervised contrastive pre-training Self-supervised contrastive learning, SSCL stage includes a feature extractor and a projection network, and the fine-tuning CD stage includes an improved UNet++ model. The improved UNet++ model uses an encoder-decoder architecture including a spatial and channel compression and excitation module, and the feature extractor uses ResNet50; S3, using the self-supervised contrast pre-training BTCD algorithm to perform self-supervised pre-training processing based on the BT algorithm on the unlabeled dual-temporal remote sensing images; S4, selecting the encoder component that provides the best features for change detection and using it in the next stage encoder; S5. Input the labeled dual-temporal remote sensing image into the improved UNet++ model, and use the prediction map and loss function output by the model for deep supervision; S6. Save the trained model weights and use the test set to test to obtain the change detection result of the test image.

2. The remote sensing image change detection method based on self-supervised feature difference learning according to claim 1, characterized in that: The operation process of the BTCD architecture of Barlow Twins self-supervision comparison pre-training and fine-tuning CD based on feature differences is as follows: S201: Definition and is at time t 1 and t 2 Bi-temporal images of the same geographic location captured separately are pre-trained on an unlabeled training set using a self-supervised algorithm so that the pre-trained encoder is fine-tuned to the labeled training set S202: In the self-supervised contrast pre-training SSCL stage, only bi-temporal images are used and For training, in the fine-tuning CD phase, bi-temporal images are used and And the labeled images Y in the training set train Conduct training; S203: After the self-supervised contrast pre-training SSCL phase is completed, a pre-trained encoder with initialization parameters is obtained; S204: In the fine-tuning CD stage, the bands of the dual-phase image are further superimposed together to create a new image with twice the number of bands, which is then input into the improved UNet++ model. The improved UNet++ model is end-to-end trained using a pre-trained encoder. The improved UNet++ model extracts features from the given dual-phase image and classifies it as "changed" or "unchanged", generating a prediction map.

3. The remote sensing image change detection method based on self-supervised feature difference learning according to claim 2, characterized in that: The self-supervised contrast pre-training SSCL stage in S202 further includes: initializing the self-supervised contrast pre-training model with supervised ImageNet weights, training an encoder E θ , where θ is a learnable parameter used to solve a pre-training task. The BTCD architecture uses three objective functions during training: BT loss function, IP loss function and CCR loss function, and performs self-supervision through hybrid loss as the loss function of the pre-training task "contrastive learning". The encoder E θ The parameter θ at the i-th iteration i The update looks like this: Among them, θ i is the iteration parameter, κ is the learning rate, represents the partial derivative, represents the loss function.

4. The remote sensing image change detection method based on self-supervised feature difference learning according to claim 3, characterized in that: The three objective functions: BT loss function, IP loss function and CCR loss function further include: The BT loss function is defined as follows: Among them, λ is a weight constant, b is the index batch sample, i and j are the vector dimensions of the index network output, C is the cross-correlation matrix between the difference representations D1 and D2, by maximizing the information of the difference representation (D1, D2) between the corresponding bi-temporal images, making them similar in the feature space, and then the model uses the modified BT loss function Loss BT Train in a self-supervised manner; The IP loss function is defined as follows: Where σ is the softmax function, DKL is the Kullback–Leibler divergence, AP1 and AP2 are the average probabilities of belonging to the corresponding features. When using bi-temporal images for change detection, it is necessary to ensure that samples with semantic similarity produce similar predictions. In order to achieve IP loss, the features between the enhanced views (X′1, X′2) and (X″1″, X″2″) of the bi-temporal images are enhanced by using the Jensen-Shannon divergence D JS As a goal, the loss function Loss is used during network training. IP , to force the model to learn features that are robust to the radiometric differences between the two-phase images; The definition of the CCR loss function is as follows: where ||.|| F Represents the Frobenius norm. The network encoder in the BTCD architecture generates activation maps from the enhanced input image X1. These activation maps are represented by E′1∈R b×c×h×w and E″1″∈R b×c×h×w , where b represents the batch size, c represents the number of output channels, h and w represent the spatial dimensions. The network encoder generates activation maps for the transformed X2 image, represented by E′2 and E″2″ respectively. In order to maintain the temporal consistency of the input images X1 and X2 in the feature space, a pairwise similarity matrix G′= between the feature maps E′ and E′2 is constructed.<E′1,E′2> , a pairwise similarity matrix G″ between feature graphs E″1″ and E″2″ is constructed.<E″1″,E″2″> , where <, > represents the dot product; In the total loss, the three loss functions (BT, IP, and CCR) are weighted and combined to derive the final network loss function: Loss Total =Loss BT +αLoss IP +βLoss CCR where α and β are loss balancing weights.

5. The remote sensing image change detection method and system based on self-supervised feature difference learning according to claim 3, characterized in that: The S204 further includes: the improved UNet++ model captures multi-scale features through skip connections and integrates them together through feature stacking, using an encoder-decoder architecture including a spatial and channel compression and excitation scSE module, and in the fine-tuning CD stage, using the pre-trained encoder E Φ The improved UNet++ model is trained end-to-end to improve the UNet++ model from the given dual-phase image and Extract features from the dataset and classify them as "changed" or "unchanged" to predict the change map The calculation formula is as follows: in Indicates the decoder and CD head, is a learnable parameter, using the parameters Φ and The improved UNet++ model is updated as follows: Among them, Loss CD represents the loss function, Y train represents the true label.

6. The remote sensing image change detection method based on self-supervised feature difference learning according to claim 1, characterized in that: The operation process of the self-supervised comparison pre-training BTCD algorithm is as follows: S301: The BTCD algorithm takes the bi-temporal image (X1, X2) as input, applies color distortion and Gaussian blur transformation to obtain an enhanced image (X1→X′1, X″1″, X2→X′2, X″2″); S302: adopting a twin network architecture, wherein the twin network architecture uses ResNet50 excluding a final classification layer as a feature extractor, followed by a projection network; S303: The input image is subjected to multi-layer convolution and pooling operations, and the receptive field is further enhanced by using dilated convolution. A convolution kernel with a step size of 16 is used to reduce the dimension of the input image, and a feature vector with a dimension of 16×16×2048 is generated to capture the basic features. S304: Apply two-dimensional adaptive average pooling to the feature vector and pass it to the projection head; the projection network contains two linear layers, each linear layer has a hidden layer size of 512 output units, and the first layer of projection is followed by a batch normalization layer and a rectified linear unit; S305: applying the absolute difference to the output embedding to obtain a difference embedding; applying the BT loss function to the difference embedding of size 1×256 to generate a cross-correlation matrix of size 256×256; S306: Using a twin encoder (E θ ), extract feature vectors E'1 and E'2 from the bi-temporal image by nonlinear projection Applied to the encoded feature vectors E'1 and E'2 to generate representations Z'1 and Z'2, the absolute feature difference is then applied to the projected output to learn representations of the changing features between the input images, as described below: D1=|Z′1-Z′2|=|G(E(X′1))–G(E(X′2))| D2=|Z"1-Z"2|=|G(E(X"1))–G(E(X"2))| When the semantic changes between the bi-temporal image features (X′1, X′2) and (X″1, X″2) are consistent, the model is constrained and ensured by applying the BT objective function on the difference feature representations D1 and D2. Achieve maximum cross-correlation of changing features; S307: The model learns specific information about the changes between the two-phase images. After the self-supervised contrast pre-training phase is completed, the encoder (E Φ ) parameters are passed to the downstream fine-tuning CD stage tasks.

7. The remote sensing image change detection method based on self-supervised feature difference learning according to claim 6, characterized in that: The testing using the test set includes: S601: Select a change detection data set to verify the performance of the BTCD algorithm, the change detection data set includes: a LEVIR-CD data set and a SYSU-CD data set; S602: the change detection datasets are uniformly cropped to 256×256, and for each dataset, samples are randomly divided into three parts: a training set, a validation set, and a test set. The training set is used for model fitting, the validation set is used for adjusting the model during training, and the test set is used for evaluating the generalization performance of the model. S603: In the LEVIR-CD dataset, the number of training sets, validation sets, and test sets are 7120, 1024, and 2048, respectively; in SYSU-CD, the number of training sets, validation sets, and test sets are 12000, 4000, and 4000, respectively; S604: In the self-supervised comparative pre-training SSCL stage, the ResNet50 model is trained using image pairs from the training set and the validation set, the batch size is set to 40, and the LARS optimizer is used to train 400 epochs, the initial learning rate is 0.003, which is adjusted by multiplying the batch size and dividing by 256, and the trade-off parameter λ of the loss function is set to λ = 5×10 -3 , using 1.5×10 -6 The weight decay parameter value of , the hyperparameters α and β are set to 100 and 2000 respectively; S605: Evaluate the performance of the BTCD algorithm in subsequent CD tasks by fine-tuning, using ResNet50 as the feature extractor, fine-tuning with the SOTAUNet++ model, and applying data augmentation techniques including random flipping, cropping, and Gaussian blurring; that is, using the AdamW optimizer, with an initial learning rate of 0.003, a batch size of 30, and training for 150 epochs; S606: Use four indicators, accuracy, recall, F1 score and intersection over union (IoU), to compare the performance of the algorithm with the SOTA method. Compare the proposed CD method with several SOTA methods. Fine-tune the improved UNet++ model on the LEVIR-CD and SYSU-CD datasets using three strategies: random initialization, supervised ImageNet pre-training, and pre-training using the BTCD algorithm. In order to evaluate the effectiveness of the proposed pre-training method, a comparative analysis is performed with several other self-supervised comparison pre-training methods. Output the change detection results of the test image.

8. A remote sensing image change detection system based on self-supervised feature difference learning, characterized by: The system comprises: Remote sensing image dataset module, used to collect dual-temporal remote sensing images and produce remote sensing image datasets; BTCD algorithm architecture module, used to build BTCD architecture for BT self-supervision contrast pre-training and fine-tuning CD based on feature differences; A pre-training processing module for pre-training an unlabeled training set using a self-supervised algorithm; UNet++ model module, which is used to perform deep supervision on labeled bi-temporal remote sensing images and loss functions based on the improved UNet++ model; The test module is used to perform testing using the test set to obtain the change detection result of the test image.

Images