A method and system for detecting incremental changes in remote sensing image domain based on frequency domain distillation and difference-guided decoupling

Abstract

This invention discloses a method and system for detecting incremental changes in remote sensing image domains based on frequency domain distillation and difference-guided decoupling. The method includes: acquiring dual-temporal remote sensing images and dividing them into old and new domain datasets; constructing a change detection network containing a teacher model and a student model, embedding a difference-guided decoupling adapter, a hollow spatial pyramid pooling module, and a multi-head task decoder; training the teacher model using the old domain data, extracting deep feature amplitude spectra, and calculating the old domain global amplitude spectrum prior; freezing the teacher model during incremental learning of the new domain, extracting the new domain features from the student model, retaining the phase spectrum and replacing it with the old domain amplitude spectrum prior, and reconstructing cross-domain synthetic features; iteratively updating the adapter parameters by jointly using cross-entropy loss, dual knowledge distillation loss, frequency domain phase alignment loss, and adapter sparsity constraint loss; and dynamically calling the corresponding decoder head to output results based on the domain label during inference. This invention effectively decouples image style and change semantics, mitigates catastrophic forgetting, and improves the robustness of cross-domain change detection.

Description

Technical Field

[0001] This invention relates to the fields of image processing and computer vision technology, specifically to a method and system for detecting incremental changes in the domain of remote sensing images based on frequency domain distillation and difference-guided decoupling. Background Technology

[0002] Remote sensing image change detection identifies dynamic changes in the land surface by analyzing multi-temporal images of the same area acquired at different times. It is widely used in areas such as urban expansion monitoring, disaster assessment, and land resource management. In recent years, deep learning technology, with its powerful nonlinear modeling and hierarchical feature extraction capabilities, has significantly improved the accuracy and automation of change detection.

[0003] Traditional deep learning change detection networks are typically trained on static, closed datasets, implicitly assuming that training and test data are independent and identically distributed. However, in real-world remote sensing applications, data often originates from different sensors, geographical regions, or seasons, leading to significant domain shifts and causing a sharp decline in the performance of existing networks in new scenarios. To improve the network's adaptability to new domains, directly mixing and jointly training new and old data is limited by factors such as storage costs, computational efficiency, and data privacy; while fine-tuning the old model using new domain data can lead to catastrophic forgetting of knowledge from the old domain. Therefore, domain incremental learning techniques have been introduced, requiring the network to effectively extract features from the new domain and firmly maintain knowledge from the old domain, even when it has no access or only access to a very small number of historical samples.

[0004] In existing domain incremental learning methods, knowledge distillation and parameter isolation mechanisms are widely used because they do not require storing historical image data. However, these methods generally lack explicit distinction between image style information and physical structure information. Their absolute feature constraints tend to bias the decision boundary towards old domain features, causing a conflict between plasticity and stability, resulting in insufficient adaptation to the new domain and blurred boundaries of the old domain. Some studies have introduced multi-scale feature alignment or dynamic routing mechanisms to improve multi-domain representation, but their single-stream processing architecture and static constraints still have limitations. When the new domain image has a strong style shift or background interference, the single-stream adapter, not guided by dual-temporal difference features, tends to fit the domain style of a single image rather than the essence of the change, resulting in structural distortion of cross-domain change detection results and a large number of false detections and false negatives.

[0005] In summary, existing domain incremental learning methods fail to effectively decouple variable domain styles from invariant spatial structures at the feature level, and lack dynamic adaptation mechanisms for dual-temporal difference features, making it difficult to achieve high-quality continuous change detection in complex and ever-changing remote sensing scenarios. Summary of the Invention

[0006] The purpose of this invention is to provide a remote sensing image domain incremental change detection method and system based on frequency domain distillation and difference-guided decoupling, so as to solve the problems of catastrophic forgetting and structural distortion caused by the coupling of image style and structural features and the lack of dual-temporal difference guidance in existing domain incremental change detection tasks.

[0007] To achieve the above objectives, the technical solution provided by this invention is: a method for detecting incremental changes in the remote sensing image domain based on frequency domain distillation and difference-guided decoupling, comprising the following steps: S1: Obtain a dataset of dual-temporal remote sensing images containing multiple types of land cover changes and pixel-level change labels. After cropping at a uniform scale, divide the dataset into an old domain dataset and a new domain dataset according to the domain attributes, and assign a corresponding domain label to each sample. S2: Construct a change detection network for domain incremental learning, including a teacher model and a student model, both with identical structures, each containing a residual backbone network for extracting multi-level bi-temporal features; a difference-guided decoupling adapter, used to generate channel attention weights based on the absolute difference map of the corresponding level bi-temporal features, and inject the recalibrated corresponding level bi-temporal features into the residual backbone network in residual form; a hollow spatial pyramid pooling module connected to the deepest layer of the residual backbone network; and a multi-head task decoder set at the end of the change detection network, consisting of multiple independent decoding heads. S3: After training the teacher model using the old domain dataset, freeze its parameters; input the old domain dataset into the teacher model, extract the dual-temporal deep features of the teacher model, obtain the amplitude spectrum through fast Fourier transform, calculate the average along the sample dimension, calculate the average amplitude spectrum as the prior of the old domain global amplitude spectrum and store it. S4: Input the new domain dataset simultaneously into the teacher model with frozen parameters and the student model to be trained; extract the bi-temporal deep features of the student model and denote them as new domain features, retain the phase spectrum of the new domain features, replace its amplitude spectrum with the prior global amplitude spectrum of the old domain, recombine the replaced amplitude spectrum and the retained phase spectrum to form a complex spectrum, and reconstruct it into cross-domain synthetic features through inverse transformation; construct a joint loss function, which includes cross-entropy loss, double knowledge distillation loss, frequency domain phase alignment loss and adapter sparsity constraint loss; iteratively update the difference-guided decoupling adapter parameters and the current new task decoder parameters in the student model through backpropagation until the model converges, and keep the remaining parameters frozen; S5: Input the dual-temporal remote sensing image data to be detected and its domain labels into the trained student model, dynamically call the corresponding decoding head in the multi-head task decoder according to the domain labels, and output the ground feature change detection results.

[0008] To optimize the above technical solution, the specific measures also include: In step S2, the difference-guided decoupling adapter is embedded in parallel after each shallow and mid-layer residual block of the residual backbone network in a bypass manner to participate in the parameter update of the student model. Specifically, it is used to: calculate the absolute difference map of the corresponding layer's dual-temporal features and use it as prior information indicating the intensity of change; extract the global context descriptor of the current single-temporal features and the absolute difference map using global average pooling, and concatenate the two in the channel dimension to obtain a joint feature vector that integrates semantic content and change prior; input the joint feature vector into the first fully connected layer, the ReLU activation function, the second fully connected layer, and the Sigmoid activation function in sequence to generate channel attention weights; use the channel attention weights to recalibrate the corresponding layer's dual-temporal features by multiplying them channel by channel, and inject the recalibrated features into the residual backbone network through residual connections after convolution.

[0009] Furthermore, in step S2, the hollow spatial pyramid pooling module is connected to the deepest layer of the residual backbone network, receives the corresponding level dual-temporal features after correction by the difference-guided decoupling adapter, and performs multi-scale sampling by setting parallel hollow convolutions with different dilation rates to enhance the context awareness capability of the change detection network for regions of different sizes of change.

[0010] In step S2, the multi-head task decoder is specifically used to: splice and fuse the high-level features output by the hollow spatial pyramid pooling module with the low-level features output by the shallow layer of the residual backbone network to gradually restore the spatial resolution; and dynamically activate the corresponding decoding head to generate a change detection probability map according to the given domain label.

[0011] Further, in step S3, the residual backbone network of the teacher model is used to extract the dual-temporal deep features of the old domain dual-temporal remote sensing image. The extracted dual-temporal deep features are applied to each channel with a two-dimensional real number fast Fourier transform to map the spatial domain features to the frequency domain, thereby obtaining the corresponding complex spectral representation. The modulus of the complex spectral representation is calculated to obtain the amplitude spectrum. All extracted amplitude spectra are averaged along the dimension of the sample set to generate and store the old domain global amplitude spectrum prior representing the global statistical law of the old domain.

[0012] In step S4, the joint loss function includes cross-entropy loss, dual-knowledge distillation loss, frequency domain phase alignment loss, and adapter sparsity constraint loss, specifically: The cross-entropy loss The expression is:

[0013] Furthermore, the loss of the dual knowledge distillation The expression is:

[0014]

[0015]

[0016] in, Indicates pixel-level change labels. This represents the predicted probability map of the student model in the current new domain task's decoding head output; This represents the new domain difference features extracted by the student model; This represents the cross-domain synthetic difference features obtained from frequency domain decoupling and reconstruction; Indicates the original characteristic distillation loss; This indicates the characteristic distillation loss in cross-domain synthesis; This represents the knowledge distillation loss calculated on the features of the new domain; Indicating cross-domain synthetic features The distillation loss calculated above; Represents the Kullback-Leibler divergence function; This indicates a normalization operation; and The student model and the teacher model output different predicted logical values ​​on the old task decoding head.

[0017] Furthermore, the frequency domain phase alignment loss The expression is:

[0018] The adapter sparse constraint loss The expression is:

[0019] in, and These represent the student model's response to dual-temporal remote sensing images. and Phase spectrum corresponding to the extracted high-level semantic features; and These represent the reference phase spectra extracted by the teacher model for the same image pair; This represents the mean square error calculation, used to quantify the degree of phase shift; The number of layers in the change detection network that embed a difference-guided decoupling adapter; For the first The number of channels in the layer; For the first Layer Attention weights for each channel.

[0020] As another important technical solution, this invention also provides a remote sensing image domain incremental change detection system based on frequency domain distillation and difference-guided decoupling, used to implement the remote sensing image domain incremental change detection method based on frequency domain distillation and difference-guided decoupling as described above, including: The domain-adaptive data preprocessing module is used to acquire a dataset containing dual-temporal remote sensing images with multiple land cover change types and pixel-level change labels. After being cropped at a uniform scale, the dataset is divided into an old domain dataset and a new domain dataset according to the domain attributes, and each sample is assigned a corresponding domain label. The module for constructing the domain incremental change detection network is used to build a change detection network for domain incremental learning. It includes a teacher model and a student model, both of which have the same structure and include a residual backbone network for extracting multi-level bi-temporal features. The module for differential guidance and decoupling is used to generate channel attention weights based on the absolute difference map of the corresponding level bi-temporal features, and injects the corresponding level bi-temporal features into the residual backbone network in the form of residuals after recalibration. The module for hollow spatial pyramid pooling is connected to the deepest layer of the residual backbone network. The module for multi-head task decoder is set at the end of the change detection network and consists of multiple independent decoder heads. The module for freezing the old domain teacher model and extracting style priors is used to freeze the parameters of the teacher model after training it with the old domain dataset. The old domain dataset is input into the teacher model to extract the dual-temporal deep features of the teacher model. The amplitude spectrum is obtained by fast Fourier transform, and the average is calculated along the sample dimension. The average amplitude spectrum is used as the old domain global amplitude spectrum prior and stored. The cross-domain synthetic feature distillation training module is used to simultaneously input the new domain dataset into the teacher model with frozen parameters and the student model to be trained; extract the bi-temporal deep features of the student model and record them as new domain features, retain the phase spectrum of the new domain features, replace its amplitude spectrum with the prior global amplitude spectrum of the old domain, and recombine the replaced amplitude spectrum and the retained phase spectrum to form a complex spectrum, which is then reconstructed into cross-domain synthetic features through inverse transformation; construct a joint loss function, which includes cross-entropy loss, dual knowledge distillation loss, frequency domain phase alignment loss, and adapter sparsity constraint loss; iteratively update the difference-guided decoupling adapter parameters and the current new task decoder parameters in the student model through backpropagation until the model converges, while the remaining parameters remain frozen; The domain-aware decoding head dynamic inference module is used to input the dual-temporal remote sensing image data to be detected and its domain labels into the trained student model, dynamically call the corresponding decoding head in the multi-head task decoder according to the domain labels, and output the ground feature change detection results.

[0021] The present invention also proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method for detecting incremental changes in the remote sensing image domain based on frequency domain distillation and difference-guided decoupling.

[0022] The present invention also proposes a computer-readable storage medium storing a computer program that enables a computer to execute a remote sensing image domain incremental change detection method based on frequency domain distillation and difference-guided decoupling as described above.

[0023] Compared with the prior art, the beneficial effects of the present invention are: This invention introduces a teacher-student model architecture and a dual knowledge distillation strategy, and applies phase alignment constraints in the frequency domain, enabling the student model to retain the spatial structure information of the old domain's features held by the teacher model when learning new domain features. Simultaneously, by extracting and utilizing the global amplitude spectrum prior of the old domain to generate cross-domain synthetic features for distillation, the change detection network's memory of historical decision boundaries is further strengthened. Experimental results show that the performance degradation of this invention in the old domain is significantly less than that of existing methods, effectively overcoming the catastrophic forgetting problem.

[0024] In this invention, the Differential Guided Decoupling Adapter (DGDA) uses the absolute difference map of the corresponding hierarchical dual-temporal features as a priori to generate channel attention weights to recalibrate the features. This enables the change detection network to focus on semantic channels related to change and suppress domain style information that is irrelevant to the task. Combined with the adapter sparse constraint loss, the change detection network is forced to represent change with the fewest key channels, thus achieving effective decoupling of change semantics from background style. This mechanism enables the change detection network to accurately identify real change areas even when faced with drastic style shifts such as illumination, season, and sensor, significantly reducing false detections and false negatives.

[0025] This invention combines the Spatial Pyramid Pooling (ASPP) module with a multi-head task decoder. While capturing multi-scale contextual information, it maintains independent decoding heads for different domain tasks, avoiding parameter interference between tasks. The frequency domain phase alignment loss ensures that the student model and the teacher model are consistent in the phase spectrum of high-level semantic features, which stabilizes the change detection network's ability to perceive structural features such as ground object edges and shapes. By combining the above mechanisms, this invention achieves excellent detection performance in new domains while maintaining high accuracy in old domains, realizing a good balance between stability and flexibility. Attached Figure Description

[0026] Figure 1 : A schematic diagram of the overall framework in an embodiment of the present invention.

[0027] Figure 2: A structural diagram of the differential guidance decoupling adapter in this embodiment of the invention.

[0028] Figure 3 : A schematic diagram of the cross-domain feature synthesis process based on frequency domain decoupling in this embodiment of the invention.

[0029] Figure 4 Visual effect diagram of incremental change detection experiment in the embodiment of the present invention. Detailed Implementation

[0030] The present invention will be further described in detail below through specific embodiments, but it should not be construed as limiting the scope of the subject matter of the present invention to the following embodiments. All technologies implemented based on the above content of the present invention fall within the scope of the present invention.

[0031] This embodiment uses the LEVIR-CD dataset as the old domain dataset and the CDD dataset as the new domain dataset to simulate the domain incremental learning process of migrating from an urban building change detection scenario to a complex land cover change detection scenario. This application scenario can correspond to the process in actual remote sensing operations where the change detection network first completes initial training on building change data acquired by a certain urban area or a certain sensor, and is then deployed to new areas, different seasons, or different imaging conditions to continuously monitor newly added buildings, road changes, vegetation cover changes, water boundary changes, and disaster-affected areas.

[0032] This invention provides a method for detecting incremental changes in the remote sensing image domain based on frequency domain distillation and difference-guided decoupling, comprising the following steps: S1: Obtain a dataset of dual-temporal remote sensing images containing multiple types of land cover changes and pixel-level change labels. After cropping at a uniform scale, divide the dataset into an old domain dataset and a new domain dataset according to the domain attributes, and assign a corresponding domain label to each sample. The collected change detection data containing various land cover change types (including T1 time-phase images, T2 time-phase images, and true pixel-level change label maps) were cropped into 256×256 pixel image patches at a uniform scale. Based on domain attributes, the data were divided into the old domain dataset (LEVIR-CD) and the new domain dataset (CDD), and each sample was assigned a corresponding domain label. After cropping, the old domain training set, validation set, and test set contained 7120, 1024, and 2048 pairs of image patches, respectively; the new domain training set contained 10000 pairs of image patches, and the validation set and test set each contained 3000 pairs of image patches.

[0033] The term "domain" refers to the data distribution formed by factors such as the geographic region, sensor type, imaging season, imaging time, spatial resolution, and background feature distribution of remote sensing imagery. The "old domain" refers to the historical data distribution that the change detection network has already learned; this old domain dataset is used for initial training of the teacher model to extract the global amplitude spectrum prior. The "new domain" refers to the data distribution newly arrived in the current incremental learning phase that requires model adaptation; this new domain dataset is used to train updatable parameters in the student model. "Old task" and "historical task" refer to change detection tasks completed based on the old domain dataset, while "new domain task" and "new task" refer to change detection tasks learned based on the current new domain dataset. Domain labels are used to identify the domain data distribution to which the input sample belongs and are used by the multi-head task decoder to call the corresponding decoding head during the inference phase.

[0034] In some implementations, to comply with the data constraints of domain incremental learning, the training set of the old domain is only used for the pre-training of the teacher model in the initial stage. When learning the new domain task, the teacher model is strictly restricted to not being able to access the training set of the old domain, and can only read the training set and validation set of the new domain to update parameters. The validation set of the old domain is only used in the preparation stage to extract the global amplitude spectrum prior of the old domain. The test set of the old domain and the test set of the new domain are jointly retained for the final evaluation of the change detection network performance.

[0035] During the network training phase, in order to improve the generalization ability of the change detection network to complex ground feature changes and effectively alleviate overfitting, data augmentation was performed on the training set data, including random horizontal flipping, random vertical flipping, random geometric rotation, and random Gaussian blurring operations performed according to probability, so as to simulate the diverse observation perspectives and imaging quality differences in remote sensing images.

[0036] To ensure the objectivity and consistency of the performance evaluation of the change detection network, the above enhancement operations are not performed on the validation and test set data; only their original spatial structure is maintained.

[0037] All spatially transformed image data are uniformly normalized, and pixel values ​​are scaled to a standard range to eliminate absolute numerical differences such as illumination and accelerate network convergence. At the same time, they are converted into floating-point multidimensional tensor data format required for deep learning framework computation, while the change label map is converted into long integer tensor format.

[0038] S2: Construct a change detection network for domain-incremental learning, such as Figure 1As shown, the model includes a teacher model and a student model. The teacher model is a change detection model whose parameters are saved and frozen after training in the old domain. It is used to provide feature references, phase references, and decoding outputs of the old task during the incremental training of the current new domain. The student model is a change detection model to be trained and updated in the current incremental learning phase. It is used to learn the semantic information of the new domain task and acquire the ability to detect changes in the new domain. The teacher model and the student model have the same network structure. Figure 1 The teacher model block represents the parameter-frozen teacher model. Apart from the teacher model block and its output, the rest of the network body represents the overall structure of the student model. Both the teacher model and the student model include a residual backbone network for extracting multi-level bi-temporal features. A difference-guided decoupling adapter is used to generate channel attention weights based on the absolute difference map of the corresponding level bi-temporal features, and injects the corresponding level bi-temporal features into the residual backbone network in residual form after recalibration. A hollow spatial pyramid pooling module connected to the deepest layer of the residual backbone network is used to extract multi-scale contextual information from high-level difference features. A multi-head task decoder set at the end of the change detection network consists of multiple independent decoding heads and dynamically calls the corresponding decoding head according to the task or domain label. In some implementations, to preserve the general visual feature representations learned by the change detection network on large-scale datasets and to fundamentally suppress the forgetting of old knowledge, this invention employs a ResNet50 pre-trained on ImageNet as the residual backbone network for feature extraction. Throughout the incremental learning process, the parameters of all convolutional layers and batch normalization layers of the ResNet50 remain frozen and do not participate in gradient updates. To enable the frozen residual backbone network to adapt to the new domain data distribution, this invention adopts a bypass injection strategy. After the residual blocks of Layer 1, Layer 2, and Layer 3 of the ResNet residual backbone network, a Differential Guided Decoupling Adapter (DGDA) module is embedded in parallel as a lightweight training plugin. This module utilizes the differential information of the corresponding layer's bi-temporal features as prior guidance to dynamically correct the residual features of the residual backbone network. Through this combination of output features and domain-adaptive residuals, the change detection network can capture the specific style and semantic changes of the current domain without destroying the original feature structure.

[0039] In some implementations, the two-temporal remote sensing images T1 and T2 are input into two shared weight feature extraction branches of the student model, respectively, and sequentially pass through Layer 1 to Layer 4 of the residual backbone network to obtain low-level and high-level semantic features at different levels. Specifically, the DGDA module embedded after Layer 1 to Layer 3 generates channel attention weights based on the differences in the corresponding two-temporal features, enhancing channels related to semantic changes and suppressing channels related to background style or domain noise, and further applying adapter sparsity constraint loss. Constrain the sparsity of channel activation.

[0040] To address the issue of dramatic changes in ground feature scale in high-resolution remote sensing imagery, this invention retains the Hollow Spatial Pyramid Pooling (ASPP) module from the DeepLabV3+ core in both the student and teacher models. For the student model, the corresponding hierarchical bi-temporal features corrected by DGDA are first subjected to absolute difference operations to generate new domain difference features. Then, the feature is input into the ASPP module, which captures multi-scale contextual information through parallel dilated convolutions with different dilation rates, enhancing the student model's ability to perceive regions of varying sizes. For the teacher model, its corresponding ASPP module and old task decoder parameters are frozen in the current incremental training phase, used to output old task reference logic values, and distilled with the student model's outputs on new domain difference features and cross-domain synthetic difference features.

[0041] To address the semantic differences and label space conflicts between different tasks in incremental learning, this invention designs a multi-head task decoder structure. The multi-head task decoder consists of multiple independent decoding heads, including a trainable decoding head for the current new domain task and a frozen old task decoding head for historical tasks. Incremental tasks represent new domain change detection data arriving continuously in a time series, with each task corresponding to a new domain data distribution. For the currently learning new domain task… The change detection network instantiates a separate new task decoding head. The decoding head fuses the new domain difference features output by ASPP with the low-level detail features output by the residual backbone network Layer 1. After convolution and upsampling operations, it finally generates a change detection probability map with the same size as the original image. Simultaneously, the old task decoding head learned from previous tasks retains its parameters frozen during subsequent training, providing response values ​​representing the historical knowledge distribution during incremental training. During the inference phase, the change detection network dynamically activates the corresponding decoding head based on the given domain label, achieving accurate predictions for specific domains.

[0042] Build a differential bootstrap decoupling adapter, such as Figure 2 As shown, the differential-guided decoupling adapter achieves effective decoupling between changing semantics and domain style by designing a differential modulation channel recalibration mechanism, explicitly utilizing dual-temporal differential features as priors, and dynamically adjusting the activation state of the feature channels.

[0043] In some implementations, in the frozen residual backbone network feature space, truly changed regions typically exhibit significant differences in feature values, while unchanged background regions show smaller differences in deep semantic features. For the first... Two-phase characteristics of the layer , Calculate its absolute difference plot As prior knowledge indicating the intensity of change, it provides key spatial clues for difference-guided decoupling adapters.

[0044]

[0045] Extract the first digit using global average pooling respectively i Temporal feature global context descriptor and the first Descriptor for global statistical information of absolute difference of layered two-phase features:

[0046]

[0047] in, Indicates the first i Global context descriptor for temporal features, Represent and Phase; Indicates the first Descriptor for global statistical information of absolute difference of layered two-temporal features Indicates the first The first layer of input Characteristics of time phase.

[0048] By concatenating and combining the two along the channel dimension, a joint feature vector integrating semantic content and prior change is obtained:

[0049] in, The symbol represents the joint feature vector that integrates semantic content and prior changes. This indicates a feature concatenation operation along the channel dimension, i.e., all dimensions are... The i The global context descriptor of temporal features is concatenated with the descriptor of the global statistical information of the absolute difference of the two-temporal deep features, generating a dimension of The joint eigenvectors.

[0050] In some implementations, a bottleneck MLP structure containing two fully connected layers is used to capture nonlinear dependencies between channels, generating difference-guided channel attention weights. :

[0051] in, It is the ReLU activation function. It is the Sigmoid activation function. and These are learnable weight parameters. This refers to the dimensionality reduction ratio.

[0052] Obtained weights As an adaptive gating system, due to With the involvement of [the adapter], the adapter tends to generate high response values ​​on channels with significant differences and low response values ​​on channels with slight differences. This weight is used to recalibrate the corresponding level of bi-temporal features, filtering out domain style information that is irrelevant to the changes.

[0053] in, This indicates the features after difference-guided recalibration. This indicates a channel-wise multiplication operation, which utilizes the generated weights. Weight the original features.

[0054] The recalibrated features are injected into the backbone network via residual connections to generate corrected bi-temporal features for the corresponding levels. :

[0055] in, This indicates that the frozen residual backbone network is in the 1st... Output characteristics of layer residual blocks.

[0056] In some implementations, the attention weights of the sensing channels are varied. The degree of activation of the adapter parameters is determined. Without constraints, the adapter tends to be active on all channels to fit the texture details of the new domain, leading to increased semantic-style coupling. The sparsity constraint mechanism, however, causes the adapter to activate on semantically strongly correlated channels (i.e.,...). ), while remaining silent on a large number of irrelevant background channels (i.e. ).

[0057] Define adapter sparse constraint loss The mean of the attention weights generated for all DGDA modules:

[0058] in, The number of layers in the change detection network that embed a difference-guided decoupling adapter; For the first The number of channels in the layer; For the first Layer Attention weights for each channel.

[0059] In some implementations, by The optimizer imposes regularization constraints on DGDA and utilizes descriptors that contain global statistical information about the absolute differences of the two-phase features. This approach provides discriminative prior guidance to the change detection network, enabling the channel attention mechanism to accurately identify and activate feature channels highly correlated with change semantics, while suppressing excessive focus on non-change regions. This mechanism completes the current task with minimal neuron activation cost, thereby achieving feature decoupling, automatically filtering out background style noise, and focusing on change semantics. It also preserves the parameters of inactive channels, reserving learning space for more complex domain tasks.

[0060] S3: After training the teacher model using the old domain dataset, freeze its parameters; input the old domain dataset into the teacher model, extract the dual-temporal deep features of the teacher model, obtain the amplitude spectrum through fast Fourier transform, calculate the average along the sample dimension, calculate the average amplitude spectrum as the prior of the old domain global amplitude spectrum and store it. The calculation of the average amplitude spectrum as the prior of the old domain global amplitude spectrum is as follows: The residual backbone network of the teacher model is used to extract the bi-temporal deep features of the old domain bi-temporal remote sensing images. A two-dimensional real-number fast Fourier transform is applied to each channel of the extracted bi-temporal deep features to map the spatial domain features to the frequency domain, obtaining the corresponding complex spectral representation. The modulus of this complex spectral representation is calculated to obtain the amplitude spectrum. All extracted amplitude spectra are averaged along the sample set dimension to generate and store the old domain global amplitude spectrum prior, representing the global statistical regularity of the old domain. .

[0061] S4: Input the new domain dataset simultaneously into the teacher model with frozen parameters and the student model to be trained; extract the bi-temporal deep features of the student model and denote them as new domain features, retain the phase spectrum of the new domain features, replace its amplitude spectrum with the prior global amplitude spectrum of the old domain, recombine the replaced amplitude spectrum and the retained phase spectrum to form a complex spectrum, and reconstruct it into cross-domain synthetic features through inverse transformation; construct a joint loss function, which includes cross-entropy loss, double knowledge distillation loss, frequency domain phase alignment loss and adapter sparsity constraint loss; iteratively update the difference-guided decoupling adapter parameters and the current new task decoder parameters in the student model through backpropagation until the student model converges, and keep the remaining parameters frozen; In some implementations, during the forward propagation process, the bi-temporal remote sensing images first pass through Layer 1 to Layer 3 and the DGDA module within the student model. Utilizing the differential modulation channel recalibration mechanism within the DGDA module, the unique semantic information of the new domain is captured in real-time during the feature extraction stage, and the attention mask weights of each adapter layer are obtained. After the bi-temporal remote sensing images flow through the residual backbone network Layer 4, high-level semantic features are extracted, and cross-domain feature synthesis operations are performed, such as... Figure 3 As shown, the specific process is as follows: Extracting dual-temporal deep features from the student model Let this be denoted as the new domain feature; then map it to the frequency domain using the two-dimensional real fast Fourier transform:

[0062] in, Representing the characteristics of the new domain In frequency coordinates Complex spectrum representation at; Representing the feature map in spatial coordinates Eigenvalues ​​at; and These are the height and width of the feature map, respectively; For spatial domain coordinate index, Frequency domain coordinate index; It is the imaginary unit.

[0063] In the frequency domain, the complex spectrum is represented as... Decomposed into amplitude spectrum Phase spectrum :

[0064] in, This represents the operation of taking the modulus of the spectrum of a complex number; This represents the operation of taking the phase angle of a complex number's spectrum.

[0065] In some implementations, for new domain features Preserving the phase spectrum of its dominant spatial structure Discard its amplitude spectrum And using the pre-stored global amplitude spectrum prior of the old domain Replace its original amplitude spectrum. Reconstruct it back to the spatial domain using inverse Fourier transform to generate cross-domain synthetic features. :

[0066] in, This represents the inverse fast Fourier transform (iRFFT). The phase spectrum represents the current input features of the new domain, used to preserve the spatial structure information of the new domain features; It represents the unit complex exponential form constructed from the phase spectrum.

[0067] Pixel-wise absolute difference operations are performed on the cross-domain synthesized features to obtain cross-domain synthesized difference features for subsequent ASPP and decoder input. :

[0068] Cross-domain synthetic differential features generated through amplitude injection mechanism It has the style of the old domain in terms of visual and statistical distribution, but retains the spatial structure of the new domain.

[0069] In the teacher model With student model Soft constraints are established at the feature and decision levels to maximize the retention of discriminative ability in the old domain while adapting to the new domain. The student model outputs the new task decoder for the current domain. The change prediction logits are calculated using pixel-level change labels. Calculate task-specific cross-entropy loss This loss function drives the DGDA module to capture domain-specific change features and update the decoder parameters, ensuring the change detection network's detection accuracy in the new domain.

[0070]

[0071] in, This represents the predicted probability map of the student model in the current new domain task decoding head output.

[0072] In some implementations, to ensure that the change detection network does not compromise its original spatial feature extraction capabilities when adapting to new domain styles, this invention designs a selective constraint mechanism during knowledge distillation: the high-level semantic features extracted by the student model and the teacher model for the same input are respectively... and The phase spectrum of the student model obtained after frequency domain decoupling Phase spectrum with teacher model ， Consistency constraints are imposed, utilizing phase information to maintain the stability of feature boundaries and spatial topology; simultaneously, constraints on the amplitude spectrum of the carrying domain style attributes are ignored, allowing the student model to adaptively fit the new domain data distribution. Based on this, through calculation... and The difference between them constructs the frequency domain phase loss The expression is:

[0073] in, and These represent the student model's response to dual-temporal remote sensing images. and Phase spectrum corresponding to the extracted high-level semantic features; and These represent the reference phase spectra extracted by the teacher model for the same image pair; This represents the mean square error calculation, used to quantify the degree of phase shift.

[0074] Selective constraints allow the amplitude spectrum of the student model to evolve freely to fit new domain data, but fix the phase spectrum to ensure that the core structural awareness capability in the change detection task is not catastrophically forgotten. To solidify the discriminative boundary of the change detection network, knowledge distillation loss is calculated not only on new domain features but also on cross-domain synthesized features. This is a dual knowledge distillation loss. The average of the two:

[0075]

[0076]

[0077] in, This represents the new domain difference features extracted by the student model; This represents the cross-domain synthetic difference features obtained from frequency domain decoupling and reconstruction; This represents the knowledge distillation loss calculated on the differential features of the new domain. This represents the distillation loss calculated on the cross-domain synthetic difference characteristics; Represents the Kullback-Leibler divergence function; This indicates a normalization operation; and These represent the change prediction logic values ​​output by the student model and the teacher model on the old task decoding head, respectively.

[0078] To improve the fidelity of the boundary of change, The generation process incorporates low-level detail difference features from the shallow output of the residual backbone network through a decoder. This enables the joint representation of high-level semantic differences and low-level spatial details.

[0079] In some implementations, the dual-track distillation mechanism enables the student model to consistently output predictions consistent with the teacher model, regardless of whether it is facing a new domain distribution or a distribution with historical domain style characteristics. This greatly enhances the robustness of the change detection network against forgetting in complex semantic scenarios.

[0080] Change detection network in the task t The total loss function includes cross-entropy loss, double knowledge distillation loss, frequency domain phase alignment loss, and adapter sparsity constraint loss:

[0081] S5: Input the new domain dual-temporal remote sensing image data to be detected and its domain labels into the trained student model, dynamically call the corresponding decoding head in the multi-head task decoder according to the domain labels, and output the ground feature change detection results.

[0082] In another embodiment of the present invention, a remote sensing image domain incremental change detection system based on frequency domain distillation and difference-guided decoupling is proposed to implement the remote sensing image domain incremental change detection method based on frequency domain distillation and difference-guided decoupling as described above, comprising: The domain-adaptive data preprocessing module is used to acquire a dataset containing dual-temporal remote sensing images with multiple land cover change types and pixel-level change labels. After being cropped at a uniform scale, the dataset is divided into an old domain dataset and a new domain dataset according to the domain attributes, and each sample is assigned a corresponding domain label. The module for constructing the domain incremental change detection network is used to build a change detection network for domain incremental learning. It includes a teacher model and a student model, both of which have the same structure and include a residual backbone network for extracting multi-level bi-temporal features. The module for differential guidance and decoupling is used to generate channel attention weights based on the absolute difference map of the corresponding level bi-temporal features, and injects the corresponding level bi-temporal features into the residual backbone network in the form of residuals after recalibration. The module for hollow spatial pyramid pooling is connected to the deepest layer of the residual backbone network. The module for multi-head task decoder is set at the end of the change detection network and consists of multiple independent decoder heads. The module for freezing the old domain teacher model and extracting style priors is used to freeze the parameters of the teacher model after training it with the old domain dataset. The old domain dataset is input into the teacher model to extract the dual-temporal deep features of the teacher model. The amplitude spectrum is obtained by fast Fourier transform, and the average is calculated along the sample dimension. The average amplitude spectrum is used as the old domain global amplitude spectrum prior and stored. The cross-domain synthetic feature distillation training module is used to simultaneously input the new domain dataset into the teacher model with frozen parameters and the student model to be trained; extract the bi-temporal deep features of the student model and record them as new domain features, retain the phase spectrum of the new domain features, replace its amplitude spectrum with the prior global amplitude spectrum of the old domain, and recombine the replaced amplitude spectrum and the retained phase spectrum to form a complex spectrum, which is then reconstructed into cross-domain synthetic features through inverse transformation; construct a joint loss function, which includes cross-entropy loss, dual knowledge distillation loss, frequency domain phase alignment loss, and adapter sparsity constraint loss; iteratively update the difference-guided decoupling adapter parameters and the current new task decoder parameters in the student model through backpropagation until the model converges, while the remaining parameters remain frozen; The domain-aware decoding head dynamic inference module is used to input the new domain dual-temporal remote sensing image data to be detected and its domain labels into the trained student model, dynamically call the corresponding decoding head in the multi-head task decoder according to the domain labels, and output the ground feature change detection results.

[0083] In another embodiment of the present invention, an electronic device is proposed, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a remote sensing image domain incremental change detection method based on frequency domain distillation and difference-guided decoupling as described above.

[0084] In another embodiment of the present invention, a computer-readable storage medium is provided storing a computer program that causes a computer to execute a remote sensing image domain incremental change detection method based on frequency domain distillation and difference-guided decoupling as described above.

[0085] In the embodiments disclosed in this application, a computer storage medium may be a tangible medium that may contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of computer storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0086] The superiority of the method of this invention is verified below through specific experiments. This experiment uses the cross-domain remote sensing change detection datasets LEVIR-CD and CDD as experimental data to simulate the incremental learning process of LEVIR... The CDD domain offset scene. After fixed-scale cropping, a standard image patch of size 256×256 is obtained. This experiment was conducted in the PyTorch deep learning framework. During the incremental training phase, the model was trained for a total of 150 epochs with a batch size of 8. Under the differential learning rate strategy, the initial learning rate of the student model's task-specific layers was set to 0.0005, and the parameters of the shared backbone layers were kept frozen. The LambdaLR learning rate decay strategy was adopted. The dynamic warm-up period was set to 20 epochs. During the training process, the F1 score on the validation set and the intersection-over-union (IoU) of change classes were used as evaluation metrics to save the best change detection network.

[0087] To verify the superiority of the proposed method in the domain incremental change detection task, it is compared with current mainstream baseline methods and continuous learning networks. The models compared include: Single-task training cap, Transformer-based change detection network (BIT), spatiotemporally robust change detection network (STRobustNet), feature extraction freezing strategy (FE), full parameter fine-tuning (FT), multi-domain incremental learning network (MDIL), multi-domain incremental change detection network (MDINet), change detection network based on feature replay and optimized channel consistency discarding (FOCUS), and domain incremental Faster-RCNN network (DI-FRCNN). The accuracy evaluation metrics for the detection results are the F1 score, IoU, and the average performance degradation rate used to quantify anti-forgetting ability in each domain. F1 and The specific quantitative test results are shown in Table 1, and the visual qualitative comparison results are as follows: Figure 4 As shown.

[0088] Table 1. Performance Comparison of Different Incremental Change Detection Methods in Incremental Experiments

[0089] As shown in Table 1, after completing the incremental learning sequence from the source domain LEVIR to the target domain CDD, traditional non-incremental learning methods exhibit severe catastrophic forgetting. The average performance degradation rate of BIT is... The F1 score is as high as -44.62%, while the F1 score of the FT in the old domain LEVIR plummeted to 12.55%. The F1 decay rate reached -43.48%, indicating that these methods completely lost their ability to extract features from the old domain without historical data playback. While methods designed for continuous learning, such as MDIL, FOCUS, and MDINet, have made some progress in mitigating forgetting, performance degradation still exists. For example, the representative MDINet method, although performing excellently on the new domain CDD, experienced a performance degradation of -4.01% on the old domain LEVIR. In contrast, the method of this invention exhibits superior stability and adaptability. After incremental learning, the F1 score of this method on the old domain LEVIR still reaches 90.40%, and the IoU remains at 82.48%, with only minor fluctuations compared to the single-task upper limit. Specifically, the average performance degradation rate of this method is... F1 is only -0.58%, With an IoU of only -1.12%, it performs best among all the comparison algorithms. The above experimental data fully verify that the difference-guided decoupling adapter and frequency domain constraint strategy proposed in this invention can accurately decouple the domain style and physical structure, enabling the change detection network to maintain the knowledge of the old domain structure while still possessing a high degree of plasticity to the style features of the new domain, thus achieving the best balance between stability and plasticity.

[0090] like Figure 4 As shown, the superiority of the method of the present invention is further verified. Figure 4 This paper demonstrates the prediction details of each comparative model in the target domain CDD (first two rows) and the old domain LEVIR (last two rows) after completing the incremental learning sequence. Conventional methods show significant shortcomings when faced with drastic background interference caused by seasonal changes in the CDD dataset (such as the snow-covered scene in the second row) and complex terrain textures. Specifically, BIT and MDIL produced some false red pixels and failed to distinguish unchanging areas covered by snow; FE, FOCUS, and DI-FRCNN exhibited severe cyan false negatives, especially in the extraction of the thin roads in the second row, where severe boundary breaks and structural defects occurred. In contrast, the method of this invention benefits from DGDA's keen perception of deep feature differences between two temporal phases, accurately decoupling seasonal style noise from real road changes, and generating prediction results consistent with the ground truth labels.

[0091] When evaluating the old domain LEVIR dataset to verify its anti-forgetting ability, BIT and STRobustNet almost completely lost their ability to recognize changed targets in the fourth row scene. While FT, MDIL, MDINet, and DI-FRCNN could maintain the discrimination stability of the old task to some extent, large areas of missed cyan detections and some false detections appeared inside large buildings in the third row and in the shadowy areas of the fourth row. This indicates that these methods destroyed the original structural representation space of the old domain during incremental adaptation. The method of this invention exhibited excellent robustness in all test scenarios: its missed and false detection pixels were minimal, demonstrating not only excellent plasticity in the new domain but also restoring the edge contours and internal details of buildings in the old domain. These results fully demonstrate the effectiveness of the proposed remote sensing image domain incremental change detection method based on frequency domain distillation and difference-guided decoupling. By fixing the spatial structural features of ground objects, it ensures that the change detection network can still stably maintain its discrimination ability for historical tasks when facing drastic domain drift, achieving high-precision domain incremental continuous ground object change detection.

[0092] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent substitutions, and improvements made by those skilled in the art to the above embodiments without departing from the scope of the technical solution of the present invention, based on the technical essence of the present invention, shall still fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for detecting incremental changes in the domain of remote sensing images based on frequency domain distillation and difference-guided decoupling, characterized in that, Includes the following steps: S1: Obtain a dataset of dual-temporal remote sensing images containing multiple types of land cover changes and pixel-level change labels. After cropping at a uniform scale, divide the dataset into an old domain dataset and a new domain dataset according to the domain attributes, and assign a corresponding domain label to each sample. S2: Construct a change detection network for domain incremental learning, which includes a teacher model and a student model. Both have the same structure and include a residual backbone network for extracting multi-level bi-temporal features. A difference-guided decoupling adapter is used to generate channel attention weights based on the absolute difference map of the corresponding level bi-temporal features, and injects the corresponding level bi-temporal features into the residual backbone network in the form of residuals after recalibrating them. Connected to the deepest void space pyramid pooling module of the residual backbone network; The multi-head task decoder, located at the end of the change detection network, consists of multiple independent decoding heads; S3: After training the teacher model using the old domain dataset, freeze its parameters; input the old domain dataset into the teacher model, extract the dual-temporal deep features of the teacher model, obtain the amplitude spectrum through fast Fourier transform, calculate the average along the sample dimension, calculate the average amplitude spectrum as the prior of the old domain global amplitude spectrum and store it. S4: Input the new domain dataset simultaneously into the teacher model with frozen parameters and the student model to be trained; extract the bi-temporal deep features of the student model and denote them as new domain features, retain the phase spectrum of the new domain features, replace its amplitude spectrum with the prior global amplitude spectrum of the old domain, recombine the replaced amplitude spectrum and the retained phase spectrum to form a complex spectrum, and reconstruct it into cross-domain synthetic features through inverse transformation; construct a joint loss function, which includes cross-entropy loss, double knowledge distillation loss, frequency domain phase alignment loss and adapter sparsity constraint loss; iteratively update the difference-guided decoupling adapter parameters and the current new task decoder parameters in the student model through backpropagation until the model converges, and keep the remaining parameters frozen; S5: Input the dual-temporal remote sensing image data to be detected and its domain labels into the trained student model, dynamically call the corresponding decoding head in the multi-head task decoder according to the domain labels, and output the ground feature change detection results.

2. The remote sensing image domain incremental change detection method based on frequency domain distillation and difference-guided decoupling according to claim 1, characterized in that: In step S2, the difference-guided decoupling adapter is embedded in parallel after each shallow and mid-layer residual block of the residual backbone network in a bypass manner to participate in the parameter update of the student model. Specifically, it is used to: calculate the absolute difference map of the corresponding layer's dual-temporal features and use it as prior information indicating the intensity of change; extract the global context descriptor of the current single-temporal features and the absolute difference map using global average pooling, and concatenate the two in the channel dimension to obtain a joint feature vector that integrates semantic content and change prior; input the joint feature vector into the first fully connected layer, the ReLU activation function, the second fully connected layer, and the Sigmoid activation function in sequence to generate channel attention weights; use the channel attention weights to recalibrate the corresponding layer's dual-temporal features by multiplying them channel by channel, and inject the recalibrated features into the residual backbone network through residual connections after convolution.

3. The remote sensing image domain incremental change detection method based on frequency domain distillation and difference-guided decoupling according to claim 1, characterized in that: In step S2, the hollow spatial pyramid pooling module is connected to the deepest layer of the residual backbone network, receives the corresponding layer dual-temporal features after correction by the differential-guided decoupling adapter, and performs multi-scale sampling by setting parallel hollow convolutions with different dilation rates to enhance the context awareness capability of the change detection network for regions of different sizes of change.

4. The remote sensing image domain incremental change detection method based on frequency domain distillation and difference-guided decoupling according to claim 1, characterized in that: In step S2, the multi-head task decoder is specifically used to: splice and fuse the high-level features output by the hollow spatial pyramid pooling module with the low-level features output by the shallow layer of the residual backbone network to gradually restore the spatial resolution; and dynamically activate the corresponding decoding head to generate a change detection probability map according to the given domain label.

5. The remote sensing image domain incremental change detection method based on frequency domain distillation and difference-guided decoupling according to claim 1, characterized in that: In step S3, the residual backbone network of the teacher model is used to extract the dual-temporal deep features of the old domain dual-temporal remote sensing image. The extracted dual-temporal deep features are applied to each channel with a two-dimensional real number fast Fourier transform to map the spatial domain features to the frequency domain and obtain the corresponding complex spectral representation. The modulus of the complex spectral representation is calculated to obtain the amplitude spectrum. All extracted amplitude spectra are averaged along the dimension of the sample set to generate and store the old domain global amplitude spectrum prior representing the global statistical law of the old domain.

6. The remote sensing image domain incremental change detection method based on frequency domain distillation and difference-guided decoupling according to claim 1, characterized in that: In step S4, the joint loss function includes cross-entropy loss, dual-knowledge distillation loss, frequency domain phase alignment loss, and adapter sparsity constraint loss, specifically: Cross-entropy loss The expression is: Double knowledge distillation loss The expression is: in, Indicates pixel-level change labels. This represents the predicted probability map of the student model in the current new domain task's decoding head output; This represents the new domain difference features extracted by the student model; This represents the cross-domain synthetic difference features obtained from frequency domain decoupling and reconstruction; This represents the knowledge distillation loss calculated on the features of the new domain; This represents the distillation loss calculated on the cross-domain synthesis feature; Represents the Kullback-Leibler divergence function; This indicates a normalization operation; and The student model and the teacher model output different predicted logical values ​​on the old task decoding head.

7. The remote sensing image domain incremental change detection method based on frequency domain distillation and difference-guided decoupling according to claim 6, characterized in that: In step S4, the frequency domain phase alignment loss The expression is: Adapter sparse constraint loss The expression is: in, and These represent the student model's response to dual-temporal remote sensing images. and Phase spectrum corresponding to the extracted high-level semantic features; and These represent the reference phase spectra extracted by the teacher model for the same image pair; This represents the mean square error calculation; The number of layers in the change detection network that embed a difference-guided decoupling adapter; For the first The number of channels in the layer; For the first Layer Attention weights for each channel.

8. A remote sensing image domain incremental change detection system based on frequency domain distillation and difference-guided decoupling, used to implement the remote sensing image domain incremental change detection method based on frequency domain distillation and difference-guided decoupling as described in any one of claims 1 to 7, characterized in that, include: The domain-adaptive data preprocessing module is used to acquire a dataset containing dual-temporal remote sensing images with multiple land cover change types and pixel-level change labels. After being cropped at a uniform scale, the dataset is divided into an old domain dataset and a new domain dataset according to the domain attributes, and each sample is assigned a corresponding domain label. The Domain Incremental Change Detection Network Construction Module is used to build a change detection network for domain incremental learning. It includes a teacher model and a student model, both of which have the same structure and include a residual backbone network for extracting multi-level bi-temporal features. The Difference-Guided Decoupling Adapter is used to generate channel attention weights based on the absolute difference map of the corresponding level bi-temporal features, and injects the corresponding level bi-temporal features into the residual backbone network in the form of residuals after recalibrating them. Connected to the deepest void space pyramid pooling module of the residual backbone network; The multi-head task decoder, located at the end of the change detection network, consists of multiple independent decoding heads; The module for freezing the old domain teacher model and extracting style priors is used to freeze the parameters of the teacher model after training it with the old domain dataset. The old domain dataset is input into the teacher model to extract the dual-temporal deep features of the teacher model. The amplitude spectrum is obtained by fast Fourier transform, and the average is calculated along the sample dimension. The average amplitude spectrum is used as the old domain global amplitude spectrum prior and stored. The cross-domain synthetic feature distillation training module is used to simultaneously input the new domain dataset into the teacher model with frozen parameters and the student model to be trained; extract the bi-temporal deep features of the student model and record them as new domain features, retain the phase spectrum of the new domain features, replace its amplitude spectrum with the prior global amplitude spectrum of the old domain, and recombine the replaced amplitude spectrum and the retained phase spectrum to form a complex spectrum, which is then reconstructed into cross-domain synthetic features through inverse transformation; construct a joint loss function, which includes cross-entropy loss, dual knowledge distillation loss, frequency domain phase alignment loss, and adapter sparsity constraint loss; iteratively update the difference-guided decoupling adapter parameters and the current new task decoder parameters in the student model through backpropagation until the model converges, while the remaining parameters remain frozen; The domain-aware decoding head dynamic inference module is used to input the new domain dual-temporal remote sensing image data to be detected and its domain labels into the trained student model, and dynamically call the corresponding decoding head in the multi-head task decoder according to the domain labels to output the ground feature change detection results.

9. An electronic device, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements a remote sensing image domain incremental change detection method based on frequency domain distillation and difference-guided decoupling as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: The computer program causes the computer to execute a remote sensing image domain incremental change detection method based on frequency domain distillation and difference-guided decoupling as described in any one of claims 1 to 7.

Images