Remote sensing image change detection method, system and device based on dual-domain learning

Abstract

The application discloses a remote sensing image change detection method, system and equipment based on dual-domain learning. The application introduces a frequency domain enhancement module and a spatial domain recovery module in the remote sensing image change detection task based on dual-domain learning. In the frequency domain enhancement module, frequency components are extracted from the input dual-time image through discrete cosine transformation, so that more accurate semantic recognition is realized to enhance the interest change. In the spatial domain recovery module, feature fusion is realized through a cross-time and cross-scale mechanism. This can guide the context modeling of low-level representations using higher-level change representations, thereby enhancing the semantic information in low-level features and improving the spatial detail recovery of change representations. The application provides a new idea for the remote sensing image change detection task from the perspective of the frequency domain and the spatial domain, and realizes the most advanced performance.

Description

Technical Field

[0001] This invention belongs to the fields of deep learning and computer vision, and particularly relates to a method for detecting changes in remote sensing images based on dual-domain learning. Background Technology

[0002] Remote sensing image change detection is a challenging and critical task that identifies changes of interest in the form of spectral variations by analyzing multi-temporal remote sensing images captured in the same geographic area. Remote sensing image change detection classifies image pixels into change maps to determine whether each part of the region represents a change of interest. This task reveals local development patterns from both natural and socio-economic perspectives and has significant implications for urban planning, environmental monitoring, and disaster assessment.

[0003] The rapid development of deep learning technology has significantly propelled progress in the field of change detection in remote sensing images, particularly with advancements in extracting hierarchical features using Convolutional Neural Networks (CNNs). For example, LGPNet employs an attention mechanism to enhance the expressive power of change representations. STANet, based on Siamese neural networks, uses a BAM module to leverage global spatiotemporal relationships for more expressive features and a PAM module to aggregate multi-scale attention representations to obtain fine-grained object details. DASNets employ a dual attention mechanism to capture long-range dependencies and obtain feature representations for change detection. SNUNet combines Siamese neural networks with the U-Net architecture, preserving high-level semantic information and detailed features through skip connections. ChangeFormer introduces a Transformer-based Siamese neural network, utilizing the Transformer model to learn spatial relationships in time series. While these methods have achieved significant performance improvements, they are still limited to addressing change detection in the spatial domain of remote sensing images, leaving room for further performance enhancement.

[0004] Recently, the frequency domain has been extensively studied in object detection and semantic segmentation. For example, merging frequency domain features into the original features after converting an image to a CNN has been shown to improve the performance of image analysis tasks. AFFormer learns local descriptive representations of clustering prototypes from a frequency perspective, eliminating the need for complex decoders. Simultaneously, it combines with the linearly complex Transformer to simplify semantic segmentation to traditional classification. These studies demonstrate that incorporating frequency information helps extract more representative features for semantic recognition.

[0005] Therefore, for remote sensing image change detection tasks with complex backgrounds, how to start from the frequency domain, introduce more frequency information, and then extract image feature information more deeply to design a change detection network with superior performance is a technical problem that urgently needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to solve the problems existing in the prior art and to provide a remote sensing image change detection method based on dual-domain learning. It aims to explore from the frequency domain, extract frequency components to enhance the changes of interest, and perform feature fusion from the temporal and spatial perspectives to better recover the spatial details of the change representation, thereby further improving the performance of current change detection.

[0007] To achieve the above-mentioned objectives, the present invention specifically adopts the following technical solution:

[0008] In a first aspect, the present invention provides a remote sensing image change detection method based on dual-domain learning, the specific steps of which are as follows:

[0009] Remote sensing images taken in the same geographical area at two different times are combined into a dual-time remote sensing image. The dual-time remote sensing image is then input into a trained change detection network to obtain the change detection results of the dual-time remote sensing image. The change detection network consists of an encoder module, a frequency domain enhancement module, a spatial domain recovery module, and a lightweight decoder module.

[0010] In the change detection network, the dual-time remote sensing image is first processed by the encoder module to extract features, resulting in dual-time feature representations at different scales. The encoder module consists of several residual blocks cascaded sequentially, with each residual block corresponding to a dual-time feature representation at one scale. Each scale's dual-time feature representation contains two single-time feature representations at the same scale. Then, each of the two single-time feature representations at the same scale is processed by the frequency domain enhancement module, which uses discrete cosine transform to extract frequency components from the input single-time feature representations, resulting in a frequency domain enhanced feature map corresponding to each scale's single-time feature representation. These frequency domain enhanced feature maps at the same scale are then combined to form a set of frequency domain enhancement maps. Enhanced feature maps: After initial feature fusion, the deepest set of frequency domain enhanced feature maps is used to obtain the deepest coarse change representation. Except for the set of frequency domain enhanced feature maps corresponding to the deepest coarse change representation, each set of frequency domain enhanced feature maps at other scales is processed by a spatial domain recovery module. At the same time, each spatial domain recovery module is additionally input with the deepest coarse change representation to obtain a change representation with the same scale as each set of input frequency domain enhanced feature maps. The number of spatial domain recovery modules is one less than the number of residual blocks. Finally, the change representation corresponding to the shallowest dual-time feature representation is input into the lightweight decoder module to obtain the change detection result of the dual-time remote sensing image.

[0011] Based on the above scheme, each step can be implemented in the following preferred manner.

[0012] As a preferred embodiment of the first aspect, the encoder module adopts a ResNet-18 backbone network, which is composed of a first residual block, a second residual block, a third residual block, and a fourth residual block cascaded in sequence. The first residual block outputs a dual-time feature representation at a first scale, the second residual block outputs a dual-time feature representation at a second scale, the third residual block outputs a dual-time feature representation at a third scale, and the fourth residual block outputs a dual-time feature representation at a fourth scale. The dual-time feature representation at the fourth scale is used as the deepest dual-time feature representation.

[0013] As a preferred embodiment of the first aspect, the specific processing flow of the frequency domain enhancement module is as follows: taking the single-time feature representation of each scale as input, firstly, the single-time feature representation of each scale is uniformly divided into a preset number of parts along the channel dimension; after pre-training the frequency domain enhancement module on the ImageNet dataset, a two-dimensional index corresponding to the two-dimensional discrete cosine transform frequency component is selected for each part, and the two-dimensional discrete cosine transform frequency component of each part is calculated according to the two-dimensional index corresponding to each part; after concatenating the two-dimensional discrete cosine transform frequency components of each part, a frequency vector is obtained; the frequency vector is passed through the second convolutional layer and the Sigmoid activation layer in sequence to obtain the frequency weights; the frequency weights are weighted and summed with the single-time feature representation of each scale to obtain the frequency domain enhanced feature map of each scale;

[0014] For the i-th part X i Two-dimensional discrete cosine transform frequency components Frep i Its functional form is:

[0015]

[0016] Among them, [u i ,v i ] indicates the i-th part X i The two-dimensional index corresponding to the frequency components of the two-dimensional discrete cosine transform, u i Indicates the first index, v i Indicates the second index; ω represents the discrete cosine transform; h∈[0,1,...,H-1] represents the height index, ω∈[0,1,...,W-1] represents the width index, and H and W represent the height and width of the single-time feature representation, respectively. Represents the i-th part X i A two-dimensional feature map with a height index of h and a width index of ω.

[0017] As a preferred embodiment of the first aspect, the specific processing flow of the initial feature fusion is as follows: After the dual-time feature representation of the fourth scale is processed by the frequency domain enhancement module, a set of frequency domain enhanced feature maps of the fourth scale is obtained. The set of frequency domain enhanced feature maps of the fourth scale is subjected to pixel-level subtraction to obtain the coarse change representation of the fourth scale. One of the frequency domain enhanced feature maps of the set of frequency domain enhanced feature maps of the fourth scale is concatenated with the coarse change representation of the fourth scale in the channel dimension to obtain the concatenated feature of the fourth scale. The concatenated feature of the fourth scale is integrated by depthwise separable convolution to obtain the deepest coarse change representation.

[0018] As a preferred option in the first aspect mentioned above, the specific processing flow for each spatial domain recovery module is as follows:

[0019] After the dual-time feature representation of the third scale is passed through the frequency domain enhancement module, a set of frequency domain enhanced feature maps of the third scale is obtained. After the dual-time feature representation of the second scale is passed through the frequency domain enhancement module, a set of frequency domain enhanced feature maps of the second scale is obtained. After the dual-time feature representation of the first scale is passed through the frequency domain enhancement module, a set of frequency domain enhanced feature maps of the first scale is obtained.

[0020] In addition to a set of frequency domain enhanced feature maps at the fourth scale, a set of frequency domain enhanced feature maps at the third scale are input into the third spatial domain recovery module. Initial feature fusion is first performed on the set of frequency domain enhanced feature maps at the third scale to obtain a coarse change representation at the third scale. The coarse change representation at the third scale is then subjected to average pooling to obtain the first feature at the third scale. The coarse change representation at the third scale is then subjected to max pooling to obtain the second feature at the third scale. The first feature at the third scale and the second feature at the third scale are concatenated to obtain the third feature at the third scale. The third feature at the third scale is then passed sequentially through the first convolutional layer and the Sigmoid activation layer to obtain the weight map corresponding to the third scale. The weight map corresponding to the third scale is downsampled. The downsampled weight map at the third scale is then weighted and summed with the deepest coarse change representation to obtain the fourth feature at the third scale. The fourth feature at the third scale and the deepest coarse change representation at the third scale are then concatenated along the channel dimension to obtain the change representation at the third scale.

[0021] A set of frequency domain enhanced feature maps at the second scale is input into the second spatial domain recovery module. Initial feature fusion is performed on the set of frequency domain enhanced feature maps at the second scale to obtain a coarse change representation at the second scale. The coarse change representation at the second scale is then subjected to average pooling to obtain the first feature at the second scale. The coarse change representation at the second scale is then subjected to max pooling to obtain the second feature at the second scale. The first feature at the second scale and the second feature at the second scale are concatenated to obtain the third feature at the second scale. The second feature at the second scale is then passed sequentially through the first convolutional layer and the Sigmoid activation layer to obtain the weight map corresponding to the second scale. The weight map corresponding to the second scale is downsampled. The downsampled weight map at the second scale is then weighted and summed with the deepest coarse change representation to obtain the fourth feature at the second scale. The fourth feature at the second scale and the change representation at the third scale are then concatenated along the channel dimension to obtain the change representation at the second scale.

[0022] A set of frequency domain enhanced feature maps at the first scale is input into the first spatial domain recovery module. Initial feature fusion is performed on the set of frequency domain enhanced feature maps at the first scale to obtain a coarse change representation at the first scale. The coarse change representation at the first scale is then subjected to average pooling to obtain the first feature at the first scale. The coarse change representation at the first scale is then subjected to max pooling to obtain the second feature at the first scale. The first feature at the first scale and the second feature at the first scale are concatenated to obtain the third feature at the first scale. The third feature at the first scale is then passed sequentially through the first convolutional layer and the Sigmoid activation layer to obtain the weight map corresponding to the first scale. The weight map corresponding to the first scale is downsampled. The downsampled first scale weight map is then weighted and summed with the deepest coarse change representation to obtain the fourth feature at the first scale. The fourth feature at the first scale and the change representation at the second scale are concatenated along the channel dimension to obtain the change representation at the first scale.

[0023] As a preferred embodiment of the first aspect, the specific processing flow of the lightweight decoder module is as follows: taking the change representation of the first scale corresponding to the shallowest dual-temporal feature representation as input, performing depthwise separable convolution on the change representation of the first scale to obtain a change detection feature map, upsampling the change detection feature map to obtain the change detection result of the dual-temporal remote sensing image.

[0024] As a preferred embodiment of the first aspect above, the first convolutional layer and the second convolutional layer have the same kernel size, which is 1×1; the depth-separable convolutional kernel size is 3×3.

[0025] As a preferred embodiment of the first aspect above, the change detection network needs to be trained in advance using a labeled training dataset before being used for actual change detection tasks, and all training data in the training dataset needs to be randomly augmented.

[0026] The loss function used in training the change detection network is a hybrid loss function, which is obtained by weighted summation of weighted cross-entropy loss function and dice loss function. The parameters of the change detection network are updated based on minimizing the hybrid loss function.

[0027] Secondly, the present invention provides a remote sensing image change detection system based on dual-domain learning, comprising:

[0028] The data acquisition module is used to acquire remote sensing images taken at two different times in the same geographical area and combine them into a dual-time remote sensing image.

[0029] The result acquisition module is used to input the dual-time remote sensing image output by the data acquisition module into the trained change detection network to obtain the change detection result of the dual-time remote sensing image. The change detection network consists of an encoder module, a frequency domain enhancement module, a spatial domain recovery module, and a lightweight decoder module.

[0030] Thirdly, the present invention provides a computer electronic device, which includes a memory and a processor;

[0031] The memory is used to store computer programs;

[0032] The processor is configured to, when executing the computer program, implement a remote sensing image change detection method based on dual-domain learning as described in any of the first aspects above.

[0033] Compared with the prior art, the present invention has the following advantages:

[0034] This invention discloses a remote sensing image change detection method based on dual-domain learning. Addressing a common problem in remote sensing image change detection—that current methods primarily focus on spatial domain exploration and optimization while neglecting the importance of frequency domain information—this invention introduces a frequency domain enhancement module and a spatial domain restoration module into the remote sensing image change detection task based on dual-domain learning. Specifically, in the frequency domain enhancement module, frequency components are extracted from the input dual-temporal images through discrete cosine transform, thereby enhancing interest changes through more accurate semantic recognition. In the spatial domain restoration module, feature fusion is achieved through cross-time and cross-scale mechanisms. This allows higher-level change representations to guide contextual modeling of lower-level representations, thereby enhancing the semantic information in lower-level features and improving the recovery of spatial details in the change representations. This invention provides a new approach to remote sensing image change detection from both frequency and spatial domain perspectives, effectively improving the performance of change detection. Attached Figure Description

[0035] Figure 1This is a model structure diagram of the change detection network DDLNet of this invention;

[0036] Figure 2 This is a flowchart illustrating the training and testing process of the change detection network DDLNet in this embodiment of the invention.

[0037] Figure 3 This is a schematic diagram of the test visualization results in an embodiment of the present invention;

[0038] Figure 4 This is a system block diagram of the present invention. Detailed Implementation

[0039] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in the various embodiments of the present invention can be combined accordingly without mutual conflict.

[0040] In the description of this invention, it should be understood that the terms "first" and "second" are used only for descriptive purposes and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first" and "second" may explicitly or implicitly include at least one of those features.

[0041] Currently, remote sensing image change detection primarily focuses on spatial domain exploration and optimization, neglecting the importance of frequency domain information, which often contains more spatial feature details. To address this issue, such as... Figure 1 As shown, in a preferred embodiment of the present invention, a remote sensing image change detection method based on dual-domain learning is proposed. Specifically, remote sensing images taken in the same geographical area at two different time points are combined into a dual-time remote sensing image. The dual-time remote sensing image is then input into a trained change detection network DDLNet to obtain the change detection result of the dual-time remote sensing image. The change detection network consists of an encoder module, a frequency-domain enhancement module (FEM), a spatial-domain recovery module (SRM), and a lightweight decoder module.

[0042] In the aforementioned change detection network, the dual-time remote sensing image is first processed by the encoder module to extract features, resulting in dual-time feature representations at different scales. The encoder module consists of several residual blocks cascaded sequentially, with each residual block corresponding to a dual-time feature representation at one scale. Each scale's dual-time feature representation contains two single-time feature representations at the same scale. Then, each of the two single-time feature representations at the same scale is processed by the frequency domain enhancement module, which uses discrete cosine transform to extract frequency components from the input single-time feature representations, resulting in frequency domain enhanced feature maps corresponding to each scale's single-time feature representation. These frequency domain enhanced feature maps at the same scale are then combined to form a set of frequency domain enhancement maps. Enhanced feature maps: After initial feature fusion, the deepest set of frequency domain enhanced feature maps is used to obtain the deepest coarse change representation. Except for the set of frequency domain enhanced feature maps corresponding to the deepest coarse change representation, each set of frequency domain enhanced feature maps at other scales is processed by a spatial domain recovery module. At the same time, each spatial domain recovery module is additionally input with the deepest coarse change representation to obtain a change representation with the same scale as each set of input frequency domain enhanced feature maps. The number of the above spatial domain recovery modules is one less than the number of the above residual blocks. Finally, the change representation corresponding to the shallowest dual-temporal feature representation is input into the above lightweight decoder module to obtain the change detection result of the dual-temporal remote sensing image.

[0043] The change detection results of the above dual-time remote sensing images are as follows: Figure 1 ( Figure 1 The result is shown in the figure below. Figure 1 The result image shows the change detection results between the two time-series remote sensing images. Black indicates that there is no change between the two time-series remote sensing images, while white indicates that there is a change.

[0044] The following describes the specific data processing procedures performed in the change detection network DDLNet in this embodiment of the invention. Figure 1 This is a diagram of the overall structure of the DDLNet model, including an encoder module, a frequency domain enhancement module, a spatial domain recovery module, and a lightweight decoder module. The encoder, frequency domain enhancement, and spatial domain recovery modules are used to obtain more expressive change representations, while the lightweight decoder module is used to recover spatial details and obtain the final change detection results. The processing flow of each module is briefly described below.

[0045] In this invention, feature extraction is first performed using an encoder module composed of residual blocks, extracting feature representations from four different scales (from shallow to deep) from the dual-time remote sensing images. Then, each scale feature representation is passed through an independent frequency domain enhancement module, which extracts frequency components from the input feature representations using discrete cosine transform, thereby achieving more accurate semantic recognition and emphasizing changes of interest. The frequency-enhanced feature map at the deepest scale, after initial feature fusion, is processed by a spatial domain restoration module. This module achieves feature fusion through a cross-time and cross-scale mechanism, using higher-level change representations to guide contextual modeling of lower-level representations, thereby enhancing the semantic information in the lower-level features and improving the spatial detail recovery of the change representations. This results in a spatially detailed reconstructed change representation, which is then used as the output of the spatial domain restoration module. Finally, the change representation output from the first spatial domain restoration module is input into a lightweight decoder module, where it undergoes depthwise separable convolution and upsampling to obtain the final dual-time remote sensing image change detection result. This yields the final change detection result.

[0046] In this invention, the encoder module employs a ResNet-18 backbone network, comprising four residual blocks cascaded sequentially: a first residual block, a second residual block, a third residual block, and a fourth residual block. The first residual block outputs a bitemporal feature representation at a first scale, the second residual block outputs a bitemporal feature representation at a second scale, the third residual block outputs a bitemporal feature representation at a third scale, and the fourth residual block outputs a bitemporal feature representation at a fourth scale. This fourth-scale bitemporal feature representation is considered the deepest bitemporal feature representation.

[0047] In each residual block, the input feature representation is sequentially passed through a third convolutional layer, a batch normalization layer, a ReLU activation function, another third convolutional layer, and another batch normalization layer to obtain an intermediate feature representation. This intermediate feature representation is then added to the input feature representation to obtain the output feature representation. The kernel size of the third convolutional layer is 3×3.

[0048] It should be noted that, due to the multi-level feature extraction layers of the encoder module, the feature representations extracted at each time point are also multi-scale. The feature representation finally output by the encoder module is the deepest feature representation, while the feature representations obtained from the remaining intermediate feature extraction layers are shallow feature representations. In the embodiments of this invention, the encoder module uses a ResNet-18 backbone network, which contains four residual blocks. For the remote sensing images at each time point, the first three residual blocks output three shallow feature representations at different scales, and the fourth residual block outputs the deepest feature representation. Therefore, for a dual-time remote sensing image composed of two remote sensing images, after processing by the ResNet-18 model backbone network, four different scales of feature representations from shallow to deep are extracted from the dual-time remote sensing image. At each of these four scales, two temporal feature representations corresponding to the two remote sensing images are formed, which is the dual-time feature representation at that scale as referred to in this invention; the temporal feature representation of each remote sensing image is a single-time feature representation.

[0049] Specifically, for the encoder module, the input consists of two remote sensing images taken at two different times within the same geographical area, each with dimensions (B×C×H×W). Here, B is the batch size of the input, and C is the number of feature channels in the input remote sensing image. During the training phase, the batch size is set according to the sample size of each batch, such as 4, 8, or 16. In this invention, C is 3, and H and W are the height and width of the input remote sensing image, respectively. First, as mentioned above, the backbone network is a ResNet-18 model. After the dual-time images undergo preliminary feature extraction using the ResNet-18 model, each image obtains three shallow feature representations and one deepest feature representation, i.e., four residual blocks output four sets of dual-time feature representations at different scales. Specifically, the first three residual blocks of the ResNet-18 model output three shallow feature representations with dimensions of (B×C×H×W). 64, 128, and 256 represent the number of feature channels in the previous layer feature representation; the deepest layer feature representation is the output of the fourth residual block of the ResNet-18 model, with a dimension of [missing value]. 512 represents the number of channels in the deepest feature representation.

[0050] In this invention, the specific processing flow of the frequency domain enhancement module is as follows:

[0051] Using the single-temporal feature representation at each scale as input, the single-temporal feature representation at each scale is first uniformly divided into a predetermined number of parts along the channel dimension. After pre-training the frequency domain enhancement module on the ImageNet dataset, a two-dimensional index corresponding to the two-dimensional discrete cosine transform frequency component is selected for each part. The two-dimensional discrete cosine transform frequency component of each part is calculated based on the two-dimensional index corresponding to each part. The two-dimensional discrete cosine transform frequency components of each part are concatenated to obtain a frequency vector. The frequency vector is then passed through a second convolutional layer and a sigmoid activation layer to obtain frequency weights. The frequency weights are then weighted and summed with the single-temporal feature representation at each scale to obtain the frequency domain enhanced feature map at each scale.

[0052] For the i-th part X i Two-dimensional discrete cosine transform frequency components Frep i Its functional form is:

[0053]

[0054] Among them, [u i ,v i ] indicates the i-th part X i The two-dimensional index corresponding to the frequency components of the two-dimensional discrete cosine transform, u i Indicates the first index, v i Indicates the second index; ω represents the discrete cosine transform; h∈[0,1,...,H-1] represents the height index, ω∈[0,1,...,W-1] represents the width index, and H and W represent the height and width of the single-time feature representation, respectively. Represents the i-th part X i A two-dimensional feature map with a height index of h and a width index of ω.

[0055] It should be noted that the purpose of the Frequency Domain Enhancement (FEM) module is to extract frequency components from the input dual-time remote sensing image through Discrete Cosine Transform (DCT), thereby achieving more accurate semantic recognition and enhancing the changes of interest. The FEM module needs to be used multiple times; each dual-time feature representation at a different scale in the encoder module needs to be input into the FEM module. Therefore, the frequency-enhanced feature maps output by the FEM module also have scale differences. When the input to this module is the fourth-scale dual-time feature representation output from the fourth residual block, the FEM module will output the deepest set of frequency-enhanced feature maps, i.e., the set of frequency-enhanced feature maps at the fourth scale. For other residual blocks, the bi-temporal feature representation of the third scale is passed through the frequency domain enhancement module to obtain a set of frequency domain enhanced feature maps of the third scale. The bi-temporal feature representation of the second scale is passed through the frequency domain enhancement module to obtain a set of frequency domain enhanced feature maps of the second scale. The bi-temporal feature representation of the first scale is passed through the frequency domain enhancement module to obtain a set of frequency domain enhanced feature maps of the first scale. The set of frequency domain enhanced feature maps of the third scale, the second scale, and the first scale all belong to shallow feature representations.

[0056] In this embodiment, for each frequency domain enhancement module (FEM), the single-time feature representation at each scale is used as input. First, the single-time feature representation X at each scale is uniformly divided into n parts [X0, X1, ..., X] along the channel dimension. n-1 ], where the i-th part X i ∈R C′×H×W And i∈[0,1,...,n-1] represents the index of each part. (Assuming C is divisible by n) represents the number of channels after uniform segmentation of the channel dimension, where n is a preset number; in this embodiment, n is 16. After pre-training the frequency domain enhancement module on the ImageNet dataset, the importance of different frequency components to classification accuracy is determined to study the impact of using different numbers of frequency components. For the i-th part X... i The selected two-dimensional index [u] corresponds to the frequency component of the two-dimensional discrete cosine transform. i ,v i Then, based on the two-dimensional index corresponding to each part, the two-dimensional discrete cosine transform frequency components of each part are calculated.

[0057] In this embodiment, for the i-th part X i Two-dimensional discrete cosine transform frequency components Frep i Its functional form is:

[0058]

[0059] Among them, Frep i ∈Rc′ It is a C'-dimensional vector, [u i ,v i ] indicates the i-th part X i The two-dimensional index corresponding to the frequency components of the two-dimensional discrete cosine transform, u i Indicates the first index, v i Indicates the second index; ω represents the discrete cosine transform; h∈[0,1,...,H-1] represents the height index, ω∈[0,1,...,W-1] represents the width index, and H and W represent the height and width of the single-time feature representation, respectively. Represents the i-th part X i A two-dimensional feature map with a height index of h and a width index of ω.

[0060] Get the i-th part X i Two-dimensional discrete cosine transform frequency components Frep i Then, the frequency components of the two-dimensional discrete cosine transform of each part are concatenated to obtain the frequency vector Frep:

[0061] Frep = concat([Frep 0 ,Frep 1 ,...,Frep n-1 ])

[0062] Here, concat represents the concatenation operation.

[0063] After obtaining the frequency vector Frep, it is then passed sequentially through a second convolutional layer and a sigmoid activation layer to obtain the frequency weights W. Frep :

[0064] W Frep =σ(fc(Frep))

[0065] Where fc represents the second convolutional layer of 1×1, and σ is the sigmoid activation layer.

[0066] Finally, the frequency weights are weighted and summed with the single-time feature representations at each scale to obtain the frequency domain enhanced feature map for each scale. Specifically, for the i-th part X at index i... i Also select the frequency weight value with index i from the frequency weights. The frequency weight value at index i And the i-th part X at index i iWeighted summation is performed until the weight values ​​for each part and each frequency have been traversed, resulting in a frequency domain enhanced feature map corresponding to the input single-time feature representation. Specifically, after the dual-time feature representations of the fourth, third, second, and first scales are processed by the frequency domain enhancement module, a set of frequency domain enhanced feature maps corresponding to the fourth, third, second, and first scales are obtained in sequence.

[0067] In this invention, the specific processing flow of the initial feature fusion is as follows: After the dual-time feature representation of the fourth scale is processed by the frequency domain enhancement module, a set of frequency domain enhanced feature maps of the fourth scale is obtained. The set of frequency domain enhanced feature maps of the fourth scale is subjected to pixel-level subtraction to obtain the coarse change representation of the fourth scale. One of the frequency domain enhanced feature maps of the set of frequency domain enhanced feature maps of the fourth scale is concatenated with the coarse change representation of the fourth scale in the channel dimension to obtain the concatenated feature of the fourth scale. The concatenated feature of the fourth scale is integrated by depthwise separable convolution to obtain the deepest coarse change representation.

[0068] It should be noted that, in this embodiment, the initial feature fusion process is as follows: for the deepest feature representation extracted by the residual block, it is processed by the frequency domain enhancement module to obtain a set of frequency domain enhanced feature maps. and By subtracting these two frequency domain enhanced feature maps, a coarse representation of the corresponding fourth-scale change, containing information about their differences, can be obtained. Then, this coarse transformation representation is concatenated with either of the two frequency domain enhanced feature maps along the channel dimension, i.e. and or The data is then concatenated, and the concatenated result is processed by depthwise separable convolution. Figure 1 Feature integration is performed on DS_Conv to obtain the deepest coarse variation representation (e.g., Figure 1 middle ).

[0069] In this invention, the spatial domain recovery module aims to achieve feature fusion through cross-time and cross-scale mechanisms, utilizing higher-level change representations to guide contextual modeling of lower-level representations, thereby enhancing the semantic information in lower-level features and better recovering the spatial details of change representations. This invention comprises three spatial domain recovery modules corresponding to the aforementioned residual blocks. The specific processing flow of each spatial domain recovery module is as follows:

[0070] After the dual-time feature representation of the third scale is processed by the frequency domain enhancement module, a set of frequency domain enhanced feature maps of the third scale is obtained. After the dual-time feature representation of the second scale is processed by the frequency domain enhancement module, a set of frequency domain enhanced feature maps of the second scale is obtained. After the dual-time feature representation of the first scale is processed by the frequency domain enhancement module, a set of frequency domain enhanced feature maps of the first scale is obtained.

[0071] In addition to a set of frequency domain enhanced feature maps at the fourth scale, a set of frequency domain enhanced feature maps at the third scale are input into the third spatial domain recovery module. Initial feature fusion is first performed on the set of frequency domain enhanced feature maps at the third scale to obtain a coarse change representation at the third scale. The coarse change representation at the third scale is then subjected to average pooling to obtain the first feature at the third scale. The coarse change representation at the third scale is then subjected to max pooling to obtain the second feature at the third scale. The first feature at the third scale and the second feature at the third scale are concatenated to obtain the third feature at the third scale. The third feature at the third scale is then passed sequentially through the first convolutional layer and the Sigmoid activation layer to obtain the weight map corresponding to the third scale. The weight map corresponding to the third scale is downsampled. The downsampled weight map at the third scale is then weighted and summed with the deepest coarse change representation to obtain the fourth feature at the third scale. The fourth feature at the third scale and the deepest coarse change representation at the third scale are then concatenated along the channel dimension to obtain the change representation at the third scale.

[0072] A set of frequency domain enhanced feature maps at the second scale is input into the second spatial domain recovery module. Initial feature fusion is performed on the set of frequency domain enhanced feature maps at the second scale to obtain a coarse change representation at the second scale. The coarse change representation at the second scale is then subjected to average pooling to obtain the first feature at the second scale. The coarse change representation at the second scale is then subjected to max pooling to obtain the second feature at the second scale. The first feature at the second scale and the second feature at the second scale are concatenated to obtain the third feature at the second scale. The second feature at the second scale is then passed sequentially through the first convolutional layer and the Sigmoid activation layer to obtain the weight map corresponding to the second scale. The weight map corresponding to the second scale is downsampled. The downsampled weight map at the second scale is then weighted and summed with the deepest coarse change representation to obtain the fourth feature at the second scale. The fourth feature at the second scale and the change representation at the third scale are then concatenated along the channel dimension to obtain the change representation at the second scale.

[0073] A set of frequency domain enhanced feature maps at the first scale is input into the first spatial domain recovery module. Initial feature fusion is performed on the set of frequency domain enhanced feature maps at the first scale to obtain a coarse change representation at the first scale. The coarse change representation at the first scale is then subjected to average pooling to obtain the first feature at the first scale. The coarse change representation at the first scale is then subjected to max pooling to obtain the second feature at the first scale. The first feature at the first scale and the second feature at the first scale are concatenated to obtain the third feature at the first scale. The third feature at the first scale is then passed sequentially through the first convolutional layer and the Sigmoid activation layer to obtain the weight map corresponding to the first scale. The weight map corresponding to the first scale is downsampled. The downsampled first scale weight map is then weighted and summed with the deepest coarse change representation to obtain the fourth feature at the first scale. The fourth feature at the first scale and the change representation at the second scale are concatenated along the channel dimension to obtain the change representation at the first scale.

[0074] It should be noted that there are three spatial domain recovery modules (SRM) mentioned above, corresponding to three sets of shallow feature representations after frequency domain enhancement (a set of frequency domain enhanced feature maps at the third scale, a set of frequency domain enhanced feature maps at the second scale, and a set of frequency domain enhanced feature maps at the first scale). Each spatial domain recovery module takes each set of frequency domain enhanced feature maps at different scales and the deepest coarse change representation after initial feature fusion as input. Taking the set of frequency domain enhanced feature maps at the third scale as an example, it first performs initial feature fusion to obtain the coarse change representation at the third scale. Then, this coarse change representation is subjected to average pooling and max pooling operations respectively. The results of average pooling and max pooling are then concatenated. The concatenated representation is passed through a first convolutional layer and a sigmoid activation layer to obtain a weight map. The downsampled weight map is then weighted and summed with the deepest coarse change representation after initial feature fusion to obtain the fourth feature at the third scale. As can be seen, this module uses higher-level change representations to guide the context modeling of lower-level representations, enhances the semantic information in the low-level features, and thus obtains the change representation after spatial detail reconstruction (i.e., the change representation at the third scale) as the output of the spatial domain recovery module.

[0075] For a set of frequency domain enhanced feature maps at the second scale and a set of frequency domain enhanced feature maps at the first scale, the weight map at each scale needs to be obtained according to the above process. The weight map is then sampled, and the sampled weight map is weighted and summed with the deepest coarseness variation representation according to the above process to obtain the fourth feature at the first scale and the fourth feature at the second scale. Figure 1 The rightmost third spatial domain recovery module needs to represent the fourth feature at the third scale with the deepest coarseness variation. Figure 1 middle By splicing along the channel dimension, a third-scale representation of the change is obtained. Figure 1 middle For the second spatial domain recovery module, after obtaining the fourth feature at the second scale, it represents the change at the third scale. Figure 1 middle By splicing along the channel dimension, a second-scale representation of the change is obtained. Figure 1 middle For the first spatial domain recovery module, after obtaining the fourth feature at the first scale, it represents the change at the second scale. Figure 1 middle The changes are concatenated along the channel dimension to obtain the first-scale representation of the changes. Figure 1 middle ).

[0076] like Figure 1 As shown, in this embodiment, for the Spatial Domain Reconstruction (SRM) module, the number of scales requiring spatial detail reconstruction is 4. For the k-th SRM module, a set of frequency domain enhanced feature maps at the k-th scale is... and The input is fed into the k-th spatial domain recovery module, where an initial feature fusion is performed on a set of frequency domain enhanced feature maps at the k-th scale to obtain a coarse variation representation at the k-th scale. Represent the roughness variation at the k-th scale After average pooling operation ( Figure 1 (Mean), to obtain the first feature of the third scale. Represent the roughness variation at the k-th scale After max pooling operation ( Figure 1 Max), to obtain the second feature at the k-th scale. By concatenating the first feature and the second feature of the third scale, we obtain the third feature at the k-th scale. The third feature at scale k is passed sequentially through the first convolutional layer and the Sigmoid activation layer to calculate the weights of the third feature at scale k, which is the weight map corresponding to scale k. in, The concatenation operation is represented by φ, where φ represents the first convolutional layer, σ is the sigmoid activation function, and the weight map corresponding to the k-th scale is downsampled. The downsampled k-th scale weight map is then weighted and summed with the deepest layer's coarse transformation representation to obtain the fourth feature at the k-th scale. By concatenating the fourth feature at the k-th scale and the deepest coarse variation representation along the channel dimension, a more refined k-th scale variation representation is obtained. Where × indicates element-wise multiplication. The deepest coarsening transformation after initial feature fusion is represented as follows: when k is 1 to 2, This represents the output of the k-th spatial domain recovery module. When k is 3, This represents the deepest level of coarseness variation.

[0077] It should be noted that the Spatial Domain Recovery (SRM) module of this invention uses depthwise separable convolution instead of ordinary convolution, in order to effectively reduce the number of parameters and computational cost of the change detection network without affecting its performance as much as possible.

[0078] In this invention, the specific processing flow of the lightweight decoder module is as follows: the change representation of the first scale corresponding to the shallowest dual-time feature representation is taken as input, the change representation of the first scale is subjected to depthwise separable convolution to obtain a change detection feature map, the change detection feature map is upsampled to obtain the change detection result of the dual-time remote sensing image (the change detection result between the two time point images).

[0079] In this embodiment, the change at the first scale represents After depthwise separable convolution, the change detection feature map is obtained. Upsampling of the change detection feature map yields the change detection results for dual-time remote sensing images. Where ε represents the depthwise separable convolution operation, Upsample 32× This indicates a 32x upsampling operation, and the dimension of the final output dual-time remote sensing image change detection result is (B×2×H×W).

[0080] It should also be noted that in this invention, the kernel size of the first convolutional layer and the second convolutional layer is the same, which is 1×1; the kernel size of the depth-separable convolution is 3×3.

[0081] It should also be noted that in this invention, the change detection network DDLNet needs to be pre-trained using a labeled training dataset before being used for actual change detection tasks. Furthermore, to improve training performance, all training data in the training dataset undergoes random data augmentation, such as rotation and flipping. Additionally, the loss function used for training the change detection network in this invention is a hybrid loss function, obtained by weighted summation of a weighted cross-entropy loss function and a dice loss function. The parameters of the change detection network are updated based on minimizing the hybrid loss function. The specific training and validation processes can follow the training and validation methods in existing change detection networks, and will not be elaborated here.

[0082] To better demonstrate the specific implementation and technical effects of the present invention, the remote sensing image change detection method based on dual-domain learning shown in the above preferred implementation is applied to a specific example below.

[0083] Example

[0084] The specific structure of the change detection network DDLNet used in this embodiment is as follows: Figure 1 As shown above, a detailed description has been provided and will not be repeated here. The specific process of this embodiment is as follows: Figure 2 As shown, the specific process of this change detection task can be divided into three stages: data preprocessing, model training and validation, and testing and visualization. The testing and visualization results are as follows: Figure 3 As shown.

[0085] 1. Data Preprocessing Stage

[0086] This embodiment uses the LEVIR-CD dataset, WHU dataset, and CLCD dataset for remote sensing image change detection as examples.

[0087] The WHU architecture dataset contains a common change detection dataset consisting of a pair of time-series aerial images of size 32507×15354. The images are cropped into non-overlapping images of size 256×256 and then randomly divided into a training set (6096 images), a validation set (762 images), and a test set (762 images).

[0088] The LEVIR-CD dataset contains 637 pairs of high-resolution dual-time-series remote sensing images for change detection. The images are cropped into non-overlapping 256×256 images and then randomly divided into a training set (7120 images), a validation set (1024 images), and a test set (2048 images).

[0089] The CLCD dataset contains 600 pairs of 512×512 sample images of farmland changes, which are randomly divided into a training set (360 images), a validation set (120 images), and a test set (120 images).

[0090] 2. Model Training and Validation

[0091] Step 1: Obtain training set data and validation set data as described above, and divide the training set and validation set into batches according to a certain batch size, with a total of N.

[0092] Step 2: Select a batch of samples with index i sequentially from the training and validation datasets, where i ∈ {0, 1, ..., N}. Train the change detection network DDLNet using each batch of training samples, and validate and evaluate it after each training round. During training, calculate the mixture loss function for each training sample. and in and These are the weighted cross-entropy loss function and the dice loss function, respectively. Based on these, the parameters of the change detection network are comprehensively adjusted according to the total loss of each batch of training samples until training ends. After completing a given number of iterations, the model converges, training is complete, and the validation results during the training process are obtained. Simultaneously, during training, the input image is randomly flipped horizontally and vertically and rotated for data augmentation.

[0093] 3. Testing and Visualization

[0094] The dual-time images of the test dataset are input into the trained change detection network DDLNet. The test output obtains the changed parts of the two images and outputs a binary classification image to complete the change detection task.

[0095] This invention evaluates the proposed change detection network DDLNet on the three aforementioned remote sensing image change detection datasets using five common metrics: F1-score (F1), recall (Rec), precision (Pre), intersection-over-union ratio (IoU), and overall accuracy (OA). Furthermore, this invention employs two metrics: the number of parameters in millions (M) and floating-point operations per second (Flops) in giga (G) to measure the network's spatial complexity and computational cost.

[0096] In this embodiment, the test visualization results are as follows: Figure 3 As shown, the test data results are presented in Tables 1 and 2 below.

[0097] Table 1 Test Data Results

[0098]

[0099]

[0100] Table 2 Parameter Quantity and Calculation Speed

[0101] DDLNet 12.67 7.35

[0102] like Figure 3As shown in the table above, the change detection network DDLNet of this invention performs well in changing remote sensing images. By extracting frequency components, it enhances interest changes through more accurate semantic recognition, while enabling feature fusion through cross-time and cross-scale mechanisms. This allows higher-level change representations to guide contextual modeling of lower-level representations, thereby enhancing semantic information in lower-level features and improving spatial detail recovery of change representations. Furthermore, the proposed method exhibits excellent performance in terms of parameter and computational costs. Therefore, this invention provides a new approach to changing remote sensing images from both frequency and spatial domain perspectives, effectively improving the performance of change detection.

[0103] It should also be noted that the remote sensing image change detection method based on dual-domain learning in the above embodiments can essentially be executed by a computer program or module. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a remote sensing image change detection system based on dual-domain learning, corresponding to the remote sensing image change detection method based on dual-domain learning provided in the above embodiments, such as... Figure 4 As shown, it includes:

[0104] The data acquisition module is used to acquire remote sensing images taken at two different times in the same geographical area and combine them into a dual-time remote sensing image.

[0105] The result acquisition module is used to input the dual-time remote sensing image output by the data acquisition module into the trained change detection network to obtain the change detection result of the dual-time remote sensing image. The change detection network consists of an encoder module, a frequency domain enhancement module, a spatial domain recovery module, and a lightweight decoder module.

[0106] Similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a computer electronic device corresponding to the remote sensing image change detection method based on dual-domain learning provided in the above embodiment, which includes a memory and a processor;

[0107] The memory is used to store computer programs;

[0108] The processor is configured to implement a remote sensing image change detection method based on dual-domain learning in the above embodiments when executing the computer program.

[0109] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0110] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0111] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the system and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.

[0112] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A remote sensing image change detection method based on dual-domain learning, characterized in that, Includes the following steps: Remote sensing images taken in the same geographical area at two different times are combined into a dual-time remote sensing image. The dual-time remote sensing image is then input into a trained change detection network to obtain the change detection results of the dual-time remote sensing image. The change detection network consists of an encoder module, a frequency domain enhancement module, a spatial domain recovery module, and a lightweight decoder module. In the change detection network, the dual-time remote sensing image is first processed by the encoder module to extract features, resulting in dual-time feature representations at different scales. The encoder module is composed of several residual blocks cascaded in sequence. Each residual block corresponds to a dual-time feature representation at one scale, and each scale dual-time feature representation contains two single-time feature representations at the same scale. Then, the two single-temporal feature representations of the same scale are each processed by the frequency domain enhancement module, and the frequency components are extracted from the input single-temporal feature representation using discrete cosine transform to obtain the frequency domain enhanced feature map corresponding to each scale of single-temporal feature representation. The frequency domain enhanced feature maps of the same scale are then combined into a set of frequency domain enhanced feature maps. After initial feature fusion, the set of frequency domain enhanced feature maps of the deepest scale is used to obtain the deepest coarse change representation. Except for the set of frequency domain enhanced feature maps corresponding to the deepest coarse change representation, each set of frequency domain enhanced feature maps of other scales is processed by a spatial domain recovery module. At the same time, each spatial domain recovery module is additionally input with the deepest coarse change representation to obtain a change representation with the same scale as each set of input frequency domain enhanced feature maps. The number of spatial domain recovery modules is one less than the number of residual blocks. Finally, the change representation corresponding to the shallowest dual-temporal feature representation is input into the lightweight decoder module to obtain the change detection result of the dual-temporal remote sensing image. The specific processing flow of the frequency domain enhancement module is as follows: Taking the single-time feature representation of each scale as input, the single-time feature representation of each scale is first uniformly divided into a preset number of parts along the channel dimension; after pre-training the frequency domain enhancement module on the ImageNet dataset, a two-dimensional index corresponding to the two-dimensional discrete cosine transform frequency component is selected for each part, and the two-dimensional discrete cosine transform frequency component of each part is calculated according to the two-dimensional index corresponding to each part; after concatenating the two-dimensional discrete cosine transform frequency components of each part, a frequency vector is obtained; the frequency vector is passed through the second convolutional layer and the sigmoid activation layer in sequence to obtain the frequency weights; the frequency weights are weighted and summed with the single-time feature representation of each scale to obtain the frequency domain enhanced feature map of each scale; For the i-th part Two-dimensional discrete cosine transform frequency components Its functional form is: ； in,[ , ] indicates the i-th part Two-dimensional indexes corresponding to the frequency components of the two-dimensional discrete cosine transform. Indicates the first index, Indicates the second index; Represents the discrete cosine transform; Indicates height index, Indicates the width index. These represent the height and width of a single temporal feature representation, respectively. Represents the i-th part Medium height index is And the width index is Two-dimensional feature map.

2. The remote sensing image change detection method based on dual-domain learning as described in claim 1, characterized in that, The encoder module adopts a ResNet-18 backbone network, which is composed of a first residual block, a second residual block, a third residual block, and a fourth residual block cascaded in sequence. The first residual block outputs a bitemporal feature representation at the first scale, the second residual block outputs a bitemporal feature representation at the second scale, the third residual block outputs a bitemporal feature representation at the third scale, and the fourth residual block outputs a bitemporal feature representation at the fourth scale. The bitemporal feature representation at the fourth scale is used as the deepest bitemporal feature representation.

3. The remote sensing image change detection method based on dual-domain learning as described in claim 2, characterized in that, The specific processing flow of the initial feature fusion is as follows: After the dual temporal feature representation of the fourth scale is passed through the frequency domain enhancement module, a set of frequency domain enhanced feature maps of the fourth scale are obtained. The set of frequency domain enhanced feature maps of the fourth scale is subjected to pixel-level subtraction to obtain the coarse change representation of the fourth scale. One of the frequency domain enhanced feature maps of the set of frequency domain enhanced feature maps of the fourth scale is concatenated with the coarse change representation of the fourth scale in the channel dimension to obtain the concatenated feature of the fourth scale. The concatenated feature of the fourth scale is integrated by depthwise separable convolution to obtain the deepest coarse change representation.

4. The remote sensing image change detection method based on dual-domain learning as described in claim 3, characterized in that, The specific processing flow for each spatial domain recovery module is as follows: After the dual-time feature representation of the third scale is passed through the frequency domain enhancement module, a set of frequency domain enhanced feature maps of the third scale is obtained. After the dual-time feature representation of the second scale is passed through the frequency domain enhancement module, a set of frequency domain enhanced feature maps of the second scale is obtained. After the dual-time feature representation of the first scale is passed through the frequency domain enhancement module, a set of frequency domain enhanced feature maps of the first scale is obtained. In addition to a set of frequency domain enhanced feature maps at the fourth scale, a set of frequency domain enhanced feature maps at the third scale are input into the third spatial domain recovery module. Initial feature fusion is first performed on the set of frequency domain enhanced feature maps at the third scale to obtain a coarse change representation at the third scale. The coarse change representation at the third scale is then subjected to average pooling to obtain the first feature at the third scale. The coarse change representation at the third scale is then subjected to max pooling to obtain the second feature at the third scale. The first feature at the third scale and the second feature at the third scale are concatenated to obtain the third feature at the third scale. The third feature at the third scale is then passed sequentially through the first convolutional layer and the Sigmoid activation layer to obtain the weight map corresponding to the third scale. The weight map corresponding to the third scale is downsampled. The downsampled weight map at the third scale is then weighted and summed with the deepest coarse change representation to obtain the fourth feature at the third scale. The fourth feature at the third scale and the deepest coarse change representation at the third scale are then concatenated along the channel dimension to obtain the change representation at the third scale. A set of frequency domain enhanced feature maps at the second scale is input into the second spatial domain recovery module. Initial feature fusion is performed on the set of frequency domain enhanced feature maps at the second scale to obtain a coarse change representation at the second scale. The coarse change representation at the second scale is then subjected to average pooling to obtain the first feature at the second scale. The coarse change representation at the second scale is then subjected to max pooling to obtain the second feature at the second scale. The first feature at the second scale and the second feature at the second scale are concatenated to obtain the third feature at the second scale. The second feature at the second scale is then passed sequentially through the first convolutional layer and the Sigmoid activation layer to obtain the weight map corresponding to the second scale. The weight map corresponding to the second scale is downsampled. The downsampled weight map at the second scale is then weighted and summed with the deepest coarse change representation to obtain the fourth feature at the second scale. The fourth feature at the second scale and the change representation at the third scale are then concatenated along the channel dimension to obtain the change representation at the second scale. A set of frequency domain enhanced feature maps at the first scale is input into the first spatial domain recovery module. Initial feature fusion is performed on the set of frequency domain enhanced feature maps at the first scale to obtain a coarse change representation at the first scale. The coarse change representation at the first scale is then subjected to average pooling to obtain the first feature at the first scale. The coarse change representation at the first scale is then subjected to max pooling to obtain the second feature at the first scale. The first feature at the first scale and the second feature at the first scale are concatenated to obtain the third feature at the first scale. The third feature at the first scale is then passed sequentially through the first convolutional layer and the Sigmoid activation layer to obtain the weight map corresponding to the first scale. The weight map corresponding to the first scale is downsampled. The downsampled first scale weight map is then weighted and summed with the deepest coarse change representation to obtain the fourth feature at the first scale. The fourth feature at the first scale and the change representation at the second scale are concatenated along the channel dimension to obtain the change representation at the first scale.

5. The remote sensing image change detection method based on dual-domain learning as described in claim 4, characterized in that, The specific processing flow of the lightweight decoder module is as follows: the change representation at the first scale corresponding to the shallowest dual-temporal feature representation is taken as input, the change representation at the first scale is subjected to depthwise separable convolution to obtain a change detection feature map, and the change detection feature map is upsampled to obtain the change detection result of the dual-temporal remote sensing image.

6. The remote sensing image change detection method based on dual-domain learning as described in claim 5, characterized in that, The first and second convolutional layers have the same kernel size, 1×1; the depth-separable convolution kernel size is 3×3.

7. The remote sensing image change detection method based on dual-domain learning as described in claim 1, characterized in that, Before being used in actual change detection tasks, the change detection network needs to be trained in advance using a labeled training dataset, and all training data in the training dataset needs to be randomly augmented. The loss function used in training the change detection network is a hybrid loss function, which is obtained by weighted summation of weighted cross-entropy loss function and dice loss function. The parameters of the change detection network are updated based on minimizing the hybrid loss function.

8. A remote sensing image change detection system based on dual-domain learning, characterized in that, To implement the remote sensing image change detection method based on dual-domain learning as described in any one of claims 1 to 7, the method includes: The data acquisition module is used to acquire remote sensing images taken at two different times in the same geographical area and combine them into a dual-time remote sensing image. The result acquisition module is used to input the dual-time remote sensing image output by the data acquisition module into the trained change detection network to obtain the change detection result of the dual-time remote sensing image. The change detection network consists of an encoder module, a frequency domain enhancement module, a spatial domain recovery module, and a lightweight decoder module.

9. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement a remote sensing image change detection method based on dual-domain learning as described in any one of claims 1 to 7.

Images