A high-resolution remote sensing image building change detection network and method

Abstract

The application provides a high-resolution remote sensing image building change detection network and method, and belongs to the field of image processing; solves the key problems of task-independent interference, sparse distribution of change areas and fuzzy boundaries existing in the existing remote sensing image change detection technology; comprises an encoder module, an interactive high-frequency enhancement module, a fuzzy-guided difference enhancement module, a hybrid attention module, a decoder module and a prediction head; the interactive high-frequency enhancement module is used for feature interaction and enhancement of multi-scale dual-time-phase features output by the encoder module; after the interactive high-frequency enhancement module, a double-path structure is adopted in the network, the fuzzy-guided difference enhancement module is arranged on a difference branch in the double-path, the standard difference and the fuzzy difference are fused, and a dynamic weight fusion strategy is combined to realize fine modeling of multi-scale feature differences; the hybrid attention module is arranged on a channel cascading branch in the double-path; and the application is applied to remote sensing image building change detection.

Description

Technical Field

[0001] This application relates to the field of remote sensing image change detection technology, and in particular to a high-resolution remote sensing image building change detection network and method. Background Technology

[0002] With continuous population growth and urban expansion, the impact of human activities on the Earth's surface environment has reached an all-time high. Remote sensing building change detection technology aims to accurately identify changes in the quantity, density, distribution, and morphology of building clusters by analyzing and comparing spatial information from multiple temporal images of the same geographical area, and has become a key technological means to support modern urban governance.

[0003] Traditional change detection methods primarily rely on spectral analysis and manual intervention, and their technical frameworks can be broadly categorized into two typical modes: image processing and manual interpretation. Image processing-based approaches typically utilize methods such as interpolation and change vector analysis to generate difference maps, then define change regions by setting thresholds. While simple to implement, this approach's over-reliance on spectral features often leads to non-semantic changes caused by environmental factors such as seasonal variations and lighting differences being misidentified as real changes. Another approach, combining manual interpretation with machine learning, requires professionals to visually interpret images, manually annotate building outlines as training samples, and then utilize shallow models such as support vector machines, K-nearest neighbors, and random forests for feature learning. While this method partially improves efficiency, its generalization ability remains limited due to the reliance on manual feature extraction design. Furthermore, model training requires a large amount of labeled data, consuming significant manpower and easily introducing subjective errors, and involves tedious parameter adjustments, limiting its practicality.

[0004] In recent years, revolutionary advancements in deep learning technology have provided a new technological paradigm for remote sensing building change detection. Compared to traditional methods that rely on manual feature design, deep learning extracts high-level features from remote sensing images through end-to-end autonomous learning, thereby improving detection accuracy and stability. Convolutional Neural Networks (CNNs), with their excellent feature extraction capabilities, achieve autonomous feature learning from pixel-level texture to object-level semantics. By introducing attention mechanisms and multi-scale feature fusion, they enhance the model's ability to capture details such as building edges and textures.

[0005] In recent years, deep learning technology has received widespread attention and application in remote sensing image building change detection tasks due to its excellent feature representation capabilities. However, existing remote sensing image change detection technologies still face many challenges and limitations, such as task-irrelevant interference, sparse distribution of changed areas, and blurred boundaries. Therefore, further research and improvement are needed. Summary of the Invention

[0006] To address the aforementioned technical issues, this application proposes a high-resolution remote sensing image building change detection network and method.

[0007] The technical solution adopted in this application is as follows: a high-resolution remote sensing image building change detection network, including an encoder module, an interactive high-frequency enhancement module, a fuzzy-guided differential enhancement module, a hybrid attention module, a decoder module, and a prediction head. The encoder module employs a twin encoder for multi-scale feature extraction of the input raw dual-temporal high-resolution remote sensing image. The interactive high-frequency enhancement module is used for feature interaction and enhancement of the multi-scale dual-temporal features output by the encoder module. After the interactive high-frequency enhancement module, the network adopts a dual-path structure, wherein the fuzzy-guided differential enhancement module is set in the dual path. In the differential branch, the fuzzy-guided differential enhancement module achieves refined modeling of multi-scale feature differences by fusing standard difference and fuzzy difference and combining dynamic weight fusion strategy. The hybrid attention module is set on the channel cascade branch in the dual path, and the channel attention module and spatial attention module are serially connected to complete the feature enhancement of the channel and spatial dimensions in turn. Then, the dual-temporal features output by the dual-path structure are upsampled and fused by the decoder module to output a dual-temporal feature map of the same size as the original dual-temporal high-resolution remote sensing image. The dual-temporal feature map is then input into the prediction head to finally output the change map.

[0008] Furthermore, the twin encoder of the encoder module uses a pre-trained CNN as the backbone network to extract multi-level features from dual-temporal high-resolution remote sensing images.

[0009] Furthermore, the interactive high-frequency enhancement module maps the features of the dual-temporal images extracted by the encoder module through a 1×1 convolution, and obtains the query, key, and value through reshaping, and then performs cross-attention mechanism operation to obtain the output after cross-attention calculation;

[0010] Furthermore, the interactive high-frequency enhancement module also introduces an adaptive Sobel operator to extract high-frequency components of the feature map and obtain edge-related information.

[0011] Furthermore, an isotropic Sobel operator with a minimum kernel size of 3 in eight directions is used to extract high-frequency components of each channel in the input feature map in eight different directions. In this process, a Top-k direction enhancement strategy is introduced to dynamically select the three directions with the most intense response in each channel and perform weighted fusion. Then, the feature map with high-frequency information enhancement generated by the adaptive Sobel operator is residually connected with the original input feature map. Finally, interactive high-frequency enhancement features with the same number of input channels are generated by 1×1 convolution.

[0012] Furthermore, the standard difference in the fuzzy-guided differential enhancement module adopts a spatial normalization strategy to enhance local contrast. It performs maximum value normalization along the channel dimension, constrains the feature values ​​to the [0,1] interval, and enhances the contrast of local regions.

[0013] Furthermore, the fuzzy difference in the fuzzy-guided difference enhancement module uses the constructed basic difference features as the basic feature expression. Then, based on fuzzy set theory, a nonlinear transformation S-shaped membership function is constructed by introducing channel-adaptive learnable parameters to obtain fuzzy difference features. The basic difference features and fuzzy difference features are then concatenated by channels and input into the dynamic fusion sub-network. A weight generator constructed by depthwise separable convolution and 1×1 convolution predicts spatially adaptive fusion weights. Finally, the enhanced hybrid difference features are obtained through weighted fusion.

[0014] Furthermore, the channel attention module in the hybrid attention module adopts the frequency domain channel attention network FCANet, and introduces a two-dimensional discrete cosine transform into the frequency domain channel attention network FCANet. The two-dimensional discrete cosine transform realizes the diversity modeling of channel attention weights by performing multi-band component decomposition and selective fusion on the feature map, and finally transforms the input features into multi-spectral attention.

[0015] Furthermore, a combination strategy of binary cross-entropy loss and Dice loss is adopted to train the network and optimize its parameters.

[0016] A method for detecting building changes in high-resolution remote sensing imagery includes the following steps:

[0017] S1: Preprocess the high-resolution remote sensing image change detection dataset;

[0018] S2: Train the model constructed using a high-resolution remote sensing image building change detection network to obtain the optimal model;

[0019] S3: Detect changes in buildings in high-resolution remote sensing images using a trained model and output a binary map to represent the changes.

[0020] A computer device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method.

[0021] A computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of the method.

[0022] The advantages of this application over the prior art are as follows:

[0023] (1) The introduction of the IHFM module enables the dynamic focusing of change-related regions, thereby reducing interference from irrelevant information and maximizing the accuracy of boundary detection.

[0024] (2) Introducing the FDEM module can effectively alleviate the problems of noise interference and blurred boundary detection in changing regions.

[0025] (3) The HAM module is introduced to capture cross-channel frequency features through frequency domain multi-spectral channel attention and focus on key areas in conjunction with spatial attention mechanism to form a channel-space collaborative feature optimization mechanism, which improves the model’s ability to perceive complex features and effectively improves the accuracy and robustness of change detection.

[0026] (4) The combination strategy of binary cross-entropy loss and Dice loss can effectively balance the contribution of changed and unchanged samples to model training, thereby improving the detection accuracy of building change edges, while avoiding overfitting or over-refinement. Attached Figure Description

[0027] The following description, in conjunction with the accompanying drawings, further illustrates this application:

[0028] Figure 1 A schematic diagram of the structure of the improved network provided in the embodiments of this application;

[0029] Figure 2 This is a schematic diagram of the structure of the interactive high-frequency enhancement module provided in the embodiments of this application;

[0030] Figure 3 A schematic diagram of an isotropic Sobel operator in eight directions provided for embodiments of this application;

[0031] Figure 4 A schematic diagram of the structure of the fuzzy-guided differential feature enhancement module provided in an embodiment of this application;

[0032] Figure 5 This is a schematic diagram of the structure of the hybrid attention module provided in an embodiment of this application;

[0033] Figure 6 This is a schematic diagram of the FcaNet network structure provided in an embodiment of this application;

[0034] Figure 7 A schematic diagram of the spatial attention structure provided in the embodiments of this application;

[0035] Figure 8 A comparison chart of the change detection results of different models provided in the embodiments of this application. Detailed Implementation

[0036] like Figures 1 to 8As shown, this application provides a high-resolution remote sensing image building change detection network, which is an interactive high-frequency enhanced dual-path differential feature extraction network (IHF-DPDNet), and its network structure is as follows. Figure 1 As shown, IHF-DPDNet is designed as a Siamese architecture, using a pre-trained CNN (ResNet18) as the backbone network to extract multi-level features from dual-temporal remote sensing images. Specifically, it employs multi-layer feature maps with spatial sizes of 64×64, 32×32, and 16×16 in the first three stages of ResNet18. Based on this, an interactive high-frequency enhancement module establishes dynamic associations between dual-temporal features, utilizes a cross-attention mechanism to achieve feature interaction, and combines an adaptive Sobel operator to enhance high-frequency boundary information, effectively improving the model's ability to perceive subtle changes. Next, the network adopts a dual-path parallel structure, where the difference branch introduces a fuzzy-guided difference enhancement module. By fusing standard difference and fuzzy difference features, a dynamic weight strategy is used to achieve multi-scale difference modeling, significantly mitigating noise interference and boundary ambiguity. To further enhance the expressive power of difference features, a hybrid attention module is designed in the channel cascade branch. In the frequency domain, multi-spectral channel attention captures cross-channel feature dependencies, combined with a spatial attention mechanism to focus on key regions, forming a channel-space co-optimization mechanism aimed at improving the model's ability to perceive complex features. Finally, a progressive strategy is used to fuse differential features at different scales to generate the final change detection results.

[0037] The specific implementation plan is as follows:

[0038] (1) An Interactive High Frequency Enhancement Module (IHFM) was proposed, with the following structure: Figure 2 As shown, interactive feature learning is performed between the two temporal images before capturing the true difference information, and a lightweight edge detection method is used to enhance the high-frequency difference information. Using IHFM, regions related to change can be dynamically focused, thereby reducing interference from irrelevant information and maximizing the accuracy of boundary detection.

[0039] The features F1 and F2 extracted from the dual-temporal images in the previous stage are mapped by a 1×1 convolution, and the query Q, key K and value V are obtained by reshaping. Then, the cross-attention mechanism is performed, and its expression is shown in equations (1) and (2):

[0040]

[0041] Ci =F i +Cross-Attention(Q i ,K j V j (2);

[0042] Where i and j represent the sequence numbers of the feature maps at different time phases, and d i Represents vector Q i Dimensions Let F represent the cross-attention probability distribution. i and C i F represents the original input feature map and the output calculated after cross-attention, respectively. i Attachment is achieved by skipping connections. Discriminative information learned through an interactive mechanism of cross-temporal images significantly improves the model's ability to perceive surface changes. This mechanism not only effectively suppresses task-irrelevant interference factors but also provides richer contextual information to help the model accurately identify areas of change.

[0043] To accurately characterize the boundary information of changing regions, especially in high-resolution remote sensing imagery where details such as building boundaries are easily affected by blurring and background interference, we designed an adaptive Sobel high-frequency enhancement component in IHFM. By introducing the Sobel operator, which is beneficial for feature extraction, we extract the high-frequency components of the feature map while acquiring edge-related information with minimal computational resource consumption, improving the model's ability to perceive the boundaries of changing regions and ensuring clearer and more stable detection results. However, since buildings in high-resolution remote sensing imagery often have varying shapes, the isotropic Sobel operator shows a greater advantage due to its robustness to shape changes. The structure of the eight-directional isotropic Sobel operator is as follows: Figure 3 As shown.

[0044] By employing the Sobel operator with a minimum kernel size of 3, high-frequency components of the input feature map in eight different directions are effectively extracted, resulting in significant enhancement of high-frequency features across various geometries and microstructures. These Sobel kernels are stored as fixed parameters, eliminating the need for learning during training, and are computed via group convolution. This feature effectively controls the number of network parameters while ensuring computational efficiency and reducing structural risks in the model. Since the edge characteristics of different land features vary, Sobel calculations in certain directions may better highlight the edges of target objects, while other directions may be affected by noise. To address this issue, we introduce a Top-k direction enhancement strategy, dynamically selecting the three most important directions in each channel and performing weighted fusion. Specifically, we perform a difference operation between C1 and C2 to obtain the difference feature C.diff Then the difference feature C diff The interactive features C1 and C2 are concatenated along the channel dimension, and the high-frequency information-enhanced feature map generated by adaptive Sobel edge detection is connected to the original input residual. Finally, the interactive high-frequency enhanced features with the same number of input channels are generated by 1×1 convolution. The calculation process is shown in equations (3) to (5):

[0045] C diff =C1-C2(3);

[0046]

[0047] in: Indicates the C after splicing i and C diff feature;

[0048] i∈{1,2} represents two different time phases;

[0049] Conv sobel (·) indicates edge detection using the Sobel operator.

[0050] IHFM effectively alleviates the problem of partial information loss caused by the increase in the number of neural network layers during training. It makes the most of the interactive differential discrimination feature information between the two temporal images, while adaptively enhancing high-frequency information, strengthening the model's sensitivity to changing regions and boundary accuracy, and ultimately improving the overall accuracy and robustness of the model.

[0051] (2) Traditional difference methods mainly rely on directly calculating the pixel-level differences between two feature maps. However, this method is sensitive to noise and easily affected by factors such as illumination and sensor noise. Furthermore, in remote sensing images, many changing regions often exhibit blurred transitions, which traditional difference methods struggle to effectively characterize, potentially leading to missed or false detections. To more accurately describe these regions, we propose a Fuzzy-guided Differential Enhancement Module (FDEM), with the structure shown below. Figure 4 As shown, this module achieves refined modeling of multi-scale feature differences by fusing standard difference and fuzzy difference and combining them with a dynamic weight fusion strategy.

[0052] The standard deviation in FDEM employs a spatial normalization strategy to enhance local contrast. Maximum value normalization is performed along the channel dimension, constraining the eigenvalues ​​to the [0,1] interval to enhance local contrast while preserving the original spatial distribution characteristics. The calculation process is shown in equation (6):

[0053]

[0054] Among them, D norm Indicates the basic standard deviation characteristics. This represents the feature map after feature interaction enhancement processing. Let B represent the set of real numbers, C represent the batch size, H represent the number of channels, W represent the feature map height, and max(·) represent the maximum value operation along the channel dimension. This operation extracts the maximum difference value of each spatial location across all channels as the local contrast benchmark in the spatial dimension. The parameter ε is a minimal constant (default value is 10). -6 This is used to avoid numerical instability when the denominator is zero, ensuring the robustness of the calculation without affecting the significance of the normalization result.

[0055] In fuzzy difference, the basic standard deviation feature D is first constructed. norm As a basic feature representation, based on fuzzy set theory, a nonlinear transformation S-shaped membership function is constructed by introducing channel-adaptive learnable parameters. The calculation process is shown in equation (7):

[0056]

[0057] Among them, D fuzzy Represents the fuzzy difference feature, θ c The learnable threshold parameter for the channel dimension determines the activation threshold of the fuzzy difference feature, and is adjusted by θ. c The model can adaptively distinguish significant differences between different channels from background noise and suppress interference from low-intensity differences. c θ is a learnable scaling factor for the channel dimension, controlling the steepness of the fuzzy gradient. A larger value results in a more pronounced fuzzy difference feature at the threshold θ. c The more sensitive the response in the vicinity, the more dramatic the gradient change; conversely, the response is smoother, suitable for modeling gradual differences. ⊙ represents element-wise multiplication (broadcasting elements in the channel dimension), mapping the absolute difference features to the (0,1) interval using the Sigmoid function. Two difference features D norm With D fuzzy After channel concatenation, the data is input into a dynamic fusion subnetwork. A weight generator constructed using depthwise separable convolutions and 1×1 convolutions predicts spatially adaptive fusion weights W'. Finally, weighted fusion yields the enhanced hybrid difference features D. hybrid The calculation process is shown in equations (8) and (9):

[0058] W'=σ(Conv 1×1 (DSConv 3×3 ([D norm ;D fuzzy ]))) (8);

[0059] D hybrid =W⊙D norm +(1-W)⊙D fuzzy (9);

[0060] Where [·;·] represents channel concatenation operation, DSConv(·) represents 3×3 depthwise separable convolution, σ represents the Sigmoid activation function, and ⊙ represents element-wise multiplication operation.

[0061] (3) Inspired by CBAM, this application designs a Hybrid Attention Module (HAM) to address the characteristics of cascaded branching channels. The structure is as follows: Figure 5 As shown, the channel attention module and the spatial attention module are executed sequentially. The channel attention module uses the frequency domain channel attention network FCANet, which is then connected to the original spatial attention module to sequentially enhance features in both the channel and spatial dimensions. Input feature F input H1 and H2 are bi-temporal interactive high-frequency enhancement features that have been spliced ​​and subjected to 3×3 convolution operations.

[0062] Traditional implementations of channel attention modules primarily rely on scalar computation for each channel, a typical example being the Global Average Pooling (GAP) strategy used in SENet. While this strategy is simple and computationally efficient, it suffers from certain feature representation bottlenecks. Because GAP is overly simplistic, it struggles to effectively capture detailed information in complex inputs, particularly the homogenization of feature distributions caused by averaging, which disrupts the diverse representation between channels. Therefore, this application introduces the FCANet framework into the channel cascade branching, extending channel attention to the frequency domain and transforming channel feature compression and reconstruction into a frequency domain decomposition problem. From a frequency domain analysis perspective, it is demonstrated that the traditional GAP method can essentially be considered a special case of frequency domain feature decomposition. This allows for adaptive selection and compression coding of multi-band features while preserving the overall representational capability of the channels, significantly improving the discriminative power of attention weights.

[0063] Within the FCANet framework, the two-dimensional discrete cosine transform (DCT) is applied to the frequency domain channel attention mechanism. By performing multi-band component decomposition and selective fusion on the feature map, it overcomes the representational limitations of the traditional GAP method while maintaining computational efficiency, and realizes the diverse modeling of channel attention weights. Specifically, the calculation process of the two-dimensional discrete cosine transform is shown in equation (10):

[0064]

[0065] Where H and W are the height and width of the input feature map, respectively. Let (i,j) be the value of the pixel at position (i,j) in the feature map. It is a two-dimensional DCT spectrum, which can be viewed as a weighted sum of each pixel in the feature map. The cosine part is equivalent to the weights. The spatial domain can be converted to the frequency domain through two-dimensional discrete cosine transform. h is the index in the height direction, and w is the index in the width direction.

[0066] The FCANet network structure is shown below. Figure 6 Its core component is the multispectral channel attention within the rectangle, and its specific implementation is as follows:

[0067] First, the input feature X is divided into several parts [X0, X1, ..., X] along the channel dimension. n-1 ],in:

[0068]

[0069] The number of channels C must be divisible by the number of regions n. C′ is the number of channels in the divided region. Each divided region is assigned a specific two-dimensional DCT frequency component. For each divided region, a corresponding two-dimensional DCT frequency component is assigned, and the corresponding frequency domain representation is obtained through the transformation process shown in equation (11):

[0070]

[0071] in, Indicates the two-dimensional DCT transform in (u i ,v i The components on ) This represents the value of the input signal at position (h, w). This indicates that the DCT basis functions are at frequency (u i ,v i and the value at position (h,w), [u i ,v i ] is with X i The corresponding two-dimensional exponent of the frequency components, through compression, makes Freq i Transform it into a vector with channel dimension C′, and finally compress the Freq. i All vectors are concatenated to obtain a multispectral vector. The calculation process is shown in equation (12):

[0072] Freq=compress(X)=cat([Freq 0 ,Freq 1 ,…,Freq n-1 ]) (12);

[0073] In summary, the process of transforming input features X into multispectral attention can be expressed as equation (13):

[0074] FcaAtt=σ(FC(Freq)) (13);

[0075] Here, FC represents fully connected, and σ represents the sigmoid activation function. Finally, FCANet applies the obtained multispectral channel attention weights to the input feature X through element-wise multiplication to obtain the output feature F′. out This achieves the goal of fully utilizing the compressed channel information.

[0076] Spatial attention mechanisms achieve explicit modeling of the relationship between local regions and global context by dynamically adjusting the spatial weight allocation of feature maps. Its structure diagram is shown below. Figure 7 The mechanism first performs global max pooling and average pooling operations on the input features to extract the saliency distribution features and mean statistical features of the spatial dimension, respectively. Subsequently, feature concatenation and convolution compression operations of the channel dimension are applied to construct the spatial weight matrix, which is then normalized by the Sigmoid function to form a spatial attention mask. The final output features are obtained by multiplying the original features and the attention mask pixel by pixel, and the calculation process is shown in equations (14) and (15):

[0077] M s =σ(Conv 7×7 ([F max ;F avg ])) (14);

[0078] F″ out =M s ×F′ out (15);

[0079] Among them, F max For the feature map after max pooling, F avg This represents the feature map after average pooling, [;] represents the channel concatenation operation, Conv 7×7 (·) denotes a 2D convolution operation with a kernel size of 7×7, where σ is the Sigmoid activation function, and M s This represents the spatial attention weights of the feature map. To maintain gradient flow stability and preserve original feature information, a learnable residual connection mechanism is introduced. After adjusting the dimensions through a 1×1 convolution, it is fused with the attention features, and the final output can be represented as:

[0080] F out =F″ out +Conv 1×1 (F input (16).

[0081] (4) Using binary cross-entropy loss L BCE With Dice loss L DiceThe combined strategy of the two loss functions has been proven to have a synergistic effect in multi-class imbalanced sample processing scenarios, effectively balancing the contributions of changed and unchanged samples to model training, thereby improving the detection accuracy of building change edges. The total loss function is shown in Equation (17):

[0082] L = L BCE +L Dice (17).

[0083] The pixel-level binary cross-entropy (BCE) loss function can effectively measure the difference between the model's prediction and the true label, promoting the model's accurate detection of changing regions, and is therefore widely used in binary classification problems. In this embodiment, BCE Loss is selected as part of the loss function, and the calculation method of BCE Loss is shown in Equation (18):

[0084]

[0085] Among them, y i This represents the variation of pixel i in the ground-based real-label image (y). i =1 indicates a change, y i =0 indicates no change); p i represents the probability of pixel i changing; N represents the total number of pixels.

[0086] Because the number of pixels in unchanged and changed areas differs significantly in dual-temporal remote sensing images, the sample distribution exhibits a highly imbalanced characteristic. When relying solely on binary cross-entropy loss, the model's attention to unchanged pixels is excessively amplified, resulting in a limitation on its ability to identify change features. To address this, a loss function based on the Dice coefficient is introduced. By measuring the spatial overlap between the predicted region and the true label, the impact of class imbalance on the training process is effectively mitigated, promoting the convergence and optimization of the network in change-sensitive regions. The calculation method of Dice Loss is shown in Equation (19):

[0087]

[0088] Among them, y i Represents the label value of pixel i; represents the predicted value of pixel i; N represents the total number of pixels.

[0089] This application also proposes a method for detecting building changes in high-resolution remote sensing images using the aforementioned network model, comprising the following steps:

[0090] S1: Preprocess the high-resolution remote sensing image change detection dataset, including geometric registration, radiometric correction, image enhancement, etc., and divide it into training set, validation set and test set;

[0091] S2: Train the high-resolution remote sensing image building change detection network model to obtain the optimal model;

[0092] S3: Detect changes in buildings in high-resolution remote sensing images using a trained model and output a binary map to represent the changes.

[0093] The network model in this application employs a training-and-evaluate approach during training. After each epoch of training, the model is saved and evaluated simultaneously, with the optimal model being retained. After training, the optimal model is used to predict the test set. This embodiment uses five metrics to evaluate inference performance: precision (Pre), recall (Rec), F1 score, intersection-over-union ratio (IoU), and overall accuracy (OA). Specifically, the metrics are defined as follows:

[0094]

[0095] Wherein, TP represents the number of pixels predicted to change and actually changing, TN represents the number of pixels predicted to remain unchanged and actually remaining unchanged, FP represents the number of pixels predicted to change but actually remaining unchanged, and FN represents the number of pixels predicted to remain unchanged but actually changing.

[0096] To objectively and comprehensively compare the performance of the IHF-DPDNet network proposed in this application, this embodiment selects mainstream and advanced change detection methods in the field of change detection for comparison. The model comparison methods include classic CNN-based algorithms in the change detection task field: FC-EF, FC-Siam-conc, and FC-Siam-diff; Unet++-based methods: SNUNet; attention mechanism-based methods: IFNet; and Transformer-based methods: BIT. Table 1 shows the quantitative evaluation results of different methods on the LEVIR-CD dataset.

[0097] Table 1. Quantitative evaluation results of different methods on the LEVIR-CD dataset.

[0098]

[0099] Table 1 presents the comparative results of seven networks in change detection on the LEVIR-CD dataset. Our application achieves the best results in all four core metrics: Recall, F1-score, IoU, and OA, at 89.15%, 90.57%, 82.93%, and 99.06%, respectively. Although IFN performs best in Precision (92.68%), its Recall value is relatively low, indicating potential missed detections. In contrast, our application maintains a near-optimal Precision of 91.61% while significantly improving Recall.

[0100] A comparison chart of change detection results from different models is shown below. Figure 8 As shown in the first, second, and fourth rows of Figure 1, the FC series models exhibit significant false positives and false negatives, especially for targets with subtle changes in class imbalance. Furthermore, for large, continuous buildings, the boundaries of changing regions are blurred and lack integrity. In contrast, the IHF-DPDNet proposed in this application effectively highlights key edge information and captures the continuity of building outlines through an interactive high-frequency enhancement module, while significantly reducing false positives and false negatives in areas of subtle change. As shown in the third row of Figure 3, thanks to the dual-branch mechanism and the fuzzy difference enhancement module's ability to model nonlinear regions of gradual changes, the IHF-DPDNet of this application suppresses blocky false alarms caused by noise. Simultaneously, for densely built areas, it effectively reduces the adhesion between buildings, enabling the model to maintain stable detection performance even in complex scenes.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A high-resolution remote sensing image building change detection network system, characterized in that: The network includes an encoder module, an interactive high-frequency enhancement module, a fuzzy-guided differential enhancement module, a hybrid attention module, a decoder module, and a prediction head. The encoder module employs a twin encoder to extract multi-scale features from the input raw dual-temporal high-resolution remote sensing image. The interactive high-frequency enhancement module is used to perform feature interaction and enhancement on the multi-scale dual-temporal features output by the encoder module. After the interactive high-frequency enhancement module, the network adopts a dual-path structure. The fuzzy-guided differential enhancement module is set on the differential branch of the dual path. By fusing standard difference and fuzzy difference and combining a dynamic weight fusion strategy, the fuzzy-guided differential enhancement module achieves refined modeling of multi-scale feature differences. The hybrid attention module is set on the channel cascade branch of the dual path, serially connecting the channel attention module and the spatial attention module to sequentially complete feature enhancement in the channel and spatial dimensions. The dual-temporal features output from the dual-path structure are then upsampled and fused by the decoder module to output a dual-temporal feature map of the same size as the original dual-temporal high-resolution remote sensing image. The dual-temporal feature map is then input into the prediction head to finally output the change map. The standard difference in the fuzzy-guided differential enhancement module uses a spatial normalization strategy to enhance local contrast. It performs maximum value normalization along the channel dimension and constrains the feature values ​​to the [0,1] interval to enhance the contrast of local areas. The fuzzy differential feature in the fuzzy-guided differential enhancement module uses the constructed basic differential features as the basic feature expression. Then, based on fuzzy set theory, a nonlinear transformation S-shaped membership function is constructed by introducing channel-adaptive learnable parameters to obtain fuzzy differential features. The basic differential features and fuzzy differential features are then concatenated by channels and input into the dynamic fusion sub-network. A weight generator constructed by depthwise separable convolution and 1×1 convolution predicts spatially adaptive fusion weights. Finally, the enhanced hybrid differential features are obtained through weighted fusion.

2. The high-resolution remote sensing image building change detection network system according to claim 1, characterized in that: The twin encoder of the encoder module uses a pre-trained CNN as the backbone network to extract multi-level features from dual-temporal high-resolution remote sensing images.

3. The high-resolution remote sensing image building change detection network system according to claim 1, characterized in that: The interactive high-frequency enhancement module maps the features of the dual-temporal images extracted by the encoder module through a 1×1 convolution, and obtains the query, key and value through reshaping, and then performs cross-attention mechanism operation to obtain the output after cross-attention calculation; The interactive high-frequency enhancement module also introduces an adaptive Sobel operator to extract high-frequency components of the feature map and obtain edge-related information.

4. The high-resolution remote sensing image building change detection network system according to claim 3, characterized in that: An isotropic Sobel operator with a minimum kernel size of 3 in eight directions is used to extract high-frequency components of each channel in the input feature map in eight different directions. In this process, a Top-k direction enhancement strategy is introduced to dynamically select the three directions with the most intense response in each channel and perform weighted fusion. Then, the feature map with high-frequency information enhancement generated by the adaptive Sobel operator is residually connected with the original input feature map. Finally, interactive high-frequency enhancement features with the same number of input channels are generated by 1×1 convolution.

5. A high-resolution remote sensing image building change detection network system according to claim 1, characterized in that: The channel attention module in the hybrid attention module adopts the frequency domain channel attention network FCANet, and introduces a two-dimensional discrete cosine transform into the frequency domain channel attention network FCANet. The two-dimensional discrete cosine transform realizes the diversity modeling of channel attention weights by performing multi-band component decomposition and selective fusion on the feature map, and finally transforms the input features into multi-spectral attention.

6. A high-resolution remote sensing image building change detection network system according to any one of claims 1-5, characterized in that: A combination strategy of binary cross-entropy loss and Dice loss is used to train the network and optimize its parameters.

7. A method for detecting building changes in high-resolution remote sensing imagery, characterized in that: Includes the following steps: S1: Preprocess the high-resolution remote sensing image change detection dataset; S2: Train the model constructed using the high-resolution remote sensing image building change detection network system as described in any one of claims 1-6 to obtain the optimal model; S3: Detect changes in buildings in high-resolution remote sensing images using a trained model and output a binary map to represent the changes.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method of claim 7.

9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that: When the computer program / instructions are executed by the processor, they implement the steps of the method of claim 7.

Images