A Deep Learning-Based Method and System for Remote Sensing Image Change Detection
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-31
- Publication Date
- 2026-08-14
AI Technical Summary
[0004]本发明的目的在于提供一种基于深度学习的遥感图像变化检测方法及系统，以解决背景技术中现有变化检测在伪变化干扰、类别不平衡和复杂地物识别等方面存在局限的问题
[0052]1、本发明设置一种兼顾时间与空间信息的变化检测网络，该网络设计了差异信息引导模块，显式建模双时相图像特征间的差异，利用坐标注意力机制聚焦变化区域，有效提升了模型对时间变化的响应能力，该网络还设计了上下文特征融合模块，该模块结合自注意力机制与通道注意力机制，动态调节多尺度特征之间的交互方式，从而增强了语义耦合性与细节保持能力，还设置有轻量级多尺度融合解码器，在保持低参数量的同时，逐层整合深层语义信息，实现变化区域的还原与识别效果；
Smart Images
Figure CN122574459A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of remote sensing image change detection technology, specifically a remote sensing image change detection method and system based on deep learning. Background Technology
[0002] With the continuous development of satellite technology, the types of remote sensing images are constantly improving, and the resolution has also increased from medium resolution to high resolution, facilitating the provision of clearer and richer information about the Earth's surface. Remote sensing image change detection involves qualitative and quantitative analysis of remote sensing images of the same area from different periods to identify areas of change on the Earth's surface. This technology is widely used in fields such as land spatial planning, urban expansion monitoring, disaster assessment, environmental monitoring, and resource surveys. High-resolution remote sensing images, in particular, focus on key changes such as building growth and land cover changes. However, in practical applications, remote sensing image change detection still faces some challenges: such as imaging differences from different sensors, seasonal variations and shadow artifacts caused by inconsistencies in time and angle, and interference from complex features in high-resolution images. These factors all place higher demands on the accuracy of change detection.
[0003] With the rapid development of deep learning, it has been widely and effectively applied in the field of remote sensing image change detection. Based on this, this paper presents a deep learning-based method and system for remote sensing image change detection, which can eliminate the drawbacks of existing technologies and improve the limitations of traditional change detection in areas such as false change interference, class imbalance, and complex feature identification. Summary of the Invention
[0004] The purpose of this invention is to provide a method and system for detecting changes in remote sensing images based on deep learning, so as to solve the limitations of existing change detection methods in the background art in terms of false change interference, class imbalance and complex ground object recognition.
[0005] To achieve the above objectives, the present invention provides the following technical solution:
[0006] A method for detecting changes in remote sensing images based on deep learning, specifically including the following steps:
[0007] Step S1: Obtain remote sensing images of the same area at different times, and preprocess the remote sensing images to obtain a standardized input feature map;
[0008] Step S2: Construct a change detection network that takes into account both temporal and spatial information. The change detection network sequentially includes a feature extraction backbone network, a difference information guidance module, a context feature fusion module, and a lightweight multi-scale fusion decoder.
[0009] Step S3: Input the standardized input feature map into the feature extraction backbone network to extract multi-scale spatial semantic features;
[0010] Step S4: The difference information guidance module is used to perform difference modeling and enhancement on the multi-scale spatial semantic features to obtain the difference enhancement features of the focused change area;
[0011] Step S5: Perform cross-level semantic fusion and alignment on the differential enhancement features through the context feature fusion module to obtain a highly coupled fusion feature map;
[0012] Step S6: Input the fused feature map into the lightweight multi-scale fusion decoder, integrate deep semantic information layer by layer, classify changed and invariant regions, and output the change detection results of the remote sensing image.
[0013] Furthermore, the preprocessing in step S1 includes sequential image cropping, normalization, denoising, and size unification operations. The number of channels and resolution of the standardized input feature map match the input requirements of the feature extraction backbone network, and the texture, contour, and semantic information of the remote sensing image are preserved.
[0014] Furthermore, the feature extraction backbone network in step S2 is a pre-trained ResNet18 network.
[0015] Furthermore, the specific implementation process of the difference information guidance module in step S4 is as follows:
[0016] Step S41: Calculate the difference map of multi-scale spatial semantic features at different time phases;
[0017] Step S42: Extract salient regions of spatial and channel dimensions from the difference map using a coordinate attention mechanism to obtain attention weights;
[0018] Step S43: Reweight the difference map using attention weights to obtain difference enhancement features;
[0019] Step S44: Fuse the difference enhancement features with the multi-scale spatial semantic features element by element to form a discriminative feature representation that combines difference information and original features, and obtain the difference enhancement feature map.
[0020] Furthermore, the coordinate attention mechanism in step S42 includes:
[0021] Global average pooling is performed on the difference map along both the vertical and horizontal spatial dimensions to obtain feature tensors in two directions. The feature tensors are concatenated along the spatial dimensions and then fed into a 1×1 convolutional layer for channel fusion. After nonlinear activation and batch normalization, the difference map is divided into two tensors again, and attention weights in the height and width directions are generated by the Sigmoid function. The attention weights are then used to perform spatial augmentation on the input difference map.
[0022] Furthermore, the specific implementation process of the coordinate attention mechanism is as follows:
[0023] ;
[0024] ;
[0025] ;
[0026] ;
[0027] ;
[0028] ;
[0029] ;
[0030] ;
[0031] in, , Let be a tensor in the vertical and horizontal directions and have dimensions satisfying , , For channel index, For height direction index, The width of the feature map. Indexed in the width direction. The input is the raw image data. For feature map height, For the concatenated feature tensor, For splicing, To share the transformed feature tensor, It is a non-linear activation function. For batch normalization, Represented as a 1×1 convolution, and These are attention in the height direction and attention in the width direction, respectively. For tensor splitting operations, Let the attention weight tensor be in the vertical direction. For the Sigmoid function, Let the attention weight tensor be in the horizontal direction. Spatial augmentation difference map output by the coordinate attention mechanism.
[0032] Furthermore, the specific implementation process of the context feature fusion module in step S5 is as follows:
[0033] Step S51: Upsample the deep features in the differential enhancement features to match the spatial scale of the shallow features;
[0034] Step S52: Perform convolution operations on the deep and shallow features after scale matching to achieve cross-level semantic alignment;
[0035] Step S53: Introduce a self-attention mechanism, construct a context association graph of semantically aligned features, generate cross-positional attention weights, and aggregate information from the features;
[0036] Step S54: The context features and shallow features after information aggregation are concatenated along the channel dimension, and after being enhanced by the channel attention mechanism, a highly coupled fusion feature map is obtained.
[0037] Furthermore, the self-attention mechanism in step S53 maps the input features into three feature matrices—query, key, and value—through linear transformation, calculates the similarity matrix between the query and the key using the Softmax function, performs element-wise multiplication of the similarity matrix with the value to achieve information aggregation, and finally outputs the aggregated context features through linear transformation.
[0038] The specific implementation process of the self-attention mechanism is as follows:
[0039] , , ;
[0040] ;
[0041] ;
[0042] ;
[0043] in, To query the feature matrix, The key feature matrix, The eigenvalue matrix, , , All are learnable weight matrices and satisfy... , Input features and dimensions satisfy , For the number of channels, For feature map height, The width of the feature map. This is the attention weight matrix. For normalized exponential functions, To query the transpose of the feature matrix, The mapped dimension, For the sequence length to satisfy , For contextual representation, This is the transpose of the attention weight matrix. The final output features of the self-attention mechanism, To output the mapping matrix and satisfy , Reshaping dimensions.
[0044] Furthermore, the specific implementation process of the lightweight multi-scale fusion decoder in step S6 is as follows:
[0045] The features at different scales in the input fused feature map are upsampled by bilinear interpolation. Then, the feature maps of different layers are concatenated along the channel dimension, and 3×3 convolution is used for dimensionality reduction. Finally, a 1×1 convolutional layer is used to classify the feature map into changed regions and invariant regions, thus obtaining the final change detection result of the remote sensing image.
[0046] A deep learning-based remote sensing image change detection system, used to execute a deep learning-based remote sensing image change detection method, includes:
[0047] The image preprocessing module is used to receive remote sensing images of the same area at different times and sequentially perform image cropping, normalization, denoising and size unification operations, and output a standardized input feature map.
[0048] The change detection module has a built-in change detection network and computational logic. It receives standardized input feature maps, performs dual-temporal feature extraction, difference enhancement, multi-scale semantic fusion and classification, and outputs the change detection results of remote sensing images.
[0049] The post-processing module is used to optimize the detection results and generate several visual schematic diagrams of the detection results.
[0050] The network training module is used to construct the dataset and perform joint training and parameter optimization on the change detection network.
[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0052] 1. This invention establishes a change detection network that takes into account both temporal and spatial information. This network is designed with a difference information guidance module, which explicitly models the differences between features in dual-temporal images and uses a coordinate attention mechanism to focus on the changing region, effectively improving the model's response to temporal changes. The network is also designed with a context feature fusion module, which combines self-attention and channel attention mechanisms to dynamically adjust the interaction between multi-scale features, thereby enhancing semantic coupling and detail preservation capabilities. It is also equipped with a lightweight multi-scale fusion decoder, which integrates deep semantic information layer by layer while maintaining a low number of parameters, achieving the restoration and recognition of changing regions.
[0053] 2. This invention provides a remote sensing image change detection method based on deep learning. By combining deep learning with difference guidance and contextual features, it enhances the ability to identify fine-grained semantic differences, solves common problems such as false differences and blurred boundaries in practical change detection tasks, enhances the response of change areas in dual-temporal images, and extracts deep correlations between multi-scale semantic features, thereby effectively addressing the challenges of complex backgrounds, diverse land features, and false changes in remote sensing scenes. Attached Figure Description
[0054] Figure 1 This is a schematic diagram of the method steps of the present invention.
[0055] Figure 2 This is a schematic diagram of the change detection network of the present invention.
[0056] Figure 3 This is a schematic diagram of the method flow of the present invention.
[0057] Figure 4 This is a schematic diagram of the coordinate attention mechanism of the present invention.
[0058] Figure 5 This is a schematic diagram of the structure of the difference information guidance fusion module of the present invention.
[0059] Figure 6 This is a schematic diagram of the context feature fusion module of the present invention.
[0060] Figure 7 This is a schematic diagram of the system structure of the present invention.
[0061] Figure 8 This is a visual comparison diagram of different methods on the LEVIR-CD dataset.
[0062] Figure 9 This is a visual comparison diagram of different methods on the SYSU-CD dataset.
[0063] Figure 10 This is a visual comparison diagram of different methods on the WHU-CD dataset.
[0064] Figure 11 This is a visual comparison diagram of ablation experiments of different modules in this invention.
[0065] Figure labeling: Feature extraction backbone network 10, difference information guidance module 20, context feature fusion module 30, lightweight multi-scale fusion decoder 40, image preprocessing module 100, change detection module 200, detection postprocessing module 300, network training module 400. Detailed Implementation
[0066] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.
[0067] Example 1
[0068] In this embodiment, as Figures 1-6 As shown, this invention provides a method for detecting changes in remote sensing images based on deep learning, specifically including the following steps:
[0069] Step S1: Acquire remote sensing images of the same area at different times, and preprocess the remote sensing images to obtain standardized input feature maps;
[0070] Specifically, the preprocessing includes sequential image cropping, normalization, denoising, and size unification operations. The number of channels and resolution of the input feature map are standardized to match the input requirements of the feature extraction backbone network 10, while preserving the texture, contour, and semantic information of the land cover in the remote sensing image.
[0071] Step S2: Construct a change detection network (DGCFF-Net) that takes into account both temporal and spatial information. The change detection network sequentially includes a feature extraction backbone network 10, a difference-guided fusion (DGF) module 20, a contextual feature fusion (CFF) module 30, and a lightweight multi-scale fusion decoder 40. The overall architecture diagram of this change detection network (DGCFF-Net) is shown below. Figure 3As shown, RestBlock is the residual block, Input is the input, Conv3x3→BN→ReLU represents the process of 3×3 convolution → batch normalization → linear activation, Downsample is downsampling, Layer1 / Layer2 / Layer3 / Layer4 represent the four feature extraction layers of the ResNet18 network, BasicBlock is the basic block, Output is the output, Element-wise Addition means element-wise addition, CAM means channel attention mechanism, F1 / F2 / F3 / F4 means features 1 / 2 / 3 / 4, corresponding to different scale feature maps output by Layer1-Layer4, Shareweight is shared weight, Coordinate Attention Module is coordinate attention mechanism, Upsample is upsampling, and Concatenation is concatenation.
[0072] Specifically, the feature extraction backbone network 10 is a pre-trained ResNet18 network. For example, the ResNet18 network contains four feature extraction layers, which output multi-scale spatial semantic features with resolutions of 1 / 4, 1 / 8, 1 / 16, and 1 / 32, respectively. The features of each layer share weights and suppress gradient vanishing through residual connections.
[0073] Step S3: Input the standardized input feature map into the feature extraction backbone network 10 to extract multi-scale spatial semantic features;
[0074] Step S4: Perform differential modeling and enhancement on multi-scale spatial semantic features through the differential information guidance module 20 to obtain differential enhancement features focusing on the change region;
[0075] Specifically, this difference guidance is used to improve the network model's responsiveness to changing regions and suppress irrelevant background interference. To avoid instability caused by using difference information alone, this module fuses the difference enhancement features with the original features to form a more discriminative representation of the changes, such as... Figure 5 As shown, Up represents upsampling, and the specific implementation process of the difference information guidance module 20 is as follows:
[0076] Step S41: Calculate the difference map (residual) of multi-scale spatial semantic features at different time phases;
[0077] Step S42: Extract salient regions of spatial and channel dimensions from the difference map using a coordinate attention mechanism to obtain attention weights;
[0078] Step S43: Reweight the difference map using attention weights to obtain difference enhancement features;
[0079] Step S44: Fuse the difference enhancement features with the multi-scale spatial semantic features element by element to form a discriminative feature representation that combines difference information and original features, and obtain the difference enhancement feature map;
[0080] Specifically, such as Figure 4 As shown, Element-wise Subtraction represents element-wise subtraction, and Element-wise Multiplication represents element-wise multiplication. The coordinate attention mechanism in step S42 includes:
[0081] Coordinate information embedding requires global average pooling of the difference map along both the vertical and horizontal spatial dimensions to obtain feature tensors in two directions (horizontal and vertical). and ):
[0082] ;
[0083] ;
[0084] The feature tensors are concatenated along the spatial dimension and then fed into a 1×1 convolutional layer for channel fusion. After nonlinear activation and batch normalization:
[0085] ;
[0086] ;
[0087] It is then split into two tensors ( and Attention weights in the height and width directions are generated using the Sigmoid function, and these weights are then used to spatially augment the input difference map.
[0088] ;
[0089] ;
[0090] ;
[0091] ;
[0092] in, , Let be a tensor in the vertical and horizontal directions and have dimensions satisfying , , For channel index, For height direction index, The width of the feature map. Indexed in the width direction. The input is the raw image data. For feature map height, For the concatenated feature tensor, For splicing, To share the transformed feature tensor, It is a non-linear activation function. For batch normalization, Represented as a 1×1 convolution, and These are attention in the height direction and attention in the width direction, respectively. For tensor splitting operations, Let the attention weight tensor be in the vertical direction. For the Sigmoid function, Let the attention weight tensor be in the horizontal direction. The spatial enhancement difference map output by the coordinate attention mechanism is obtained by encoding the spatial information of the feature map along the horizontal and vertical directions respectively. After the coarse difference map is processed by coordinate attention, the response of the change region is enhanced, while the background interference is suppressed, and the change region can be focused more accurately.
[0093] Step S5: Perform cross-level semantic fusion and alignment on the differential enhancement features through the context feature fusion module 30 to obtain a highly coupled fusion feature map;
[0094] Specifically, the context feature fusion module 30 is used to perform guided fusion of cross-level features during the network encoding stage, such as... Figure 6 As shown, CA is the channel attention mechanism, and Liner is the linear transformation. The specific implementation process is as follows:
[0095] Step S51: Upsample the deep features in the differential enhancement features to match the spatial scale of the shallow features;
[0096] Step S52: Perform convolution operations on the deep and shallow features after scale matching to achieve cross-level semantic alignment;
[0097] Step S53: Introduce a self-attention mechanism to construct a context association graph of semantically aligned features, generate cross-positional attention weights, and aggregate information of features to enhance semantic consistency across scales.
[0098] Step S54: The context features and shallow features after information aggregation are concatenated along the channel dimension. After being enhanced by the channel attention mechanism, a highly coupled fusion feature map is obtained to further highlight key change areas and suppress background interference.
[0099] The self-attention mechanism maps input features into three feature matrices—query, key, and value—through a linear transformation. It then uses the Softmax function to calculate the similarity matrix between the query and the key, and finally compares the similarity matrix with the value... ) Perform element-wise multiplication to aggregate information, and finally output the aggregated context features through linear transformation;
[0100] The basic idea of the self-attention mechanism is to dynamically generate attention weights through the relationship between queries, keys, and values, thereby achieving cross-location information aggregation. The specific implementation process is as follows:
[0101] , , ;
[0102] ;
[0103] ;
[0104] ;
[0105] in, To query the feature matrix, The key feature matrix, The eigenvalue matrix, , , All are learnable weight matrices and satisfy... , Input features and dimensions satisfy , For the number of channels, For feature map height, The width of the feature map. This is the attention weight matrix. For normalized exponential functions, To query the transpose of the feature matrix, The mapped dimension, For the sequence length to satisfy , For contextual representation, This is the transpose of the attention weight matrix. The final output features of the self-attention mechanism, To output the mapping matrix and satisfy , Reshaping dimensions;
[0106] Step S6: Input the fused feature map into the lightweight multi-scale fusion decoder 40, integrate deep semantic information layer by layer, classify changed and invariant regions, and output the change detection results of the remote sensing image.
[0107] Specifically, the implementation process of the lightweight multi-scale fusion decoder 40 is as follows:
[0108] The features at different scales in the input fused feature map are upsampled by bilinear interpolation. Then, the feature maps of different layers are concatenated along the channel dimension, and 3×3 convolution is used for dimensionality reduction. Finally, a 1×1 convolutional layer is used to classify the feature map into changed regions and invariant regions, thus obtaining the final change detection result of the remote sensing image.
[0109] Example 2
[0110] Among them, such as Figure 7 As shown, the present invention also provides a deep learning-based remote sensing image change detection system for executing the deep learning-based remote sensing image change detection method shown in Embodiment 1 above, comprising:
[0111] The image preprocessing module 100 is used to receive remote sensing images of the same area at different times and sequentially perform image cropping, normalization, denoising and size unification operations, output a standardized input feature map, output a standardized input feature map that is completely matched with the input parameters (number of channels, resolution) of the DGCFF-Net change detection network, and retain the texture, contour and semantic core information of the remote sensing images, providing qualified input data;
[0112] The change detection module 200 has a built-in change detection network and computational logic. It is used to receive standardized input feature maps, complete dual-temporal feature extraction, difference enhancement, multi-scale semantic fusion and classification, and output the change detection results of remote sensing images. It integrates a feature extraction backbone network 10, a difference information guidance module 20, a context feature fusion module 30 and a lightweight multi-scale fusion decoder 40. All computational logic related to feature extraction, difference modeling, feature fusion and change detection is completed within this module.
[0113] The post-processing module 300 is used to post-process and optimize the detection results, generate several visualization diagrams of the detection results, receive the detection results of the DGCFF-Net change detection network, and can overlay visualization based on the original remote sensing image to generate a comparison map with highlighted change areas, meeting the needs of visualization and data reuse in practical applications.
[0114] The network training module 400 is used to build the dataset and perform joint training and parameter optimization of the change detection network;
[0115] In this embodiment, the above-mentioned change detection method is fully implemented through the collaborative work of the image preprocessing module 100, the change detection module 200, the detection postprocessing module 300, and the network training module 400.
[0116] To further demonstrate the practical effect of the deep learning-based remote sensing image change detection method in this invention, a visual representation of the deep learning-based remote sensing image change detection scheme is provided, as shown below. Figures 8-11 As shown;
[0117] Figures 8-10 This diagram illustrates the training results of different models. Three common datasets were used, each divided into training, testing, and validation portions according to a specific ratio. Model training was performed using these three datasets to obtain the model's accuracy in detecting changed regions. Figures 8-10 In the diagram, columns T1 and T2 represent remote sensing images from different time phases, column GT represents the ground condition, column a represents FC-EF, column b represents FC-siam-conc, column c represents SNUNet, column d represents BIT, column e represents ChangeFormer, column f represents A2Net, column g represents DMINet, column h represents SEIFNet, column i represents DGCFF-Net, and column ai represents different network models. Column i represents the change detection network in this invention.
[0118] Figure 8 This paper presents a visual comparison between the proposed method and the state-of-the-art (SOTA) method on the LEVIR-CD dataset. The background of this dataset consists of simple objects such as roads and vegetation. Based on the images, it can be seen that DGCFF-Net has no missed detections of small-area building changes, while some models (such as FC-EF and A2Net) have obvious missed detections. The building boundaries of DGCFF-Net are clear and unblurred, while the boundaries of some models (such as FC-siam-conc and DMINet) have jagged distortion. DGCFF-Net has no obvious false detections of the background and does not identify invariant areas such as roads and vegetation as changes, while some models (such as BIT) have a small number of false detections of the background.
[0119] Figure 9 This paper presents a visual comparison between the proposed method and the state-of-the-art (SOTA) method on the SYSU-CD dataset, which is a complex scene with multiple objects. Based on the images, it can be seen that DGCFF-Net failed to detect false changes caused by shadow and lighting differences as real changes, while some models (such as SNUNet and ChangeFormer) showed obvious false changes and cluttered backgrounds. DGCFF-Net clearly classified and detected mixed changes of multiple ground objects, such as mixed change areas of buildings and vegetation, and could accurately distinguish and completely detect them, while some models (such as SEIFNet) missed ground object changes.
[0120] Figure 10 This paper presents a visual comparison between the proposed method and state-of-the-art (SOTA) methods on the WHU-CD dataset, which contains large-scale complex scenes with wide coverage. Based on the images, it can be seen that DGCFF-Net has no missed detections of large-scale dispersed change regions, such as small buildings and land cover changes in remote areas, while some models (such as FC-EF and DMINet) have missed detections of dispersed regions. DGCFF-Net has the best boundary fitting for change regions, especially for irregularly shaped change regions (such as curved water bodies and irregular buildings), while some models (such as BIT and SEIFNet) have obvious boundary offsets.
[0121] The dual-temporal remote sensing images are input into the corresponding network model. In this invention, the pre-trained RestNet18 is used as the basic skeleton. The images of different temporal phases are processed into different scales and spatial features of different scales are extracted. Then, the difference information guidance module 20, the context feature fusion module 30 and the lightweight multi-scale fusion decoder 40 are used to perform the change region detection operation. The experimental results obtained by the change detection network in this invention are compared with the results obtained by other models under the same experimental conditions and using the same dataset. The comparison results are shown in Table 1.
[0122] Table 1 - Schematic diagram of quantitative comparison of experimental results
[0123]
[0124] Table 1 shows the results of different network models trained on the same dataset under the same experimental conditions. The best results are indicated in bold. LEVIR-CD, SYSU-CD, and WHU-CD are the datasets. FLOPS(G) is the number of floating-point operations per second, Params(M) is the number of model parameters, OA is the overall precision, F1 is the F1 score, IoU is the intersection-over-union ratio, Rec is the recall rate, Pre is the precision rate, FC-EF, FC-siam-conc, SNUNet, BIT, Changeformer, A2Net, DMINet, and SEIFNet are eight mainstream remote sensing change detection models, and DGCFF-Net is the network model used in this invention. Based on Table 1, the network model shows that compared to lightweight models (such as FC-EF, FC-siam-conc, and A2Net), DGCFF-Net has slightly higher parameter and computational costs, but achieves significantly higher accuracy (e.g., LEVIR-CD's F1 score improved from 87.99% to 91.09%). Compared to high-precision heavy models (such as SNUNet, BIT, and ChangeFormer), DGCFF-Net has lower parameter and computational costs (e.g., SNUNet has 12.035M parameters but achieves 54.83G FLOPS), but achieves the best accuracy in the main metrics. The final conclusion is that DGCFF-Net balances lightweight design with low computational cost, achieving a balance between accuracy and lightweight design.
[0125] By combining different modules in the change detection network of this invention, and comparing the results obtained by different groups trained on the same dataset under the same experimental conditions, a visual diagram is obtained as shown below. Figure 11 As shown in Table 2, the performance comparison results are obtained. Figure 11 In the image, columns T1 and T2 represent remote sensing images from different time phases, column GT represents the ground reality, column a represents BasiclineLine (basic model, retaining only the feature extraction backbone network 10 and the lightweight multi-scale fusion decoder 40), column b represents BasiclineLine+DGF (basic model + differential information guidance module 20), and column c represents BasiclineLine+DGF+CFF (basic model + differential information guidance module 20 + context feature fusion module 30, i.e., DGCFF-Net).
[0126] Table 2 - Schematic diagram of ablation experimental results for different modules on the LEVIR-CD dataset
[0127]
[0128] Table 2 shows the results of ablation experiments on different modules, verifying the roles of the differential information guidance module 20 and the context feature fusion module 30. The ablation experiments were used to remove model modules individually or in combination. According to Table 2, after adding DGF, F1 improved from 0.9029 to 0.9069, and IoU improved from 0.8229 to 0.8296, proving that DGF can effectively model the differences in features across two time periods, focus on changing regions, and improve the model's detection capabilities. After adding CFF on top of DGF, all indicators improved, proving that CFF can effectively fuse multi-scale context features, enhance semantic coupling, and further improve detection accuracy and boundary accuracy. Therefore, the final conclusion is that both DGF and CFF modules can improve model performance, and the combined use is the most effective.
[0129] In summary, the remote sensing image change detection method and system based on deep learning proposed in this invention effectively improves the accuracy and robustness of change detection by designing a difference information guidance module 20 and a context feature fusion module 30. Furthermore, experimental results on multiple public datasets demonstrate that this invention has good application prospects.
[0130] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for detecting changes in remote sensing images based on deep learning, characterized in that, Specifically, the following steps are included: Step S1: Obtain remote sensing images of the same area at different times, and preprocess the remote sensing images to obtain a standardized input feature map; Step S2: Construct a change detection network that takes into account both temporal and spatial information. The change detection network sequentially includes a feature extraction backbone network, a difference information guidance module, a context feature fusion module, and a lightweight multi-scale fusion decoder. Step S3: Input the standardized input feature map into the feature extraction backbone network to extract multi-scale spatial semantic features; Step S4: The difference information guidance module is used to perform difference modeling and enhancement on the multi-scale spatial semantic features to obtain the difference enhancement features of the focused change area; Step S5: Perform cross-level semantic fusion and alignment on the differential enhancement features through the context feature fusion module to obtain a highly coupled fusion feature map; Step S6: Input the fused feature map into the lightweight multi-scale fusion decoder, integrate deep semantic information layer by layer, classify changed and invariant regions, and output the change detection results of the remote sensing image.
2. The remote sensing image change detection method and system based on deep learning according to claim 1, characterized in that, The preprocessing in step S1 includes sequential image cropping, normalization, denoising, and size unification operations. The number of channels and resolution of the standardized input feature map match the input requirements of the feature extraction backbone network, and the texture, contour, and semantic information of the remote sensing image are preserved.
3. The remote sensing image change detection method and system based on deep learning according to claim 1, characterized in that, The feature extraction backbone network in step S2 is a pre-trained ResNet18 network.
4. The remote sensing image change detection method and system based on deep learning according to claim 1, characterized in that, The specific implementation process of the difference information guidance module in step S4 is as follows: Step S41: Calculate the difference map of multi-scale spatial semantic features at different time phases; Step S42: Extract salient regions of spatial and channel dimensions from the difference map using a coordinate attention mechanism to obtain attention weights; Step S43: Reweight the difference map using attention weights to obtain difference enhancement features; Step S44: Fuse the difference enhancement features with the multi-scale spatial semantic features element by element to form a discriminative feature representation that combines difference information and original features, and obtain the difference enhancement feature map.
5. The remote sensing image change detection method and system based on deep learning according to claim 4, characterized in that, The coordinate attention mechanism in step S42 includes: Global average pooling is performed on the difference map along both the vertical and horizontal spatial dimensions to obtain feature tensors in two directions. The feature tensors are concatenated along the spatial dimensions and then fed into a 1×1 convolutional layer for channel fusion. After nonlinear activation and batch normalization, the difference map is divided into two tensors again, and attention weights in the height and width directions are generated by the Sigmoid function. The attention weights are then used to perform spatial augmentation on the input difference map.
6. The remote sensing image change detection method and system based on deep learning according to claim 5, characterized in that, The specific implementation process of the coordinate attention mechanism is as follows: ； ； ； ； ； ； ； ； in, , Let be a tensor in the vertical and horizontal directions and have dimensions satisfying , , For channel indexing, For height direction index, The width of the feature map. Indexed in the width direction. The input is the raw image data. For feature map height, For the concatenated feature tensor, For splicing, To share the transformed feature tensor, It is a non-linear activation function. For batch normalization, Represented as a 1×1 convolution, and These are attention in the height direction and attention in the width direction, respectively. For tensor splitting operations, Let the attention weight tensor be in the vertical direction. For the Sigmoid function, Let the attention weight tensor be in the horizontal direction. Spatial augmentation difference map output by the coordinate attention mechanism.
7. The remote sensing image change detection method and system based on deep learning according to claim 1, characterized in that, The specific implementation process of the context feature fusion module in step S5 is as follows: Step S51: Upsample the deep features in the differential enhancement features to match the spatial scale of the shallow features; Step S52: Perform convolution operations on the deep and shallow features after scale matching to achieve cross-level semantic alignment; Step S53: Introduce a self-attention mechanism, construct a context association graph of semantically aligned features, generate cross-positional attention weights, and aggregate information from the features; Step S54: The context features and shallow features after information aggregation are concatenated along the channel dimension, and after being enhanced by the channel attention mechanism, a highly coupled fusion feature map is obtained.
8. The remote sensing image change detection method and system based on deep learning according to claim 7, characterized in that, The self-attention mechanism in step S53 maps the input features into three feature matrices—query, key, and value—through linear transformation. It then uses the Softmax function to calculate the similarity matrix between the query and the key, and performs element-wise multiplication of the similarity matrix with the value to achieve information aggregation. Finally, it outputs the aggregated context features through linear transformation. The specific implementation process of the self-attention mechanism is as follows: ， ， ； ； ； ； in, To query the feature matrix, The key feature matrix, The eigenvalue matrix, , , All are learnable weight matrices and satisfy... , Input features and dimensions satisfy , For the number of channels, For feature map height, The width of the feature map. This is the attention weight matrix. For normalized exponential functions, To query the transpose of the feature matrix, The mapped dimension, For the sequence length to satisfy , For contextual representation, This is the transpose of the attention weight matrix. The final output features of the self-attention mechanism, To output the mapping matrix and satisfy , Reshaping dimensions.
9. The remote sensing image change detection method and system based on deep learning according to claim 1, characterized in that, The specific implementation process of the lightweight multi-scale fusion decoder in step S6 is as follows: The features at different scales in the input fused feature map are upsampled by bilinear interpolation. Then, the feature maps of different layers are concatenated along the channel dimension, and 3×3 convolution is used for dimensionality reduction. Finally, a 1×1 convolutional layer is used to classify the feature map into changed regions and invariant regions, thus obtaining the final change detection result of the remote sensing image.
10. A deep learning-based remote sensing image change detection system, used to execute the deep learning-based remote sensing image change detection method according to any one of claims 1-9, characterized in that, include: The image preprocessing module is used to receive remote sensing images of the same area at different times and sequentially perform image cropping, normalization, denoising and size unification operations, and output a standardized input feature map. The change detection module has a built-in change detection network and computational logic. It receives standardized input feature maps, performs dual-temporal feature extraction, difference enhancement, multi-scale semantic fusion and classification, and outputs the change detection results of remote sensing images. The post-processing module is used to optimize the detection results and generate several visual schematic diagrams of the detection results. The network training module is used to construct the dataset and perform joint training and parameter optimization on the change detection network.