A method for detecting changes in remote sensing images based on cross-scale guidance and enhancement

Abstract

This invention discloses a remote sensing image change detection method based on cross-scale guidance and enhancement. The method includes the following steps: extracting multi-level features from dual-temporal remote sensing images using a weight-shared dual encoder, and improving the perception and representation capabilities of multi-scale features using a high-order feature interaction module; improving the skip connection operation between the decoder and encoder, proposing a cross-scale guidance enhancement module to enhance features of interest and filter irrelevant background interference; and using a dual-temporal feature alignment and fusion module to fuse the change semantic information reconstructed by the multi-level decoder, extracting the change features of interest, and avoiding spurious changes caused by dual-temporal feature matching errors. This method utilizes low-level spatial information to guide and constrain the reconstruction process of deep semantic information, enhances features of interest, filters irrelevant background interference, and takes into account the registration error problem between multi-scale feature characteristics and dual-temporal features, thereby improving the accuracy of change detection.

Description

Technical Field

[0001] This invention belongs to the field of optical image processing and relates to an optical image change detection method, specifically a remote sensing image change detection method based on cross-scale guidance and enhancement dual encoder-decoder. Background Technology

[0002] High-resolution optical remote sensing images can be acquired using satellites, drones, and other detection platforms, offering advantages such as space-based / air-based observation perspectives and high revisitability. Change detection technology aims to identify differences in the state of surface objects or phenomena within the same geographical area at different times. Monitoring changes in surface objects of interest using dual-temporal remote sensing image change detection technology is a research hotspot in the field of remote sensing and has been widely applied in urban planning, agricultural pest and disease monitoring, and natural disaster assessment. Furthermore, it shows broad application prospects in emerging industries such as digital urban and rural planning, smart agricultural IoT, and natural resource utilization planning.

[0003] Deep learning has demonstrated outstanding performance in computer vision and has been introduced into the field of change detection in remote sensing images, attracting increasing attention. Deep learning-based change detection methods can be mainly categorized into single encoder-decoder (SED) methods, dual encoder-single decoder (DESD) methods, and dual encoder-decoder (DED) methods. High-resolution optical remote sensing in two time phases suffers from inherent differences in observation perspective, illumination intensity, and seasonal variations in ground features, leading to more diverse changes in multi-scale ground feature characteristics across different time phases. Furthermore, phenomena such as shadow changes and ground feature occlusion exist. The rich textural details in high-resolution remote sensing images also exacerbate the interference of irrelevant changes in extracting changes of interest. Feature fusion strategies in SED and DESD methods suffer from the loss of change information and inaccurate edge extraction of changed ground features due to the aliasing of target and background information. In contrast, DED-based change detection methods utilize a weight-sharing two-time phase encoder-decoder design to extract and reconstruct change information from both time phases separately, effectively filtering out irrelevant background interference. Therefore, DED-based change detection methods have become a research hotspot in this field.

[0004] Current change detection methods based on DED lack spatiotemporal guidance and constraints in the extraction of change information and do not fully explore the interrelationships between dual-temporal features. Therefore, when faced with complex backgrounds and multi-scale changing objects in remote sensing images, they are prone to missed detections and false alarms, making it difficult to meet the practical application needs of optical remote sensing data processing. Summary of the Invention

[0005] To address the aforementioned problems of current DED-based change detection methods, this invention, targeting the application needs of change detection in optical remote sensing images, introduces cross-scale guidance and enhancement strategies into the change detection method, thus proposing a remote sensing image change detection method based on cross-scale guidance and enhancement. This method utilizes low-level spatial information to guide and constrain the reconstruction process of deep semantic information, enhancing the change features of interest while filtering out irrelevant background interference. It also considers the registration error problem between multi-scale ground feature characteristics and dual-temporal features, thereby improving the accuracy of change detection.

[0006] The objective of this invention is achieved through the following technical solution:

[0007] A remote sensing image change detection method based on cross-scale guidance and enhancement includes the following steps:

[0008] Step 1: Use a weight-sharing dual encoder to extract multi-level features from dual-temporal remote sensing images, and utilize a high-order feature interaction module to improve the perception and representation capabilities of multi-scale features. The specific steps are as follows:

[0009] Step 11: Input the dual-temporal remote sensing images into a weight-shared dual encoder, extract the dual-temporal image information through a stacked 4-layer encoder module and map it to the same spatiotemporal feature space;

[0010] Steps 1 and 2: After the weight-sharing dual encoder, a high-order feature interaction module (HFIM) is introduced, with the input feature map F... in After a splitting operation, it is divided into 5 parts along the channel dimension, with one part remaining unchanged, denoted as F. ori For the other four, perform translation operations in the left, right, up, and down directions respectively, denoted as [F]. c1 ,F c2 ,F c3 ,F c4 After translation, zeros are filled into the empty positions and out-of-focus pixels are removed. The above operations are described as follows:

[0011] F ori =F[0:H,0:W,0:C id ]←F in [0:H,0:W,0:C id ]

[0012] F c1 =F[0:H,1:W,C id :C id +C g ]←F in [0:H,0:W-1,C id :C id +C g ]

[0013] F c2 =F[0:H,0:W-1,C id +C g :C id +2C g ]←F in [0:H,1:W,C id +C g :C id +2C g ]

[0014] F c3 =F[0:H-1,0:W,C id +2C g :C id +3C g ]←F in [1:H,0:W,C id +2C g :C id +3C g ]

[0015] F c4 =F[1:H,0:W,C id +3C g :C id +4C g ]←F in [0:H-1,0:W,C id +3C g :C id +4C g ]

[0016] Where F represents the translated feature map, and H, W, and C represent F respectively. in Height, width, and number of channels, C id C represents the number of channels that remain unchanged. g This represents the number of channels in the translation section, and C id +4×C g =C;

[0017] Step 13: After the grouping and translation operations, a convolutional layer with a 1×1 kernel and element-wise multiplication are used to integrate the information from the four translation operation groups. Multi-channel interaction is achieved by sequentially performing element-wise multiplication with the subsequent translation groups.

[0018] F c1 =Conv 1×1 (F c1 )

[0019]

[0020]

[0021]

[0022] Among them, Conv 1×1 (·) indicates a convolutional layer with a kernel size of 1×1. Indicates element-wise multiplication;

[0023] Step 14: Concatenate and convolve the multi-channel interactive features with the invariant feature parts to obtain the result of high-order feature interaction fusion:

[0024] F out =Conv 1×1 (Concat[F ori ,F c1 ,F c2 ,F c3 ,F c4 ])+F in

[0025] Wherein, Concat(·) represents series connection along the channel axis;

[0026] Step 2: Improve the skip connection operation between the decoder and encoder using the channel attention concept. Propose a cross-scale guided enhancement module to enhance features of interest and filter out irrelevant background interference. The specific steps are as follows:

[0027] Step 21: CGEM first processes the low-level feature map Perform global average pooling to compress it to 1×1×C. i A vector is used to characterize the importance of each channel in the feature map, where H is a vector. i =H0 / 2 i W i =W0 / 2 i C i =2 i+4 These represent the size of the corresponding feature map and the number of feature channels, respectively.

[0028] Step 22: In order to achieve deep feature map The reconstruction guides the F-axis through convolution operations. d(i+1) The number of channels is compressed to C i And multiply it pixel by pixel with the guiding feature vector;

[0029] Steps two and three: Process the deep feature map F d(i+1) Channel compression and upsampling operations are performed, and the upsampled feature map A is concatenated along the channel dimension to obtain a concatenated feature map that retains both spatial and semantic information.

[0030] Step 24: Use a 1×1 convolutional layer to compress the number of channels in the stitched feature map to C. i Output deep decoding feature map The above process can be expressed using mathematical formulas:

[0031]

[0032] Among them, f GAP (·) denotes adaptive average pooling, f FC (·) represents a fully connected layer, Up(·) represents upsampling, and Baseconv(·) represents a combination of convolutional layers, batch normalization layers, and non-linear activation functions;

[0033] Step 3: Use a series of dual-temporal feature alignment and fusion modules to fuse the change semantic information reconstructed by the multi-level decoder, extract the change features of interest, and avoid spurious changes caused by dual-temporal feature matching errors. The specific steps are as follows:

[0034] Step 31: Convert the feature maps output by the corresponding layers of the dual decoders. and The data is concatenated along the channel dimension, and then fed into a semantic flow field extraction subnetwork containing two convolutional layers with a kernel size of 5×5. The output of the subnetwork is the semantic flow field information. Mathematically, this can be expressed as:

[0035]

[0036] Among them, Conv flow (·) denotes the semantic flow field extraction subnetwork;

[0037] Step 32: After obtaining the flow field information, calculate Δf i By splitting the material evenly along the channel dimension, we obtain... and

[0038] Step 33: Use semantic flow field information to perform spatial position mutual correction on the two temporal features respectively, and use bilinear difference to obtain the corrected feature map;

[0039] Steps 3 and 4: Calculate the Euclidean distance between the two temporal feature maps and the other temporal corrected feature map, and then stitch the two distance maps together along the channel dimension;

[0040] Step 35: Obtain the alignment and fusion feature map using a forward propagation convolutional layer with a 1×1 kernel, represented as:

[0041]

[0042] Here, wrap(·) represents the operation of calculating the corrected feature map based on bilinear interpolation.

[0043] Compared with the prior art, the present invention has the following advantages:

[0044] (1) A dual encoder-decoder change detection network is proposed. By using cross-scale guided and enhanced decoders, the spatial information of low-level features guides the fine reconstruction process of deep semantic features, thereby improving the change detection accuracy of high-resolution optical remote sensing images.

[0045] (2) A high-order feature interaction module is introduced at the center of the backbone network. It uses splitting and translation operations to efficiently obtain hierarchical receptive fields, and then extracts and fuses multi-level features to enhance the feature extraction module’s ability to perceive multi-scale ground objects in complex scenes.

[0046] (3) A dual-temporal feature alignment and fusion module was designed. It perceives spatiotemporal differences through feature interaction and decouples the background pseudo-change interference problem based on optical flow to extract more accurate target change information. Attached Figure Description

[0047] Figure 1 This is a schematic diagram of a remote sensing image change detection process based on cross-scale guidance and enhancement.

[0048] Figure 2 Flowchart of the high-order feature interaction module;

[0049] Figure 3 A schematic diagram of the cross-scale guided enhanced decoding module;

[0050] Figure 4 This is a dual-temporal feature alignment and fusion module;

[0051] Figure 5 The WHU-CD dataset test visualization results are shown in the example (the gray area in the detection results graph represents false alarms or missed detections, and the white area represents correctly detected changes).

[0052] Figure 6 The LEVIR-CD dataset test results are shown in the example (the gray area in the detection results graph represents false alarms or missed detections, and the white area represents correctly detected changes). Detailed Implementation

[0053] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.

[0054] This invention provides a method for change detection in remote sensing images based on cross-scale guidance and enhancement. The method is grounded in the target change detection mechanism in optical remote sensing images and proposes a cross-scale guided dual encoder-decoder change detection network. This network proposes a cross-scale guided enhanced decoder to constrain the semantic information reconstruction process and enhances the feature representation of changed objects by guiding deep feature fusion. Simultaneously, to obtain representative features from complex scenes, the network designs an efficient high-order feature interaction module, utilizing multi-layered receptive fields to enhance the perception of multi-scale changed object features. Finally, the network constructs a dual-temporal feature alignment fusion module, which interactively perceives spatiotemporal differences, decoupling and filtering out interference from background pseudo-changes. Figure 1 As shown, the specific steps include the following:

[0055] Step 1: Use a weight-sharing dual encoder to extract multi-level features from dual-temporal remote sensing images, and utilize a high-order feature interaction module to improve the perception and representation capabilities of multi-scale features. The specific steps are as follows:

[0056] Step 11: Input the dual-temporal remote sensing images into a weight-shared dual encoder. A stacked 4-layer encoder module extracts the dual-temporal image information and maps it to the same spatiotemporal feature space. This facilitates the alignment and comparison of remote sensing images from different times and imaging conditions along the feature dimension, obtaining information on changes in features of interest. The encoder module adopts a ResNet-18 network structure. As the number of channels in the encoder's output feature map increases (64 / 128 / 256 / 512), the feature map size is downsampled sequentially to (H0 / 2, W0 / 2), (H0 / 4, W0 / 4), (H0 / 8, W0 / 8), and (H0 / 8, W0 / 8), where H0 and W0 represent the height and width of the input image, respectively.

[0057] Steps 1 and 2: After the weight-sharing dual encoder, a High-Order Feature Interactions Module (HFIM) is introduced. The structure of the HFIM is as follows: Figure 2 As shown, the input feature map F in After a splitting operation, it is divided into 5 parts along the channel dimension, with one part remaining unchanged, denoted as F. ori For the other four, perform translation operations in the left, right, up, and down directions respectively, denoted as [F]. c1 ,F c2 ,F c3 ,F c4 After translation, zeros are filled into the empty positions and out-of-focus pixels are removed. This operation can be described as follows:

[0058]

[0059] Where F represents the translated feature map, and H, W, and C represent F respectively. in Height, width, and number of channels, C id C represents the number of channels that remain unchanged. g This represents the number of channels in the translation section, and C id +4×C g =C.

[0060] Step 13: After the grouping and translation operations, a convolutional layer with a 1×1 kernel and element-wise multiplication are used to integrate the information from the four translation operation groups. Multi-channel interaction is achieved by sequentially performing element-wise multiplication with the subsequent translation groups.

[0061]

[0062] Among them, Conv 1×1 (·) indicates a convolutional layer with a kernel size of 1×1. This represents element-wise multiplication.

[0063] Step 14: Concatenate and convolve the multi-channel interactive features with the invariant feature parts to obtain the result of high-order feature interaction fusion:

[0064] F out =Conv 1×1 (Concat[F ori ,F c1 ,F c2 ,F c3 ,F c4 ])+F in (3)

[0065] Concat(·) represents series connection along the channel axis.

[0066] Step 2: Improve the skip connection operation between the decoder and encoder using the channel attention concept, and propose a cross-scale guided enhancement module to enhance features of interest and filter out irrelevant background interference.

[0067] Introducing the Cross-Scale Guided Enhancement Module (CGEM), such as... Figure 3 As shown, this method aims to obtain multi-scale target spatial domain enhancement features without edge aliasing. CGEM borrows the idea of ​​channel attention mechanism, utilizing more low-level features of ground objects of interest to guide the spatial reconstruction process of deep high-level semantic features, filtering out background noise in low-level features from interfering with target features.

[0068] CGEM first processes the low-level feature map Perform global average pooling to compress it to 1×1×C. i A vector is used to characterize the importance of each channel in the feature map, where H is a vector. i =H0 / 2 i W i =W0 / 2 i C i =2 i+4 These represent the size and number of feature channels of the corresponding feature map, respectively; in order to achieve deep feature map... The reconstruction guides the F-axis through convolution operations. d(i+1) The number of channels is compressed to C i And multiply it pixel-by-pixel with the guiding feature vector, and then multiply the deep feature map F. d(i+1) Channel compression and upsampling operations are performed, and the concatenated feature map A is then concatenated along the channel dimension to obtain a concatenated feature map that preserves both spatial and semantic information. Finally, a 1×1 convolutional layer is used to compress the number of channels in the concatenated feature map to C. i Output deep decoding feature map The above process can be expressed using mathematical formulas:

[0069]

[0070] Among them, f GAP (·) denotes adaptive average pooling, f FC (·) represents a fully connected layer, Up(·) represents upsampling, and Baseconv(·) represents a combination of convolutional layers, batch normalization layers, and non-linear activation functions.

[0071] Step 3: Use a series of dual-temporal feature alignment and fusion modules to fuse the change semantic information reconstructed by the multi-level decoder, extract the change features of interest, and avoid spurious changes caused by dual-temporal feature matching errors.

[0072] To avoid the influence of spurious changes caused by matching errors or side-look problems during the dual-temporal feature fusion process, this invention introduces a dual-temporal feature alignment fusion module (BAFM) based on optical flow, such as... Figure 4 As shown. First, the feature maps output by the corresponding layers of the dual decoders are... and The data is concatenated along the channel dimension, and then fed into a semantic flow field extraction subnetwork containing two convolutional layers with a kernel size of 5×5. The output of the subnetwork is the semantic flow field information. Mathematically, this can be expressed as:

[0073]

[0074] Among them, Conv flow (·) denotes the semantic flow field extraction subnetwork. After calculating the flow field information, Δf i By splitting the material evenly along the channel dimension, we obtain... and Then, semantic flow field information is used to perform spatial location mutual correction on the two temporal features, and the corrected feature maps are obtained using bilinear difference. Next, the Euclidean distance between each of the two temporal feature maps and the corrected feature map of the other temporal phase is calculated, and the two distance maps are concatenated along the channel dimension. Finally, a forward propagation convolutional layer with a 1×1 kernel is used to obtain an aligned and fused variation feature map, represented as:

[0075]

[0076] Here, wrap(·) represents the operation of calculating the corrected feature map based on bilinear interpolation.

[0077] Furthermore, the loss functions used in the training of the change detection model in this invention are all hybrid loss functions, namely the sum of the binary cross-entropy loss function and the dice loss function. Before applying the above-mentioned change detection steps to actual change detection tasks, the change detection model must be trained in advance using labeled training data. The specific training and verification processes can follow the training and verification methods in existing change detection models, and therefore will not be elaborated upon here.

[0078] The change detection method of the present invention will be applied to a specific embodiment below to demonstrate its technical effect.

[0079] Example

[0080] The specific network structure of the change detection model 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 can be divided into three stages: dataset preparation, model training and validation, and testing and visualization. The testing and visualization results are as follows. Figure 5 and Figure 6 As shown.

[0081] 1. Dataset Preparation

[0082] This embodiment uses the WHU-CD dataset and the LEVIR-CD dataset for remote sensing image change detection as examples.

[0083] The WHU-CD dataset contains over 10,000 buildings within a 20.5 square kilometer area. Each image in the dataset is 15354×32507 pixels in size, and the geographically corrected dataset has a resolution of 1.6 pixels. In this embodiment, each image pair is segmented into non-overlapping images of size 256×256 pixels, and further divided into training, validation, and test sets, containing 6096, 760, and 760 image pairs, respectively. The LEVIR-CD dataset images have a spatial resolution of 0.5 meters. The images in the dataset cover different regions and were captured at different times, including the effects of seasonal variations and changes in sunlight. In this embodiment, each image pair is segmented into non-overlapping images of size 256×256 pixels, and further divided into training, validation, and test sets, containing 7120, 1024, and 2048 image pairs, respectively.

[0084] 2. Model Training and Validation

[0085] 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.

[0086] 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 model of this invention using each batch of training samples, and validate and evaluate it after each training round. During training, calculate the mixed loss function for each training sample, which is the sum of the binary cross-entropy loss function and the dice loss function. Based on this, adjust the parameters of the network comprehensively 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, random horizontal and vertical flipping and rotation of the input image are performed during training to augment the data.

[0087] 3. Testing and Visualization

[0088] After the change detection model of the present invention is trained, it takes the dual-temporal images of the test dataset as its input and outputs a binary map of the change region of the dual-temporal image. The image size is the same as the input image size. The region with a pixel value of 1 represents the region that has changed, and the region with a pixel value of 0 represents the region that has not changed, thereby completing the change detection task.

[0089] This invention uses five of the most commonly used evaluation metrics to quantitatively evaluate the aforementioned remote sensing image change detection WHU-CD and LEVIR-CD datasets, namely F1-score (F1), recall (Rec), precision (Pre), intersection-over-union ratio (IoU), and overall precision (OA).

[0090] In this embodiment, some test visualization results are as follows: Figure 5 and Figure 6 As shown in Table 1, the quantitative results of the test are presented.

[0091] Table 1 Quantitative Test Results

[0092]

[0093] Combination Figure 5 , Figure 6 As shown in the quantitative results in Table 1, the change detection method of this invention can effectively perform change detection tasks on remote sensing images. It effectively integrates spatial information from low-level features during semantic reconstruction and introduces HFIM to improve the perception of multi-scale targets, thus exhibiting good performance even in complex scenes. Figure 5 The first line and Figure 6 As can be seen in the third line, the inconsistency between the illumination angle and observation angle in the preceding and following phases altered the shadows around the target, resulting in partial occlusion. However, the change detection method of this invention reduces the interference of these factors through dual-temporal feature alignment and fusion, thereby improving detection accuracy.

Claims

1. A method for detecting changes in remote sensing images based on cross-scale guidance and enhancement, characterized in that... The method includes the following steps: Step 1: Use a weight-sharing dual encoder to extract multi-level features from dual-temporal remote sensing images, and utilize a high-order feature interaction module to improve the perception and representation capabilities of multi-scale features. Step 2: Improve the skip connection operation between the decoder and encoder using the channel attention concept. Propose the cross-scale guided enhancement module CGEM to enhance features of interest and filter out irrelevant background interference. The specific steps are as follows: Step 21: CGEM first processes the low-level feature map Perform global average pooling to compress it to 1×1×C. i A vector is used to characterize the importance of each channel in the feature map, where H... i =H0 / 2 i W i =W0 / 2 i C i =2 i+4 These represent the size of the corresponding feature map and the number of feature channels, respectively. Step 22: In order to achieve deep feature map The reconstruction guides F through convolution operations. d(i+1) The number of channels is compressed to C i And multiply it pixel by pixel with the guiding feature vector; Steps two and three: Process the deep feature map F d(i+1) Channel compression and upsampling operations are performed, and the upsampled feature map A is concatenated along the channel dimension to obtain a concatenated feature map that retains both spatial and semantic information. Step 24: Use a 1×1 convolutional layer to compress the number of channels in the stitched feature map to C. i Output deep decoding feature map The above process can be expressed using mathematical formulas: Among them, f GAP (·) denotes adaptive average pooling, f FC (·) represents a fully connected layer, Up(·) represents upsampling, and Baseconv(·) represents a combination of convolutional layers, batch normalization layers, and non-linear activation functions; Step 3: Use a series of dual-temporal feature alignment and fusion modules to fuse the change semantic information reconstructed by the multi-level decoder, extract the change features of interest, and avoid spurious changes caused by dual-temporal feature matching errors. The specific steps are as follows: Step 31: Convert the feature maps output by the corresponding layers of the dual decoders. and The data is concatenated along the channel dimension, and then fed into a semantic flow field extraction subnetwork containing two convolutional layers with a kernel size of 5×5. The output of the subnetwork is the semantic flow field information. Mathematically, this can be expressed as: Among them, Conv flow (·) denotes the semantic flow field extraction subnetwork; Step 32: After obtaining the flow field information, calculate Δf i By splitting the material evenly along the channel dimension, we obtain... and Step 33: Use semantic flow field information to perform spatial position mutual correction on the two temporal features respectively, and use bilinear difference to obtain the corrected feature map; Steps 3 and 4: Calculate the Euclidean distance between the two temporal feature maps and the other temporal corrected feature map, and then stitch the two distance maps together along the channel dimension; Step 35: Obtain the alignment and fusion feature map using a forward propagation convolutional layer with a 1×1 kernel, represented as: Here, wrap(·) represents the operation of calculating the corrected feature map based on bilinear interpolation.

2. The remote sensing image change detection method based on cross-scale guidance and enhancement according to claim 1, characterized in that... The specific steps of step one are as follows: Step 11: Input the dual-temporal remote sensing images into a weight-shared dual encoder, extract the dual-temporal image information through a stacked 4-layer encoder module and map it to the same spatiotemporal feature space; Steps 1 and 2: After the weight-sharing dual encoder, a high-order feature interaction module (HFIM) is introduced, with the input feature map F... in After a splitting operation, it is divided into 5 parts along the channel dimension, with one part remaining unchanged, denoted as F. ori For the other four, perform translation operations in the left, right, up, and down directions respectively, denoted as [F]. c1 ,F c2 ,F c3 ,F c4 After translation, zeros are filled into the empty positions and out-of-focus pixels are removed. The above operations are described as follows: F ori =F[0:H,0:W,0:C id ]←F in [0:H,0:W,0:C id ] F c1 =F[0:H,1:W,C id :C id +C g ]←F in [0:H,0:W-1,C id :C id +C g ] F c2 =F[0:H,0:W-1,C id +C g :C id +2C g ]←F in [0:H,1:W,C id +C g :C id +2C g ] F c3 =F[0:H-1,0:W,C id +2C g :C id +3C g ]←F in [1:H,0:W,C id +2C g :C id +3C g ] F c4 =F[1:H,0:W,C id +3C g :C id +4C g ]←F in [0:H-1,0:W,C id +3C g :C id +4C g ] Where F represents the translated feature map, and H, W, and C represent F respectively. in Height, width, and number of channels, C id C represents the number of channels that remain unchanged. g This represents the number of channels in the translation section, and C id +4×C g =C; Step 13: After the grouping and translation operations, a convolutional layer with a 1×1 kernel and element-wise multiplication are used to integrate the information from the four translation operation groups. Multi-channel interaction is achieved by sequentially performing element-wise multiplication with the subsequent translation groups. F c1 =Conv 1×1 (F c1 ) Among them, Conv 1×1 (·) indicates a convolutional layer with a kernel size of 1×1. Indicates element-wise multiplication; Step 14: Concatenate and convolve the multi-channel interactive features with the invariant feature parts to obtain the result of high-order feature interaction fusion: F out =Conv 1×1 (Concat[F ori ,F c1 ,F c2 ,F c3 ,F c4 ])+F in Concat(·) represents series connection along the channel axis.

3. The remote sensing image change detection method based on cross-scale guidance and enhancement according to claim 2, characterized in that... The encoder module adopts a ResNet-18 network structure design. As the number of channels in the output feature map of the encoder increases to 64 / 128 / 256 / 512, the size of the feature map is downsampled to (H0 / 2,W0 / 2), (H0 / 4,W0 / 4), (H0 / 8,W0 / 8) and (H0 / 8,W0 / 8) respectively, where H0 and W0 represent the height and width of the input image, respectively.

Images