A lightweight remote sensing image change detection method under spatial misalignment

Abstract

The application discloses a lightweight remote sensing image change detection method under spatial misalignment. The application comprises the following steps: firstly, a MobileNet V3 Large is used to extract multi-scale features for registration and change detection; secondly, a spatial consistency module is used to realize semi-dense feature point matching, and a cross-scale spatial transformation model is established to align different scale feature maps with each other; then, different scale feature maps pass through a space-time difference coordination module to enhance the space-time heterogeneity of double-time-phase features; finally, multi-scale difference features are fused to generate a change detection result. The application selects SVCD, SYSU-CD and SECOND data sets for experiments, and compares with current mainstream change detection networks. The results show that the application can effectively construct a spatial transformation relationship between images to be registered, is significantly better than other methods in quantitative analysis and qualitative analysis, and has certain advantages in network complexity.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing, and specifically relates to a lightweight method for detecting changes in remote sensing images under spatial misalignment. Background Technology

[0002] Remote sensing image registration and change detection are key tasks in multi-temporal analysis. Remote sensing image registration involves pixel-by-pixel spatial alignment of images of the same area from different imaging mechanisms, sensors, time periods, and perspectives. It is a crucial step in downstream tasks such as multi-temporal image change detection, image fusion, and image analysis. Change detection analyzes remote sensing images of a specific area from different time periods to identify and quantify surface changes in the same area. It has wide applications in fields such as natural resource monitoring, ecological environment monitoring, and national defense and security.

[0003] In the field of traditional remote sensing image registration, scholars have proposed numerous methods, mainly categorized into region-based, feature-based, and learning-based methods. Region-based methods, relying on similarity measures such as mutual information and cross-correlation, directly construct the optimal geometric transformation model by comparing shallow image information, achieving pixel-level alignment between images to be registered. However, region-based methods are sensitive to nonlinear radiometric (intensity) differences, and the low-level representation of images suffers from severe feature redundancy, posing significant challenges to constructing robust feature detection and description models between images. Feature-based registration involves four steps: feature extraction, feature description, feature matching, and spatial geometric transformation. It utilizes significant geometric features such as points, lines, and surfaces in the image, estimating geometric transformation parameters based on local correspondences between features, as exemplified by representative methods like SIFT, ORB, and SURF. However, due to accumulated computational errors and the heterogeneity of local image deformations (such as terrain undulations and sensor distortion), feature-based methods struggle to accurately represent nonlinear pixel-level deformations, leading to distorted modeling of pixel spatial correspondences.

[0004] Thanks to the rapid development of artificial intelligence technology, end-to-end remote sensing image registration based on deep learning has shown great potential. Simultaneously, numerous change detection methods based on deep learning have emerged, achieving significant breakthroughs in detection accuracy and robustness in complex change scenarios compared to traditional methods. However, with the rapid development of Earth observation technology, the heterogeneity of bi-temporal images under different sensors and imaging conditions leads to differences in the geometric morphology and spectral response of the same ground feature in bi-temporal images. Therefore, how to effectively construct substantial spatial-spectral differences between bi-temporal images has received continuous attention from researchers. Although existing methods have made significant progress in both remote sensing image registration and change detection, they are usually treated as two independent tasks, failing to effectively construct a unified framework for registration and change detection. Furthermore, existing change detection methods lack the synergistic interaction between difference feature space and semantic information, making it difficult to effectively construct substantial spatial-spectral differences between bi-temporal images. Summary of the Invention

[0005] Current deep learning methods typically treat registration and change detection as independent tasks, failing to unify their processing. Furthermore, existing change detection methods lack the collaborative interaction between spatial and semantic information of differential features, making it difficult to effectively construct substantial spatial-spectral differences between two temporal images. Therefore, this invention proposes a lightweight network that jointly performs remote sensing image registration and change detection. First, MobileNet V3 Large is used to extract multi-scale features for registration and change detection. Second, a spatial consistency module is used to achieve semi-dense feature point matching, establishing a cross-scale spatial transformation model that aligns feature maps at different scales. Then, a spatiotemporal difference collaboration module enhances the spatiotemporal heterogeneity of the two-temporal features. Finally, the multi-scale differential features are fused to generate the change detection results. Experiments were conducted using the SVCD, SYSU-CD, and SECOND datasets, and comparisons were made with current mainstream change detection networks. The results show that the proposed method can effectively construct the spatial transformation relationship between images to be registered, significantly outperforming other methods in both quantitative and qualitative analysis, and also exhibiting advantages in network complexity.

[0006] The technical solution of this invention is a lightweight remote sensing image change detection method for spatial misalignment, comprising the following steps:

[0007] For images from different time phases, a weight-sharing backbone network is used to extract multi-scale feature maps at different time phases;

[0008] Encoding stage: A cross-scale spatial transformation model is established using the spatial consistency module to align feature maps at different scales and to construct a semi-dense correspondence between preceding and following temporal images at a certain scale, in order to supervise the training of the spatial consistency module.

[0009] The aligned feature maps at different scales are then processed through a spatiotemporal difference coordination module to obtain substantial spatiotemporal difference features between the two temporal images.

[0010] Decoding stage: Multi-scale spatial spectrum difference features are fused to generate change detection results.

[0011] Furthermore, the backbone network is MobileNet V3 Large. Before feature extraction, a random affine transformation is performed on the previous temporal image T1 to obtain T1'. Based on the matching relationship between the previous temporal image T1 and the subsequent temporal image T2, a semi-dense correspondence between T1' and T2 at a 1 / 8 scale is constructed.

[0012] Furthermore, the spatial consistency module includes a feature point detection branch and a local descriptor branch;

[0013] In the feature point detection branch: First, a sliding window expansion operation is used to flatten the local regions of the input grayscale image into features along the channel dimension. Each channel corresponds to a local, non-overlapping 8×8 grid region of the original image; then, four 1×1 convolution pairs are used. The number of channels was adjusted; finally, the lightweight feature point detection framework ALIKE was selected as the teacher network to supervise the feature point heatmap. For the later time-phase image T2, the same method was used to obtain the feature point heatmap. The difference is that: first, a random affine transformation is performed on the previous time phase image T1 to obtain T1', and T1' is converted into a grayscale image, while the subsequent time phase image T2 is directly converted into a grayscale image.

[0014] In the descriptor branch: First, using 1×1 convolution and bilinear upsampling operations, the extracted multi-scale feature maps are... To maintain a consistent spatial dimension, calculate the sum of all feature maps. Then, by using successive convolution operations with different kernel sizes, semi-dense local feature descriptors are calculated. Finally, the confidence scores of the feature points are obtained using three consecutive 1×1 convolutions and the sigmoid function. For the later temporal image T2, the same method was used to obtain semi-dense local feature descriptors. and feature point confidence The subscripts 1 and 2 are used to identify the preceding time phase image T1 and the following time phase image T2.

[0015] Furthermore, the spatiotemporal difference coordination module includes spatial coordination perception and semantic coordination perception. Spatial coordination perception and semantic coordination perception calculate the absolute difference between aligned bi-temporal feature maps at different scales i. Secondly, semantic difference features between aligned dual-temporal feature maps are constructed by alternating channel splicing. ∈ Subsequently, an efficient channel attention mechanism and 1×1 convolution are used to obtain the semantic difference features after channel dimensionality reduction. ∈ H and W The height and width are defined separately, and C represents the number of channels. Then, in the semantic collaborative perception branch, a multi-channel cross-attention mechanism is used to calculate the semantic collaborative differential features. ∈ In the spatial collaborative perception branch, feature maps under different receptive fields are calculated using four parallel convolutions with different dilation rates, and spatial collaborative differential features are calculated using a spatial cross-attention mechanism. ∈ Finally, regarding and Channel splicing is performed, followed by 1×1 convolution, batch normalization, and ReLU activation to obtain spatiotemporal difference collaborative features. ∈ .

[0016] Furthermore, in spatial collaborative perception, a spatial cross-attention mechanism is used to construct the spatial dependency relationship between spatial difference features and semantic difference features; given a spatial difference feature map... ∈ Right now semantic difference feature map ∈ Right now First, feature maps with different receptive fields are generated using 3×3 depthwise separable convolutions with different dilation rates. ; then on Summation is performed, and the ReLU activation function is used to calculate the feature map after fusion of different receptive fields. ∈ ;

[0017] (1)

[0018] In the formula: This represents a kernel size of 3 and a dilatation rate of 1. d Depth-separable convolution;

[0019] Secondly, in terms of channel dimension, for Average pooling and max pooling are performed separately, then the two are concatenated, and finally a 7×7 convolution is applied. Calculating spatial feature maps using the Sigmoid function ∈ ;

[0020] (2)

[0021] In the formula: AvGpool and Maxpool represent average pooling and max pooling, respectively, and Concat represents channel concatenation. Represents the Sigmoid function;

[0022] Then, in the spatial dimension of any channel, for Average pooling is performed, followed by calculation of channel feature maps using 1×1 ordinary convolution, ReLU activation function, and Sigmoid function. ∈ ;

[0023] (3)

[0024] In the formula: This represents a convolution kernel size of 1, with input and output channels of [missing information]. C 1. C Convolution of 2;

[0025] Finally, and , The result of multiplication and Adding them together yields a spatial cooperative attention graph. ∈ ;

[0026] (4).

[0027] Furthermore, in semantic collaborative perception, a multi-channel cross-attention mechanism is used to construct the channel dependency relationship between spatial difference features and semantic difference features; given a spatial difference feature map... ∈ Right now semantic difference feature map ∈ Right now First, regarding , Perform average pooling with a 5×5 window, followed by layer normalization and 1×1 convolution. The spatial pooling feature map is calculated. Channel pooling feature map and ∈ ;

[0028] (5)

[0029] (6)

[0030] In the formula: The window size is 5, and the input and output space sizes are respectively... H × W , H × W Average pooling, Representative layer normalization;

[0031] Then, and Perform matrix multiplication and divide by the scaling factor. Then, the attention weight matrix is ​​obtained by normalization using the Softmax function, and finally, it is combined with... Multiplying yields the features associated with the global context. ∈ ;

[0032] (7)

[0033] Finally, for Perform global average pooling, then combine with Perform matrix multiplication to obtain a semantic collaborative attention graph. ∈ ;

[0034] (8).

[0035] Furthermore, the joint local descriptor loss in the spatial consistency module Feature point confidence loss Feature point detection loss Pixel-level offset loss Training is conducted; the specific calculation process for each loss is as follows:

[0036] 1) Local descriptor loss

[0037] First, from the semi-dense descriptor of the spatial transformation model , Mid-sampling yielded two sets of descriptors. , ∈ Each group contains N A 64-dimensional descriptor, N First, calculate the number of semi-dense corresponding points; second, calculate the descriptor. , Two-way similarity matrix between S∈ Then, respectively for S and S T Each row is normalized using the softmax function to obtain the matching probability matrix. P= softmax( S )and Q= softmax( S T ),matrix P The i-th row represents The i-th point and The matching probability of all points. Q Similarly, ideally, the correct matching correspondence should be located in the matrix. P and Q On the diagonal; therefore, the local descriptor loss is the sum of the negative log-likelihoods of the correct match probabilities:

[0038] (9)

[0039] 2) Feature point confidence loss

[0040] First, from the confidence plot in the spatial transformation model , Mid-sampling yields two sets of confidence scores that correspond to the sampling point locations of the local descriptors. , ∈ Then, using the matching probability matrix P and Q Calculation from The i-th point is Maximum matching confidence in And its reverse maximum matching confidence. The confidence level of the bidirectional matching at the i-th point is Finally, the feature point confidence loss is constructed using the L1 norm:

[0041] (10)

[0042] 3) Feature point detection loss

[0043] The feature point detection branch employs knowledge distillation for supervised training; the supervisory signal provided by the teacher network ALIKE is the set of feature point locations at the original image resolution. A ={( x , y Feature point detection loss Defined as:

[0044] (11)

[0045] In the formula: and It is a feature point heatmap predicted by the network. H and W These are the height and width of the heatmap, respectively. yes At position ( i , j The corresponding ALIKE predicted label at ) Indicates the position ( i , j At point ), the model predicts that the label belongs to the tag. The probability of the category, The calculation formula is as follows:

[0046] (12)

[0047] In the formula: This represents rounding down, for heatmaps. Each position on ( i , j ), which corresponds to an 8×8 grid region in the grayscale image; if a certain feature point exists ( x m , y m ) ∈A If the data is downsampled into the grid, its relative position within the grid is encoded as an integer between 0 and 63; otherwise, the label is... Setting it to 64 indicates that there are no feature points at this location;

[0048] 4) Pixel-level offset loss

[0049] The sampled descriptor and The data is stitched along the channel dimension and then input into a multilayer perceptron (MLP) to calculate the prediction offset confidence matrix. O∈ ;

[0050] (13)

[0051] Then, based on the prediction offset confidence matrix O With true offset label l∈ Construct the offset loss function for {0,1,2,...,63}, and use the bidirectional matching confidence of the matched pairs. M i Weighted;

[0052] (14)

[0053] In the formula: Let the set of semi-dense points of T1´ be... P ={( x , y )}, l The calculation method is as follows:

[0054] (15)

[0055] Ultimately, the overall matching loss function defines the weighted average of the losses for each item.

[0056] Furthermore, in the decoding stage, a scale-adaptive perception module is used to fuse differential feature maps at different scales using a dynamic weight fusion method that combines top-down and bottom-up approaches. Finally, bilinear interpolation is used to upsample the feature maps by a factor of 4, and then a 1×1 convolution is used to adjust the number of channels to the number of categories, thereby outputting the change detection results.

[0057] The present invention also provides a lightweight remote sensing image change detection system for spatial misalignment, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute the lightweight remote sensing image change detection method for spatial misalignment as described in the above technical solution.

[0058] The present invention also provides a computer-readable storage medium, including a readable storage medium on which a computer program is stored, wherein when the computer program is executed, it implements a lightweight remote sensing image change detection method for spatial misalignment as described in the above technical solution.

[0059] This invention proposes a lightweight change detection network, LJRCDNet, that combines remote sensing image registration and change detection. Model comparison experiments show that, even in complex change scenarios involving spatial misalignment and pseudo-spatial-spectral differences between two temporal images, this method can capture the true change information of ground features effectively with relatively low network complexity. Ablation experiments fully demonstrate the coupling effect between the SCM and STDCM modules. The SCM module, by establishing a spatial transformation model between the two temporal images, provides a foundation for the STDCM module to capture substantial spatial-spectral differences; the joint embedding of both improves the change detection accuracy. Attached Figure Description

[0060] Figure 1 This is a diagram illustrating the overall network architecture of an embodiment of the present invention.

[0061] Figure 2 This is a structural diagram of the spatial consistency module in an embodiment of the present invention;

[0062] Figure 3 This is a structural diagram of the spatiotemporal difference coordination module according to an embodiment of the present invention;

[0063] Figure 4 This is a schematic diagram of the semi-dense correspondence of dual-temporal images in an embodiment of the present invention;

[0064] Figure 5 These are visualization results of different models on the SVCD dataset in embodiments of the present invention;

[0065] Figure 6 These are visualization results of different models on the SYSU-CD dataset in embodiments of the present invention;

[0066] Figure 7 This is a visualization of the different models on the SECOND dataset in the embodiments of the present invention. Detailed Implementation

[0067] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0068] This invention proposes a lightweight joint registration and change detection network (LJRCDNet): (1) a spatial consistency module (SCM) is designed to construct a spatial transformation model between two temporal images; (2) a lightweight spatio-temporal difference collaboration module (STDCM) is designed to promote the collaborative interaction of difference features at the spatial and semantic levels.

[0069] Example 1

[0070] This invention provides a lightweight remote sensing image change detection method for spatial misalignment, comprising the following steps:

[0071] For images from different time phases, a weight-sharing backbone network is used to extract multi-scale feature maps at different time phases;

[0072] Encoding stage: A cross-scale spatial transformation model is established using the spatial consistency module to align feature maps at different scales and to construct a semi-dense correspondence between preceding and following temporal images at a certain scale, in order to supervise the training of the spatial consistency module.

[0073] The aligned feature maps at different scales are then processed through a spatiotemporal difference coordination module to obtain substantial spatiotemporal difference features between the two temporal images.

[0074] Decoding stage: Multi-scale spatial spectrum difference features are fused to generate change detection results.

[0075] The specific implementation of each part of the present invention will be described in detail below:

[0076] 1. Overall network structure

[0077] LJRCDNet adopts an end-to-end encoding and decoding architecture, such as Figure 1 As shown. Before model training, a randomized affine transformation (RHT) is performed on the previous temporal image T1 to obtain T1´. ∈ , H and W Here, T1 represents the height and width of the image, and 3 represents the number of channels in the image. Meanwhile, T1 and T2 are spatially aligned bi-temporal images. Based on the true matching relationship between the preceding temporal image T1 and the following temporal image T2, a semi-dense correspondence between T1' and T2 at a 1 / 8 scale is constructed as ground truth for subsequent network training. Encoder stage: First, the backbone network MobileNet V3 Large is used to extract four shared feature maps at different scales for image registration and change detection. , ( The subscripts 1 and 2 are used to identify T1 and T2, respectively, and their dimensions are as follows: H / 4× W / 4×64、 H / 8× W / 8×128、 H / 16× W / 16×256 and H / 32× W / 32×512; then, , ( The spatial transformation relationship between the two temporal images is constructed using the Spatial Consistency Module (SCM). H (Including semi-dense local feature descriptors and feature point confidence), and simultaneously performing affine transformation (HT) on feature maps at different scales in previous time phases to obtain... Feature map of spatial alignment ( );at last, , ( The substantial spatial-spectral differences between two temporal images are constructed using the Spatiotemporal Difference Coordination Module (STDCM). Decoder stage: First, the Scale Adaptive Perception Module (SAPM) is used to fuse difference feature maps at different scales using a top-down and bottom-up dynamic weight fusion method. The output feature map size is [size missing]. H / 4× W / 4×128; Finally, bilinear interpolation is used to upsample the feature map by 4 times, and then 1×1 convolution is used to adjust the number of channels to the number of categories, thereby outputting the change detection result.

[0078] 1.1 Spatial Consistency Module

[0079] Currently, image matching methods based on deep learning can be broadly categorized into detector-based and detectorless methods. Detectorless methods typically utilize rich contextual information within images, exhibiting good matching performance in scenarios with weak or repetitive textures. However, these dense matching methods often suffer from high network complexity, limiting the image resolution they can process. To address this, this invention proposes a detector-based semi-dense matching method, namely the Spatial Consistency Module (SCM). This module utilizes self-supervised training to construct spatial transformation relationships between two temporal images. The SCM module consists of a feature point detection branch and a local descriptor branch, such as... Figure 2 As shown.

[0080] In the feature point detection branch: First, through a sliding window expansion operation, the local regions of the input grayscale image (the grayscale images corresponding to T1´ and T2) are flattened into features in the channel dimension. ∈ Each channel corresponds to a local, non-overlapping 8×8 grid region of the original image. H' and W' The values ​​are respectively H / 8 and W / 8. Then, use four 1×1 convolutions to... The number of channels was adjusted to 65, and an additional "no feature point" trash can channel was added to distinguish between the presence and absence of feature points. Finally, the lightweight feature point detection framework ALIKE was selected as the teacher network to supervise the feature point heatmap. ∈ .

[0081] For the later phase image T2, the same method was used to obtain the feature point heatmap. .

[0082] In the descriptor branch: First, using 1×1 convolution and bilinear upsampling operations, ... ( To maintain a consistent spatial dimension, calculate the sum of all feature maps. ∈ Then, by using successive convolution operations with different kernel sizes, semi-dense local feature descriptors are calculated. ∈ Finally, the feature point confidence scores are obtained using three consecutive 1×1 convolutions and the sigmoid function. ∈ .

[0083] For the later temporal image T2, the same method was used to obtain semi-dense local feature descriptors. and feature point confidence .

[0084] 1.2, Spatiotemporal Difference Collaboration Module

[0085] To address the significant pseudo-spatial-spectral difference variations in dual-temporal remote sensing images, current mainstream methods often employ a dual-branch architecture based on subtraction and channel stitching to construct spatial and spectral difference variations. However, these methods lack modeling of the spatial-spectral coupling mechanism during the difference feature enhancement stage, failing to establish spatial-spectral interaction constraints between difference features. Therefore, a lightweight spatiotemporal difference coordination module (STDCM) is designed. Utilizing spatial cross-attention and multi-channel cross-attention mechanisms, it constructs a synergistic relationship between spatial and semantic difference features, thereby enhancing the spatiotemporal heterogeneity of dual-temporal image features. STDCM consists of spatial co-sensing and semantic co-sensing, such as... Figure 3 As shown.

[0086] First, spatial collaborative sensing and semantic collaborative sensing compute aligned bi-temporal feature maps at the same scale. and absolute difference between Secondly, it is constructed by alternating the splicing of channels. and Semantic difference features between ∈ Subsequently, efficient channel attention (ECA) and 1×1 convolution are used to obtain semantic difference features after channel dimensionality reduction. ∈ Then, in the semantic collaborative perception branch, a multi-channel cross-attention mechanism is used to calculate semantic collaborative differential features. ∈ In the spatial collaborative perception branch, feature maps under different receptive fields are calculated using four parallel convolutions with different dilation rates, and spatial collaborative differential features are calculated using a spatial cross-attention mechanism. ∈ Finally, regarding and Channel splicing is performed, followed by 1×1 convolution, batch normalization, and ReLU activation to obtain spatiotemporal difference collaborative features. ∈ .

[0087] 1.2.1 Spatial Collaborative Perception

[0088] In spatial collaborative perception, a spatial cross-attention mechanism is used to construct the spatial dependencies between spatial and semantic difference features. Given a spatial difference feature map... ∈ (Right now ) and semantic difference feature map ∈ (Right now First, feature maps with different receptive fields are generated using 3×3 depthwise separable convolutions with different dilation rates. ( ). Then on Summation is performed, and the ReLU activation function is used to calculate the feature map after fusion of different receptive fields. ∈ :

[0089] (1)

[0090] In the formula: This represents a kernel size of 3 and a dilatation rate of 1. d Depth-separable convolutions.

[0091] Secondly, in terms of channel dimension, for Average pooling and max pooling are performed, then the two are concatenated, and finally a spatial feature map is calculated using a 7×7 convolution and a sigmoid function. ∈ :

[0092] (2)

[0093] In the formula: AvGpool and Maxpool represent average pooling and max pooling, respectively, and Concat represents channel concatenation. This represents the Sigmoid function.

[0094] Then, in the spatial dimension of any channel, for Average pooling is performed, followed by calculation of channel feature maps using 1×1 ordinary convolution, ReLU activation function, and Sigmoid function. ∈ :

[0095] (3)

[0096] In the formula: This represents a convolution kernel size of 1, with input and output channels of [missing information]. C 1. C 2 convolutions.

[0097] Finally, and , The result of multiplication and Adding them together yields a spatial cooperative attention graph. ∈ :

[0098] (4)

[0099] 1.2.2 Semantic Collaborative Perception

[0100] In semantic collaborative perception, a multi-channel cross-attention mechanism is used to construct the channel dependency relationship between spatial and semantic difference features. Given a spatial difference feature map... ∈N H×W×C semantic difference feature map ∈ First, regarding , A 5×5 window average pooling process is performed, followed by layer normalization and 1×1 convolution to calculate the spatial pooling feature map. Channel pooling feature map and ∈ :

[0101] (5)

[0102] (6)

[0103] In the formula: The window size is 5, and the input and output space sizes are respectively... H ×W , H × W Average pooling, Representative level normalization.

[0104] Then, and Perform matrix multiplication and divide by the scaling factor. Then, the attention weight matrix is ​​obtained by normalization using the Softmax function, and finally, it is combined with... Multiplying yields the features associated with the global context. ∈ :

[0105] (7)

[0106] Finally, for Perform global average pooling, then combine with Perform matrix multiplication to obtain a semantic collaborative attention graph. ∈ :

[0107] (8)

[0108] 2. Model training dataset and training strategy

[0109] 2.1, Dataset

[0110] The effectiveness of the model was tested using the SVCD, SYSU-CD, and SECOND datasets.

[0111] 1) The SVCD dataset focuses on the dynamic changes of land features at multiple scales and the seasonal changes of natural land features. It includes 7 sets of remote sensing images with a resolution of 0.03-1 meter (4725×2200 pixels) and 4 sets of RGB images (1900×1000 pixels). The preprocessing in this experiment is consistent with that of Lebedev et al. (2018). After image cropping and standardization, 16,000 pairs of images with a pixel size of 256×256 are generated, including 10,000 pairs of training samples, 3,000 pairs of validation samples, and 3,000 pairs of test samples.

[0112] 2) The SYSU-CD dataset is a remote sensing image dataset for binary change classification, containing 20,000 pairs of 0.5-meter high-resolution aerial images acquired between 2007 and 2014, covering changes in various land cover types such as buildings, vegetation, roads, and water bodies. The images in this dataset are 256×256 pixels, and the ratio of the training set, validation set, and test set is 6:2:2, with 12,000, 4,000, and 4,000 image pairs, respectively.

[0113] 3) The SECOND dataset focuses on multi-category land surface changes, involving six typical land features: non-vegetated land surfaces, forests, low vegetation, water bodies, building structures, and recreational facilities. This dataset is labeled with both semantic and binary classification changes. This experiment uses binary classification, with an image resolution of 512×512 pixels. The original data contains 4662 image pairs, of which 2968 pairs are currently publicly available and labeled. During the experiment, the images were cropped to 256×256 pixels and non-overlapping, then divided into 8312 training pairs, 1188 validation pairs, and 2372 test pairs in a 7:1:2 ratio.

[0114] 2.2 Training Strategy

[0115] This invention employs the PyTorch 1.10.2 deep learning framework and trains and validates the proposed network model on an NVIDIA GeForce RTX4090 GPU. A two-stage frozen training strategy is used, with a total of 300 training epochs. Specifically, the parameters of the fourth layer of the backbone network, the spatiotemporal difference co-processing module, the scale-adaptive perception module, and the 1×1 convolutional layer are frozen in the first 150 epochs; the parameters of the aforementioned frozen modules are unfrozen in the last 150 epochs, while the parameters of the first three stages of the backbone network and the spatial consistency module are frozen. Detailed parameter configurations are as follows: batch size is 8, optimizer momentum is 0.9, weight decay coefficient is 0.001, and the initial learning rate is 0.0001 for both the first 150 epochs and the last 150 epochs. The image size of all datasets is 256×256, and the model's floating-point operations (FLOPs) and total number of parameters are statistically analyzed through independent modules.

[0116] 2.3 Loss Function

[0117] 2.3.1 Matching Loss

[0118] In the first 150 rounds of network training, the semi-dense correspondence between T1´ and T2 at a 1 / 8 scale (e.g.) is utilized. Figure 4 As shown), the local descriptor loss is jointly optimized using supervised learning. Feature point confidence loss Feature point detection loss Pixel-level offset loss .

[0119] 1) Local descriptor loss

[0120] First, from semi-dense descriptors , Mid-sampling yielded two sets of descriptors. , ∈ Each group contains N A 64-dimensional descriptor, N This represents the number of semi-dense corresponding points. Next, the descriptor is calculated. , Two-way similarity matrix between S∈ Then, respectively for S and S T Each row is normalized using the softmax function to obtain the matching probability matrix. P= softmax( S )and Q= softmax( S T ),matrix P The i-th row represents The i-th point and The matching probability of all points, Q Similarly, ideally, the correct matching correspondence should be located in the matrix. P and Q On the diagonal. Therefore, the local descriptor loss is the sum of the negative log-likelihoods of the correct match probabilities:

[0121] (9)

[0122] 2) Feature point confidence loss

[0123] First, from the confidence plot , Mid-sampling yields two sets of confidence scores that correspond to the sampling point locations of the local descriptors. , ∈ Then, using the matching probability matrix P and Q Calculation from The i-th point is Maximum matching confidence in And its reverse maximum matching confidence. The confidence level of the bidirectional matching at the i-th point is Finally, the feature point confidence loss is constructed using the L1 norm:

[0124] (10)

[0125] 3) Feature point detection loss

[0126] The feature point detection branch employs knowledge distillation for supervised training. The supervisory signal provided by the teacher network (ALIKE) is the set of feature point locations at the original image resolution. A ={( x , y Feature point detection loss Defined as:

[0127] (11)

[0128] In the formula: and It is a feature point heatmap predicted by the network. H and W These are the height and width of the heatmap, respectively. yes At position ( i , j The corresponding ALIKE predicted label at ) Indicates the position ( i , j At point ), the model predicts that the label belongs to the tag. The probability of the category, The calculation formula is as follows:

[0129] (12)

[0130] In the formula: This represents rounding down; specifically, for heatmaps... Each position on ( i , j This corresponds to an 8×8 grid region in the grayscale image. If a certain feature point exists ( x m , y m ) ∈A If a value is downsampled into the grid, its relative position within the grid is encoded as an integer between 0 and 63; otherwise, the label is... Setting it to 64 indicates that there are no feature points at this location.

[0131] 4) Pixel-level offset loss

[0132] To further improve the sub-pixel accuracy of feature matching, a pixel-level offset loss was designed. First, the sampled descriptors... and The data is concatenated along the channel dimension and then input into a multilayer perceptron (MLP) to calculate the prediction offset confidence matrix. O∈ ;

[0133] (13)

[0134] Then, based on the prediction offset confidence matrix O With true offset label l∈ Construct the offset loss function for {0,1,2,...,63}, and use the bidirectional matching confidence of the matching pair. M i Weighting allows the network to focus on learning offsets for high-quality matches;

[0135] (14)

[0136] In the formula: Let the set of semi-dense points of T1´ be... P ={( x , y )}, l The calculation method is as follows:

[0137] (15)

[0138] Finally, the total matching loss function is defined as:

[0139] (16)

[0140] The coefficients for each loss can be set according to requirements.

[0141] 2.3.2, Change Detection Loss

[0142] To address the class imbalance problem in remote sensing change detection, this invention employs a composite loss function combining BCE loss (BL) and Diceloss (DL) to mitigate the impact of uneven distribution between changed and unchanged regions, improve the accuracy of change region detection, and enhance model robustness.

[0143] (17)

[0144] (18)

[0145] (19)

[0146] In the formula: p To predict probabilities, y∈ {0,1} represents the true category label.

[0147] 3. Model Evaluation and Experimental Results

[0148] 3.1 Evaluation Indicators

[0149] To quantitatively evaluate the model's performance, precision (P), recall (R), intersection over union (IoU), F1 score (F1), and overall accuracy (OA) are used. The specific calculation formulas are as follows:

[0150] (20)

[0151] (twenty one)

[0152] (twenty two)

[0153] (twenty three)

[0154] (twenty four)

[0155] In the formula: TP (True Positive) represents the number of correctly classified changed categories; TN (True Negative) represents the number of correctly classified unchanged categories; FP (False Positive) represents the number of incorrectly classified changed categories; and FN (False Negative) represents the number of incorrectly classified unchanged categories. Higher values ​​for P, R, IoU, F1 score, and OA indicate better model performance.

[0156] 3.2 Model Comparison Experiment

[0157] To verify the effectiveness of LJRCDNet, 11 different network structures were selected for comparative experiments from the perspectives of quantitative analysis, qualitative analysis, and network complexity analysis:

[0158] 1) SNUNet: The feature extraction backbone only uses multiple 3×3 ordinary convolutions, without designing a differential feature enhancement module, and only uses a dense connection structure to fuse differential feature semantic information at different scales.

[0159] 2) RFANet: MobileNet V2 is selected as the feature extraction backbone. In a dense connection mode, multi-scale fusion of different levels of features of single temporal images is performed. The semantic segmentation and aggregation module is used to construct high-level semantic change information of multi-scale ground objects, thereby dynamically adjusting the channel weights of low-level differential features.

[0160] 3) ChangeForme: It adopts a weight-sharing Siamese hierarchical Transformer encoder to construct long-distance dependencies of single-phase multi-scale feature maps and uses the feature difference module to construct global spatial relationships of differential features.

[0161] 4) TinyCD (Codegoni et al., 2023): It uses the pre-trained EfficientNet-b4 as the feature extraction backbone and designs a hybrid and attention masking module. First, the hybrid module performs channel-alternating splicing of dual-temporal features, and then captures the confidence of the changing region through a pixel-level multilayer perceptron.

[0162] 5) CDNeXt: ConvNeXt was selected as the main feature extraction backbone, and a spatiotemporal interactive attention module was designed. By constructing the spatial semantic invariance of dual-temporal features, the interference of viewpoint changes and background was suppressed. Then, the original features were fused through residual connections to generate a spatiotemporal interactive attention map.

[0163] 6) TFI-GR: ResNet18 is used as the backbone for feature extraction. A differential feature extraction module is designed to enhance the interaction between features in two time phases. Then, the detailed information and semantic information of the differential features are fused through the guided refinement module.

[0164] 7) DMINet: ResNet18 was selected as the backbone for feature extraction. A joint attention module that integrates self-attention and cross-attention was designed. Then, a dual-branch structure of feature subtraction and channel concatenation was used to construct multi-level spatial and semantic difference features.

[0165] 8) AMTNet: ResNet18 or ResNet50 is selected as the backbone for feature extraction. Differentiated fusion strategies are adopted for features of different scales, namely, spatial dimension exchange for high-resolution features and channel dimension exchange for low and medium-resolution features. Spatial attention module, channel attention module and self-attention mechanism are used to construct differential features with long-distance dependencies.

[0166] 9) SEIFNet: ResNet18 was selected as the backbone for feature extraction. A spatiotemporal difference enhancement module was designed. A dual-branch structure of feature subtraction and channel concatenation was adopted to capture the global and local information of the difference features respectively, and then they were added together to generate difference enhancement features.

[0167] 10) A2Net: The pre-trained MobileNet V2 was selected as the backbone for feature extraction. A neighborhood aggregation module and a progressive change recognition module were designed to fuse features of different scales in a single time phase and to construct the spatial relationship of multi-scale differential features, respectively.

[0168] 11) EfficientCD: EfficientNet is selected as the backbone for feature extraction. Based on the hierarchical exchange strategy, adjacent-level features of the feature pyramid of the dual-temporal images are exchanged. The differences between the dual-temporal features are quantified by Euclidean distance. Finally, the difference features are fused using a multi-level decoder module.

[0169] 3.3.1 Quantitative Analysis

[0170] Table 1 shows the quantitative comparison results between our proposed method and 11 other models. Overall, all change detection methods rely too heavily on spatially aligned bi-temporal images, resulting in less than ideal performance across all datasets. Specifically:

[0171] 1) SNUNet's feature extraction backbone is too simple, lacking a corresponding differential feature enhancement module. It only utilizes a dense connection structure to fuse semantic aggregate features at different scales, without constructing the spatial-spectral difference relationship between bi-temporal features, resulting in the worst overall performance across all datasets. ChangeFormer utilizes a self-attention mechanism in the feature extraction stage to construct the global spatial relationship between bi-temporal features, and then calculates the spatial difference of features at the same level, achieving better overall performance than SNUNet. TinyCD introduces an SE (squeeze-and-excitation) module in the feature extraction stage, effectively constructing the semantic association between bi-temporal features. It uses a hybrid module to construct the semantic difference between features at different scales, and then uses an attention mask module to quantify the confidence of changing pixels, achieving better overall performance than ChangeFormer. RFANet not only utilizes a dense connection structure to fuse bi-temporal features at different scales, but also constructs the relationship between high-level semantic information and low-level detail information of differential features, improving the expressive power of differential features, and achieving better overall performance than TinyCD.

[0172] 2) CDNeXt, TFI-GR, and DMINet all designed effective differential feature enhancement modules, and their overall performance is better than the above methods. Their overall performance is comparable. AMTNet first performs differential fusion of dual-temporal features, and then uses spatial attention mechanism, channel attention mechanism, and self-attention mechanism to effectively construct the global relationship of multi-scale changing land features. Its overall performance is slightly better than the above three methods. SEIFNet designed a dual-branch structure spatiotemporal differential enhancement module with feature subtraction and channel concatenation, and constructed the global-local spatial-spectral relationship of differential features. It effectively suppressed the pseudo-changes caused by interference from seasons, illumination, shadows, etc. Its overall performance is better than AMTNet.

[0173] 3) A2Net first integrates features from adjacent phases through a domain aggregation module, enhancing the semantic and detailed information of single-temporal features at different scales. Then, it fuses differential features at different scales through multi-branch dilated convolutions with decreasing dilation rates, effectively enhancing the information of ground features changing at different scales. Finally, it solves the spatial and semantic misalignment problem of differential features at different levels through a supervised attention module, and its overall performance is better than SEIFNet. EfficientCD adopts a different feature exchange strategy than TinyCD and AMTNet, namely a parameterless dual-temporal feature hierarchical exchange strategy, which not only preserves the semantic integrity of features at different scales, but also improves the model's ability to represent differential features. Then, by combining a dual-branch structure layer-by-layer decoder with Euclidean distance, it quantifies the pixel-level differences of dual-temporal images and achieves the progressive fusion of differential features, and its overall performance is better than the above methods.

[0174] 4) Although the above methods achieved good results in the change detection scenario of dual-temporal image alignment, the experimental results show that if there is a certain degree of geometric shift in the images, the performance of the above methods drops significantly. The proposed method utilizes a lightweight spatial consistency module (SCM) and a spatiotemporal difference coordination module (STDCM) to effectively construct the spatial transformation relationship and spatiotemporal difference feature spatial-spectral correlation of dual-temporal images. Except for the accuracy performance on the SYSU-CD and SECOND datasets, it achieved the best performance on all other datasets. LJRCDNet increased the IoU ratio of the second-best values ​​on the SVCD, SYSU-CD, and SECOND test datasets by 6.59%, 6.21%, and 5.66%, respectively.

[0175] Table 1. Comparison of metrics for different models on different datasets

[0176]

[0177] Note: Bold text indicates the best value in each column; underlined text indicates the second best value in each column.

[0178] 3.3.2 Qualitative Analysis

[0179] Figure 5-7 The visualization results of various models on the SVCD, SYSU-CD, and SECOND test datasets are shown, including: Figure 5-7 The first to third rows are the change detection scenario under spatial misalignment; Figure 5-7 The fourth row depicts a large-scale change in terrain features. Figure 5-7Rows 5 and 6 depict complex changing scenarios, where spatial misalignment and pseudo-spatial-spectral differences coexist. It can be seen that, compared to other methods, LJRCDNet can perform change detection tasks well even with spatial misalignment in dual-temporal images, despite interference from pseudo-spatial-spectral differences such as ground texture, shadows, and seasonal variations.

[0180] (1) SVCD dataset

[0181] from Figure 5 As shown in rows 1-3, although all methods can detect changed features well, such as vehicle changes in rows 2 and 3 and hardened site changes in row 2, the false detection rate of all comparison methods is significantly higher than that of the proposed method due to the geometric offset of internal roads in the dual-temporal images. In the large-scale feature change scene in row 4, it can be seen that even when spatial alignment between dual-temporal images is not possible, the edge structure of large-scale features extracted by the proposed method is relatively complete. In the complex change scene with spatial misalignment and shadow interference in rows 5-6, LJRCDNet performs best overall compared to all comparison methods, with no significant missed or false detections.

[0182] (2) SYSU-CD dataset

[0183] from Figure 6 As can be seen from rows 1 to 3, LJRCDNet did not exhibit significant false negatives. For example, in the newly built park (row 1), newly added cargo ship (row 2), and newly added farmland (row 3), although LJRCDNet showed slight false positives and false negatives, its overall performance remained the best. In the large-scale land cover change scenario (row 4), it can be seen that the internal structure of the large-scale land cover extracted by LJRCDNet is relatively complete. In the pseudo-spatial difference change scenario caused by vegetation canopy density changes (row 5), and the pseudo-spectral difference change scenario caused by illumination changes (row 6), except for the method presented in this paper, the other methods all showed varying degrees of false positives and false negatives.

[0184] (3) SECOND dataset

[0185] from Figure 6 Rows 1-3 show that LJRCDNet effectively detected changed ground features without significant false negatives. In the newly added large building scene in row 4, ChangeFormer and LJRCDNet extracted relatively complete large-scale ground feature internal structures. In the newly added building shadow interference scene in row 5, except for CDNeXt and LJRCDNet, the other models exhibited significant false positives. In the newly added scene where the building's spectrum is similar to the surrounding forest in row 6, except for the method presented in this paper, the other methods failed to effectively detect the building change.

[0186] In practice, the above process can be automated using computer software technology.

[0187] Example 2

[0188] This invention also provides a lightweight remote sensing image change detection system for spatial misalignment, including a processor and a memory. The memory stores program instructions, and the processor calls the program instructions in the memory to execute the lightweight remote sensing image change detection method for spatial misalignment as described in the above technical solution.

[0189] Example 3

[0190] This invention also provides a computer-readable storage medium, including a readable storage medium on which a computer program is stored. When the computer program is executed, it implements a lightweight remote sensing image change detection method for spatial misalignment as described in the above technical solution.

[0191] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A lightweight change detection method for spatial misalignment of remote sensing images, characterized in that, Includes the following steps: For images from different time phases, a weight-sharing backbone network is used to extract multi-scale feature maps at different time phases; Encoding stage: A cross-scale spatial transformation model is established using the spatial consistency module to align feature maps at different scales and to construct a semi-dense correspondence between preceding and following temporal images at a certain scale, in order to supervise the training of the spatial consistency module. The spatial consistency module includes a feature point detection branch and a local descriptor branch; In the feature point detection branch: First, a sliding window expansion operation is used to flatten the local regions of the input grayscale image into features along the channel dimension. Each channel corresponds to a local, non-overlapping 8×8 grid region of the original image; then, four 1×1 convolution pairs are used. The number of channels was adjusted; finally, the lightweight feature point detection framework ALIKE was selected as the teacher network to supervise the feature point heatmap. For the later time-phase image T2, the same method was used to obtain the feature point heatmap. The difference is that: first, a random affine transformation is performed on the previous time phase image T1 to obtain T1', and T1' is converted into a grayscale image, while the subsequent time phase image T2 is directly converted into a grayscale image. In the descriptor branch: First, using 1×1 convolution and bilinear upsampling operations, the extracted multi-scale feature maps are... To maintain a consistent spatial dimension, calculate the sum of all feature maps. Then, by using successive convolution operations with different kernel sizes, semi-dense local feature descriptors are calculated. Finally, the confidence scores of the feature points are obtained using three consecutive 1×1 convolutions and the sigmoid function. For the later temporal image T2, the same method was used to obtain semi-dense local feature descriptors. and feature point confidence , where subscripts 1 and 2 are used to identify the preceding time phase image T1 and the following time phase image T2; The aligned feature maps at different scales are then processed through a spatiotemporal difference coordination module to obtain substantial spatiotemporal difference features between the two temporal images. The spatiotemporal difference coordination module includes spatial coordination perception and semantic coordination perception. Spatial coordination perception and semantic coordination perception calculate the absolute difference between aligned bi-temporal feature maps at different scales i. Secondly, semantic difference features between aligned dual-temporal feature maps are constructed by alternating channel splicing. ∈ Subsequently, an efficient channel attention mechanism and 1×1 convolution are used to obtain the semantic difference features after channel dimensionality reduction. ∈ H and W Height and width are respectively, and C is the number of channels; Then, in the semantic collaborative perception branch, a multi-channel cross-attention mechanism is used to calculate semantic collaborative differential features. ∈ In the spatial collaborative perception branch, feature maps under different receptive fields are calculated using four parallel convolutions with different dilation rates, and spatial collaborative differential features are calculated using a spatial cross-attention mechanism. ∈ Finally, regarding and Channel splicing is performed, followed by 1×1 convolution, batch normalization, and ReLU activation to obtain spatiotemporal difference collaborative features. ∈ ; In spatial collaborative perception, a spatial cross-attention mechanism is used to construct the spatial dependency relationship between spatial difference features and semantic difference features; given a spatial difference feature map... ∈ Right now semantic difference feature map ∈ Right now First, feature maps with different receptive fields are generated using 3×3 depthwise separable convolutions with different dilation rates. ; then on Summation is performed, and the ReLU activation function is used to calculate the feature map after fusion of different receptive fields. ∈ ; (1) In the formula: This represents a kernel size of 3 and a dilatation rate of 1. d Depth-separable convolution; Secondly, in terms of channel dimension, for Average pooling and max pooling are performed separately, then the two are concatenated, and finally a 7×7 convolution is applied. Calculating spatial feature maps using the Sigmoid function ∈ ; (2) In the formula: AvGpool and Maxpool represent average pooling and max pooling, respectively, and Concat represents channel concatenation. Represents the Sigmoid function; Then, in the spatial dimension of any channel, for Average pooling is performed, followed by calculation of channel feature maps using 1×1 ordinary convolution, ReLU activation function, and Sigmoid function. ∈ ; (3) In the formula: This represents a convolution kernel size of 1, with input and output channels of [missing information]. C 1. C Convolution of 2; Finally, and , The result of multiplication and Adding them together yields a spatial cooperative attention graph. ∈ ; (4) In semantic collaborative perception, a multi-channel cross-attention mechanism is used to construct the channel dependency relationship between spatial difference features and semantic difference features; given a spatial difference feature map... ∈ Right now semantic difference feature map ∈ Right now First, regarding , Perform average pooling with a 5×5 window, followed by layer normalization and 1×1 convolution. The spatial pooling feature map is calculated. Channel pooling feature map and ∈ ; (5) (6) In the formula: The window size is 5, and the input and output space sizes are respectively... H × W , H × W Average pooling, Representative layer normalization; Then, and Perform matrix multiplication and divide by the scaling factor. Then, the attention weight matrix is ​​obtained by normalization using the Softmax function, and finally, it is combined with... Multiplying yields the features associated with the global context. ∈ ; (7) Finally, for Perform global average pooling, then combine with Perform matrix multiplication to obtain a semantic collaborative attention graph. ∈ ; (8) Decoding stage: Multi-scale spatial spectrum difference features are fused to generate change detection results.

2. The lightweight remote sensing image change detection method for spatial misalignment as described in claim 1, characterized in that: The backbone network is MobileNet V3 Large. Before feature extraction, a random affine transformation is performed on the previous temporal image T1 to obtain T1'. Based on the matching relationship between the previous temporal image T1 and the subsequent temporal image T2, a semi-dense correspondence between T1' and T2 at a 1 / 8 scale is constructed.

3. The lightweight remote sensing image change detection method for spatial misalignment as described in claim 1, characterized in that: Joint local descriptor loss in spatial consistency module Feature point confidence loss Feature point detection loss Pixel-level offset loss Training is conducted; the specific calculation process for each loss is as follows: 1) Local descriptor loss First, from the semi-dense descriptor of the spatial transformation model , Mid-sampling yielded two sets of descriptors. , ∈ Each group contains N A 64-dimensional descriptor, N First, calculate the number of semi-dense corresponding points; second, calculate the descriptor. , Two-way similarity matrix between S∈ Then, respectively for S and S T Each row is normalized using the softmax function to obtain the matching probability matrix. P= softmax( S )and Q= softmax( S T ),matrix P The i-th row represents The i-th point and The matching probability of all points. Q Similarly, ideally, the correct matching correspondence should be located in the matrix. P and Q On the diagonal; therefore, the local descriptor loss is the sum of the negative log-likelihoods of the correct match probabilities: (9) 2) Feature point confidence loss First, from the confidence plot in the spatial transformation model , Mid-sampling yields two sets of confidence scores that correspond to the sampling point locations of the local descriptors. , ∈ ; Then, using the matching probability matrix P and Q Calculation from The i-th point is Maximum matching confidence in and its reverse maximum matching confidence. The confidence level of the bidirectional matching at the i-th point is Finally, the feature point confidence loss is constructed using the L1 norm: (10) 3) Feature point detection loss The feature point detection branch employs knowledge distillation for supervised training; the supervisory signal provided by the teacher network ALIKE is the set of feature point locations at the original image resolution. A ={( x , y Feature point detection loss Defined as: (11) In the formula: and It is a feature point heatmap predicted by the network. H and W These are the height and width of the heatmap, respectively. yes At position ( i , j The corresponding ALIKE predicted label at ) Indicates the position ( i , j At point ), the model predicts that the label belongs to the tag. The probability of the category, The calculation formula is as follows: (12) In the formula: This represents rounding down, for heatmaps. Each position on ( i , j This corresponds to an 8×8 grid area in the grayscale image; If a certain feature point exists ( x m , y m ) ∈A If the data is downsampled into the grid, its relative position within the grid is encoded as an integer between 0 and 63; otherwise, the label is... Setting it to 64 indicates that there are no feature points at this location; 4) Pixel-level offset loss The sampled descriptor and The data is stitched along the channel dimension and then input into a multilayer perceptron (MLP) to calculate the prediction offset confidence matrix. O∈ ; (13) Then, based on the prediction offset confidence matrix O With true offset label l∈ Construct the offset loss function for {0,1,2,...,63}, and use the bidirectional matching confidence of the matched pairs. M i Weighted; (14) In the formula: Let the set of semi-dense points of T1´ be... P ={( x , y )}, l The calculation method is as follows: (15) Ultimately, the overall matching loss function defines the weighted average of the losses for each item.

4. The lightweight remote sensing image change detection method for spatial misalignment as described in claim 1, characterized in that: In the decoding stage, a scale-adaptive perception module is used to fuse differential feature maps at different scales using a dynamic weight fusion method that combines top-down and bottom-up approaches. Finally, bilinear interpolation is used to upsample the feature maps by a factor of 4, and then a 1×1 convolution is used to adjust the number of channels to the number of classes, thereby outputting the change detection results.

5. A lightweight remote sensing image change detection system for spatial misalignment, characterized in that: It includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the program instructions in the memory to execute the lightweight remote sensing image change detection method for spatial misalignment as described in any one of claims 1-4.

6. A computer-readable storage medium, characterized in that, The method includes a readable storage medium on which a computer program is stored, which, when executed, implements a lightweight remote sensing image change detection method for spatial misalignment as described in any one of claims 1-4.

Images