Remote sensing image change detection method based on Siamese network and SAM model

Abstract

The invention discloses a remote sensing image change detection method based on a Siamese network and an SAM model, and belongs to the technical field of remote sensing image change detection. The remote sensing image change detection method based on the Siamese network and the SAM model comprises the following steps: selecting two remote sensing images in a standard data set, and preprocessing the two remote sensing images; inputting the preprocessed remote sensing image into an SAM model for feature extraction, and generating a high-dimensional image; global feature information and difference feature information of the high-dimensional image are enhanced for the image after feature extraction through a global information enhancement module and a difference information enhancement module; features are extracted through the SAM model, global feature information and difference feature information of an image are respectively enhanced in combination with the global information enhancement module and the difference information enhancement module, so that the sensitivity of the model to changes is enhanced, and then the detection capability of subtle changes is improved through fusion of the global feature information and the difference feature information.

Description

Technical Field

[0001] The present invention relates to the technical field of remote sensing image change detection, in particular to a remote sensing image change detection method based on a Siamese network and a SAM model. Background Art

[0002] With the rapid development of remote sensing technology, change detection in remote sensing images has become an important research topic in fields such as environmental monitoring, urban planning, and disaster assessment. Traditional remote sensing image change detection methods typically rely on pixel-based or object-based comparison techniques. While these methods can detect changes to a certain extent, they generally suffer from several issues, such as sensitivity to noise, insufficient processing capabilities for large-scale data, and an inability to effectively capture complex change details.

[0003] In recent years, the rise of deep learning technology has provided new directions for change detection in remote sensing imagery. In particular, deep learning models such as convolutional neural networks (CNNs) and generative adversarial networks (GANs) have achieved remarkable progress in remote sensing image processing. However, most of these methods rely on large amounts of annotated data, and change detection in complex backgrounds remains challenging. Furthermore, existing methods often focus on extracting local features while ignoring the global contextual information in remote sensing imagery. This results in poor robustness and accuracy in complex scenes.

[0004] In recent years, the Siamese network and the Segment Anything Model (SAM), two advanced model structures, have achieved outstanding results in the fields of change detection and image segmentation, respectively. Siamese networks can learn differences between images by sharing parameters and have demonstrated excellent performance in multiple change detection tasks. The SAM model, with its performance in large-scale image segmentation tasks, has successfully achieved automatic segmentation of various objects and possesses strong global context understanding capabilities.

[0005] However, although existing methods have made some progress in remote sensing image change detection, the following shortcomings still exist: 1) Most existing methods do not make sufficient comprehensive use of the global information and difference features in the image; 2) Existing change detection methods lack accuracy and robustness when dealing with complex environmental changes, especially detail changes and edge detection; 3) Existing methods have poor adaptability to the diverse types of remote sensing images and multi-source data, and have not yet achieved effective processing of cross-domain data. Summary of the Invention

[0006] In view of the above problems existing in the existing technology, the present invention is proposed.

[0007] To achieve the above object, the present invention provides the following technical solution: a remote sensing image change detection method based on Siamese network and SAM model, comprising:

[0008] Select two remote sensing images from the standard data set and preprocess the two remote sensing images;

[0009] The preprocessed remote sensing image is input into the SAM model for feature extraction and generates a high-dimensional image;

[0010] The global feature information and difference feature information of the high-dimensional image are enhanced by the global information enhancement module and the difference information enhancement module.

[0011] The global feature information and difference feature information of high-dimensional images are fused using the difference global information fusion module;

[0012] Change detection is performed on the image after fusion of features through the shared weight mechanism in the Siamese network architecture.

[0013] As a further solution of the present invention: preprocessing includes cropping, resizing and standardizing the remote sensing image.

[0014] As a further solution of the present invention: the two remote sensing images are two sets of pictures of the same area but collected at different time points.

[0015] As a further solution of the present invention, the steps of inputting the pre-processed remote sensing image into the SAM model for feature extraction and generating a high-dimensional image include:

[0016] Input the remote sensing image into the convolutional neural network module in the SAM model;

[0017] The convolutional neural network module encodes the spatial information in the image at multiple levels and extracts features at different scales;

[0018] By utilizing the adaptive characteristics of the SAM model, the feature extraction strategy is dynamically adjusted according to different areas of the remote sensing image to capture different types of change information in the image;

[0019] Perform difference enhancement on the extracted feature information to highlight the change areas in the comparison of different images;

[0020] Generate change detection results, mark the changed areas, and construct high-dimensional images.

[0021] As a further solution of the present invention: the difference information enhancement module enhances the accuracy of the image difference feature information after the fusion feature through the following steps, specifically,

[0022] Calculate the pixel-level difference of the two fused features of the remote sensing image to generate a change layer that highlights the changed area;

[0023] Apply spatial smoothing algorithm to smooth the boundaries of the detected change area;

[0024] Combine multi-scale change detection to enhance the extraction of image details.

[0025] As a further solution of the present invention: the step of fusing the global feature information and the difference feature information of the high-dimensional image using the difference global information fusion module includes:

[0026] The difference global information fusion module stitches the global feature information and difference feature information of the high-dimensional image;

[0027] The concatenated features are processed by convolution with three different dilation rates;

[0028] The feature combination images after convolution processing with three different dilation rates are analyzed at multiple scales;

[0029] The residual connection is used to add the global feature information and difference feature information of the dimensional image to the feature information after convolution;

[0030] The feature information after residual connection is optimized through additional convolutional layers, and an image with fused feature information is output.

[0031] As a further solution of the present invention, the three different expansion ratios are 1, 2 and 3 respectively.

[0032] As a further solution of the present invention, the step of performing change detection on the image after fusion of features by using the shared weight mechanism in the Siamese network architecture includes:

[0033] The Siamese network uses a shared weight mechanism to calculate the similarity and difference between image pairs after inputting fusion features;

[0034] According to the calculated similarity value, the change category is determined and the location, type and intensity of the changed area are output;

[0035] Through comparative analysis, the weights and biases in the Siamese network are adjusted.

[0036] As a further solution of the present invention: the image change detection includes change type classification, area type classification and change intensity classification.

[0037] As a further solution of the present invention: change detection is optimized by the following steps, specifically,

[0038] Based on the detection results of the changed areas, the regions are filtered and merged to remove redundant detection results and retain the most relevant changed areas;

[0039] Generate a visualization layer and highlight the changed areas;

[0040] Produce detailed statistical reports of changes.

[0041] Compared with the existing technology, the beneficial effect of the present invention is that: the remote sensing image change detection method based on the Siamese network and the SAM model extracts features through the SAM model, and combines the global information enhancement module and the difference information enhancement module to enhance the global feature information and difference feature information of the image respectively, thereby enhancing the model's sensitivity to changes, and then through the fusion of global feature information and difference feature information, improving the detection ability of subtle changes, and thus being able to perform well on various remote sensing data sets, significantly improving the accuracy of change detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be derived from these drawings without inventive effort. Among them:

[0043] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0044] Figure 2 Schematic diagram of the structure of the global information enhancement module of the present invention.

[0045] Figure 3 Schematic diagram of the structure of the difference information enhancement module of the present invention.

[0046] Figure 4 Schematic diagram of the structure of the difference global information fusion module of the present invention. DETAILED DESCRIPTION

[0047] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0048] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0049] Next, the present invention is described in detail with reference to schematic diagrams. For ease of illustration, when describing embodiments of the present invention, cross-sectional views illustrating device structures may be partially enlarged and not to scale. Furthermore, the schematic diagrams are merely illustrative and should not limit the scope of protection of the present invention. Furthermore, in actual production, three-dimensional dimensions, including length, width, and depth, should be included.

[0050] Furthermore, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it necessarily refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0051] Example 1

[0052] like Figure 1 As shown, the present invention provides a technical solution: a remote sensing image change detection method based on Siamese network and SAM model, comprising:

[0053] S1: Select two remote sensing images from the standard dataset (LEVIR-CD) and preprocess the two remote sensing images;

[0054] It should be noted that the two remote sensing images are two sets of pictures of the same area but collected at different time points; preprocessing includes cropping, resizing and standardization of remote sensing images. Preprocessing is to ensure the consistency of input data and meet the requirements of model training; remote sensing image preprocessing mainly uses Python-related image processing libraries (OpenCV) and transformation functions provided by the deep learning framework (PyTorch) for processing.

[0055] For example, in preprocessing, the image standard is to resize large images to the corresponding size of 256*256, with no overlapping parts between the images; the adjustment is to make the image usable for network training.

[0056] S2: The preprocessed remote sensing image is input into the SAM model for feature extraction and generates a high-dimensional image;

[0057] It should be noted that the SAM model (Segment Anything Model, image feature extraction module) includes the following features: 1. The SAM model can automatically adapt to different types of remote sensing image data, including high-resolution images, low-resolution images, optical images, radar images, etc.; 2. The SAM model can adaptively process different types of geographic area data, such as cities, villages, forests, agriculture, wetlands, etc.; the SAM model is used to extract features from two input remote sensing images to capture global and local detail difference information; the SAM model can extract rich semantic information from images, making subsequent change detection more accurate.

[0058] S3: enhancing the global feature information and difference feature information of the high-dimensional image by a global information enhancement module and a difference information enhancement module;

[0059] It should be noted that the global information enhancement module (GIEM) is used to enhance the global features of the image, while the difference information enhancement module (DIEM) is used to strengthen the difference features in the image; these two modules improve the global perception of the image and the sensitivity of the changed areas respectively.

[0060] S4: using the difference global information fusion module to fuse the global feature information and difference feature information of the high-dimensional image;

[0061] It should be emphasized that the global information is fused with the difference information through the difference global information fusion module (DGIF), further enhancing the ability to detect subtle changes in remote sensing images; this module can improve the recognition accuracy of complex changes, especially in complex backgrounds.

[0062] S5: Change detection is performed on the image after fusion of features through the shared weight mechanism in the Siamese network architecture.

[0063] It should be noted that the fused information is subjected to change classification, and final change detection is performed based on the Siamese network architecture. Through this step, the model can accurately locate the changed areas in the remote sensing image and output change detection results. The Siamese network architecture fully utilizes the global and differential information of the image. After processing, the resulting high-level features are a combination of image differences and global information.

[0064] Furthermore, the pre-processed remote sensing image is input into the SAM model for feature extraction, and the steps of generating a high-dimensional image include:

[0065] S21: Input the remote sensing image into the convolutional neural network module in the SAM model;

[0066] S22: Convolutional neural network module (CNN) performs multi-level encoding on the spatial information in the image and extracts features of different scales;

[0067] S23: Utilizing the adaptive characteristics of the SAM model, dynamically adjusting the feature extraction strategy according to different regions of the remote sensing image to capture different types of change information in the image; wherein, the SAM model processes these regions in an adaptive manner, so as to accurately capture the differences between global changes and local details in the image;

[0068] S24: performing difference enhancement on the extracted feature information to highlight the changed areas in the comparison of different images;

[0069] S25: Generate change detection results, mark the changed areas, construct a high-dimensional image and return to step S21 to continue processing the next pair of input images.

[0070] Furthermore, the steps of performing change detection on the image after fusion of features through the shared weight mechanism in the Siamese network architecture include:

[0071] S51: Using the shared weight mechanism through the Siamese network, the similarity and difference between the image pairs after inputting the fusion features are calculated;

[0072] It should be noted that the shared weight machine means that the two sub-networks in the Siamese network use the same weight parameters when processing two input images. This means that no matter how the two input images are processed in the network, they use the same parameters for feature extraction and processing. Its functions are: 1. Improve computational efficiency; 2. Enhance the generalization ability of learning.

[0073] S52: Determine the change category based on the calculated similarity value and output the location, type and intensity of the changed area;

[0074] S53: Through comparative analysis, adjust the weights and biases in the Siamese network to improve the accuracy of the model in identifying changed areas and ensure that changed areas are not misdetected or missed.

[0075] It should be noted that image change detection includes change type classification, area type classification and change intensity classification; specifically classification by change type, such as: building changes, vegetation changes, water body changes, etc.; classification by area type, such as: urban areas, forest areas, agricultural areas, etc.; classification by change intensity, such as: large-scale changes, local changes, small changes, etc.

[0076] Furthermore, change detection is optimized through the following steps:

[0077] Based on the detection results of the changed areas, the regions are filtered and merged to remove redundant detection results and retain the most relevant changed areas;

[0078] Generate a visualization layer to highlight the changed areas to facilitate subsequent analysis and decision-making by users;

[0079] Produce detailed statistical reports on changes, including information such as change type, change area, and change intensity.

[0080] The specific process is as follows: First, the user prepares two remote sensing images from different time points as input images. These images are processed in parallel using a Siamese network architecture, using SAM as the encoder for feature extraction. The SAM model is frozen during this step to preserve its pre-trained feature extraction capabilities. Through powerful global and local feature extraction, it extracts key features from both images, providing a foundation for subsequent change detection. Next, the model enhances the global information and local difference information in the images using the Global Information Enhancement Module (GIEM) and the Difference Information Enhancement Module (DIEM), respectively. GIEM focuses on capturing large-scale background changes in the image, while DIEM emphasizes detailed changes, providing higher sensitivity in areas with smaller changes. In this process, GIEM and DIEM improve image change detection capabilities from different perspectives. The enhanced global and difference information are then fused using the Difference Global Information Fusion Module (DGIF), integrating global background and subtle changes to produce a more comprehensive and accurate representation of changes. Finally, the fused feature map is input into the Change Classification Module, which outputs a change detection map that identifies the changed areas between the two images. Through visual output, users can intuitively see the areas of change in remote sensing images and their specific features, facilitating further analysis and decision-making. This method is particularly suitable for detecting detailed changes in complex backgrounds, and can provide strong support for applications such as urban expansion, environmental monitoring, and post-disaster assessment. In practical applications, users can adjust the weights of the enhancement module or optimize network parameters according to their needs to improve the accuracy of change detection. Compared with existing methods, the present invention has high efficiency, accuracy, and stronger robustness, and can significantly improve the performance of remote sensing image change detection, especially in subtle changes and complex scenes.

[0081] In summary, compared with the prior art, the present invention has the following advantages:

[0082] 1) Enhanced perception of details and global information: By extracting global and local information through SAM and combining the enhancement characteristics of GIEM and DIEM modules, the present invention can better identify subtle changes in remote sensing images and is particularly suitable for change detection in complex backgrounds.

[0083] 2) Improving the accuracy and robustness of change detection: By integrating global information with difference information, the model not only takes into account large-scale global information during change detection, but also captures local changes in detail, significantly improving the accuracy of change detection. In particular, it can provide high-precision change identification results in monitoring complex scenarios (such as urban expansion and environmental changes).

[0084] 3) Efficient change classification capability: By matching and classifying the extracted features through a Siamese network, the present invention can accurately classify the changed areas in remote sensing images, reduce the false detection rate, and improve the recall rate of change detection.

[0085] 4) Wide applicability and excellent performance: This method can be effectively applied to various remote sensing datasets, and performs well in urban construction, post-disaster assessment, and environmental monitoring, especially in detecting changes in details and low-contrast changes.

[0086] 5) Efficient application in practical scenarios: The present invention is applicable to change monitoring of various remote sensing image data and has broad application prospects, including but not limited to urban expansion, environmental change monitoring, post-disaster assessment and other fields, and can significantly improve the accuracy and reliability of change detection in these scenarios.

[0087] Example 2

[0088] like Figure 2 and 3 As shown, the difference from the previous embodiment is that this embodiment further records a global information enhancement module (GIEM) and a difference information enhancement module (DIEM) to enhance the global feature information and difference feature information of the high-dimensional image.

[0089] Specifically in this embodiment: the Global Information Enhancement Module (GIEM) is designed to comprehensively enhance the global information in the feature representation, thereby improving the robustness of the model under large-scale changes, especially in application scenarios such as urban expansion and natural disaster assessment. Through three attention mechanisms - pixel attention (Pixel Attention), channel attention (Channel Attention) and simple pixel attention (Simple Pixel Attention), the global contextual information in the image is effectively enhanced; these attention mechanisms work together to help the network further enhance the global information of the image on the basis of feature extraction, thereby enhancing the model's ability to perceive changes.

[0090] Among them, the Pixel Attention mechanism analyzes the spatial information in the image to highlight the areas most sensitive to changes. This mechanism emphasizes areas of greater spatial importance by assigning different weights to each pixel. This method can effectively identify areas with potential changes in the image, especially in complex backgrounds, and can pay more attention to the details of the changes. The main goal of this mechanism is to utilize the local correlation of spatial features and, through a weighted approach, to give more attention to key pixels, thereby helping to detect more subtle changes.

[0091] Among them, the goal of the Channel Attention mechanism is to strengthen the most discriminative channel features by evaluating the importance of each feature channel; each channel will automatically adjust its weight according to its contribution to the global feature, suppressing the interference of feature channels irrelevant to change detection; in this way, the network can extract the most important feature information from multiple channels, effectively filter out redundant information, and enhance the accuracy and reliability of change detection.

[0092] Among them, the Simple Pixel Attention mechanism is a low-complexity pixel-level feature enhancement method that aims to accurately adjust the feature representation of each pixel with less computational overhead. It enhances its distinguishing ability in change detection tasks by weighting the features of each pixel. Although the mechanism is relatively simple, it can effectively improve the expressive power of feature maps, ensuring improved pixel-level accuracy while maintaining computational efficiency.

[0093] To further optimize the efficiency of the global information enhancement module, point convolution (PWConv) was introduced. By performing 1×1 convolution, PWConv can significantly reduce computational costs while maintaining the ability to process feature maps across channels. This operation not only reduces the amount of computation but also enhances the module's efficiency when processing high-dimensional features, making it particularly suitable for large-scale remote sensing image change detection tasks. The introduction of PWConv enables GIEM to perform global feature enhancement in a lightweight manner while avoiding excessive consumption of computing resources.

[0094] The GIEM module combines the three aforementioned attention mechanisms with PWConv to enhance global information while ensuring efficient computational performance. By strengthening global context, GIEM improves the network's sensitivity to subtle changes in remote sensing imagery, thereby enhancing the accuracy and robustness of change detection. GIEM is particularly effective in dealing with complex backgrounds and large-scale changes, providing more accurate detection results.

[0095] Furthermore, the difference information enhancement module (DIEM) enhances the accuracy of the image difference feature information after the fusion feature through the following steps, specifically,

[0096] Calculate the pixel-level difference of the two fused features of the remote sensing image to generate a change layer that highlights the changed area;

[0097] Apply spatial smoothing algorithm to smooth the boundaries of the detected change areas to avoid over-segmentation or omission;

[0098] Combined with multi-scale change detection methods, it enhances the extraction of image details, especially when dealing with complex changes, while maintaining high precision and high recall.

[0099] It should be noted that DIEM (Difference Information Enhancement Module) aims to enhance the expression of difference features by capturing and emphasizing the differences in the horizontal and vertical directions of the input feature map; this goal is achieved by adopting the Coordinate Attention Mechanism. The specific method is as follows Figure 3 shown.

[0100] 31) Input Feature Splitting and Processing: DIEM first uses Adaptive Average Pooling to split the input feature map into horizontal and vertical components. This step ensures that the module focuses on local information in the image in different directions, rather than just changes in a single direction. After pooling these components, feature representations in two directions are obtained, which facilitates the subsequent operation to better capture changes in both directions.

[0101] 32) Feature splicing and convolution processing: The split horizontal and vertical components are spliced together and processed through shared convolution and activation layers. The convolution operation can extract higher-level spatial features, and the activation layer further enhances the nonlinear expression capability. Through this processing, DIEM can learn more detailed difference information from features in different directions while ensuring efficient processing.

[0102] 33) Generation and Application of Directional Differences: After feature processing, the module re-splits the processed features into horizontal and vertical representations. By generating directional attention weights, DIEM weights the features in each direction to highlight significant spatial differences. In this way, the module can effectively enhance the difference information of the input image in different directions, thereby improving the performance of change detection.

[0103] 34) Dynamic Adjustment and Precise Representation: The advantage of the coordinate attention mechanism is that it can dynamically adjust the feature response, focusing on significant changes in the horizontal or vertical direction according to the specific conditions of the input feature map; this approach enables DIEM to flexibly respond to different types of changes while maintaining high computational efficiency.

[0104] 35) Through this coordinate attention mechanism, DIEM not only enhances the ability to distinguish differential features but also improves the network's sensitivity to subtle changes in the image. In the task of change detection in remote sensing images, DIEM effectively emphasizes the differences between images at two time points, enabling the network to identify and locate the changed areas. This approach is particularly suitable for processing remote sensing images with complex backgrounds or detailed changes, and can improve the accuracy of change detection.

[0105] In summary, DIEM, by focusing on horizontal and vertical differences, not only enhances the model's discriminative capabilities but also ensures efficient computation through a coordinate attention mechanism. This module accurately captures significant changes in different directions, improving overall performance in remote sensing image change detection tasks. Whether it's subtle changes in static backgrounds or significant changes in dynamic scenes, DIEM plays a key role in helping the network better extract valuable change information from the input image.

[0106] Difference Information Enhancement Module (DIEM): DIEM focuses on the details of changing regions in an image, enabling the model to accurately identify subtle changes. This module can effectively reduce false detections caused by complex backgrounds in environmental monitoring and post-disaster assessments.

[0107] Example 3

[0108] The difference from the previous embodiment is that the difference global information fusion module (DGIF) is further described to effectively combine global features and difference features, capturing a wide range of contextual information while highlighting local changes. This module ensures the capture of fine-grained image changes and large-scale patterns through a multi-scale convolutional structure, while avoiding a significant increase in computational costs. Its overall design is as follows Figure 4 As shown;

[0109] Specifically, the steps of using the difference global information fusion module (DGIF) to fuse the global feature information and difference feature information of the high-dimensional image are as follows:

[0110] S41: The difference global information fusion module (DGIF) stitches the global feature information and difference feature information of the high-dimensional image;

[0111] S42: The concatenated features are processed by convolution with three different dilation rates;

[0112] Among them, three different expansion rates are set to 1, 2 and 3 respectively; the introduction of dilated convolution enables the convolution layer to expand the receptive field and capture features at different scales without significantly increasing the amount of computation; convolution with a dilation rate of 1 focuses on local features within a smaller range, while convolutions with dilation rates of 2 and 3 capture larger-scale patterns and more macroscopic global features, respectively; this multi-scale convolution design can simultaneously extract fine-grained changes and extensive background information, thereby enhancing the model's perception of changes at different scales.

[0113] S43: Analyzing the feature combination images after convolution processing with three different expansion rates at multiple scales;

[0114] It should be noted that three convolutions with different expansion rates combine the processed features and analyze the input image at multiple scales. This processing method can effectively capture subtle changes in a small range, while also capturing structural changes at larger scales. Through parallel convolution, the DGIF module not only improves the accuracy of capturing changes, but also enhances the global understanding of the change pattern.

[0115] S44: using residual connection to add the global feature information and difference feature information of the dimensional image to the feature information after convolution;

[0116] Specifically, to further optimize the feature fusion process, the DGIF module employs residual connections, which add the original global and differential features (including their scaled versions) to the convolved features. Residual connections are useful for: first, preserving the original global and differential information, ensuring that the network does not lose critical information; second, the residual structure helps maintain gradient flow during training, making model training more stable. This design effectively avoids information loss or vanishing gradients during training, improving the module's stability and effectiveness.

[0117] S45: The feature information after residual connection is optimized through an additional convolutional layer, and an image of the fused feature information is output.

[0118] It should be noted that the features after residual connection are further optimized through additional convolutional layers to ensure that the fused features can fully adapt to downstream tasks; through these optimization steps, the final output generated by the DGIF module has higher quality, providing more accurate feature representation for subsequent change detection.

[0119] Among them, the DGIF module ensures the full utilization of global and differential information by combining multi-scale convolution, residual connection and efficient feature fusion methods; compared with traditional single-scale processing methods, the DGIF module performs better in capturing subtle changes and large-scale patterns, making the network more robust and accurate in remote sensing image change detection tasks; ultimately, the DGIF module can not only capture changes at different scales, but also effectively fuse global and differential information, providing stronger feature expression capabilities for change detection.

[0120] In summary, the Differential Global Information Fusion Module (DGIF) maximizes the use of global and differential information through an innovative combination of multi-scale convolutions, residual connections, and an efficient fusion strategy, providing a more robust and efficient feature representation for remote sensing image change detection. Through optimization of this module, the network excels in a variety of remote sensing image change scenarios, accurately detecting both subtle image changes and large-scale pattern changes.

[0121] The DGIF module improves the accuracy of distinguishing changed areas by weighted fusion of global information and difference information, and can effectively handle complex and detailed changes in remote sensing images.

[0122] Example 4

[0123] Different from the previous embodiment, this embodiment provides specific experiments to illustrate the effect of this solution. The specific experiments are as follows:

[0124] The experiments were conducted using an NVIDIA GeForce RTX 3090 GPU produced in Taiwan, China; the implementation was based on PyTorch (version 2.2.1) with CUDA 12.1 as the underlying framework.

[0125] During training, the batch size is set to 4, the initial learning rate is set to 0.00005, and the minimum learning rate is set to 0.000001; the model is trained for a total of 100 epochs, and the learning rate is updated using the cosine annealing scheme.

[0126] Comparative experiments were conducted on the LEVIR-CD dataset, comparing the proposed Siamese-SAM model with multiple other models. All methods were trained using the same method to ensure fairness and reliability of the results. The following table lists the detailed quantification of the model evaluation metrics.

[0127]

[0128]

[0129] From the results, we can see that among the deep learning-based change detection methods, FC-EF performs the worst, with F1 and IoU scores of only 80.98% and 68.03% respectively. Compared with SAM-CD, the F1 score of Siamese-SAM in our scheme is improved by 1.22% and the IoU is improved by 2.1%.

[0130] Through the above comparative experiments, it is not difficult to find that the remote sensing image change detection method based on the Siamese network and the SAM model of the present invention extracts features through the SAM model, and combines the global information enhancement module and the difference information enhancement module to enhance the global feature information and difference feature information of the image respectively, thereby enhancing the model's sensitivity to changes, and then through the fusion of global feature information and difference feature information, improves the ability to detect subtle changes, and thus can perform well on various remote sensing data sets, significantly improving the accuracy of change detection.

[0131] It is important to note that the configuration and arrangement of the present application, as illustrated in various exemplary embodiments, are illustrative only. Although only a few embodiments are described in detail in this disclosure, those reading this disclosure will readily appreciate that numerous modifications are possible without materially departing from the novel teachings and advantages of the subject matter described herein. For example, variations in the size, dimensions, structure, shape, and proportions of various components, as well as parameter values such as temperature, pressure, mounting arrangements, use of materials, color, and orientation, are possible. For example, components shown as integrally formed may be constructed from multiple parts or components, the positions of components may be inverted or otherwise altered, and the nature, number, or position of discrete components may be modified or changed. Therefore, all such modifications are intended to be encompassed within the scope of this invention. The order or sequence of any process or method steps may be altered or reordered according to alternative embodiments. In the claims, any "means-plus-function" clause is intended to cover structures that perform the functions described herein, and not only structural equivalence but also equivalent structures. Other substitutions, modifications, changes, and omissions may be made in the design, operating conditions, and arrangement of the exemplary embodiments without departing from the scope of this invention. Therefore, the invention is not limited to the specific embodiments, but extends to various modifications that still fall within the scope of the appended claims.

[0132] Furthermore, in order to provide a concise description of the exemplary embodiments, all features of the actual embodiment may not be described, i.e., those features that are not relevant to the best mode presently contemplated for carrying out the invention or those that are not relevant to implementing the invention.

[0133] It should be understood that in the development of any actual embodiment, as in any engineering or design project, numerous implementation-specific decisions may be made. Such a development effort may be complex and time-consuming, but for those of ordinary skill having the benefit of this disclosure, the development effort will be a routine task of design, fabrication, and production without undue experimentation.

[0134] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A remote sensing image change detection method based on Siamese network and SAM model, characterized by: include, Select two remote sensing images from the standard data set and preprocess the two remote sensing images; The preprocessed remote sensing image is input into the SAM model for feature extraction and generates a high-dimensional image; The global feature information and difference feature information of the high-dimensional image are enhanced by the global information enhancement module and the difference information enhancement module. The global feature information and difference feature information of high-dimensional images are fused using the difference global information fusion module; Change detection is performed on the image after fusion of features through the shared weight mechanism in the Siamese network architecture.

2. The remote sensing image change detection method based on Siamese network and SAM model according to claim 1, characterized in that: Preprocessing includes cropping, resizing and normalization of remote sensing images.

3. The remote sensing image change detection method based on Siamese network and SAM model according to claim 2, characterized in that: The two remote sensing images are two sets of pictures of the same area but collected at different time points.

4. The remote sensing image change detection method based on Siamese network and SAM model according to any one of claims 1 to 3, characterized in that: The pre-processed remote sensing image is input into the SAM model for feature extraction, and the steps of generating a high-dimensional image include: Input the remote sensing image into the convolutional neural network module in the SAM model; The convolutional neural network module encodes the spatial information in the image at multiple levels and extracts features at different scales; By utilizing the adaptive characteristics of the SAM model, the feature extraction strategy is dynamically adjusted according to different areas of the remote sensing image to capture different types of change information in the image; Perform difference enhancement on the extracted feature information to highlight the change areas in the comparison of different images; Generate change detection results, mark the changed areas, and construct high-dimensional images.

5. The remote sensing image change detection method based on Siamese network and SAM model according to claim 4, characterized in that: The difference information enhancement module enhances the accuracy of the image difference feature information after fusion features through the following steps: Calculate the pixel-level difference of the two fused features of the remote sensing image to generate a change layer that highlights the changed area; Apply spatial smoothing algorithm to smooth the boundaries of the detected change area; Combine multi-scale change detection to enhance the extraction of image details.

6. The remote sensing image change detection method based on Siamese network and SAM model according to claim 5, characterized in that: The steps of fusing the global feature information and the difference feature information of the high-dimensional image using the difference global information fusion module include: The difference global information fusion module stitches the global feature information and difference feature information of the high-dimensional image; The concatenated features are processed by convolution with three different dilation rates; The feature combination images after convolution processing with three different dilation rates are analyzed at multiple scales; The residual connection is used to add the global feature information and difference feature information of the dimensional image to the feature information after convolution; The feature information after residual connection is optimized through additional convolutional layers, and an image with fused feature information is output.

7. The remote sensing image change detection method based on Siamese network and SAM model according to claim 6, characterized in that: The three different expansion ratios are 1, 2 and 3.

8. The remote sensing image change detection method based on Siamese network and SAM model according to any one of claims 5 to 7, characterized in that: The steps of detecting changes in the image after fusion of features through the shared weight mechanism in the Siamese network architecture include: The Siamese network uses a shared weight mechanism to calculate the similarity and difference between image pairs after inputting fusion features; According to the calculated similarity value, the change category is determined and the location, type and intensity of the changed area are output; Through comparative analysis, the weights and biases in the Siamese network are adjusted.

9. The remote sensing image change detection method based on Siamese network and SAM model according to claim 8, characterized in that: The image change detection includes change type classification, area type classification and change intensity classification.

10. The remote sensing image change detection method based on Siamese network and SAM model according to claim 9, characterized in that: Change detection is optimized through the following steps, specifically, Based on the detection results of the changed areas, the regions are filtered and merged to remove redundant detection results and retain the most relevant changed areas; Generate a visualization layer and highlight the changed areas; Produce detailed statistical reports of changes.

Images