Remote sensing change detection method and system based on foreground and background separation and historical feature multiplexing

Abstract

The present application relates to the technical field of remote sensing image change detection, and provides a remote sensing change detection method and system based on foreground and background separation and historical feature reuse. The method comprises the following steps: based on a pair of preprocessed double-time-phase remote sensing images, using a ChangeMamba encoder, four features of different preset scales are obtained for each double-time-phase remote sensing image; based on the fourth layer foreground feature and the fourth layer background feature, using a foreground and background separation module, after fusion, the foreground and background are separated by inputting a hollow space pyramid pooling module to generate a foreground mask; based on the foreground mask, the third layer foreground feature and the third layer background feature, using an attention module, two parallel operations are performed to respectively generate foreground enhancement features and background suppression features, and weighted fusion is performed to obtain a foreground change map and a background change map; based on the foreground change map and the background change map, using a ChangeMamba decoder, change information of a to-be-detected region is obtained.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image change detection technology, and in particular to a remote sensing change detection method and system based on foreground and background separation and historical feature reuse. Background Technology

[0002] With increasingly stringent requirements for ecological environment monitoring in nature reserves, remote sensing image change detection technology has become an indispensable tool for ecological monitoring due to its advantages of large-scale and periodic observation. This technology accurately identifies land cover changes by analyzing remote sensing images of the same area at different times, providing a scientific basis for ecological environment assessment, human activity monitoring, and evaluation of ecological restoration effectiveness.

[0003] In recent years, deep learning technology has greatly promoted the development of remote sensing change detection. From early methods based on convolutional neural networks to the Transformer-based architecture in recent years, change detection models have continuously improved in accuracy and efficiency. However, these methods still have limitations in dealing with the unique challenges of nature reserves: the changed areas within nature reserves are usually small in proportion and scattered, contrasting sharply with the large areas of stable background; at the same time, the historical data accumulated over long-term monitoring contains rich temporal variation patterns, but existing methods have failed to effectively mine and utilize this information.

[0004] In current technological developments, visual backbone networks based on state-space models have demonstrated significant advantages. Among them, the VMamba (Visual State Space Model) is a representative work, providing a new solution for visual tasks through an innovative two-dimensional selective scanning mechanism. The core of this approach lies in processing two-dimensional image data through a multi-directional scanning strategy, achieving effective spatial modeling while maintaining linear computational complexity. Specifically, this approach employs a hierarchical network architecture, constructing multi-level feature representations by stacking multiple visual state space modules. Within each module, the core two-dimensional selective scanning mechanism handles spatial features: first, the input feature map is unfolded into sequence data along four different directions (including horizontal, vertical, and two diagonal directions), ensuring that each location can establish connections with neighboring regions in each direction. During the sequence data processing stage, the sequences generated in each direction are independently analyzed by a dedicated S6 module. The S6 module can dynamically adjust its internal parameters according to the input content, thereby flexibly capturing important information features in the sequence. After processing, the sequence results from each direction are recombine into a two-dimensional feature map, achieving effective fusion of multi-directional information. This scheme performs well in visual tasks such as image classification and object detection, especially in processing high-resolution images, demonstrating good computational efficiency and providing a new technical path for visual representation learning.

[0005] While existing solutions perform well in general vision tasks, they still have significant technical shortcomings in the specific application scenario of remote sensing change detection in nature reserves. First, existing solutions lack a foreground-background differentiation processing mechanism specifically for change detection. In nature reserve monitoring, the actual change areas are often small and scattered, while existing methods perform indiscriminate feature processing on the entire image, making it difficult to focus on subtle changes. This makes the model susceptible to interference from large areas of stable background, leading to false positives and false negatives. Second, existing solutions fail to effectively utilize the temporal change patterns contained in historical remote sensing data. The large amount of historical imagery accumulated from nature reserve monitoring contains valuable information on land cover evolution, such as typical change patterns and the distribution characteristics of change intensity. However, existing methods typically employ an isolated analysis mode, processing only the current two-phase images and ignoring the reference information provided by historical data. This limitation not only wastes information resources but also prevents the model from identifying false negatives due to incomplete observation time, such as missing drastic changes at the beginning of a monitoring period. Summary of the Invention

[0006] To address the technical problems existing in the background art, this invention provides a remote sensing change detection method and system based on foreground / background separation and historical feature reuse. This invention employs a change detection scheme capable of achieving refined foreground / background separation processing. By designing a dedicated foreground / background separation module, combined with multi-scale feature fusion and spatial context modeling, adaptive and accurate separation of the foreground and background regions is achieved. Furthermore, foreground masks guide the change detection process, and attention mechanisms are used to differentially enhance foreground features and suppress background interference, significantly improving the accuracy of identifying truly changed regions.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] The first aspect of the present invention provides a remote sensing change detection method based on foreground and background separation and historical feature reuse.

[0009] A remote sensing change detection method based on foreground / background separation and historical feature reuse includes:

[0010] Acquire a pair of two-temporal remote sensing images of the area to be detected; the pair of two-temporal remote sensing images includes a foreground remote sensing image and a background remote sensing image;

[0011] Based on preprocessed dual-temporal remote sensing image pairs, a ChangeMamba encoder is used to obtain four features at different preset scales for each dual-temporal remote sensing image. Based on the fourth-layer foreground and fourth-layer background features, a foreground-background separation module is used. After fusion, the features are input into a hollow spatial pyramid pooling module for foreground-background separation to generate a foreground mask. Based on the foreground mask, the third-layer foreground features, and the third-layer background features, an attention module is used to perform two parallel operations to generate foreground enhancement features and background suppression features, which are then weighted and fused to obtain foreground change maps and background change maps. Based on the foreground change maps and background change maps, a ChangeMamba decoder is used to obtain change information of the region to be detected. Based on the fourth-layer foreground and fourth-layer background features, a feature storage module is used to calculate the difference vector, and then perform centering and dimensionality reduction processing to obtain dimensionality-reduced temporal features and dimensionality-reduced difference features. Clustering is performed on the dimensionality-reduced temporal features and dimensionality-reduced difference features to obtain classification results. The classification results are then used to determine if any features are missed.

[0012] Furthermore, each dual-temporal remote sensing image yields four features at different preset scales, including: based on the foreground remote sensing image, a first layer of foreground features, a second layer of foreground features, a third layer of foreground features, and a fourth layer of foreground features; based on the background remote sensing image, a first layer of background features, a second layer of background features, a third layer of background features, and a fourth layer of background features.

[0013] Furthermore, based on the fourth-layer foreground features and the fourth-layer background features, a foreground-background separation module is used to perform fusion, and then the fusion is input into a hollow space pyramid pooling module for foreground-background separation to generate a foreground mask; the method includes:

[0014] The fourth-layer foreground features and the fourth-layer background features are input into the foreground-background separation module to calculate the absolute difference, squared difference, and cosine similarity. The absolute difference, squared difference, and cosine similarity are then concatenated and compressed to obtain the first compressed feature. The first compressed feature is then input into the feature fusion module for residual connection to obtain the fused feature.

[0015] The fused features are input into the hollow spatial pyramid pooling module. Through four hollow convolutional layers with different dilation rates, features at four different scales of the fused features are extracted. These four features at different scales are concatenated to obtain concatenated features. The concatenated features are compressed to obtain second compressed features. The second compressed features are input into the upsampling layer and the segmentation head to obtain the foreground mask.

[0016] Furthermore, during the training phase, after obtaining the foreground mask, a total loss function is used to compare the foreground mask with the ground truth labels to optimize the parameters of the ChangeMamba encoder; the total loss function is expressed by the following formula:

[0017]

[0018] in, Indicates focal loss; Indicates die loss; Indicates boundary loss; , and These are adjustable weighting coefficients.

[0019] Furthermore, based on the foreground mask, the third-layer foreground features, and the third-layer background features, an attention module is used to perform two parallel operations to generate foreground enhancement features and background suppression features, which are then weighted and fused to obtain foreground change maps and background change maps; the method includes:

[0020] The parallel operation is as follows: input the foreground mask, the third-layer foreground features, and the third-layer background features into the attention module, perform feature concatenation and channel dimensionality reduction, and perform residual connection between the obtained features and the foreground mask to obtain foreground enhancement features; based on the foreground mask, use the adaptive weight module to calculate the foreground weights;

[0021] Another parallel operation is as follows: Based on the foreground mask, calculate the background mask; input the background mask, the third-layer foreground features and the third-layer background features into the attention module, perform feature concatenation and channel dimensionality reduction, and perform residual connection between the obtained features and the background mask to obtain the background suppression features;

[0022] The foreground enhancement features, foreground weights, and background suppression features are weighted and fused to obtain weighted fused features. The weighted fused features are then residually connected with the third-layer foreground features and the third-layer background features to obtain foreground change maps and background change maps.

[0023] Furthermore, based on the fourth-layer foreground and fourth-layer background features, a feature storage module is used to calculate the difference vector, and then centering and dimensionality reduction are performed to obtain the dimensionality-reduced temporal features and dimensionality-reduced difference features; the method includes:

[0024] The large storage module stores the fourth-layer background features and the feature difference between the fourth-layer background features and the fourth-layer foreground features for each input.

[0025] The PCA model was initially trained using a public dataset to obtain a fixed projection matrix and mean.

[0026] Based on the difference between the fourth-layer background features and the mean, the centered post-temporal features are obtained; based on the difference between the feature differences and the mean, the centered difference features are obtained.

[0027] The centered post-temporal features and difference features are multiplied by the projection matrix to obtain the dimensionality-reduced post-temporal features and dimensionality-reduced difference features.

[0028] Furthermore, clustering is performed on the dimensionality-reduced temporal features and dimensionality-reduced difference features to obtain classification results. The method includes: inputting the dimensionality-reduced temporal features and dimensionality-reduced difference features into the clustering module, performing clustering calculations respectively, and obtaining two types of memory databases, each containing several pre-classified clusters; the database classified using the dimensionality-reduced temporal features is the state database, and the database classified using the dimensionality-reduced difference features is the change database, which is classified according to the stable state and change pattern of the subsequent temporal phases respectively; thus, the classification results are obtained.

[0029] After obtaining the classification results, the background features of the fourth layer and the feature differences between the background features of the fourth layer and the foreground features of the fourth layer are stored in their respective positions, and relevant information is calculated.

[0030] Furthermore, the classification results are used to determine if any false negatives have been detected; methods include:

[0031] Obtain the cluster ID from the change pattern memory generated by the clustering module, and use this cluster ID to find the cluster to which the input feature pair belongs. Then, obtain the cluster ID from the state memory generated in the clustering module and use this cluster ID to find the cluster to which the input feature pair belongs. Find The distribution of changes, through the distribution of changes and The degree of suspicion is calculated based on the drastic change attribute of the cluster. If the threshold is reached, it is judged as a missed problem; if it is not reached, it is normal.

[0032] A second aspect of the present invention provides a remote sensing change detection system based on foreground and background separation and historical feature reuse.

[0033] A remote sensing change detection system based on foreground / background separation and historical feature reuse includes:

[0034] The data acquisition unit is used to acquire a pair of dual-temporal remote sensing images of the area to be detected; the pair of dual-temporal remote sensing images includes a foreground remote sensing image and a background remote sensing image;

[0035] The image analysis and processing unit, based on preprocessed dual-temporal remote sensing image pairs, employs a ChangeMamba encoder to obtain four features at different preset scales for each dual-temporal remote sensing image. Based on the fourth-layer foreground and background features, a foreground / background separation module is used to perform fusion, and the resulting data is input into a hollow spatial pyramid pooling module for foreground / background separation, generating a foreground mask. Based on the foreground mask, the third-layer foreground features, and the third-layer background features, an attention module performs two parallel operations to generate foreground enhancement features and background suppression features, which are then weighted and fused to obtain foreground and background change maps. Based on the foreground and background change maps, a ChangeMamba decoder is used to obtain change information in the region to be detected. Based on the fourth-layer foreground and background features, a feature storage module calculates the difference vector, performs centering and dimensionality reduction processing, obtaining dimensionality-reduced temporal features and dimensionality-reduced difference features. Clustering is performed on the dimensionality-reduced temporal features and dimensionality-reduced difference features to obtain classification results. Finally, the classification results are used to determine if any false negatives have been detected.

[0036] A third aspect of the present invention provides a computer device comprising:

[0037] A processor, adapted to execute computer programs;

[0038] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in the remote sensing change detection method based on foreground / background separation and historical feature reuse as described in the first aspect above.

[0039] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and to execute steps in the remote sensing change detection method based on foreground-background separation and historical feature reuse as described in the first aspect above.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] This invention provides a remote sensing change detection method and system based on foreground / background separation and historical feature reuse. It mainly includes two core innovations: First, a remote sensing image change detection network based on foreground / background separation is proposed. Building upon the ChangeMamba model, accurate foreground / background separation is achieved through a fusion module, a Spatial Pyramid Pooling (ASPP) module, and dual prediction heads. Furthermore, a specially designed composite loss function of Focal Loss, Dice Loss, and Boundary Loss is used for coordinated optimization, effectively improving the separation accuracy of the foreground (changed areas) and background (unchanged areas). Second, a method using foreground masks to guide the generation of change maps is designed. Through foreground attention, background attention, and adaptive weight modules, the generated foreground mask is used to differentially enhance foreground features and suppress background features, thereby significantly improving the accuracy of change detection.

[0042] This invention establishes a complete system for storing and intelligently reusing historical features. By designing efficient feature dimensionality reduction and clustering methods, combined with a hybrid storage strategy, it achieves effective organization and rapid retrieval of historical features. A reasoning mechanism based on historical change patterns is constructed, which intelligently compares the current detection task with the historical feature database to identify abnormal change patterns, warn of potential missed detection risks, and improve the ability to analyze changes over time. Attached Figure Description

[0043] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0044] Figure 1 This is a flowchart illustrating a remote sensing change detection method based on foreground and background separation and historical feature reuse, as shown in an embodiment of the present invention.

[0045] Figure 2 This is a flowchart of another embodiment of the remote sensing change detection method based on foreground and background separation and historical feature reuse shown in the embodiments of the present invention;

[0046] Figure 3 This is a schematic diagram of one embodiment of the remote sensing image change detection network shown in this invention.

[0047] Figure 4 This is a schematic diagram of the foreground and background separation module in the remote sensing image change detection network shown in an embodiment of the present invention;

[0048] Figure 5 This is a schematic diagram of the attention module in a remote sensing image change detection network according to an embodiment of the present invention;

[0049] Figure 6This is a schematic diagram of the storage module in the remote sensing image change detection network shown in an embodiment of the present invention;

[0050] Figure 7 This is a schematic diagram of an embodiment of a remote sensing change detection system based on foreground and background separation and historical feature reuse, as shown in this invention.

[0051] Figure 8 This is an example diagram of the first phase remote sensing image in a spatially registered dual-temporal remote sensing image pair, as shown in an embodiment of the present invention.

[0052] Figure 9 Is with Figure 8 The corresponding post-temporal remote sensing images together constitute a two-temporal image pair that completes spatial registration;

[0053] Figure 10 It is Figure 8 and Figure 9 The diagram shows the result of image change in the area to be detected after the two-temporal image pairs are input into the remote sensing image change detection network.

[0054] Figure 11 This is a structural diagram of a computer device shown in an embodiment of the present invention. Detailed Implementation

[0055] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0056] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0057] To facilitate understanding of the technical solutions of this invention, some technical terms involved in this invention will be introduced below.

[0058] Dual-temporal remote sensing image pairs / image pairs: These are two remote sensing images / images of the same geographic area at two different times. They consist of a first-temporal remote sensing image / image and a second-temporal remote sensing image / image, with the first-temporal remote sensing image / image being captured earlier than the second-temporal remote sensing image / image.

[0059] Spatial registration refers to the alignment of remote sensing images / images from different time periods of the same geographic area to ensure that images / images from different time periods of the same geographic area are aligned.

[0060] ChangeMamba: This refers to the existing remote sensing change detection network upon which this invention is based. It captures long-range spatiotemporal dependencies through a four-layer feature extraction layer of the encoder, a Cross-Scan mechanism, and the S6 operator, thereby achieving dual-temporal feature difference modeling.

[0061] Foreground mask: A probability map corresponding to the spatial size of the input image, used to mark regions in the image that may change.

[0062] Dilated convolution injects zero elements into a standard convolution, expanding the receptive field of the convolution kernel without increasing the number of parameters or sacrificing spatial resolution. The effective receptive field of an X×X dilated convolution with a dilation rate r=k is... ,in For example, a 3×3 dilated convolution with a dilation rate r=6 has a receptive field of 13×13.

[0063] WHU-CD is a building change detection dataset for a certain region, spanning from 2012 to 2016, primarily covering the reconstruction process following a major earthquake in the city. The dataset contains approximately 15,000 image pairs with a spatial resolution of 0.075 m / pixel.

[0064] LEVIR-CD is a large-scale, high-quality change detection dataset from Google Earth satellite imagery, covering multiple cities in Texas, USA, spanning 5 to 14 years. It consists of 637 1024×1024 image pairs with a resolution of 0.5m / pixel.

[0065] The neighborhood radius is a value used to define the neighborhood in the DBSCAN clustering algorithm. It is one of the core parameters of this algorithm. If the Euclidean distance between feature A and feature B is within the neighborhood radius, then feature B is said to be within the neighborhood of feature A. The value of this parameter differs in the "second clustering stage" and the "third clustering stage".

[0066] MinPts, the minimum number of cluster points, is a parameter used to define cluster points in the DBSCAN clustering algorithm. It is one of the core parameters of the algorithm. If the number of features in the neighborhood of a certain feature is greater than or equal to MinPts, then that feature is called a cluster point. The value of this parameter is different in the "second clustering stage" and the "third clustering stage".

[0067] Figure 1 This is a flowchart illustrating a remote sensing change detection method based on foreground / background separation and historical feature reuse, as shown in an embodiment of the present invention; see reference. Figure 1 The method includes:

[0068] Acquire a pair of two-temporal remote sensing images of the area to be detected; the pair of two-temporal remote sensing images includes a foreground remote sensing image and a background remote sensing image;

[0069] Based on preprocessed dual-temporal remote sensing image pairs, a ChangeMamba encoder is used to obtain four features at different preset scales for each dual-temporal remote sensing image. Based on the fourth-layer foreground and fourth-layer background features, a foreground-background separation module is used. After fusion, the features are input into a hollow spatial pyramid pooling module for foreground-background separation to generate a foreground mask. Based on the foreground mask, the third-layer foreground features, and the third-layer background features, an attention module is used to perform two parallel operations to generate foreground enhancement features and background suppression features, which are then weighted and fused to obtain foreground change maps and background change maps. Based on the foreground change maps and background change maps, a ChangeMamba decoder is used to obtain change information of the region to be detected. Based on the fourth-layer foreground and fourth-layer background features, a feature storage module is used to calculate the difference vector, and then perform centering and dimensionality reduction processing to obtain dimensionality-reduced temporal features and dimensionality-reduced difference features. Clustering is performed on the dimensionality-reduced temporal features and dimensionality-reduced difference features to obtain classification results. The classification results are then used to determine if any features are missed.

[0070] In the specific application scenario of nature reserves, this invention effectively suppresses background noise interference and enhances the focus on areas of real change through the synergy of a foreground / background separation module and an attention guidance mechanism. Simultaneously, by efficiently storing and reusing historical features, a memory mechanism for continuously accumulating and utilizing change detection knowledge is established, significantly improving the ability to analyze long-term temporal changes. This solution organically combines refined foreground / background processing with the reuse of historical experience, providing a complete and efficient technical solution for addressing false change interference and long-term monitoring omissions in change detection in nature reserves.

[0071] Figure 2 This is a flowchart of another embodiment of the remote sensing change detection method based on foreground and background separation and historical feature reuse, as shown in this invention; see also... Figure 2 The remote sensing change detection method based on foreground and background separation and historical feature reuse adopts... Figure 3 The remote sensing image change detection network shown includes: a ChangeMamba encoder, a foreground / background separation module, an attention module, a feature storage module, and a ChangeMamba decoder.

[0072] First, a training dataset needs to be constructed. The WHU-CD, LEVIR-CD, and SYSU-CD seasonal change detection datasets are used as the base datasets for training. For nature reserves, remote sensing data from two time phases of the same nature reserve area can be selected from Gaofen-1 (GF-1), Gaofen-2 (GF-2), and Gaofen-6 (GF-6). These satellite data have high spatial resolution and wide coverage, making them very suitable for change detection tasks in nature reserves. After acquiring image pairs, spatial registration and image preprocessing are performed on the dataset. Spatial registration and image enhancement of the image pairs are achieved through steps such as feature point extraction, feature matching, geometric transformation estimation, radiometric normalization, and resampling.

[0073] The network is then initialized. The ChangeMamba encoder is initialized using VMamba model weights pre-trained on the ImageNet large image dataset. Specifically, the pre-trained weight file is loaded, and the pre-trained parameters are assigned to each layer of the encoder. Since the foreground / background separation and attention modules are new modules, no pre-trained weights are available; therefore, the convolutional layer weights are initialized using Kaiming normal distribution. The ChangeMamba decoder is initialized using Kaiming normal distribution, consistent with the encoder's initial strategy. The dimensionality reduction module in the history storage module pre-trains a PCA model using a publicly available dataset, obtaining a fixed projection matrix P and mean μ, which will not change during subsequent use. The parameters ε and MinPts in the clustering module of the history storage module will be set during system debugging based on the feature distribution; the initial values ​​are determined based on data statistical analysis. The storage module in the history storage module creates an SQLite index table and HDF5 data file upon initial run, initially empty, and is continuously updated as features are stored.

[0074] The ChangeMamba encoder is used to extract dual-temporal remote sensing image pairs for remote sensing image change detection of the input region to be detected. , Each dual-temporal remote sensing image Feature maps of four different preset scales ( ). Bi-temporal remote sensing images extracted by the backbone network The Each feature map , . These are remote sensing images from previous time periods. This is a post-temporal remote sensing image.

[0075] In this embodiment, the image The four different preset scales used for feature map extraction are related to the image. The four different preset scales used for feature map extraction are the same. That is... and Same scale, and Same scale, and Same scale, and The scales are the same. In this embodiment, , , , The scale decreases sequentially. . For deep features, and This is a mid-level feature. These are shallow features. The input image resolution is... , The resolution is , The resolution is , The resolution is , The resolution is .

[0076] Figure 4 This is a schematic diagram of the foreground and background separation module in a remote sensing image change detection network according to an embodiment of the present invention; see reference. Figure 4 The foreground and background separation module receives the deepest features output from the ChangeMamba encoder. This data is then input into the feature difference calculation section of the fusion module. In this section, the system calculates the feature difference based on the input data. To calculate the absolute difference difference of squares and cosine similarity As shown below:

[0077]

[0078]

[0079]

[0080] in, The resolution is absolute difference difference of squares and cosine similarity The resolution is Then, the three values ​​are input into the feature concatenation part, where the system performs a concat operation on the three values ​​at the channel level, concatenating them into a single feature. As shown below:

[0081]

[0082] in, The resolution is Therefore, this feature The input is fed into the channel compression section to compress it at the channel level, resulting in compressed features. As shown below:

[0083]

[0084] in, The resolution is Finally, the compressed features The input is fed into the feature fusion part and residual connections are used to obtain the final output fused features. As shown below:

[0085]

[0086] in, The resolution is After feature fusion is performed by the fusion module, the output of the fusion module is... Continue inputting into the void space pyramid pooling module to capture multi-scale information. They will be placed into four dilated convolutional layers with different dilation rates. The kernels of all four convolutional layers are 3×3, but the dilation rates are different from those of the first layer. Second layer void ratio The third layer void ratio and the fourth layer void ratio The receptive fields of each layer are 13×13 for the first layer, 25×25 for the second, 37×37 for the third, and 49×49 for the fourth. After passing through four dilated convolutional layers in parallel, the system fuses the features. All multi-layer features were extracted and output as follows: , , and Each layer has a different receptive field, coverage, and the extracted details. The four extracted features are then input into the feature concatenation part, where they are concatenated to obtain... As shown below:

[0087]

[0088] The features are then concatenated and fed into the compressed channel section to obtain the features after compression at the channel level. As shown below:

[0089]

[0090] in, The resolution is Because the attention mechanism uses the third-layer output of the ChangeMamba encoder. The resolution of the output of this layer is... Therefore, the system will then... The input is increased in resolution by upsampling the layer. The upsampling layer uses bilinear interpolation to obtain... The resolution is The attention component uses The resolution is consistent. A segmentation head is used at the end to receive the output of the upsampled layer. And after passing through a 1×1 convolutional layer, the output becomes a foreground mask. .

[0091] Output foreground mask Then, the foreground mask It will be passed into the total loss function, which will then convert the foreground mask into a foreground mask. By comparing the data with real, manually labeled tags, a loss value and a gradient of trainable parameters are generated through a designed total loss function. In this invention, the total loss function... for , and The weighted sum is as follows:

[0092]

[0093] in, This loss function is used to optimize the recognition of small objects. In nature reserve scenarios, the foreground area is usually small and scattered. This loss function can enhance the learning of sparse foreground and improve the anti-interference ability. Used to maintain the integrity of the foreground region and reduce fragmented predictions; Used to impose additional constraints on the boundary, clear boundaries help subsequent change detection focus on the real change area and reduce false positives caused by vague boundaries; , and These are adjustable weighting coefficients used to balance the contribution of each loss term to the total loss; their values ​​typically range from [value range missing]. , , .

[0094] The loss items are as follows:

[0095] (1) The formula is as follows:

[0096]

[0097] Where N is the total number of samples; This represents the model's predicted probability that a sample belongs to the true class t. These are the category balancing weights, used to balance the loss contributions of positive and negative samples; To focus on parameters, control the degree of weight decay for easily classified samples.

[0098] (2) The formula is as follows:

[0099]

[0100] Where M is the total number of pixels in the image; This is the predicted value for the m-th pixel; This is the true label value of the m-th pixel; This is a smoothing factor to prevent the denominator from being zero.

[0101] (3) The formula is as follows:

[0102] in, The boundary region contains the set of pixels near the actual boundary. This represents the total number of pixels within the boundary region; For pixels within the boundary region; This is the distance from the model at pixel p to the nearest ground truth boundary; This is the probability prediction value of the model at pixel p.

[0103] The calculated gradients are backpropagated, passing through the modules described above to the ChangeMamba encoder to guide parameter changes. After one training iteration, the network regenerates a foreground mask. The process involves regenerating the loss value and gradient, repeating the above steps, until the generated loss value decreases to a certain level and stabilizes, no longer decreasing significantly. At this point, the model is considered "trained" and the loop ends; otherwise, the loop continues indefinitely. The output of the trained model is the final foreground mask. .

[0104] After obtaining the final foreground mask Then, the system will use the foreground mask. The input to the attention module guides the ChangeMamba decoder to generate a more accurate change map. Figure 5 This is a schematic diagram of the attention module in a remote sensing image change detection network according to an embodiment of the present invention; see reference. Figure 5 The attention module received the input foreground mask. and the third-layer output features of the ChangeMamba encoder Then, two parallel operations will be performed:

[0105] One is to use a foreground mask. The input is fed to the foreground attention module and the adaptive weight module. In the foreground attention module, the system will use the third layer output features of the ChangeMamba encoder. With foreground mask Feature concatenation and channel dimensionality reduction are performed, and then the fused features are combined with the foreground mask. Residual connections are used to obtain features that enhance the foreground. As shown below:

[0106]

[0107] In the adaptive weight module, the system uses a foreground mask. Calculate weights To determine whether the current region leans more towards the foreground or the background, as shown below:

[0108] like If the area is large, it tends to be more foreground; otherwise, it tends to be more background. The system will obtain... and Input into the fusion attention module.

[0109] Secondly, through foreground masking Calculate the background mask The calculation formula is:

[0110] Calculate the background mask The same operation in the foreground attention module is then used to generate features that suppress the background. As shown below:

[0111] The system will receive Input into the fusion attention module.

[0112] In the fusion attention module, the system will process the input... , and By performing a weighted summation, the characteristics of the fusion are obtained. As shown below:

[0113] Finally, the weighted and fused features are... The third layer output features of the ChangeMamba encoder The final output of the attention module is obtained by performing residual connections separately. The formula is as follows:

[0114] The system will use this output in the ChangeMamba decoder section to guide the decoder in generating the final change graph.

[0115] The method for obtaining the remote sensing image change detection network includes: building a network model of the remote sensing image change detection network; constructing a training dataset for the change detection task; initializing the built network model, wherein the encoder part uses pre-trained weights on a large image dataset for parameter initialization; iteratively optimizing the initialized network model using an end-to-end training method. During training, the predicted output of the network is calculated through forward propagation, and the difference between the predicted output and the true label is calculated using the composite loss function described in this invention. Then, the trainable parameters of the network are updated through the backpropagation algorithm until the model converges, and finally the trained remote sensing image change detection network is obtained.

[0116] On the other hand, this invention provides a module for detecting whether a detection was missed due to a late observation start point. Firstly, this module receives the deepest features from the ChangeMamba encoder portion. The module calculates the difference vector between two features. Since the system only uses the later-phase features in this module. Sum and difference vectors Therefore, only these two features will be stored in this module.

[0117] Figure 6 This is a schematic diagram of the storage module in a remote sensing image change detection network according to an embodiment of the present invention; see reference. Figure 6 Before actually running this module, the system pre-trains a PCA model for dimensionality reduction using publicly available remote sensing image change detection datasets. This pre-trained model is then used in the storage module without changing its parameters, thus improving efficiency and standardization. The system first performs spatial registration and resolution unification on publicly available remote sensing image change detection datasets (such as WHU-CD and LEVIR-CD datasets); then it loads the ChangeMamba encoder to batch extract post-temporal features and difference vectors from the datasets; finally, all features are synthesized into a large feature matrix. Where K is the total number of samples (twice the number of original image pairs), and 512 is the dimension of the last layer of features in the encoder; then the mean is calculated. As shown below:

[0118]

[0119] in, It is the k-th row of the characteristic matrix. Then, the centered feature matrix is ​​calculated. Covariance Matrix As shown below:

[0120]

[0121]

[0122] Then, the covariance matrix... Perform eigenvalue decomposition:

[0123]

[0124] in, It is a diagonal eigenvalue matrix. The matrix consists of eigenvectors, with each column representing an eigenvector. To determine the compression target dimension, the variance contribution rate of each principal component must first be calculated. and cumulative variance contribution rate :

[0125]

[0126]

[0127] Then, the target dimension is determined based on the calculation results. Dimensions that retain more than 95% of the variance:

[0128]

[0129] After finding the target dimension, in the feature vector matrix Find the first d eigenvectors to form the projection matrix. :

[0130]

[0131] After training the PCA model, the mean vector is... Projection matrix and the target dimension of dimensionality reduction Store these trained parameters for later use in subsequent module runs. While training the PCA model, keep the weights of the ChangeMamba encoder unchanged to ensure they are not updated during the training phase.

[0132] The system will initially accept output features from the deepest layer of the ChangeMamba encoder. The difference vector between the two features is calculated using this feature pair. Then the system will read the stored mean vector. Projection matrix and the target dimension of dimensionality reduction And calculate the features after dimensionality reduction. and :

[0133]

[0134]

[0135]

[0136]

[0137] in, , This indicates a centralized characteristic.

[0138] This allows the features to be reduced to a fixed dimension using fixed parameters. dimension.

[0139] In the feature of dimensionality reduction and Before inputting into the clustering module, the module initializes a dynamic buffer, a candidate noise pool, and a stable noise pool. The dynamic buffer and the candidate noise pool are two independent collections, while the stable noise pool is a database. The dynamic buffer is used when the amount of data in the database is small to reduce errors caused by the small data volume; the candidate noise pool is used when the amount of data is large and cluster classification is used exclusively to store features detected as noise; and the stable noise pool stores stable noise that persists in the candidate noise pool.

[0140] After obtaining the dimensionality reduction features and Then, the system inputs these two features into the clustering module. This module classifies the two input features separately, forming two memory banks, and then uses the dimensionality-reduced post-temporal features... The memory formed by classification is called the state database, which is then processed by the difference vector after dimensionality reduction. The memory formed by classification is called the change library. The physical storage pool of this invention is the same; the "state library" and "memory library" mentioned above refer to SQLite indexes. This invention achieves searching the same database based on different standard classifications by establishing two different SQLite indexes.

[0141] Initially, the clustering module has no data; data is automatically stored with each input. In the early stages of operation, the data volume is extremely small (approximately 1-30 data pairs), a period referred to as the "first clustering stage." During this stage, the system uses only a dynamic buffer for each cluster, storing each feature pair input into the module in the dynamic buffer. As the data grows, the number of feature pairs in the dynamic programming region also increases until it reaches a maximum value, which is set to 30 in this invention.

[0142] When the number of feature pairs in the dynamic buffer reaches 30, the module enters the "second clustering stage." In this stage, the input features are still placed in the dynamic buffer, and every time the number of feature pairs in the dynamic buffer reaches 30, the module re-clusters all feature pairs from the database and all feature pairs from the dynamic buffer. In the clustering algorithm, the module uses the post-time features after dimensionality reduction... The difference vector after dimensionality reduction Let's categorize them separately. For example, the module will treat all features as a set and first calculate the Euclidean distance between each feature in the set. :

[0143]

[0144] in, and These are two features in the feature set. It is the target dimension for dimensionality reduction in the dimensionality reduction module. It is the value of the c-th dimension of the u-th feature; then the number of neighborhood features for each point is calculated. And used to determine the core point As shown below:

[0145]

[0146]

[0147] Among them, those that satisfy the second formula These are all core points. After identifying the core points, a new cluster is created. and a core point Place it into the cluster; then retrieve it. All points within the neighborhood of a given point are called the "seed set"; then, a loop operation is performed: a point is selected from the "seed set". If the point is unvisited, mark it as visited. If the point is a core point, add all points in its neighborhood to the "seed set". If the point does not yet belong to any cluster, mark it as visited. Join the cluster After the loop completes, a complete cluster is obtained. Extending this method to the entire feature set will divide all sets into several clusters and several noise clusters whose density is unreachable for any core point. For a given cluster... The module will calculate the cluster center of the cluster. :

[0148]

[0149] in, It is a cluster The number of features in the dataset. It's important to note that the above classification only generates an index for that classification using one feature from each feature pair, without involving a physical storage pool. Each run of the clustering algorithm uses all features to re-divide into several clusters. After classification, the module generates new cluster indices and deletes all features stored in the dynamic buffer. Noise generated during classification is also removed and will not be placed into any cluster. Input features are not only placed in the dynamic buffer but also attempted to be assigned to existing clusters. The module calculates the Euclidean distance, finds the cluster closest to the feature, and then determines whether the distance between the feature and this cluster is less than a threshold (set to 1.5 in this invention). If the distance is less than the threshold, the feature is assigned to the cluster and the cluster information is updated; if the distance is greater than the threshold, no assignment is made, and the feature is stored in the candidate noise pool.

[0150] When the number of feature pairs stored in the dynamic buffer and database reaches 120, the module enters the "third clustering stage." In this stage, the dynamic buffer is deprecated, and the module will only use clusters for classification. Simultaneously, the module enables a candidate noise pool and a stable noise pool to receive temporary and long-term noise, respectively. The module's input is the dimensionality-reduced post-temporal features. The difference vector after dimensionality reduction For each feature: the input is used to directly calculate the Euclidean distance to each cluster center, and the feature pair is merged into the nearest cluster with a distance less than a threshold, and the information of that cluster is updated. If the distance between the feature pair and all cluster centers is greater than the threshold, the feature pair is placed in the candidate noise pool and its noise count is set. Add 1. The candidate noise pool has a storage limit. When the candidate noise pool is full, the module will re-cluster all feature pairs in the database and all feature pairs in the candidate noise pool. The clustering method is the same as the clustering method in the "second clustering stage". After re-clustering, the cluster index will be regenerated. The noise generated in the re-clustering will be stored in the candidate noise pool and its noise count will be added. Add 1. If there is noise in the candidate noise pool after re-clustering, add 1. If the value is 3, the noise is considered stable noise and is permanently moved to the stable noise pool, no longer participating in subsequent re-clustering and comparison.

[0151] Furthermore, if, during the classification process, a feature pair is assigned to a corresponding cluster in the change library but is judged as noise in the state library, then improvements are needed. Since this system is change detection, the change library has higher priority. The module will use the change library's judgment to assign the feature pair to a cluster within the change library and attach a state anomaly marker. This state anomaly marker will then play a role in the calculations of subsequent modules.

[0152] Furthermore, if, during the classification process, a feature pair is assigned to a corresponding cluster in the state library but is judged as noise in the variation library, improvement is also needed. Consistent with the reasons mentioned above, the variation library has higher priority, so the module will use the variation library's judgment to create a special new variation candidate label, which will also play a role in subsequent module calculations.

[0153] In the storage module, this invention uses a combination of SQLite and HDF5 storage to address the efficient organization, fast indexing, and compressed storage of massive amounts of data. The storage structure is as follows:

[0154] 1. SQLite index tables, including: main feature pair index table, state cluster statistics table, and change cluster statistics table. The core fields of the main feature pair index table are: feature pair ID, HDF5 storage path, state cluster ID, change cluster ID, and change intensity (the 2-norm of the difference vector, quantifying the degree of change, used for severity classification). The core fields of the state cluster statistics table are: cluster ID, cluster center feature vector, number of feature pairs in the cluster, sum of change intensities of the cluster, distribution of change patterns in the cluster, and average change intensity of the cluster. The core fields of the change cluster statistics table are: cluster ID, cluster center difference vector, average change intensity of the cluster, qualitative classification of the cluster, and number of feature pairs in the cluster.

[0155] 2. The HDF5 storage structure includes: system information, feature pair data set, clustering data set, and noise pool set. The system information includes: a) PCA parameter set: target dimension d for dimensionality reduction, PCA projection matrix P, and feature mean vector μ; b) clustering parameter set: neighborhood radius of the state clusters. Neighborhood radius of variation clustering The minimum number of samples (MinPts) for state clustering and change clustering are defined as follows: c) Qualitative hierarchical nodes. The feature pair data set includes: several feature pairs; each feature pair includes: the post-temporal feature after dimensionality reduction, the difference vector after dimensionality reduction, and precise storage of the change intensity values. The clustering data set includes: state clustering and change clustering; the state clustering includes several clusters, each including: the cluster's central feature vector, an array of IDs for all members within the cluster, the average change intensity within the cluster, and the maximum change intensity within the cluster; the change clustering includes several clusters, each including: the cluster's central feature vector, an array of IDs for all members within the cluster, the average change intensity within the cluster, and a histogram of change intensity variations. The noise pool group includes: the ID array of stable noise pool feature pairs and the current total number of stable noise pairs.

[0156] After the module classifies the input feature pairs, the feature reuse module uses the classification results to determine if there are any missed detections. The core logic of this determination is: if the later phase state of a certain region is usually caused by "drastic changes" (for example, 75% of bare land is transformed from forest), but the input difference vector shows "minor changes" (for example, only 7% of bare land is transformed from bare land by minor changes), then this combination of "state-change" is abnormal. It is very likely that the missed detection is due to the late observation time, missing the earlier drastic change process from forest to bare land and only seeing the later minor change process from bare land to bare land.

[0157] Each time a feature pair is input, the system calculates the intensity of change for that feature pair. :

[0158]

[0159] in, It is the dimension of the feature vector. This is the value of the difference vector in the i-th dimension. After cluster reallocation due to the initial use of the clustering algorithm or reclassification operations in the "second clustering stage" and "third clustering stage," the module automatically calculates the portion of data that needs to be stored in the index table:

[0160] (1) Distribution of change patterns of state clusters :

[0161] For each state cluster The module will read all feature pairs in that cluster and their corresponding change clusters. Then calculate the conditional probability. :

[0162]

[0163] in, It is a state It is due to change Number of samples generated It is a state The total number of samples. The purpose of this conditional probability is to reveal the probability of the change that leads to the current state, and to identify the most likely change that leads to the current state.

[0164] (2) Qualitative grading in variation clusters

[0165] This classification is not based on a fixed threshold, but rather on a global distribution. First, after classification, the average intensity of change within each variation cluster is calculated. :

[0166]

[0167] Then, by statistically analyzing the intensity of change across all feature pairs, the quartiles are identified. 25th percentile: The upper limit for slight variation; 50th percentile: median; 75th percentile: The lower limit of dramatic changes.

[0168] Qualitative classification is based on the average intensity of change in each cluster:

[0169]

[0170] The above calculations are the preparatory work for this module. After inputting feature pairs, there are four possible outcomes:

[0171] (1) When a new feature pair is input and the subsequent phase features of the feature pair Sum and difference vectors When none of the elements are noise, the first step the module will take is to obtain the state cluster to which the feature pair is assigned. and change clusters and the intensity of change of this feature pair Then find the state cluster in the state database table. Distribution of change patterns Calculate low-frequency outliers :

[0172]

[0173] The physical meaning of this outlier is whether the "state-change" combination is an anomaly with extremely low probability; the second step is to identify the state cluster. The highest proportion of change clusters in the distribution of change intensity and the hierarchical classification of the cluster Then calculate the severity contradiction value. :

[0174]

[0175] in, For variation clusters Qualitative classification, The highest level of qualitative classification (3 in this invention) represents the physical meaning of this outlier, indicating whether the change is significantly different from the mainstream change that led to the state. The third step is to calculate the change suppression value. :

[0176]

[0177] in, For state clusters The average intensity of change is used, and the physical meaning of this outlier is to detect a change that is below the average level of change; finally, the suspiciousness is calculated. :

[0178]

[0179] in, In this invention, it is set Therefore, the final formula for the degree of suspicion is:

[0180]

[0181] (2) When a new feature pair is input, the difference vector of the feature pair Normally assigned to a change cluster However, the characteristics of the later phase This is identified as noise. At this point, the module will accept the allocation of the variable clusters and calculate... :

[0182]

[0183] in, It is in state The following observations The proportion, It is a state The probability of appearing in all data. It is a change The probability of occurrence in all variations. Then select the probability. Largest state cluster As a means of inference :

[0184]

[0185] Then use the variation cluster and state clusters The suspiciousness result is obtained by performing calculation (1). However, since this level of suspicion is obtained through reasoning, its credibility should be lower than that of direct reasoning. Therefore, a discount factor (defined as 0.7 in this invention) is multiplied by the obtained level of suspicion to obtain the final level of suspicion. :

[0186]

[0187] (3) When a new feature pair is input, the subsequent phase features of that feature pair The state cluster was assigned normally. But the difference vector This is identified as noise. At this point, the module accepts the assignment of a state cluster and finds the average change intensity of that state cluster. Calculate the strength ratio :

[0188] If the strength ratio Below the threshold (An empirical threshold, set at 0.4 in this invention) is then considered abnormal, and... The value is assigned to 0.8; if If the value is greater than or equal to the threshold, it is considered normal. The value is assigned to 0.

[0189] (4) When a new feature pair is input, but the subsequent phase features Sum and difference vectors If both are judged as noise, the feature pair is placed in the candidate noise pool and no comparison is performed.

[0190] After obtaining the final level of suspicion Then the judgment can be made:

[0191]

[0192] Output the results after obtaining them.

[0193] To address the issue of existing technologies failing to fully utilize historical information, this invention constructs a complete historical remote sensing data feature storage and reuse module. This module employs PCA dimensionality reduction technology to compress high-dimensional features extracted by the ChangeMamba encoder, utilizes the DBSCAN clustering algorithm to classify the dimensionality-reduced post-temporal features and difference vectors into state clusters and change pattern clusters, respectively, and innovatively adopts a hybrid storage strategy combining HDF5 and SQLite to achieve efficient organization and management of historical features. Through the feature reuse module, the system can respond to user selections, comparing newly input feature pairs with the historical memory, effectively identifying and alerting to potential missed detections due to late observation start times.

[0194] The above combination Figure 1 The remote sensing change detection method based on foreground and background separation and historical feature reuse provided in the embodiments of the present invention has been described in detail. Next, the remote sensing change detection system based on foreground and background separation and historical feature reuse provided in the embodiments of the present invention will be described in conjunction with the accompanying drawings.

[0195] Figure 7 This is a schematic diagram of an embodiment of a remote sensing change detection system based on foreground and background separation and historical feature reuse, as shown in this invention; see reference. Figure 7 The system includes a data acquisition unit and an image analysis and processing unit. The data acquisition unit periodically acquires remote sensing images of nature reserves and constructs dual-temporal remote sensing image pairs for change detection through temporal registration, ensuring the temporal consistency of the data. The image analysis and processing unit integrates an image preprocessing unit and a remote sensing image change detection network constructed according to any of the aforementioned embodiments to complete subsequent change identification tasks. The image analysis and processing unit establishes a connection with the data acquisition unit. Its workflow includes calling the image preprocessing unit to perform spatial registration on the input target dual-temporal remote sensing image pairs, then sending the registered image pairs into the remote sensing image change detection network for analysis and processing, and finally outputting image change information of the area to be detected.

[0196] In practical applications, the data acquisition unit first acquires a pair of dual-temporal remote sensing images of the area to be detected, i.e., a pair of dual-temporal remote sensing images of the target, and transmits them to the image preprocessing unit within the image analysis and processing unit. This preprocessing unit performs spatial registration on the received image pairs, and then inputs the registered images into the remote sensing image change detection network for calculation, thereby identifying image changes in the area to be detected. For example... Figure 8 This presents a set of pre-temporal remote sensing images from a pair of spatially registered target dual-temporal remote sensing images. Figure 9 This is the corresponding later-phase remote sensing image in the same image pair. Figure 8 and Figure 9The registered images shown are input into the remote sensing image change detection network of this invention for processing, and the results are as follows: Figure 10 The diagram shows the result of image changes in the region to be detected. Figure 10 In the image analysis unit, black areas represent the image background, white areas represent the image foreground, and the area corresponding to the foreground is the detected change portion of the image. Preferably, the image analysis unit is equipped with a graphical user interface (GUI), allowing users to visually view the obtained image change results for the detected area.

[0197] Optionally, the graphical user interface (GUI) integrates an API interface, allowing users to upload custom remote sensing image change detection networks. The interface also features a dedicated image change detection network display area, showcasing all uploaded network models to the user for easy selection. The image analysis and processing unit integrates a network loading module that automatically loads the user-selected remote sensing image change detection network.

[0198] Optionally, the data transmission unit supports interface types including, but not limited to: USB interface, Wi-Fi wireless module, and Ethernet interface. The data acquisition unit typically refers to a device capable of acquiring remote sensing images, such as various remote sensing image sensors. The image analysis and processing unit is usually performed by a computer.

[0199] See Figure 11 The diagram shows the structure of a computer device, which includes a processor, a communication interface, and a computer-readable storage medium. The processor, communication interface, and computer-readable storage medium are connected via a bus or other means. The communication interface is used to receive and send data. The computer-readable storage medium can be stored in the computer device's memory. The computer-readable storage medium stores computer programs, including program instructions, and the processor executes the program instructions stored in the computer-readable storage medium. The processor (or CPU, Central Processing Unit) is the computing and control core of the computer device, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions to implement the corresponding steps in the remote sensing change detection method embodiment based on foreground / background separation and historical feature reuse.

[0200] This embodiment provides a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the processing system of the computer device.

[0201] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A remote sensing change detection method based on foreground / background separation and historical feature reuse, characterized in that, include: Acquire a pair of two-temporal remote sensing images of the area to be detected; the pair of two-temporal remote sensing images includes a foreground remote sensing image and a background remote sensing image; Based on the preprocessed dual-temporal remote sensing image pairs, the ChangeMamba encoder is used to obtain four features at different preset scales for each dual-temporal remote sensing image; Based on the fourth-layer foreground features and the fourth-layer background features, a foreground-background separation module is used. After fusion, the mixture is input into the hollow space pyramid pooling module to separate the foreground and background and generate a foreground mask. Based on the foreground mask, the third-layer foreground features, and the third-layer background features, an attention module is used to perform two parallel operations to generate foreground enhancement features and background suppression features, which are then weighted and fused to obtain foreground change maps and background change maps. Based on the foreground change maps and background change maps, a ChangeMamba decoder is used to obtain change information of the region to be detected. Based on the fourth-layer foreground features and the fourth-layer background features, a feature storage module is used to calculate the difference vector, and then perform centering and dimensionality reduction processing to obtain dimensionality-reduced temporal features and dimensionality-reduced difference features. Clustering is performed on the temporal features and dimensionality difference features after dimensionality reduction to obtain the classification results; Determine if any classification results are missed.

2. The remote sensing change detection method based on foreground / background separation and historical feature reuse according to claim 1, characterized in that, Each dual-temporal remote sensing image yields four features at different preset scales, including: based on the foreground remote sensing image, a first layer of foreground features, a second layer of foreground features, a third layer of foreground features, and a fourth layer of foreground features; based on the background remote sensing image, a first layer of background features, a second layer of background features, a third layer of background features, and a fourth layer of background features.

3. The remote sensing change detection method based on foreground / background separation and historical feature reuse according to claim 1, characterized in that, The method, based on fourth-layer foreground and fourth-layer background features, employs a foreground / background separation module. After fusion, the data is input into a hollow spatial pyramid pooling module for foreground / background separation, generating a foreground mask. The method includes: The fourth-layer foreground features and the fourth-layer background features are input into the foreground-background separation module to calculate the absolute difference, squared difference, and cosine similarity. The absolute difference, squared difference, and cosine similarity are then concatenated and compressed to obtain the first compressed feature. The first compressed feature is then input into the feature fusion module for residual connection to obtain the fused feature. The fused features are input into the hollow spatial pyramid pooling module. Through four hollow convolutional layers with different dilation rates, features at four different scales of the fused features are extracted. These four features at different scales are concatenated to obtain concatenated features. The concatenated features are compressed to obtain second compressed features. The second compressed features are input into the upsampling layer and the segmentation head to obtain the foreground mask.

4. The remote sensing change detection method based on foreground / background separation and historical feature reuse according to claim 1, characterized in that, During the training phase, after obtaining the foreground mask, the total loss function is used to compare the foreground mask with the ground truth labels to optimize the parameters of the ChangeMamba encoder; the total loss function is expressed by the following formula: in, Indicates focal loss; Indicates die loss; Indicates boundary loss; , and These are adjustable weighting coefficients.

5. The remote sensing change detection method based on foreground / background separation and historical feature reuse according to claim 1, characterized in that, Based on foreground masks, third-layer foreground features, and third-layer background features, an attention module is used for two-way parallel operations to generate foreground enhancement features and background suppression features, which are then weighted and fused to obtain foreground and background change maps. The method includes: The parallel operation is as follows: input the foreground mask, the third-layer foreground features, and the third-layer background features into the attention module, perform feature concatenation and channel dimensionality reduction, and perform residual connection between the obtained features and the foreground mask to obtain foreground enhancement features; based on the foreground mask, use the adaptive weight module to calculate the foreground weights; Another parallel operation is as follows: Based on the foreground mask, calculate the background mask; input the background mask, the third-layer foreground features and the third-layer background features into the attention module, perform feature concatenation and channel dimensionality reduction, and perform residual connection between the obtained features and the background mask to obtain the background suppression features; The foreground enhancement features, foreground weights, and background suppression features are weighted and fused to obtain weighted fused features. The weighted fused features are then residually connected with the third-layer foreground features and the third-layer background features to obtain foreground change maps and background change maps.

6. The remote sensing change detection method based on foreground / background separation and historical feature reuse according to claim 1, characterized in that, Based on the fourth-layer foreground and fourth-layer background features, a feature storage module is used to calculate the difference vector, followed by centering and dimensionality reduction processing to obtain the dimensionality-reduced temporal features and dimensionality-reduced difference features; the method includes: The large storage module stores the fourth-layer background features and the feature difference between the fourth-layer background features and the fourth-layer foreground features for each input. The PCA model was initially trained using a public dataset to obtain a fixed projection matrix and mean. Based on the difference between the fourth-layer background features and the mean, the centered post-temporal features are obtained; based on the difference between the feature differences and the mean, the centered difference features are obtained. The centered post-temporal features and difference features are multiplied by the projection matrix to obtain the dimensionality-reduced post-temporal features and dimensionality-reduced difference features.

7. The remote sensing change detection method based on foreground / background separation and historical feature reuse according to claim 1, characterized in that, Clustering is performed on the dimensionality-reduced temporal features and dimensionality-reduced difference features to obtain classification results. The method includes: inputting the dimensionality-reduced temporal features and dimensionality-reduced difference features into the clustering module, performing clustering calculations separately, and obtaining two types of memory databases, each containing several pre-classified clusters; the database classified using the dimensionality-reduced temporal features is the state database, and the database classified using the dimensionality-reduced difference features is the change database, which is classified according to the stable state and change pattern of the subsequent temporal phases, respectively; and the classification results are obtained. After obtaining the classification results, the background features of the fourth layer and the feature differences between the background features of the fourth layer and the foreground features of the fourth layer are stored in their respective positions, and relevant information is calculated. Alternatively, methods for determining missed detections in the classification results include: Obtain the cluster ID from the change pattern memory generated by the clustering module, and use this cluster ID to find the cluster to which the input feature pair belongs. Then, obtain the cluster ID from the state memory generated in the clustering module and use this cluster ID to find the cluster to which the input feature pair belongs. Find The distribution of changes, through the distribution of changes and The degree of suspicion is calculated based on the drastic change attribute of the cluster. If the threshold is reached, it is judged as a missed problem; if it is not reached, it is normal.

8. A remote sensing change detection system based on foreground / background separation and historical feature reuse, characterized in that: include: The data acquisition unit is used to acquire a pair of dual-temporal remote sensing images of the area to be detected; the pair of dual-temporal remote sensing images includes a foreground remote sensing image and a background remote sensing image; The image analysis and processing unit is used to obtain features at four different preset scales for each dual-temporal remote sensing image based on the preprocessed dual-temporal remote sensing image pairs using the ChangeMamba encoder. Based on the fourth-layer foreground features and the fourth-layer background features, a foreground-background separation module is used. After fusion, the mixture is input into the hollow space pyramid pooling module to separate the foreground and background and generate a foreground mask. Based on the foreground mask, the third-layer foreground features, and the third-layer background features, an attention module is used to perform two parallel operations to generate foreground enhancement features and background suppression features, which are then weighted and fused to obtain foreground change maps and background change maps. Based on the foreground change maps and background change maps, a ChangeMamba decoder is used to obtain change information of the region to be detected. Based on the fourth-layer foreground features and the fourth-layer background features, a feature storage module is used to calculate the difference vector, and then perform centering and dimensionality reduction processing to obtain dimensionality-reduced temporal features and dimensionality-reduced difference features. Clustering is performed on the temporal features and dimensionality difference features after dimensionality reduction to obtain the classification results; Determine if any classification results are missed.

9. A computer device, characterized in that, A processor, adapted to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, implements the steps of the remote sensing change detection method based on foreground / background separation and historical feature reuse as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and to execute the steps of the remote sensing change detection method based on foreground / background separation and historical feature reuse as described in any one of claims 1-7.

Images