A multispectral image change detection method and system

Abstract

The present application discloses a multispectral image change detection method and system, which relates to the field of image processing technology. The method includes: obtaining an image at a first moment and an image at a second moment; extracting multi-layer features of the image at the first moment and the image at the second moment; aggregating the first feature and the second feature of the current layer with the first feature and the second feature of the adjacent layer respectively to obtain a first aggregate feature and a second aggregate feature; extracting the mutual attention distribution between the first aggregate feature and the second aggregate feature to obtain a first change attention feature map and a second change attention feature map, and determining the corresponding differential feature map; extracting multi-scale change information in the differential feature map to obtain a scale fusion feature map; fusing the scale fusion feature map to obtain a change detection map. The present application adopts a cross-temporal fusion attention unit, a change location recognition module, and a supervisory attention module to improve the ability to accurately focus on the area of ​​interest, thereby improving the accuracy of multi-scale change detection.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a multispectral image change detection method and system. Background Art

[0002] Multispectral image change detection refers to the process of extracting changed areas from multispectral images of the same scene at different time stages. Multispectral remote sensing images contain key information such as color, texture, and material of objects, and do not significantly increase the data dimension. Therefore, change detection of multispectral remote sensing images is widely used in urban planning, disaster assessment, and resource survey. At present, a large number of methods for remote sensing image change detection have been studied. Early classic change detection methods include algebra-based detection methods, image transformation methods, and classification-based methods. However, these traditional methods usually rely on manual feature extraction, resulting in limited detection accuracy.

[0003] Deep learning technology has become the mainstream method in increasingly complex change detection scenarios due to its powerful feature extraction capabilities and efficient detection performance. The powerful expression ability of convolutional neural networks has significantly improved the detection performance of the model, especially in exploring richer spatiotemporal change information in early feature interactions. In addition, the attention mechanism introduced in recent years can automatically focus on relevant information and filter out irrelevant data, making feature extraction more effective.

[0004] However, since remote sensing images are usually acquired at different time stages, in practical applications, a large number of unexpected changes will be introduced due to factors such as seasonal changes, changes in lighting conditions, and building renovations, which will interfere with the detection results. Current algorithms mainly focus on extracting features from the channel or spatial dimensions of dual-phase images, while ignoring the rich spatiotemporal change information. In addition, early feature interactions are not sufficient to convey their respective spatial information, making it difficult for the model to focus on relevant change areas, and the accuracy of detecting multi-scale change areas after feature encoding needs to be improved. Summary of the invention

[0005] The embodiments of the present application provide a multispectral image change detection method and system to solve the problem that the image change detection method in the prior art has low accuracy.

[0006] On the one hand, an embodiment of the present application provides a multispectral image change detection method, comprising:

[0007] Acquire an image at a first moment and an image at a second moment, wherein the image at the first moment and the image at the second moment are multispectral images acquired at different moments in the same scene;

[0008] Extracting multi-layer features of the image at the first moment and the image at the second moment respectively, features of the same layer constitute a feature pair, and the feature pair includes the first feature and the second feature;

[0009] Aggregate the first feature of the current layer and the first feature of the adjacent layer to obtain a first aggregate feature corresponding to the feature pair, and aggregate the second feature of the current layer and the second feature of the adjacent layer to obtain a second aggregate feature corresponding to the feature pair;

[0010] Extracting the mutual attention distribution between the first aggregated feature and the second aggregated feature of the same layer, respectively obtaining a first change attention feature map and a second change attention feature map, and determining a corresponding differential feature map according to the first change attention feature map and the second change attention feature map;

[0011] Extract the multi-scale change information in each layer of the differential feature map, encode the multi-scale change information in different directions, and fuse the encoding results in different directions to obtain a scale fusion feature map;

[0012] The scale fusion feature maps of different layers are fused sequentially to obtain the change detection map.

[0013] On the other hand, the embodiment of the present application further provides a multispectral image change detection system, including:

[0014] An encoder is used to obtain an image at a first moment and an image at a second moment, wherein the image at the first moment and the image at the second moment are multispectral images collected at different moments in the same scene; extract multi-layer features of the image at the first moment and the image at the second moment respectively, wherein the features of the same layer constitute a feature pair, wherein the feature pair includes a first feature and a second feature; aggregate the first feature of the current layer and the first feature of the adjacent layer to obtain a first aggregate feature corresponding to the feature pair, and aggregate the second feature of the current layer and the second feature of the adjacent layer to obtain a second aggregate feature corresponding to the feature pair; extract the mutual attention distribution between the first aggregate feature and the second aggregate feature of the same layer to obtain a first change attention feature map and a second change attention feature map respectively, and determine the corresponding differential feature map according to the first change attention feature map and the second change attention feature map;

[0015] The decoder is used to extract the multi-scale change information in each layer of the differential feature map, encode the multi-scale change information in different directions, fuse the encoding results in different directions to obtain a scale fusion feature map; and fuse the scale fusion feature maps of different layers in turn to obtain a change detection map.

[0016] A multispectral image change detection method and system in this application has the following advantages:

[0017] 1. The introduction of cross-temporal fusion attention units reduces the interference of irrelevant changes, enhances early feature interactions, and improves the ability to accurately focus on areas of interest.

[0018] 2. A change location and recognition module is introduced. The change features are gradually searched on the feature map through dilated convolution, and the coordinate attention mechanism is used to accurately locate the change area, thereby improving the accuracy of multi-scale change detection.

[0019] 3. A supervised attention module is introduced to achieve multi-level feature aggregation from high-level to low-level. Experimental results show that the method of this application performs well on LEVIR-CD (Large-Scale Remote Sensing Building Change Detection) and GVLM (Global Very-High-Resolution Landslide Mapping) datasets, effectively improving the change detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0021] Figure 1 An architectural diagram of a multispectral image change detection method provided in an embodiment of the present application.

[0022] Figure 2 An architectural diagram of the layer neighborhood and cross-time aggregation modules provided for an embodiment of the present application.

[0023] Figure 3 A flowchart of the cross-temporal fusion attention unit provided in an embodiment of the present application.

[0024] Figure 4 This is an architecture diagram of the change positioning identification module provided in an embodiment of the present application.

[0025] Figure 5 An architectural diagram of the coordinate attention mechanism provided in an embodiment of the present application. DETAILED DESCRIPTION

[0026] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0027] Figure 1 The present invention provides a method for detecting changes in multispectral images, including the following steps:

[0028] S100, acquiring an image at a first moment and an image at a second moment, where the image at the first moment and the image at the second moment are multispectral images acquired at different moments in the same scene.

[0029] S110, extracting multi-layer features of the image at the first moment and the image at the second moment respectively, features of the same layer form a feature pair, and the feature pair includes a first feature and a second feature.

[0030] Exemplarily, the MobileNetV2 network may be used to extract multi-layer features of the image at the first moment and the image at the second moment.

[0031] MobileNetV2 is a lightweight deep convolutional neural network, mainly used for efficient image classification, object detection, semantic segmentation and other tasks in mobile devices and embedded systems. Compared with traditional deep networks such as ResNet or VGG (Visual Geometry Group), MobileNetV2 is designed to greatly reduce the amount of computation and model size while ensuring accuracy, which is especially suitable for resource-constrained scenarios. MobileNetV2 significantly reduces the amount of model computation through deep separable convolution. Compared with other deep networks, MobileNetV2 has a much smaller model size and is suitable for mobile and embedded devices.

[0032] exist Figure 1 In the above figure, there are four layers of multi-layer features extracted, among which f 2 1 and f 2 2 is the feature pair of the first layer, f 3 1 and f 3 2 is the feature pair of the second layer, f 4 1 and f 4 2 is the feature pair of the third layer, f 5 1 and f 5 2 is the feature pair of the fourth layer. Among these feature pairs, f 2 1 , f 3 1 , f4 1 and f 5 1 All are first characteristics, f 2 2 , f 3 2 , f 4 2 and f 5 2 Both are second features. Among the first and second features, the features of the latter layer are extracted based on the features of the previous layer. Therefore, the size of the features of the latter layer is half of the size of the features of the previous layer.

[0033] S120, aggregate the first feature of the current layer and the first feature of the adjacent layer to obtain a first aggregate feature corresponding to the feature pair, and aggregate the second feature of the current layer and the second feature of the adjacent layer to obtain a second aggregate feature corresponding to the feature pair.

[0034] For example, Figure 2 As shown, the first feature and the second feature will be input into the layer neighborhood and cross-time aggregation module, which includes a layer neighborhood aggregation unit, a cross-time fusion attention unit and a difference unit. The first features of two or three adjacent layers are aggregated by the layer neighborhood aggregation unit to obtain the first aggregated feature. Correspondingly, the second features of two or three adjacent layers are aggregated by the layer neighborhood aggregation unit to obtain the second aggregated feature. The first aggregated feature and the second aggregated feature will be simultaneously input into the cross-time fusion attention unit. After the fusion processing of the adjacent layer features, the expression ability of the temporal feature can be effectively improved.

[0035] S130, extract the mutual attention distribution between the first aggregate feature and the second aggregate feature of the same layer, obtain the first change attention feature map and the second change attention feature map respectively, and determine the corresponding differential feature map according to the first change attention feature map and the second change attention feature map.

[0036] Exemplarily, the process of generating the first change attention feature map and the second change attention feature map by the cross-temporal fusion attention unit is as follows: Figure 3 As shown in Figure 2, the two changing attention feature maps generated by the cross-time fusion attention unit contain the features of the current time point and partial expressions of other time points, thereby suppressing the interference of external noise. The first aggregate feature and the second aggregate feature of size C×H×W The input cross-temporal fusion attention unit is first processed by convolution, and then the change attention feature map of feature interaction is constructed through query, key and value matrices. The formula is as follows:

[0037]

[0038] in, represents the changing attention feature map of the output of the attention unit fused across time, j =1 or 2, in j =1, Represents the first change attention feature map, j =2, Represents the second change attention feature map. Indicates that the softmax function is used for activation processing. Q 1. K 1 and j =1 Vj Respectively represent the query, key, and value matrices after the first aggregated feature is convolved, Q 2. K 2 and j =2 Vj They represent the query, key, and value matrices after the convolution of the second aggregated feature, respectively. Indicates the absolute value. reshape indicates the result after reshaping. j =1 represents the first aggregate feature, j =2 Represents the second aggregate feature.

[0039] Furthermore, after obtaining the first change attention feature map and the second change attention feature map, the difference unit will perform difference processing on the first change attention feature map and the second change attention feature map to obtain a difference feature map. In the case of generating four layers of feature pairs, each layer of feature pairs will generate a difference feature map after being processed by the layer neighborhood and cross-time aggregation modules, so the four layers of feature pairs will correspondingly obtain four difference feature maps. d ,in i =2, 3, 4 and 5.

[0040] S140, extracting multi-scale change information from each layer of the differential feature map, encoding the multi-scale change information in different directions, and fusing the encoding results in different directions to obtain a scale fusion feature map.

[0041] For example, the four differential feature maps generated are d The input is sent to the change location recognition module, which is used to enhance the exploration of spatiotemporal change features and extract the location information of the changes through multi-scale learning.

[0042] The method for extracting multi-scale change information in each layer of differential feature map by the positioning recognition module includes:

[0043] S141, performing multiple residual convolution processing on the differential feature map to obtain corresponding residual features. S142, summing the differential feature map and the residual features of the same layer in sequence to obtain multi-scale change information.

[0044] Specifically, the positioning and recognition module will first perform dilated convolution on the differential feature map and the summed feature map before summing the differential feature map and the residual feature map in sequence. During the dilated convolution, the dilated rate used for the differential feature map and the summed feature map gradually decreases. By using convolution operations with different dilated rates, the change characteristics of each layer can be gradually learned, and a more accurate change area can be located.

[0045] like Figure 4 As shown in , the positioning recognition module also performs residual convolution on the differential feature map four times. Since each residual convolution is based on the same differential feature map, the results of multiple residual convolutions are the same. d 2 as an example, the result obtained after residual convolution is Figure 4 From bottom to top, they are the residual features of the first scale d 2 2 , the residual features of the second scale d 2 3 , the residual features of the third scale d 2 4 and the residual features of the fourth scale d 2 5 .

[0046] For the first differential feature map d 2. Void rate d =7 after the dilated convolution, and the residual feature of the first scale d 2 2 The sum is calculated and the result is then passed through the void rate d = 5, and pass it upwards until the hole rate is completed. d = 1 dilated convolution result and the residual feature of the fourth scale d 2 5 The above calculation can be expressed by the following formula:

[0047]

[0048] in, , , and represents the result after each summation, Represents the dilated convolution processing, and 3×3 is the convolution kernel size of the dilated convolution.

[0049] The positioning recognition module obtains multi-scale change information of each layer , , and After that, it will be further processed by the coordinate attention mechanism to achieve flexible positioning of the change area by integrating direction perception and position sensitive information. The coordinate attention mechanism encodes the multi-scale change information in different directions, fuses the encoding results in different directions, and obtains the scale fusion feature map, which specifically includes the following steps:

[0050] S143, perform average pooling processing on the multi-scale change information in the horizontal direction and the vertical direction respectively to obtain a horizontal vector and a vertical vector; S144, concatenate the horizontal vector and the vertical vector in the spatial dimension and perform two-dimensional convolution processing to obtain a concatenated vector; S145, perform layer normalization processing and nonlinear processing on the concatenated vector in turn to obtain a coding result; S146, split the coding result into two coding vectors, perform two-dimensional convolution and Sigmoid activation processing on each coding vector in turn, and re-weight the results of the two convolution and activation processing in the channel dimension to obtain a scale fusion feature map, such as Figure 5 shown.

[0051] S150, the scale fusion feature maps of different layers are sequentially fused to obtain a change detection map.

[0052] Exemplarily, after re-weighting processing using the coordinate attention mechanism, the re-weighted result is upsampled twice using bilinear interpolation to obtain a scale fusion feature map, and each layer of the differential feature map is processed in the same way by the change location recognition module to obtain a scale fusion feature map with the same number of layers as the differential feature map. The supervised attention module is first processed in order from large to small, and then the scale fusion feature maps are fused in sequence to obtain a change detection map, the size of which is the same as the size of the input image at the first moment and the image at the second moment. The supervised attention module adjusts the weight of the scale fusion feature map by combining context information to achieve aggregation of multi-level features. During the aggregation process, the size of the image is gradually restored to the resolution of the input image, and finally a change detection map is generated.

[0053] The present application also provides a multispectral image change detection system, which includes:

[0054] An encoder is used to obtain an image at a first moment and an image at a second moment, wherein the image at the first moment and the image at the second moment are multispectral images collected at different moments in the same scene; extract multi-layer features of the image at the first moment and the image at the second moment respectively, wherein the features of the same layer constitute a feature pair, wherein the feature pair includes a first feature and a second feature; aggregate the first feature of the current layer and the first feature of the adjacent layer to obtain a first aggregate feature corresponding to the feature pair, and aggregate the second feature of the current layer and the second feature of the adjacent layer to obtain a second aggregate feature corresponding to the feature pair; extract the mutual attention distribution between the first aggregate feature and the second aggregate feature of the same layer to obtain a first change attention feature map and a second change attention feature map respectively, and determine the corresponding differential feature map according to the first change attention feature map and the second change attention feature map;

[0055] The decoder is used to extract the multi-scale change information in each layer of the differential feature map, encode the multi-scale change information in different directions, fuse the encoding results in different directions to obtain a scale fusion feature map; and fuse the scale fusion feature maps of different layers in turn to obtain a change detection map.

[0056] Exemplarily, before detecting the image at the first moment and the image at the second moment, the encoder and the decoder perform structural optimization on the encoder and the decoder, and the loss function used in the optimization process adopts a composite loss function composed of a cross entropy loss and a block loss.

[0057] The composite loss function can be written as follows:

[0058]

[0059] in, L all represents the composite loss function, L CE represents the cross entropy loss, L Dice represents the block loss, λ 1 and λ 2 are the weights of cross entropy loss and block loss respectively.

[0060] In classification tasks, cross entropy loss usually measures the difference between the predicted class distribution and the true class distribution, effectively optimizing pixel-level classification accuracy as follows:

[0061]

[0062] in, N represents the number of samples, M represents the number of categories, p tc It is a samplet Belongs to category c The predicted probability of y tc is a sign function, if the sample t The true category is c ,but y tc =1, otherwise y tc =0.

[0063] The block loss is a regional loss, as shown in the following formula, which is used to evaluate the similarity between the predicted segmentation image and the true segmentation image in the segmentation task. This loss has better performance in the case of imbalanced positive and negative samples, and can make up for the disadvantage of poor performance of cross entropy loss when processing unbalanced data. Therefore, the combination of the two losses helps to improve the overall performance of the change detection model.

[0064]

[0065] in, TP Represents the number of samples that are actually positive and correctly predicted as positive. FP Indicates the number of samples that are actually negative but are mistakenly predicted as positive. FN Indicates the number of samples that are actually positive but are incorrectly predicted as negative.

[0066] Simulation experiment

[0067] In order to prove the effectiveness of the method proposed in this application, five currently popular methods were used for comparison in the experiment. The quantitative detection accuracy Kappa, IoU, F1, Recall and Precision values ​​of the method proposed in this application and the other five comparison algorithms on the datasets LEVIR-CD and GVLM are shown in Table 1. Among them:

[0068] (1) BIT (Bitemporal Image Transformer) is introduced in detail in “Chen, Z. Qi, and Z. Shi, “Remote sensing image change detection with transformers,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1–14, 2021. 3, 7”.

[0069] (2) USSFC-Net (Ultralightweight Spatial Spectral Feature Cooperation Network) is introduced in detail in “T. Lei, X. Geng, H. Ning, Z. Lv, M.Gong, Y. Jin, and AKNandi, “Ultralightweight spatial–spectral feature cooperation network for change detection in remote sensing images,” IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1–14, 2023. 7”.

[0070] (3) SNUNet (Siamese Nested UNet) is introduced in detail in “S. Fang, K. Li, J. Shao, and Z. Li, “Snunet-cd: A densely connected siamese network for change detection of vhr images,” IEEE Geoscience and Remote Sensing Letters, vol. 19, pp. 1–5, 2021. 8”.

[0071] (4) DMINet (Dual-branch Multilevel Intertemporal Network) is introduced in detail in “Y. Feng, J. Jiang, H. Xu, and J. Zheng, “Change detection on remote sensing images using dual-branch multilevel intertemporal network,” IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp.1–15, 2023. 8”.

[0072] (5) C2FNet (Coarse-to-Fine Network) is introduced in detail in “C. Han, C. Wu, M.Hu, J. Li, and H. Chen, “C2f-semicd: A coarse-to-fine semi-supervised change detection method based on consistency regularization in high-resolution remote sensing images,” IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–21, 2024.8”.

[0073] Table 1 Evaluation indicators of different algorithms on the data set (value: %)

[0074]

[0075] As can be seen from Table 1, compared with the lightweight networks BIT and USSFC-Net, the method of the present application has fewer FLOPs (Floating Point Operations Per second), which is only 3.17G. In short, the method of the present application uses the smallest FLOPs and fewer parameters to obtain better detection results. These experimental results prove that the method of the present application has a good balance between accuracy and efficiency.

[0076] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0077] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. A multispectral image change detection method, characterized in that: include: Acquire a first moment image and a second moment image, wherein the first moment image and the second moment image are multispectral images acquired at different moments in the same scene; Extracting multi-layer features of the image at the first moment and the image at the second moment respectively, wherein the features at the same layer constitute a feature pair, and the feature pair includes a first feature and a second feature; Aggregate the first feature of the current layer and the first feature of the adjacent layer to obtain a first aggregate feature corresponding to the feature pair, and aggregate the second feature of the current layer and the second feature of the adjacent layer to obtain a second aggregate feature corresponding to the feature pair; Extracting the mutual attention distribution between the first aggregated features and the second aggregated features of the same layer, respectively obtaining a first change attention feature map and a second change attention feature map, and determining a corresponding differential feature map according to the first change attention feature map and the second change attention feature map; Extracting multi-scale change information from the differential feature map of each layer, encoding the multi-scale change information in different directions, and fusing the encoding results in different directions to obtain a scale fusion feature map; The scale fusion feature maps of different layers are fused in sequence to obtain a change detection map.

2. A multispectral image change detection method according to claim 1, characterized in that: The method for extracting the multi-scale change information in the differential feature map of each layer includes: Performing multiple residual convolution processes on the differential feature maps to obtain corresponding residual features; The differential feature map and the residual feature of the same layer are summed up in sequence to obtain the multi-scale change information.

3. A multispectral image change detection method according to claim 2, characterized in that: Before the differential feature map and the residual feature are sequentially summed, a dilated convolution process is performed on the differential feature map and the summed feature map.

4. The multispectral image change detection method according to claim 3, characterized in that: During the dilated convolution process, the dilation rate used for the differential feature map and the summed feature map gradually decreases.

5. The multispectral image change detection method according to claim 1, characterized in that: The multi-scale change information is encoded in different directions, and the encoding results in different directions are fused to obtain a scale fusion feature map, including: Performing average pooling processing on the multi-scale change information in the horizontal direction and the vertical direction respectively to obtain a horizontal vector and a vertical vector; Splicing the horizontal vector and the vertical vector in the spatial dimension to obtain a spliced ​​vector; Performing layer normalization processing and nonlinear processing on the splicing vector in sequence to obtain the encoding result; The encoding result is split into two encoding vectors, the encoding vectors are sequentially subjected to convolution and activation processing, and the results after the two convolution and activation processing are re-weighted in the channel dimension to obtain the scale fusion feature map.

6. A multispectral image change detection method according to claim 5, characterized in that: After the re-weighted processing, the re-weighted result is up-sampled twice using bilinear interpolation to obtain the scale fusion feature map. After the same processing is performed on each layer of the differential feature map, the scale fusion feature map with the same number of layers as the differential feature map is obtained. The scale fusion feature maps are fused in order from large to small in number of layers to obtain the change detection map.

7. The multispectral image change detection method according to claim 1, characterized in that: The MobileNetV2 network is used to extract the multiple layers of features of the image at the first moment and the image at the second moment.

8. The multispectral image change detection method according to claim 1, characterized in that: The first change attention feature map and the second change attention feature map are differentially processed to obtain the differential feature map.

9. A multispectral image change detection system, the system applying the multispectral image change detection method according to any one of claims 1 to 8, characterized in that: include: An encoder is used to obtain an image at a first moment and an image at a second moment, wherein the image at the first moment and the image at the second moment are multispectral images collected at different moments in the same scene; extract multi-layer features of the image at the first moment and the image at the second moment respectively, wherein the features of the same layer constitute a feature pair, wherein the feature pair includes a first feature and a second feature; aggregate the first feature of the current layer and the first feature of the adjacent layer to obtain a first aggregate feature corresponding to the feature pair, and aggregate the second feature of the current layer and the second feature of the adjacent layer to obtain a second aggregate feature corresponding to the feature pair; extract the mutual attention distribution between the first aggregate feature and the second aggregate feature of the same layer to obtain a first change attention feature map and a second change attention feature map respectively, and determine a corresponding differential feature map according to the first change attention feature map and the second change attention feature map; The decoder is used to extract the multi-scale change information in the differential feature map of each layer, encode the multi-scale change information in different directions, fuse the encoding results in different directions to obtain a scale fusion feature map; and fuse the scale fusion feature maps of different layers in sequence to obtain a change detection map.

10. The multispectral image change detection system according to claim 9, characterized in that: Before detecting the image at the first moment and the image at the second moment, the encoder and the decoder perform structural optimization on the encoder and the decoder, and the loss function used in the optimization process adopts a composite loss function composed of a cross entropy loss and a block loss.

Images