Multi-scale nested remote sensing image change detection method, system and computer terminal

Through the attention mechanism and multi-scale nested network structure, the problems of loss of feature information of small objects and large amount of calculation are solved, and the accuracy and efficiency of remote sensing image change detection are improved.

CN114743110BActive Publication Date: 2024-12-03SHAANXI WASU INFORMATION TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210195921.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-01
Publication Date
2024-12-03
Estimated Expiration
2042-03-01

AI Technical Summary

Technical Problem

In neural networks, small objects are prone to lose their main feature information during downsampling due to less pixel information, and the existing multi-scale change enhancement method is computationally large.

Method used

A multi-scale nested remote sensing image change detection method based on attention mechanism is adopted, and high-level semantic features are extracted through twin networks. After the characteristics are fused, deep feature extraction is performed in a multi-scale superimposed network, and change detection is performed based on feature distance calculation and threshold determination.

Benefits of technology

The correlation of unchanged pixels is improved, the correlation of changing pixels is reduced, the calculation amount is reduced, and the accuracy and running speed of detection results are enhanced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114743110B_ABST
    Figure CN114743110B_ABST
Patent Text Reader

Abstract

The present invention belongs to the technical field of remote sensing image processing, and discloses a multi-scale nested remote sensing image change detection method, system and computer terminal. The multi-scale nested remote sensing image change detection method comprises: selecting a public data set and preprocessing the data set; constructing a network model based on the ECA attention mechanism to extract single-phase image features; designing and constructing a network model based on multi-scale double superposition; selecting training samples to train the network model and selecting a test set to test the network model. The present invention uses feature pyramids, twin networks and attention mechanisms to detect changes in remote sensing images, and at the same time weights two loss functions to solve the problem of class imbalance, considers multi-scale superposition to enhance the extraction of network fusion features and the impact of class imbalance on detection results, and greatly speeds up the running speed while reducing resource consumption.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image processing, and in particular relates to a multi-scale nested remote sensing image change detection method, system and computer terminal. Background Art

[0002] At present, change detection of remote sensing images is a process of quantitatively analyzing and determining the characteristics of ground object changes from remote sensing images obtained in different time periods. This technology has extremely important computer terminals in resource and environmental detection, natural disaster monitoring and evaluation, surveying and mapping, and military. For example, dynamic change detection of forests or vegetation, real-time monitoring of crop growth, urban changes (streets, buildings, etc.), dynamic monitoring of military strategic targets, etc.

[0003] In the early stage of remote sensing image change detection research, traditional change detection technology was more commonly used. In reality, the images obtained are often affected by some uncertain factors, such as lighting, season, shooting angle, etc. Therefore, traditional detection methods require preprocessing operations such as geometric correction and image registration to align the position, which is cumbersome and the accuracy of the detection results is low. Deep learning-based methods are increasingly used in the field of change detection. Inputting dual-phase or multi-phase remote sensing images with time series into the end-to-end model structure directly obtains the final change detection result map. The trained model structure has strong robustness and good adaptability. Different types of remote sensing image data can be input into the network, and the dependence of the detection result map on the difference map is eliminated.

[0004] Feature pyramids are mainly used in target detection. In neural networks, many networks use single high-level feature information, but this has a big flaw in detecting small objects, that is, small objects themselves have less pixel information, and it is very easy to lose their main feature information during the downsampling process. In order to solve the problem caused by the size difference of this object, the classic method is to use image pyramids to enhance multi-scale changes, but this method will bring a huge amount of calculation. However, the network structure of the feature pyramid can solve the problem of multi-scale changes while increasing the amount of calculation very little.

[0005] The basic idea of ​​the attention mechanism is to let the network learn to pay attention to important feature information, so that it can ignore irrelevant information and focus on key information. Most attention mechanisms are implemented through masks. The principle of masking is to represent the key features in the data through another layer of new weights, and to enable the network to learn the key features through learning and training. The channel attention mechanism focuses on the resource allocation between each convolution channel in the neural network. It learns the degree of dependence of each channel and adjusts different feature maps according to the degree of dependence.

[0006] The twin network measures the similarity between two inputs. The two inputs are placed in two networks, and the two networks share weights. Then the two networks map the inputs to a new space, form representations in the new space, and finally calculate the similarity of the two inputs through the loss function.

[0007] Through the above analysis, the problems and defects of the existing technology are as follows: in the neural network, small objects themselves have less pixel information, and it is very easy to lose their main feature information during the downsampling process; the existing method of using image pyramid for multi-scale change enhancement will bring a huge amount of calculation. Summary of the invention

[0008] In view of the problems existing in the prior art, the present invention provides a multi-scale nested remote sensing image change detection method, system and computer terminal, and more particularly, relates to a multi-scale nested remote sensing image change detection method, system and computer terminal based on an attention mechanism.

[0009] The present invention is implemented as follows: a multi-scale nested remote sensing image change detection method, the multi-scale nested remote sensing image change detection method comprising: firstly, two single-phase images are placed in a first-stage twin network to extract their respective high-level semantic features on the basis of shared parameters, and then the extracted features are fused and mapped to a new space; a second-stage multi-scale overlay network is constructed in the space, and the deep-level features of the fused features are extracted and similarity judgment is performed after the similarity judgment is performed, so that the correlation of unchanged pixels is increased and the correlation between changed pixels is reduced; in the third stage, the feature distance of the features output by the network is calculated, and finally the result graph of the change detection is obtained through the judgment of the threshold.

[0010] Furthermore, the multi-scale nested remote sensing image change detection method comprises the following steps:

[0011] Step 1: Select a public data set and preprocess it. Select a public data set and preprocess it. Preprocessing the data set makes the training data samples more complex, making the trained network structure more robust.

[0012] Step 2: construct a network model based on the ECA attention mechanism to extract single-phase image features; construct a network model based on the ECA attention mechanism to extract single-phase image features;

[0013] Step three, design and construct a network model based on multi-scale double superposition; design and construct a network model based on multi-scale double superposition to extract semantic features of fusion features, increase the correlation between unchanged pixels, and reduce the correlation between changed pixels;

[0014] Step 4: Select training samples to train the network model and select test sets to test the network model. Select training samples to train the network model and select test sets to test the network model to verify the effectiveness of the proposed model.

[0015] Furthermore, the preprocessing of the data set in step 1 includes cropping and image enhancement.

[0016] Furthermore, the feature extraction of the single-phase image in step 2 includes:

[0017] (1) In the stage of extracting single-phase image features, ResNet18 is used as the basic network to extract features; the feature map of the last layer of each residual module is upsampled through a 1×1 convolution operation, and then the upsampled features are merged by channel to fuse low-level features and high-level features to obtain richer features; the residual block is initialized with pre-trained weight values, and the parameters are adaptively adjusted based on the extracted general features to make the model suitable for specific data sets.

[0018] (2) The ECA attention mechanism is added to the ResNet18 network. After global average pooling without dimensionality reduction, the convolution kernel size is adaptively selected for one-dimensional convolution to obtain the attention to different channels. The twin network structure model is adopted to input two images of different phases into the network separately for feature extraction. In the last layer, the similarity loss function is calculated for the two feature vectors, and then the network as a whole is updated and trained. When the network structure of the two image inputs is exactly the same, it is also called weight sharing, which is actually to use the same structural framework to extract the feature information of each image.

[0019] Furthermore, the multi-scale network superposition in step three includes:

[0020] (1) The different features extracted by the twin network are fused and put into the multi-scale feature fusion network to extract fusion features of different scales; a new module is redesigned in the scale fusion network to extract information of different scales from the fused features again; a pyramid structure with three convolutions of 3×3, 5×5, and 7×7 is used, and a global pooling branch is introduced.

[0021] (2) The features output by the network are used to calculate the feature distance, and then the final change detection map is obtained through threshold determination.

[0022] Furthermore, the step 4 of selecting training samples to train the network model and selecting a test set to test the network model includes:

[0023] By training the network to learn the nonlinear transformation from input to embedding space, the embedding vectors corresponding to samples with high similarity are close, and vice versa. Contrastive loss is used to make each pixel pair in the embedding space have a larger distance if it changes, and a smaller distance if it does not change. The contrastive loss function is used in the training stage to learn the parameters of the network by reducing the loss function:

[0024]

[0025] Where d represents the Euclidean distance between the features of two samples; y represents whether the two samples are matching labels, y = 1 means the two samples are similar or matching, y = 0 means they are not matching; margin is the set threshold. In the test phase, the change map p is obtained by fixed threshold segmentation:

[0026]

[0027] Use the focalloss loss function to solve the problem of severe imbalance in the ratio of positive and negative samples:

[0028]

[0029] Among them, y is the label value; y' is the predicted value; α is the balance factor, which is used to balance the uneven proportion of positive and negative samples; γ is used to adjust the rate at which the weight of simple samples is reduced.

[0030] The total loss function is:

[0031] L=L cl +λL fl .

[0032] Among them, λ is the adjustment parameter of the two loss functions.

[0033] Another object of the present invention is to provide a multi-scale nested remote sensing image change detection system for implementing the multi-scale nested remote sensing image change detection method, the multi-scale nested remote sensing image change detection system comprising:

[0034] Dataset selection module, used to select public datasets;

[0035] The dataset preprocessing module is used to crop the dataset and perform image enhancement preprocessing;

[0036] Image feature extraction module, used to extract single-phase image features by building a network model based on the ECA attention mechanism;

[0037] Network model building module, used to design and build a network model based on multi-scale double superposition;

[0038] The network model testing module is used to select training samples to train the network model and select test sets to test the network model.

[0039] Another object of the present invention is to provide a computer device, the computer device comprising a memory and a processor, the memory storing a computer program, and when the computer program is executed by the processor, the processor performs the following steps:

[0040] First, the high-level features of the two single-phase images are extracted, and then the features are fused and put into a multi-scale superposition network to extract the fused features again, and finally the fused features are separated.

[0041] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the following steps:

[0042] First, the high-level features of the two single-phase images are extracted, and then the features are fused and put into a multi-scale superposition network to extract the fused features again, and finally the fused features are separated.

[0043] Another object of the present invention is to provide an information data processing terminal, which is used to implement the multi-scale nested remote sensing image change detection system.

[0044] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are analyzed from the following aspects:

[0045] First, in view of the technical problems existing in the above-mentioned prior art and the difficulty of solving the problems, the technical solutions to be protected by the present invention and the results and data during the research and development process are closely combined to analyze in detail and deeply how the technical solutions of the present invention solve the technical problems, and some creative technical effects brought about after solving the problems. The specific description is as follows:

[0046] The present invention discloses a method for change detection of dual-phase remote sensing images. The method introduces an attention mechanism to extract features from a single-phase image, fuses two sets of features, maps the fused features to a specific feature space, extracts deep features of the fused features in the space through a multi-scale nested network, and then performs similarity judgment again, so that the correlation of unchanged pixels is increased and the correlation between changed pixels is reduced. At the same time, two loss functions are weighted to solve the problem of category imbalance.

[0047] Second, considering the technical solution as a whole or from the perspective of the product, the technical effects and advantages of the technical solution to be protected by the present invention are described in detail as follows:

[0048] The present invention mainly utilizes feature pyramid, twin network and attention mechanism to perform change detection on remote sensing images, considers multi-scale superposition to enhance the extraction of network fusion features, considers the impact of class imbalance on detection results, and greatly speeds up the operation speed while reducing resource consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0050] Figure 1 It is a flow chart of a multi-scale nested remote sensing image change detection method provided by an embodiment of the present invention.

[0051] Figure 2 It is a schematic diagram of a multi-scale nested remote sensing image change detection method provided by an embodiment of the present invention.

[0052] Figure 3 It is a structural block diagram of a multi-scale nested remote sensing image change detection system provided by an embodiment of the present invention;

[0053] In the figure: 1. Data set selection module; 2. Data set preprocessing module; 3. Image feature extraction module; 4. Network model construction module; 5. Network model testing module.

[0054] Figure 4 It is an overall network framework diagram provided by an embodiment of the present invention.

[0055] FIG. 5( a ) is a framework diagram of a single-phase image feature extraction method according to an embodiment of the present invention.

[0056] FIG5( b ) is a framework diagram of multi-scale superposition provided by an embodiment of the present invention.

[0057] FIG. 6( a ) and FIG. 6( b ) are schematic diagrams of experimental results provided by an embodiment of the present invention.

[0058] Figure 7 It is a schematic diagram of experimental results provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0060] 1. Explanatory Examples In order to enable those skilled in the art to fully understand how to implement the present invention, this section provides an illustrative example that expands and describes the technical solution of the claims.

[0061] In view of the problems existing in the prior art, the present invention provides a multi-scale nested remote sensing image change detection method, system and computer terminal, and the present invention is described in detail below in conjunction with the accompanying drawings.

[0062] like Figure 1 As shown, the multi-scale nested remote sensing image change detection method provided by the embodiment of the present invention includes the following steps:

[0063] S101, selecting a public data set and preprocessing the data set;

[0064] S102, constructing a network model based on the ECA attention mechanism to extract single-phase image features;

[0065] S103, design and construct a network model based on multi-scale double superposition;

[0066] S104, selecting training samples to train the network model and selecting a test set to test the network model.

[0067] The schematic diagram of the multi-scale nested remote sensing image change detection method provided by the embodiment of the present invention is shown in Figure 2 .

[0068] like Figure 3 As shown, the multi-scale nested remote sensing image change detection system provided by the embodiment of the present invention includes:

[0069] Data set selection module 1, used to select public data sets;

[0070] Dataset preprocessing module 2, used for cropping the dataset and preprocessing image enhancement;

[0071] Image feature extraction module 3, used to extract single-phase image features by building a network model based on the ECA attention mechanism;

[0072] The network model construction module 4 is used to design and construct a network model based on multi-scale double superposition;

[0073] The network model testing module 5 is used to select training samples to train the network model and select test sets to test the network model.

[0074] The present invention proposes a multi-scale feature superposition network structure. First, high-level features of two single-phase images are extracted, and then the features are fused and put into a multi-scale superposition network to extract the fused features again, and finally the fused features are separated.

[0075] This embodiment provides a mapping method for remote sensing image change detection, the flow chart is as follows: Figure 2 As shown, the overall framework diagram is as follows Figure 4 As shown, the following steps are included:

[0076] Step 1, selection of data set;

[0077] In order to verify the robustness of the proposed model structure, two datasets were selected to test the performance of the network model.

[0078] Dataset 1: LEVIR-CD dataset is used for change detection. This dataset collects 637 high-resolution image pairs with a size of 1024*1024 pixels. These dual-phase images come from 20 different areas in several cities in Texas, USA. The collection time is from 2002 to 2018.

[0079] Dataset 2: The CDD dataset is used for change detection. The dataset contains 16,000 sets of images that change with the seasons obtained from Google Earth. The image size is 256×256 pixels and the spatial resolution is 3 to 100 cm / px.

[0080] Step 2: Preprocess, crop and enhance the dataset.

[0081] Step 3: extract the features of the single-phase image, as shown in Figure 5(a).

[0082] Step 31, in the stage of extracting single-phase image features, ResNet18 is used as the basic network to extract features. Since the high-level semantic features extracted by the neural network are accurate but the position information is rough, and the position information extracted by the low-level is fine but lacks semantic information, the feature map of the last layer of each residual module is upsampled by a 1×1 convolution operation, and then the upsampled features are merged by channels to fuse the low-level features and high-level features to obtain richer features. The residual block is initialized with pre-trained weight values, and the parameters are adaptively adjusted based on the extracted general features to make the model suitable for a specific data set. This can reduce the number of network training times.

[0083] Step 32: Reducing the dimension will lose some potentially important feature information, so the ECA attention mechanism is added to the ResNet18 network.

[0084] The structure of the ECA module is shown in Figure 5(a). The extracted high-dimensional features I∈R H×W×C , after global average pooling, it is transformed into I′∈R 1×C, and then adaptively select a suitable k value so that I′ can be operated with the adjacent k feature vectors to regain the assigned weight value, and finally act on the original feature map.

[0085]

[0086] where w i is the weight value of each channel, is the weight value of the jth channel adjacent to the i channel, is the element value of the jth channel adjacent to the i channel.

[0087] w=σ(C1D K (y))

[0088] The strategy expressed in this formula can be implemented through convolution operations, that is, using 1D convolution.

[0089] Where C1D represents a 1D convolution operation of size k.

[0090]

[0091]

[0092] The selection of the adaptive k value can be obtained according to the above formula, where b and γ are adjustable parameters.

[0093] Without dimensionality reduction, it first undergoes global average pooling, and then adaptively selects the convolution kernel size for one-dimensional convolution to obtain the degree of attention to different channels. This part adopts the structural model of the twin network, and inputs two images of different phases into the network separately for feature extraction. Then, in the last layer, the similarity loss function is calculated for the two feature vectors, and the entire network is updated and trained. When the network structure of the two image inputs is exactly the same, it is also called weight sharing, which actually uses the same structural framework to extract the feature information of each of the two images.

[0094] Step 4: Multi-scale network superposition, as shown in Figure 5(b).

[0095] Step 41, fuse the different features extracted by the twin network and put them into the multi-scale feature fusion network to extract fusion features of different scales. In the scale fusion network, a new module is redesigned to extract information of different scales from the fused features again. Due to the low resolution of high-level features, the use of large convolution kernels does not cause too much computational burden. A pyramid structure with three convolutions of 3×3, 5×5, and 7×7 is used. At the same time, a global pooling branch is introduced to further improve performance;

[0096] The specific model structure is shown in the PFM module in Figure 5(b). The fused feature F in At the same time, it passes through three branches. The first branch passes through two layers of 7×7 convolution, 3×3 and 5×5 convolution, and 4×4 deconvolution; the second branch directly takes the fusion feature F in After 1×1 convolution, the two branch features are fused and then merged into the third fusion feature F in After the global pooling branch layer, the final output feature F is obtained out .

[0097] This module first obtains the features of larger objects through 7×7 convolution at the same layer, and then obtains the features of smaller objects through 3×3 and 5×5 kernel convolutions of two different scales. At the same time, it fuses the global information of the initial features to capture more receptive field information.

[0098] Step 42, calculate the feature distance of the features output by the network, and then obtain the final change detection map through threshold determination.

[0099] Step 5, this method is a method based on deep metrics. This method trains the network to learn the nonlinear transformation from input to embedding space. The embedding vectors corresponding to samples with large similarity are close, and vice versa. Therefore, contrast loss is used here to make each pixel pair in the embedding space have a larger distance if it changes, and a smaller distance if it does not change. The contrast loss function is used in the training stage to learn the parameters of the network by reducing the loss function.

[0100]

[0101] Where d represents the Euclidean distance between the features of two samples, y represents whether the two samples are matching labels, y=1 means the two samples are similar or matching, y=0 means they are not matching, and margin is the set threshold. In the test phase, the change map p is obtained by fixed threshold segmentation:

[0102]

[0103] In order to reduce the imbalance between the changed and unchanged sample categories in remote sensing images, the focal loss function is used to solve the problem of severe imbalance in the ratio of positive and negative samples.

[0104]

[0105] Among them, y is the label value, y' is the predicted value, α is the balance factor, which is used to balance the uneven ratio of positive and negative samples. γ is used to adjust the rate at which the weight of simple samples is reduced.

[0106] Therefore, the total loss function is:

[0107] L=L cl +λL fl

[0108] Among them, λ is the adjustment parameter of the two loss functions.

[0109] The change detection results on the last two datasets are shown in Figure 6(a) and Figure 6(b).

[0110] 2. Computer terminal implementation examples In order to prove the creativity and technical value of the technical solution of the present invention, this section provides computer terminal implementation examples on specific products or related technologies for the technical solution of the claims.

[0111] The present invention is implemented using software, and the experiment is implemented on an Ubuntu system with a system memory of 64 GB, a graphics card of GeForce RTX 2080 Ti, and a graphics card memory of 11 GB.

[0112] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. It can be understood by a person of ordinary skill in the art that the above-mentioned devices and methods can be implemented using computer executable instructions and / or contained in a processor control code, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. Such code is provided on the carrier medium. The device and its modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, and can also be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0113] 3. Evidence of the effects of the embodiments. The embodiments of the present invention have achieved some positive effects during the development or use process, and indeed have great advantages over the prior art. The following content is described in conjunction with the data, charts, etc. of the test process.

[0114] In order to verify the overall effectiveness of the proposed network structure, the proposed network is compared with STANet, DSMSNNet, and Siam-NestedUnet on the same dataset. The experimental results are shown in Figure 7 shown.

[0115] like Figure 7As shown in the figure, the first and second columns are remote sensing images taken in two time periods, the third column is the label map, the fourth column is the result map of STANet, the fifth column is the result map of DSMSCNNet, the sixth column is the result map of Siam-NestedUNet, and the seventh column is the result map of the proposed method. From the third set of data, it can be seen that the general contours of the objects detected by STANet and Siam-NestedUNet are not complete. In addition, in the remaining data, all methods have detected the basic contours of the changing objects, and the difference lies in the accuracy of the shape contours of the detected objects. From the fourth set of data, it can be seen that the proposed method has fewer detected error points than STANet and Siam-NestedUNet. In summary, the proposed method is more detailed than other methods, the contours are clearer, and it is more resistant to noise points.

[0116] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A multi-scale nested remote sensing image change detection method, characterized in that: The multi-scale nested remote sensing image change detection method comprises: Step 1: Select a public data set and preprocess the data set; Step 2: Build a network model based on ECA attention mechanism to extract single-phase image features; Step 3: Design and construct a network model based on multi-scale double superposition; Step 4: Select training samples to train the network model and select test sets to test the network model; The feature extraction of the single-phase image in step 2 includes: (1) In the stage of extracting single-phase image features, ResNet18 is used as the basic network to extract features; the feature map of the last layer of each residual module is upsampled through a 1×1 convolution operation, and then the upsampled features are merged by channel to fuse low-level features and high-level features to obtain richer features; the residual block is initialized with pre-trained weight values, and the parameters are adaptively adjusted based on the extracted general features to make the model suitable for specific data sets; (2) Add the ECA attention mechanism to the ResNet18 network. After global average pooling without dimensionality reduction, the convolution kernel size is adaptively selected for one-dimensional convolution to obtain the attention of different channels. The twin network structure model is adopted to input two images of different phases into the network separately for feature extraction. The similarity loss function is calculated for the two feature vectors in the last layer, and then the network is updated and trained as a whole. When the network structure of the two image inputs is exactly the same, it is also called weight sharing, which is actually to use the same structural framework to extract the feature information of the two images. The multi-scale network superposition in step 3 includes: (a) The different features extracted by the twin network are fused and put into the multi-scale feature fusion network to extract fusion features of different scales; a new module is redesigned in the scale fusion network to extract information of different scales from the fused features again; a pyramid structure with three convolutions of 3×3, 5×5, and 7×7 is used, and a global pooling branch is introduced; (b) The features output by the network are used to calculate the feature distance, and then the final change detection map is obtained through threshold determination.

2. The multi-scale nested remote sensing image change detection method according to claim 1, characterized in that: The preprocessing of the data set in step 1 includes cropping and image enhancement.

3. The multi-scale nested remote sensing image change detection method according to claim 1, characterized in that: The step 4 of selecting training samples to train the network model and selecting a test set to test the network model includes: By training the network to learn the nonlinear transformation from input to embedding space, the embedding vectors corresponding to samples with high similarity are close, and vice versa. Contrastive loss is used to make each pixel pair in the embedding space have a larger distance if it changes, and a smaller distance if it does not change. The contrastive loss function is used in the training stage to learn the parameters of the network by reducing the loss function: , in, d Represents the Euclidean distance between two sample features; y Indicates whether two samples are matching labels, y=1 means the two samples are similar or matching, y=0 means they do not match; margin is the set threshold; In the test phase, the change map is obtained by fixed threshold segmentation. p : ; Use the focalloss loss function to solve the problem of severe imbalance in the ratio of positive and negative samples: , in, y is the tag value; is the predicted value; It is a balancing factor, which is used to balance the uneven ratio of positive and negative samples. Used to adjust the rate at which the weight of simple samples is reduced; The total loss function is: , Among them, λ is the adjustment parameter of the two loss functions.

4. A multi-scale nested remote sensing image change detection system implementing the multi-scale nested remote sensing image change detection method according to any one of claims 1 to 3, characterized in that: The multi-scale nested remote sensing image change detection system comprises: Dataset selection module, used to select public datasets; The dataset preprocessing module is used to crop the dataset and perform image enhancement preprocessing; Image feature extraction module, used to extract single-phase image features by building a network model based on the ECA attention mechanism; Network model building module, used to design and build a network model based on multi-scale double superposition; The network model testing module is used to select training samples to train the network model and select test sets to test the network model.

5. A computer device, characterized in that: The computer device includes a memory and a processor, the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the multi-scale nested remote sensing image change detection method as described in any one of claims 1 to 3.

6. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the steps of the multi-scale nested remote sensing image change detection method as described in any one of claims 1 to 3.

7. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the multi-scale nested remote sensing image change detection system as described in claim 4.