Remote Sensing Image Semantic Change Detection Method for Nature Reserves Based on Coordinate Attention and Dual-Temporal Semantic Reasoning

Abstract

The present invention relates to a method for semantic change detection of remote sensing images of nature reserves based on coordinate attention and dual-time semantic reasoning, including: acquiring remote sensing image data; performing data preprocessing and data augmentation; constructing a residual coordinate attention module, constructing a dual-time semantic reasoning module, and constructing a CAB-SRNet network model; based on a joint loss function, training the CAB-SRNet network model using a training set, and after training is completed, evaluating the recognition accuracy and extraction effect of the CAB-SRNet network model using a test set; performing semantic change recognition: after preprocessing the remote sensing image of the nature reserve to be extracted, inputting it into the trained CAB-SRNet network model, and the CAB-SRNet network model outputs a prediction image to obtain the semantic change recognition result of the land use of the nature reserve. The present invention can effectively identify the position information and boundary information of the changed area, improve the change recognition accuracy; and can better distinguish land use types in key areas to obtain good semantic change recognition effects.

Description

Technical Field

[0001] The present invention relates to the technical field of image change recognition, and in particular to a method for semantic change detection of remote sensing images of nature reserves based on coordinate attention and dual-time semantic reasoning. Background Technique

[0002] Nature reserves are the core carriers of ecological construction and play a primary role in maintaining national ecological security. With the rapid development of industrialization and urbanization, the contradiction between protection and development has become increasingly prominent. The fragmentation of environmental patches caused by human activities poses a great threat to biodiversity. Nature reserves are numerous, large in area, wide in distribution, complex in geographical environment, complex in construction projects, and short of supervision personnel. It is difficult to detect and supervise various illegal activities in the nature reserves in a timely manner relying on traditional ground survey methods. Satellite remote sensing has the characteristics of macroscopicity, timeliness, objectivity, etc., and is the best means to carry out the supervision of human activities in nature reserves.

[0003] With the increasing popularity of high-resolution remote sensing images, the potential application scope of change detection in high-resolution dual-temporal images has been expanded. Change detection based on high-resolution remote sensing images is still a challenging task in the following two aspects: the complexity of objects in the scene and different imaging conditions, both of which result in objects with the same semantic concept showing different spectral characteristics at different times and different spatial positions (spatiotemporal). Therefore, there is an urgent need to develop a better change detection model that can identify these relevant changes in complex scenes while avoiding complex irrelevant changes caused by seasonal changes, building shadows, atmospheric changes, and changes in lighting conditions. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for semantic change detection of remote sensing images of nature reserves based on coordinate attention and dual-time semantic reasoning, which realizes automatic semantic change detection of remote sensing images, represents rich change information of the images at the same time, improves the supervision efficiency and accuracy of nature reserves, and greatly alleviates the shortage of human resources.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions: A method for semantic change detection of remote sensing images of nature reserves based on coordinate attention and dual-time semantic reasoning, the method includes the following steps in sequence:

[0006] (1) Obtain remote sensing image data: Download high-resolution remote sensing image data of Beijing-2 and Gaojing-1 in two periods of nature reserve areas;

[0007] (2) Perform data preprocessing and data augmentation: Screen, splice, and crop the acquired remote sensing images according to the vector data of nature reserves; Semantically annotate the changed areas in two periods of remote sensing images to obtain label images L1 and L2, where L1 is the semantic change map before the change and L2 is the semantic change map after the change; Crop and perform data augmentation on the remote sensing images and label images to obtain a nature reserve semantic change detection dataset; Divide the nature reserve semantic change detection dataset into a training set, a test set, and a validation set according to the ratio of 7:2:1;

[0008] (3) Construct the CAB-SRNet network model: Add a coordinate attention mechanism to the residual block to construct a residual coordinate attention module, construct a dual-temporal semantic reasoning module, and add a residual coordinate attention module and a dual-temporal semantic reasoning module to the BI-SRNet network to construct the CAB-SRNet network model;

[0009] (4) Perform model training and evaluation: Based on the semantic class loss L sem , binary change loss L change , and semantic change loss L sc combined joint loss function, use the training set to train the CAB-SRNet network model, and use the test set to evaluate the recognition accuracy and extraction effect of the CAB-SRNet network model after training;

[0010] (5) Perform semantic change recognition: After preprocessing the remote sensing images of nature reserves to be extracted, input them into the trained CAB-SRNet network model, and the CAB-SRNet network model outputs a prediction image to obtain the recognition result of the land use semantic change in nature reserves.

[0011] The specific steps of step (2) include the following steps:

[0012] (2a) Through the Arcgis10.5 software, screen, splice, and crop the remote sensing images of nature reserves; Semantically annotate the changed areas in two periods of remote sensing images of nature reserves to obtain label images L1 and L2. L1 is the semantic change map before the change and L2 is the semantic change map after the change. In the label images, no change is white, water is blue, bare land is brown, vegetation is green, buildings are red, roads are gray, and greenhouses are black; Through the sliding window method, crop the remote sensing images and label images into images of 256×256 size;

[0013] (2b) Perform data augmentation on the cropped remote sensing images and label images to expand the data volume and obtain a nature reserve semantic change detection dataset. The data augmentation includes:

[0014] Horizontal flipping: Use the image processing library OpenCV to perform horizontal flipping on the remote sensing image and the label image respectively;

[0015] Vertical flipping: Use the image processing library OpenCV to perform vertical flipping on the remote sensing image and the label image respectively;

[0016] Horizontal and vertical flipping: Use the image processing library OpenCV to perform horizontal flipping first and then vertical flipping on the remote sensing image and the label image respectively;

[0017] Shifting, scaling, random cropping, and adding noise: Perform shifting, scaling, random cropping, and adding noise on the remote sensing image and the label image respectively;

[0018] (2c) Divide the natural protected area semantic change detection dataset into a training set, a test set, and a validation set according to the ratio of 7:2:1; the training set is used to directly participate in the training of the CAB-SRNet network model for feature extraction; the validation set is used for validation after each training stage; the test set is used to test the recognition accuracy and extraction effect of the CAB-SRNet network model after training.

[0019] The specific steps of step (3) are as follows:

[0020] (3a) First, perform feature extraction on the natural protected area remote sensing image. Use the fully convolutional network to perform semantic feature extraction on the two-phase remote sensing images I1 and I2 to obtain semantic features X1 and X2;

[0021] (3b) Construct a residual coordinate attention module: Embed the coordinate attention CA module into the residual block of the ResNet network to obtain a residual coordinate attention module;

[0022] Use the residual coordinate attention module to merge the dual-temporal features for binary change detection. Use the ResNet network combined with the coordinate attention mechanism to change the information through the unaligned semantic features X1 and X2, and then map it to the binary change map C. All mappings are performed through a 1×1 convolutional layer with non-shared weights;

[0023] (3c) Construct a dual-temporal semantic reasoning module: Model the temporal correlation and consistency between the semantic features X1 and X2, and generate a query vector q i ∈R c×h×w , a key vector k i ∈R H×c′ , and a value vector v i ∈R c'×H from the input feature X i ∈R c×H , where i = 1, 2, c is the number of channels, h and w are the spatial sizes, H = hw, c' = c / r, r is the channel reduction coefficient, qi 、k i 、v i are obtained by feeding into different convolutional layers from X i Note that matrices A1, A2, A 11 、A 22 ∈R H×H The calculation formula is:

[0024]

[0025] Among them, is a softmax normalization function along the row dimension. A1, A2, A 11 、A 22 are used to record the correlation between each pair of spatial positions to obtain an enhanced feature P1, P2:

[0026] P1 = X1 + v1 × A1 + v1 × A 11

[0027] P2 = X2 + v2 × A2 + v2 × A 22

[0028] The semantic maps Q1, Q2 are obtained by mapping through a 1×1 convolutional layer with non-shared weights;

[0029] (3d) The semantic change maps S1, S2 are obtained by masking the semantic maps Q1, Q2 with the binary change map C:

[0030] S1, S2 = C · (P1, P2);

[0031] (3e) Add a residual coordinate attention module and a dual-time semantic reasoning module to the BI-SRNet network to construct the CAB-SRNet network model.

[0032] The specific steps of step (4) include the following steps:

[0033] (4a) Combine the semantic class loss L sem , the binary change loss L change and the semantic change loss L sc to form a joint loss function L scd ;

[0034] The semantic class loss L sem refers to the multi-class cross-entropy loss between the semantic maps Q1, Q2 and the label images L1 and L2. The calculation formula of the semantic class loss L sem for each pixel is:

[0035]

[0036] Among them, N is the number of semantic classes, and N is set according to the number of classes in the dataset; yi represents the predicted probability of the true value label, p i represents the predicted probability of the i-th class;

[0037] Binary change loss L change is the predicted binary change map C and the reference change map L c The binary cross entropy loss between, refer to the change graph L c Generated using L1 or L2, the binary change loss L for each pixel change The calculation formula is:

[0038] L change =-yclog(p c )-(1-y c )log(1-p c )

[0039] Among them, y c Represents the probability of change of the true value label, p c represents the predicted probability of change;

[0040] Use the cosine loss function to calculate the predicted semantic graphs Q1, Q2 and the reference change graph L c The semantic change loss L between sc , the calculation formula is as follows:

[0041]

[0042] Among them, x1 and x2 are the feature vectors of a pixel on Q1 and Q2 respectively;

[0043] The training of the two feature embedding branches in the dual-temporal semantic reasoning module is directly supervised by L1 and L2, and c assisted, and the training of the residual coordinate attention module is performed by L c Direct supervision, the formula is as follows:

[0044]

[0045] in, and is the semantic loss of each temporal branch;

[0046] (4b) The CAB-SRNet network model was trained: the pre-trained weights of the ResNet34 network were used during training, the momentum was set to 0.99, and the weight decay was set to 0.0005; the learning rate was initially set to 0.01 and then linearly decayed to 0 until 200 epochs were trained; verification was performed after each training stage;

[0047] (4c) Model accuracy evaluation: The overall accuracy OA and intersection over union IoU are used as evaluation metrics to evaluate the accuracy. The calculation formulas of the evaluation metrics are as shown in the formula:

[0048]

[0049] In the formula, TP represents the number of true positives, FP represents the number of false positives, FN represents the number of false negatives, and TN represents the number of true negatives.

[0050] The specific steps of step (5) include the following steps:

[0051] (5a) After preprocessing the remote sensing image of the nature reserve to be identified for semantic changes, resize it to a size of 256×256;

[0052] (5b) Input the adjusted image into the trained CAB-SRNet network model. After passing through the CAB-SRNet network model, a predicted image is output, and the predicted images are stitched to obtain the semantic change recognition result of the nature reserve. In the predicted image, no change is white, water is blue, bare land is brown, vegetation is green, buildings are red, roads are gray, and greenhouses are black.

[0053] From the above technical solutions, the beneficial effects of the present invention are as follows: First, the designed network of the present invention adds a coordinate attention mechanism to the residual block of the ResNet network, effectively combines semantic features for binary change detection, can effectively identify the position information and boundary information of the change area, and improves the change recognition accuracy; Second, the constructed dual-time semantic reasoning module of the present invention effectively models the temporal and spatial correlations of dual-time remote sensing images of nature reserves, improves the temporal consistency, and can better distinguish land use types in key areas; Third, aiming at the problems of scattered change areas and unbalanced change type samples in the nature reserve dataset, it is feasible to use the present invention for semantic change detection of land use in nature reserves. This method effectively utilizes the advantage of rich spatial information in high-resolution data, and at the same time models the temporal correlation of land use types with dual-time image changes to obtain good semantic change recognition effects. Description of the Drawings

[0054] Figure 1 is the flowchart of the method of the present invention;

[0055] Figure 2 is the block diagram of the CAB-SRNet network model;

[0056] Figure 3 is the block diagram of the residual coordinate attention module;

[0057] Figure 4 is the block diagram of the dual-time semantic reasoning module;

[0058] Figure 5 Schematic diagram of the multi-scale feature processing process of the dual-time semantic reasoning module. Specific implementation manner

[0059] As Figure 1 shown, a semantic change detection method for remote sensing images of nature reserves based on coordinate attention and dual-time semantic reasoning, the method includes the following steps in sequence:

[0060] (1) Obtain remote sensing image data: Download high-resolution remote sensing image data of Beijing-2 and Gaojing-1 in two periods of nature reserve areas;

[0061] (2) Perform data preprocessing and data augmentation: Screen, splice and crop the obtained remote sensing images according to the vector data of nature reserves; Semantically annotate the changed areas of the two periods of remote sensing images to obtain label images L1 and L2, where L1 is the semantic change map before change, and L2 is the semantic change map after change; Crop and perform data augmentation on the remote sensing images and label images to obtain a nature reserve semantic change detection data set; Divide the nature reserve semantic change detection data set into a training set, a test set and a validation set according to the ratio of 7:2:1;

[0062] (3) Construct the CAB-SRNet network model: Add a coordinate attention mechanism to the residual block to construct a residual coordinate attention module, construct a dual-time semantic reasoning module, add a residual coordinate attention module and a dual-time semantic reasoning module to the BI-SRNet network to construct the CAB-SRNet network model;

[0063] (4) Perform model training and evaluation: Based on the combined joint loss function of the semantic class loss L sem , the binary change loss L change , and the semantic change loss L sc , use the training set to train the CAB-SRNet network model, and use the test set to evaluate the recognition accuracy and extraction effect of the CAB-SRNet network model after training;

[0064] (5) Perform semantic change recognition: After preprocessing the remote sensing image of the nature reserve to be extracted, input it into the trained CAB-SRNet network model, and the CAB-SRNet network model outputs a prediction image to obtain the recognition result of the land use semantic change of the nature reserve.

[0065] The step (2) specifically includes the following steps:

[0066] (2a) Screen, splice and crop the remote sensing images of nature reserves through the Arcgis 10.5 software; perform semantic annotation on the changed areas of the two-phase remote sensing images of nature reserves to obtain label images L1 and L2. L1 is the semantic change map before the change, and L2 is the semantic change map after the change. In the label images, no change is white, water is blue, bare land is brown, vegetation is green, buildings are red, roads are gray, and greenhouses are black; crop the remote sensing images and label images in the form of a sliding window into images of 256×256 size;

[0067] (2b) Perform data augmentation on the cropped remote sensing images and label images to expand the data volume and obtain a semantic change detection dataset for nature reserves. The data augmentation includes:

[0068] Horizontal flipping: Use the image processing library OpenCV to perform horizontal flipping on the remote sensing images and label images respectively;

[0069] Vertical flipping: Use the image processing library OpenCV to perform vertical flipping on the remote sensing images and label images respectively;

[0070] Horizontal and vertical flipping: Use the image processing library OpenCV to perform horizontal flipping first and then vertical flipping on the remote sensing images and label images respectively;

[0071] Translation, scaling, random cropping and adding noise: Perform translation, scaling, random cropping and adding noise on the remote sensing images and label images respectively;

[0072] (2c) Divide the semantic change detection dataset for nature reserves into a training set, a test set and a validation set according to the ratio of 7:2:1 respectively; the training set is used to directly participate in the training of the CAB-SRNet network model for feature extraction; the validation set is used for validation after each training stage; the test set is used to test the recognition accuracy and extraction effect of the CAB-SRNet network model after training.

[0073] The specific steps of step (3) include the following steps:

[0074] (3a) First, perform feature extraction on the remote sensing images of nature reserves. Use a fully convolutional network to perform semantic feature extraction on the two-phase remote sensing images I1 and I2 to obtain semantic features X1 and X2;

[0075] (3b) Construct a residual coordinate attention module: Embed the coordinate attention CA module into the residual block of the ResNet network to obtain a residual coordinate attention module;

[0076] Using a residual coordinate attention module to merge dual-temporal features for binary change detection, the ResNet network is combined with the coordinate attention mechanism to change information through unaligned semantic features X1 and X2, and then mapped into the binary change map C. All mappings are performed through 1×1 convolutional layers with non-shared weights;

[0077] (3c) Construct a dual-temporal semantic reasoning module: Model the temporal correlation and consistency between semantic features X1 and X2, and generate query vector q i ∈R c×h×w from the input feature X i ∈R H×c′ , key vector k i ∈R c'×H , and value vector v i ∈R c×H , where i = 1, 2, c is the number of channels, h and w are the spatial sizes, H = hw, c′ = c / r, r is the channel reduction coefficient, and q i , k i , and v i are obtained by feeding X i into different convolutional layers. Note that the calculation formulas for matrices A1, A2, A 11 , and A 22 ∈R H×H are:

[0078]

[0079] where is a softmax normalization function along the row dimension. A1, A2, A 11 , and A 22 are used to record the correlation between each pair of spatial positions, and enhanced features P1 and P2 are obtained:

[0080] P1 = X1 + v1 × A1 + v1 × A 11

[0081] P2 = X2 + v2 × A2 + v2 × A 22

[0082] Semantic maps Q1 and Q2 are obtained through mapping by 1×1 convolutional layers with non-shared weights;

[0083] (3d) Semantic change maps S1 and S2 are obtained by masking semantic maps Q1 and Q2 with the binary change map C:

[0084] S1, S2 = C · (P1, P2);

[0085] (3e) Add a residual coordinate attention module and a dual-temporal semantic reasoning module to the BI-SRNet network to construct the CAB-SRNet network model.

[0086] Step (4) specifically includes the following steps:

[0087] (4a) Combine the semantic class loss L sem , the binary change loss L change and the semantic change loss L sc to form a joint loss function L scd ;

[0088] The semantic class loss L sem refers to the multi-class cross-entropy loss between the semantic graphs Q1, Q2 and the label images L1 and L2. The semantic class loss L sem for each pixel is calculated as follows:

[0089]

[0090] where N is the number of semantic classes, and N is set according to the number of classes in the dataset; y i represents the predicted probability of the true label, and p i represents the predicted probability of the i-th class;

[0091] The binary change loss L change is the binary cross-entropy loss between the predicted binary change graph C and the reference change graph L c . The reference change graph L c is generated using L1 or L2. The binary change loss L change for each pixel is calculated as follows:

[0092] L change = -y c log(p c ) - (1 - y c ) log(1 - p c )

[0093] where y c represents the change probability of the true label, and p c represents the predicted change probability;

[0094] Use the cosine loss function to calculate the semantic change loss L c between the predicted semantic graphs Q1, Q2 and the reference change graph L sc , and the calculation formula is as follows:

[0095]

[0096] where x1 and x2 are the feature vectors of a pixel on Q1 and Q2 respectively;

[0097] The training of the two feature embedding branches in the dual-time semantic reasoning module is directly supervised by L1 and L2, and by Lc is assisted, while the training of the residual coordinate attention module is supervised by L c directly, and the formula is as follows:

[0098]

[0099] where, and are the semantic losses of each time branch;

[0100] (4b) Training the CAB-SRNet network model: During training, the pre-trained weights of the ResNet34 network are used, the momentum is set to 0.99, and the weight decay is set to 0.0005; the learning rate is initially set to 0.01 and then linearly decays to 0 until 200 epochs are trained; validation is performed after each training stage;

[0101] (4c) Model accuracy evaluation: The overall accuracy OA and intersection over union IoU are used as evaluation metrics to evaluate the accuracy, and the calculation formulas of the evaluation metrics are as shown in the formula:

[0102]

[0103]

[0104] In the formula, TP represents the number of true positives, FP represents the number of false positives, FN represents the number of false negatives, and TN represents the number of true negatives.

[0105] The specific steps of step (5) include the following steps:

[0106] (5a) After preprocessing the remote sensing image of the nature reserve to be identified for semantic changes, resize it to a size of 256×256;

[0107] (5b) Input the adjusted image into the trained CAB-SRNet network model, and output a predicted image through the CAB-SRNet network model. The predicted images are stitched to obtain the semantic change recognition result of the nature reserve. In the predicted image, no change is white, water is blue, bare land is brown, vegetation is green, buildings are red, roads are gray, and greenhouses are black.

[0108] Example 1

[0109] To verify the effectiveness of the present invention, BI-SRNet is selected as a comparative example, and the results are compared for the nature reserve semantic change detection dataset. The overall accuracy and intersection over union of the comparison algorithms are shown in Table 1 below:

[0110] Table 1

[0111] Method OA IOU BI-SRNet 94.9 78.94 CAB-SRNet 96.47 83.12

[0112] As Figure 2 , 3 , as shown in Figure 4, in the CAB-SRNet network model of the present invention, a Siamese network is adopted for semantic feature extraction of dual-temporal remote sensing images I1 and I2; the Resca_CD module, i.e., the residual coordinate attention module, merges dual-temporal features for binary change detection; meanwhile, the TS_SR module, i.e., the dual-temporal semantic reasoning module, models the temporal correlation and consistency between semantic features X1 and X2; finally, the semantic change maps S1 and S2 are obtained by masking the semantic maps Q1 and Q2 with the binary change map C.

[0113] As Figure 3 shown, a coordinate attention mechanism CA module is added to the residual block in the ResNet network to construct a residual coordinate attention module, which is applied to change detection.

[0114] The specific process of the dual-temporal semantic reasoning module for modeling the temporal correlation and consistency between multi-scale feature extraction branches is as Figure 5 shown.

[0115] In summary, the network designed in the present invention adds a coordinate attention mechanism to the residual block of the ResNet network, effectively merges semantic features for binary change detection, can effectively identify the location information and boundary information of the change region, and improves the change recognition accuracy; the dual-temporal semantic reasoning module constructed in the present invention effectively models the temporal and spatial correlation of dual-temporal remote sensing images of nature reserves, improves the temporal consistency, and can better distinguish land use types in key areas; aiming at the problems of scattered change regions and unbalanced change type samples in the nature reserve dataset, it is feasible to use the present invention for semantic change detection of land use in nature reserves. This method effectively utilizes the advantage of rich spatial information in high-resolution data, and at the same time models the temporal correlation of the land use types of dual-temporal image changes to obtain good semantic change recognition effects.

Claims

1. A method for semantic change detection of remote sensing images in nature reserves based on coordinate attention and dual-time semantic reasoning, characterized in that: The method includes the following steps in sequence: (1) Obtain remote sensing image data: Download high-resolution remote sensing image data of Beijing-2 and Gaojing-1 for two periods in the nature reserve area; (2) Conduct data preprocessing and data augmentation: Screen, splice, and crop the obtained remote sensing images according to the vector data of nature reserves; Semantically annotate the changed areas of the two-period remote sensing images to obtain label images L1 and L2, where L1 is the semantic change map before change, and L2 is the semantic change map after change; Crop and perform data augmentation on the remote sensing images and label images to obtain a nature reserve semantic change detection dataset; Divide the nature reserve semantic change detection dataset into a training set, a test set, and a validation set according to the ratio of 7:2:1; (3) Construct the CAB-SRNet network model: Add a coordinate attention mechanism to the residual block to construct a residual coordinate attention module, construct a dual-temporal semantic reasoning module, and add a residual coordinate attention module and a dual-temporal semantic reasoning module to the BI-SRNet network to construct the CAB-SRNet network model; Construct a dual - time semantic reasoning module: Model the temporal correlation and consistency between the semantic features X1 and X2 of two - period remote sensing images, and generate query vectors q i ∈R c×h×w from the input feature X i ∈R H×c′ , key vectors k i ∈R c′×H , and value vectors v i ∈R c ×H , where i = 1, 2, c is the number of channels, h and w are the spatial sizes, H = h×w, c′ = c / r, r is the channel reduction coefficient, and q i , k i , v i are obtained by feeding Xi into different convolutional layers. Note that the calculation formulas for matrices A1, A2, A 11 , A 22 ∈R H×H are as follows: Among them, is a softmax normalization function along the row dimension, A1, A2, A 11 , A 22 are used to record the correlation between each pair of spatial positions, obtaining an enhanced feature P1, P2: P1 = X1 + v1 × A1 + v1 × A 11 P2 = X2 + v2 × A2 + v2 × A 22 Obtain semantic maps Q1 and Q2 through a 1×1 convolutional layer with non-shared weights; The semantic change maps S1 and S2 are obtained by masking the semantic maps Q1 and Q2 with the binary change map C: S1, S2 = C·(P1, P2); Add a residual coordinate attention module and a dual-temporal semantic reasoning module to the BI-SRNet network to construct the CAB-SRNet network model; (4) Perform model training and evaluation: Based on the semantic class loss L sem , binary change loss L change , and semantic change loss L sc , combine the joint loss function, use the training set to train the CAB-SRNet network model, and after training, use the test set to evaluate the recognition accuracy and extraction effect of the CAB-SRNet network model; (5) Conduct semantic change recognition: After preprocessing the remote sensing image of the nature reserve to be extracted, input it into the trained CAB-SRNet network model. The CAB-SRNet network model outputs a prediction image to obtain the recognition result of the land use semantic change in the nature reserve.

2. The semantic change detection method for remote sensing images of nature reserves based on coordinate attention and dual-temporal semantic reasoning according to claim 1, characterized in that: The specific steps of step (2) include the following: (2a) Through the Arcgis10.5 software, screen, splice, and crop the remote sensing images of the nature reserve; Semantically annotate the changed areas of the two-period remote sensing images of the nature reserve to obtain label images L1 and L2. L1 is the semantic change map before change, and L2 is the semantic change map after change. In the label image, no change is white, water is blue, bare land is brown, vegetation is green, buildings are red, roads are gray, and greenhouses are black; Crop the remote sensing images and label images in the form of a sliding window into images of 256×256 size; (2b) Perform data augmentation on the cropped remote sensing images and label images to expand the data volume and obtain a nature reserve semantic change detection dataset. The data augmentation includes: Horizontal flipping: Use the image processing library OpenCV to perform horizontal flipping on the remote sensing images and label images respectively; Vertical flipping: Use the image processing library OpenCV to perform vertical flipping on the remote sensing images and label images respectively; Horizontal and vertical flipping: Use the image processing library OpenCV to perform horizontal flipping first and then vertical flipping on the remote sensing images and label images respectively; Shifting, scaling, random cropping, and adding noise: Shift, scale, randomly crop, and add noise to the remote sensing image and the label image respectively; (2c) Divide the natural protected area semantic change detection dataset into a training set, a test set, and a validation set according to the ratio of 7:2:1 respectively; the training set is used to directly participate in the training of the CAB-SRNet network model for feature extraction; the validation set is used for validation after each training stage; the test set is used to test the recognition accuracy and extraction effect of the CAB-SRNet network model after training is completed.

3. The semantic change detection method for remote sensing images of nature reserves based on coordinate attention and dual-temporal semantic reasoning according to claim 1, characterized in that: The specific steps of step (3) are as follows: (3a) First, perform feature extraction on the remote sensing image of the natural protected area. Use the fully convolutional network to perform semantic feature extraction on the two-phase remote sensing images I1 and I2 to obtain semantic features X1 and X2; (3b) Construct a residual coordinate attention module: Embed the coordinate attention CA module into the residual block of the ResNet network to obtain the residual coordinate attention module; Use the residual coordinate attention module to merge the dual-temporal features for binary change detection. Use the ResNet network combined with the coordinate attention mechanism to change the information through the unaligned semantic features X1 and X2, and then map it to the binary change map C. All mappings are performed through a 1×1 convolutional layer with non-shared weights.

4. The semantic change detection method for remote sensing images of nature reserves based on coordinate attention and dual-temporal semantic reasoning according to claim 1, characterized in that: The specific steps of step (4) are as follows: (4a) It is composed of the semantic class loss L sem , the binary change loss L change and the semantic change loss L sc to form a combined loss function L scd ; Semantic class loss L sem refers to the multi-class cross-entropy loss between semantic maps Q1, Q2 and label images L1 and L2. The semantic class loss L sem is calculated as follows: Among them, N is the number of semantic classes, and N is set according to the number of classes in the dataset; y i represents the predicted probability of the true label, p i represents the predicted probability of the i-th class; Binary change loss L change is the binary cross-entropy loss between the predicted binary change map C and the reference change map L c The reference change map L c is generated using L1 or L2, and the binary change loss L for each pixel change is calculated as follows: L change = -y c log(p c ) - (1 - y c )log(1 - p c ) Among them, y c represents the change probability of the true value label, and p c represents the predicted change probability; Calculate the semantic change loss \(L\) between the predicted semantic maps \(Q_1\), \(Q_2\) and the reference change map \(L\) using the cosine loss function c between the semantic maps sc , and the calculation formula is as follows: Among them, x1 and x2 are the feature vectors of a pixel on Q1 and Q2 respectively; The training of the two feature embedding branches in the dual-time semantic reasoning module is directly supervised by L1 and L2 and assisted by L c while the training of the residual coordinate attention module is directly supervised by L c as follows: Among them, and are the semantic losses of each temporal branch; (4b) Train the CAB-SRNet network model: Use the pre-trained weights of the ResNet34 network during training, set the momentum to 0.99, and the weight decay to 0.0005; the learning rate is initially set to 0.01 and then linearly decays to 0 until 200 epochs are trained; perform validation after each training stage; (4c) Model accuracy evaluation: Use the evaluation metrics overall accuracy OA and intersection over union IoU to evaluate the accuracy. The calculation formulas of the evaluation metrics are shown in the formula: In the formula, TP represents the number of true positives, FP represents the number of false positives, FN represents the number of false negatives, and TN represents the number of true negatives.

5. The semantic change detection method for remote sensing images of nature reserves based on coordinate attention and dual-time semantic reasoning according to claim 1, characterized in that: The specific steps of step (5) are as follows: (5a) After preprocessing the remote sensing image of the natural protected area to be identified for semantic change, resize it to a size of 256×256; (5b) Input the adjusted image into the trained CAB-SRNet network model, and output a prediction image through the CAB-SRNet network model. Stitch to obtain the natural protected area semantic change recognition result. In the prediction image, no change is white, water is blue, bare land is brown, vegetation is green, buildings are red, roads are gray, and greenhouses are black.

Images