Remote sensing image change detection method based on neural network structure search
By constructing a remote sensing image change detection model based on neural network structure search, the problems of low quality of change feature extraction and reliance on human experience are solved, and more efficient change feature extraction and detection results are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUANGDONG SOUTH DIGITAL TECH
- Filing Date
- 2022-10-26
- Publication Date
- 2026-08-04
AI Technical Summary
Existing methods for detecting changes in remote sensing images suffer from low quality of change feature extraction and excessive reliance on human experience in module design, resulting in insufficient automation and data processing capabilities.
A remote sensing image change detection model is constructed using a neural network structure search method. Through image feature extraction, change feature extraction, and feature fusion modules, a chain structure and a bidirectional pyramid network are used for feature extraction and fusion. A fully connected network is combined for classification. The change feature extraction module is automatically designed to reduce manual intervention.
It improves the ability to extract and distinguish change features, reduces the manual input in model design, and enhances the automation and generalization ability of remote sensing image change detection.
Smart Images
Figure CN115601660B_ABST
Abstract
Description
Technical Field
[0001] This invention designs a remote sensing image change detection method based on neural network structure search, belonging to the field of remote sensing interpretation technology. Background Technology
[0002] Change detection is a crucial technology in remote sensing imagery, with wide applications in land use, urban design, resource exploration, agricultural surveys, and disaster assessment. Traditional change detection methods have undergone three main development stages. The earliest methods, proposed in the 1980s, included statistical direct comparison, image transformation-based methods, and post-classification comparison methods. These methods used single pixels as the detection unit, acquiring change features through pixel-by-pixel comparison. In the 1990s, machine learning methods such as decision trees, random forests, and support vector machines were introduced into remote sensing image processing, significantly improving change detection performance. In the 21st century, the basic detection unit shifted from pixels to objects. Object-level change detection methods effectively combined spectral and spatial information, enhancing the certainty of detection results. While these traditional methods have achieved considerable results, their modeling process suffers from low automation, weak data processing capabilities, and limited generalization ability.
[0003] Deep learning-based remote sensing change detection methods offer high automation, strong big data execution capabilities, and broad applicability. Currently, mainstream remote sensing change detection models adopt an end-to-end structure, where the encoder extracts change features from preceding and following time phases, which are then processed by the decoder to obtain the output. The encoder typically consists of three parts: an image feature extraction module, a change feature extraction module, and a feature fusion module. In the change feature extraction module, which is specific to change detection tasks, designers often only perform basic operations such as adding corresponding scale features from preceding and following time phases and then convolving them. Such designs have relatively low depth and complexity, and limited ability to extract change features. To obtain higher-quality change features, this invention investigates the change feature extraction module and proposes a remote sensing image change detection method based on neural network structure search. Summary of the Invention
[0004] This invention provides a remote sensing image change detection method based on neural network structure search, which can alleviate the problems of low quality of change feature extraction and excessive reliance on human experience in module design. While improving the model's ability to extract and distinguish change features, it reduces the human input in model design.
[0005] The technical solution of this invention is: a remote sensing image change detection method based on neural network structure search, the method comprising the following steps:
[0006] Step 1: Construct the overall framework of the remote sensing image change detection model: This framework includes an image feature extraction module, a change feature extraction module, a change feature fusion module, and a classifier;
[0007] Step 2: Define the search space for the change feature extraction module and construct a hybrid model: The search space is a chain structure with n1 paths, n2 nodes in each path, and n3 operations that can be performed between any two nodes. The operations between nodes are assigned different structural weights, where n1 is the number of features extracted by the image feature extraction module.
[0008] Step 3: Perform structure search on the existing dataset. Divide the existing dataset into two parts and use them to train the structure weights α and model weights w respectively, to obtain several different hybrid models. Select the structure with the largest F1 score on the validation set during the search process as the independent model.
[0009] Step 4: Repeat Step 3 to obtain several independent models. Then, substitute the independent models into the overall framework of the remote sensing image change detection model and evaluate them on the change detection task. On the existing dataset, evaluate the several independent models obtained in the search phase in a short time to obtain the optimal independent model.
[0010] Step 5: Substitute the optimal independent model into the overall framework of the remote sensing image change detection model to achieve remote sensing image change detection.
[0011] Furthermore, the specific steps of Step 1 are as follows:
[0012] Step 1.1: The image feature extraction module is used to extract image features at different scales;
[0013] Step 1.2: The change feature extraction module is constructed based on the neural network structure search. This module is used to extract change features from the same scale of two images to obtain change features Di at different scales, where the value of i is n1.
[0014] Step 1.3, Feature Fusion Module: Input the variation feature Di, and obtain the fused feature Pi through the bidirectional pyramid network BiFPN. BiFPN contains bidirectional fusion paths from bottom to top and from top to bottom, which enables semantic information and receptive field information to flow between different scales.
[0015] Step 1.4: The classifier is used to stitch the Pi images together and feed them into a classifier composed of a fully connected network to obtain the output image.
[0016] Furthermore, the specific implementation method of Step 1.1 is as follows;
[0017] Step 1.1.1 Multi-scale image features based on ResNet50 residual network: Input two images, ResNet50 residual network extracts four sets of image features at different scales. Specifically, the features of the first image are {C1i|(i=1,2,3,4)} and the features of the second image are {C2i|(i=1,2,3,4)} obtained by ResNet50 Layer2, Layer3, Layer4, and Layer5. 1, 2, 3, and 4 represent 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original image size, respectively.
[0018] Step 1.1.2: Image features based on dilated convolution pooling pyramid (ASPP): Input C14 and C24 features at 1 / 32 of the original image size, and use dilated convolution to upsample their size to 1 / 16 of the original image size, thereby improving semantic information while maintaining the receptive field.
[0019] Furthermore, the specific implementation method of Step 2 is as follows;
[0020] Step 2.1, Chain Path: The search space contains four search paths {Li|(i=1,2,3,4)}, which respectively process the n1 sets of image features of the same scale extracted by the feature extraction module to obtain n1 different scale variation features, i.e. n1=4;
[0021] Step 2.2, Number of Nodes: Set the number of nodes to 6 for each path, while maintaining the depth and complexity of the module. The nodes are divided into two groups. Node 1 of each path is defined as the starting node, and nodes 2-6 are defined as other nodes. The input of the starting node is the two image features of the current path, and the input of other nodes is the two image features of the current path and the output of the previous node.
[0022] Step 2.3, Available Operations: For the starting node, there are 5 available operations: add1, sub1, abs1, cat, and g_cat. add1 adds two image features and then performs a convolution; sub1 subtracts two image features and then performs a convolution; abs1 subtracts two image features, takes the absolute value, and then performs a convolution; cat concatenates two image features and then performs a convolution; g_cat concatenates two image features along their corresponding channels and then performs a grouped convolution. For other nodes, there are also 5 available operations: add2, sub2, abs2, cat, and g_cat. add2 subtracts two image features along their corresponding channels and then performs a grouped convolution. The operations that can be performed between nodes are: add1, sub1, abs1, add2, sub2, abs2, cat, and g_cat. The subtraction of each image feature with the output of the preceding node is performed, followed by concatenation of the corresponding channels and grouped convolution. The operations that can be performed between nodes are: add1, sub1, abs1, add2, sub2, abs2, cat, and g_cat.
[0023] Step 2.4, Hybrid Model Construction: Multiply the five operations between every two nodes by different structural weights. i represents a node, and the sum of the five weights is 1, constructing a hybrid model.
[0024] Furthermore, the specific steps of Step 3 are as follows:
[0025] Step 3.1, Training set partitioning: Divide the CDD training set into train1 and train2, which are used for training the structural weights α and model weights w, respectively;
[0026] Step 3.2, Structure Weight Optimization: Fix the model weights w, train the hybrid model on train1, and obtain the structure weights that minimize the loss. The optimized algorithm is Adam, with a learning rate of 3e-3;
[0027] Step 3.3, Model Weight Optimization: Fixed Structure Weights The hybrid model was trained on train2 to obtain the model weights w that minimized the loss. The optimization algorithm was stochastic gradient descent (SGD) with an initial learning rate of 0.025.
[0028] Step 3.4, Independent Model Determination: The structure search continues for several epochs. After each epoch, based on... The largest value in the search determines the independent model structure, and the independent model is evaluated on the CDD validation set. The model with the highest F1 score is the independent model obtained in this structure search.
[0029] Furthermore, the F1 value in Step 3 is calculated as follows;
[0030] Let TP represent the number of correctly predicted positive samples, FP represent the number of incorrectly predicted positive samples, TN represent the number of correctly predicted negative samples, and FN represent the number of incorrectly predicted negative samples. The evaluation metrics are calculated as follows:
[0031]
[0032]
[0033]
[0034] Furthermore, in Step 4, on the CDD dataset, the independent models obtained in the search phase are trained for 35 epochs and their F1 scores are evaluated. The model with the highest F1 score is the optimal independent model. Then, the optimal independent model is trained for 100 epochs and its F1 score is evaluated. The specific implementation method is as follows.
[0035] Step 4.1 Independent Model Evaluation: The parameters are set as follows: the optimization algorithm is selected as stochastic gradient descent (SGD), the learning rate is set to 0.05, each independent model is trained for 35 epochs, and evaluated on the CDD validation set every 5 epochs. The independent model with the largest F1 score is selected as the final model.
[0036] Step 4.2 Final Model Evaluation: The parameters are set as follows: the optimization algorithm is selected as stochastic gradient descent (SGD), the learning rate is set to 0.05, the training time is 100 epochs, and the model is evaluated on the validation set every 5 epochs. Finally, the model with the highest F1 score is selected to validate the results on the CDD test set.
[0037] The beneficial effects of this invention are:
[0038] To address the issues of low quality in remote sensing image change feature extraction and over-reliance on human experience in model design, this invention proposes a remote sensing image change detection method based on neural network structure search. This method utilizes a gradient descent-based neural network structure search approach to automatically design the change feature extraction module, saving manual input during model design. Furthermore, this invention designs a dedicated search space for change detection tasks, enhancing the ability of the change feature extraction module to extract change features. This approach has significant theoretical and practical application value for remote sensing image change detection. Attached Figure Description
[0039] Figure 1 This is a diagram illustrating the overall framework of the remote sensing image change detection model based on neural network structure search in this invention.
[0040] Figure 2 A search space chain structure diagram is proposed for this invention;
[0041] Figure 3 This is the F1 score result of the candidate structure searched by this invention after training for 35 epochs;
[0042] Figure 4 This is the topology diagram of the optimal feature extraction module structure found in this invention. Detailed Implementation
[0043] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0044] like Figure 1-4 As shown, a remote sensing image change detection method based on neural network structure search includes:
[0045] Step 1: Construct the overall framework of the remote sensing image change detection model: This framework consists of an image feature extraction module, a change feature extraction module, a change feature fusion module, and a classifier, such as... Figure 1 As shown;
[0046] As a preferred embodiment of the present invention, the specific steps of Step 1 are as follows:
[0047] Step 1.1 Image Feature Extraction Module: In order to reduce the manual input when extracting features, this invention designs an image feature extraction module to automatically obtain image features. Based on the network of different depths in the module, image features at various scales are extracted. Shallow image features have rich semantic information, while high-level networks have a larger receptive field.
[0048] As a preferred embodiment of the present invention, the specific steps of Step 1.1 are as follows:
[0049] Step 1.1.1 Multi-scale image features based on ResNet50 residual network: Input two images, ResNet50 residual network extracts four sets of image features at different scales. Specifically, the features of the first image are {C1i|(i=1,2,3,4)} and the features of the second image are {C2i|(i=1,2,3,4)} obtained from ResNet50 Layer2, Layer3, Layer4, and Layer5. 1, 2, 3, and 4 represent 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original image size, respectively.
[0050] Step 1.1.2: Image features based on dilated convolution pooling pyramid (ASPP): Input C14 and C24 features at 1 / 32 of the original image size, and use dilated convolution to upsample their size to 1 / 16 of the original image size, thereby improving semantic information while maintaining the receptive field.
[0051] Step 1.2, Change Feature Extraction Module: This module is based on neural network structure search and construction, which can break away from the excessive reliance on experience when manually designing networks. This module is used to extract change features from the first image features C11, C12, C13 and the upsampled C14, and the second image features C21, C22, C23 and the upsampled C24, to obtain {Di|i=1,2,3,4}.
[0052] Step 1.3, Feature Fusion Module: Input the variation features {Di|i=1,2,3,4}, and obtain the fused features {Pi|i=1,2,3,4} through the Bidirectional Pyramid Network (BiFPN). The BiFPN contains bidirectional fusion paths from bottom to top and from top to bottom, enabling semantic information and receptive field information to flow between different scales.
[0053] Step 1.4, Classifier: The Pi images are spliced together and fed into a classifier composed of a fully connected network to obtain the output image.
[0054] Step 2: Define the search space for the change feature extraction module and construct a hybrid model: The space is a chain structure with 4 paths, each path has 6 nodes, and there are 8 operations that can be performed between nodes;
[0055] As a preferred embodiment of the present invention, the specific steps of Step 2 are as follows:
[0056] Step 2.1, Chain Path: such as Figure 2As shown, the search space contains four search paths {Li|(i=1,2,3,4)}, which process four sets of image features of the same scale C11, C21, C12, C22, C13, C23, and upsampled C14, C24 respectively, to obtain four variation features D1, D2, D3, D4 of different scales.
[0057] Step 2.2, Number of Nodes: Set the number of nodes to 6 for each path, while maintaining the depth and complexity of the module. The nodes are divided into two groups. Node 1 of each path is defined as the starting node, and nodes 2-6 are defined as other nodes. The input of the starting node is the two image features of the current path, and the input of other nodes is the two image features of the current path and the output of the previous node.
[0058] Step 2.3, Available Operations: For the starting node, there are 5 possible operations: add1 (add two image features and then perform convolution), sub1 (subtract two image features and then perform convolution), abs1 (subtract two image features and take the absolute value, then perform convolution), cat (concatenate two image features and then perform convolution), and g_cat (concatenate two image features in their corresponding channels and then perform grouped convolution). For other nodes, there are also 5 possible operations: add2 (add two image features to the output of the previous node respectively, then concatenate the corresponding channels and perform grouped convolution), s The operations that can be performed between nodes include: ub2 (subtracting the output of the previous node from each of the two image features, then concatenating the corresponding channels and performing grouped convolution), abs2 (subtracting the absolute values of the outputs of the previous node from each of the two image features, then concatenating the corresponding channels and performing grouped convolution), cat (concatenating the outputs of the previous node from each of the two image features, then performing convolution), and g_cat (concatenating the outputs of the previous node from each of the two image features in the corresponding channels, then performing grouped convolution). Therefore, there are a total of 8 operations that can be performed between nodes: add1, sub1, abs1, add2, sub2, abs2, cat, and g_cat.
[0059] Step 2.4, Hybrid Model Construction: Multiply the five operations between every two nodes by different structural weights. (i represents a node, and the sum of the five weights is 1), construct a hybrid model.
[0060] Step 3: Perform structure search on the CDD dataset to obtain the best-performing independent model from the hybrid models: Divide the CDD training set into train1 and train2, and train the model for 30 epochs. In each epoch, train on train1 first to optimize the structure weights α, and then train on train2 to optimize the model weights w. After each epoch, evaluate on the CDD validation set, and determine the independent model structure based on the structure weights α that have the largest F1 score.
[0061] As a preferred embodiment of the present invention, the specific steps of Step 3 are as follows:
[0062] Step 3.1 Training set partitioning: The CDD dataset training set size is 10000. It is randomly divided into two equal parts, train1 and train2, with a size of 5000, which are used for training the structure weights α and model weights w, respectively.
[0063] Step 3.2, Structure Weight Training: Fix the model weights w, train the hybrid model on train1, and obtain the structure weights that minimize the loss. The optimization algorithm is Adam, with a learning rate of 3e-3, beats range (0.9, 0.999), and weight_decay of 1e-3.
[0064] Step 3.3, Model Weight Training: Fixed Structure Weights The hybrid model was trained on train2 to obtain the model weights w that minimized the loss. The optimization algorithm was stochastic gradient descent (SGD) with an initial learning rate of 0.025, and weight_decay and momentum were 1e-3 and 0.9, respectively.
[0065] Step 3.4, Independent Model Determination: The structure search lasts for 30 epochs. After each epoch, based on... The largest value in the search determines the independent model structure, and the model is evaluated on the CDD validation set. The model with the highest F1 score in 30 epochs is the independent model obtained in this structure search.
[0066] During model training, the structure weights α are all initialized. To reduce the impact of initialization on structure search, step 3 can be repeated multiple times (5 times in this embodiment) to obtain multiple independent models. To ensure the fairness of the results, all experiments are implemented based on the PyTorch framework, and a V100 GPU with 32GB of video memory is selected. The main evaluation metrics of the experiments are F1, Recall, and Precision. Let TP represent the number of correctly predicted positive samples, FP represent the number of incorrectly predicted positive samples, TN represent the number of correctly predicted negative samples, and FN represent the number of incorrectly predicted negative samples. The calculation methods for the three evaluation metrics are as follows:
[0067]
[0068]
[0069]
[0070] Table 1 shows the F1 scores of the independent models obtained from the five structural searches of this invention.
[0071] Table 1: F1 scores of independent models obtained by the five-stage structure search method of the present invention
[0072] Independent Model first The second The third Fourth Fifth F1 score (%) 71.9 68.53 67.64 70.11 70.68
[0073] Step 4: Perform structural evaluation on the change detection task (i.e., evaluate the independent models by substituting them into the overall framework of the remote sensing image change detection model): On the CDD dataset, train the independent models obtained in the search phase for 35 epochs and evaluate their F1 scores. The model with the highest F1 score is the optimal independent model. Then train the optimal independent model for 100 epochs and evaluate its F1 score.
[0074] As a preferred embodiment of the present invention, the specific steps of Step 4 are as follows:
[0075] Step 4.1 Independent Model Evaluation: The parameters are set as follows: the optimization algorithm is Stochastic Gradient Descent (SGD), the learning rate is set to 0.05, weight_decay is 1e-3, and momentum is 0.9. Each independent model is trained for 35 epochs, and evaluated on the CDD validation set every 5 epochs. The independent model with the highest F1 score is selected as the final model. The evaluation results of the five independent models are as follows. Figure 3 As shown, the highest F1 score is 93.86%, and the corresponding structure is as follows. Figure 4 As shown.
[0076] Step 4.2 Final Model Evaluation: The parameters are set as follows: the optimization algorithm is stochastic gradient descent (SGD), the learning rate is set to 0.05, the weight_decay is 1e-3, the momentum is 0.9, the training time is 100 epochs, and the model is evaluated on the validation set every 5 epochs. Finally, the model with the highest F1 score is selected to validate the results on the CDD test set. Table 2 shows the test results of the method of this invention and other methods on the CDD dataset.
[0077] Table 2: Comparison of F1 scores of the method of this invention with other methods on the CDD test set
[0078]
[0079] Compared to other methods, the method of this invention achieved the best results in all three evaluation metrics, with F1 score, recall, and precision improved by at least 3.13%, 2.22%, and 2.45%, respectively.
[0080] Step 5: To illustrate the universality of this invention, we conducted further structural transfer experiments. We selected the LEVIR-CD dataset, which contains 637 image pairs with a resolution of 1024*1024. The dataset was divided into 256*256 segments and randomly divided into a training set (7120), a validation set (1024), and a test set (2048) in a ratio of 7:1:2. The training parameters were set as follows: the optimization algorithm was selected as stochastic gradient descent (SGD), the learning rate was set to 0.025, the weight_decay was 1e-3, the momentum was 0.9, the training time was 100 epochs, and the model was evaluated on the validation set every 5 epochs. Finally, the model with the highest F1 score was selected to validate the results on the LEVIR-CD test set. Table 3 shows the test results of the method of this invention and other methods on the LEVIR-CD dataset.
[0081] Table 3: Comparison of F1 scores of the method of the present invention with other methods on the LEVIR-CD test set
[0082]
[0083] Compared to other methods, the method of this invention achieved the highest F1 score, the second highest Recall, and the third highest Precision.
[0084] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.
Claims
1. A method for detecting changes in remote sensing images based on neural network structure search, characterized in that: The specific steps of the method are as follows: Step 1: Construct the overall framework of the remote sensing image change detection model: This framework includes an image feature extraction module, a change feature extraction module, a change feature fusion module, and a classifier; Step 2: Define the search space for the change feature extraction module and construct a hybrid model: The search space is a chain structure with n1 paths, n2 nodes in each path, and n3 operations that can be performed between any two nodes. The operations between nodes are assigned different structural weights, where n1 is the number of features extracted by the image feature extraction module. The specific implementation method of Step 2 is as follows; Step 2.1, Chain Path: The search space contains four search paths {Li|(i=1,2,3,4)}, which respectively process the n1 sets of image features of the same scale extracted by the feature extraction module to obtain n1 different scale variation features, i.e., n1=4; Step 2.2, Number of Nodes: Set the number of nodes to 6 for each path, while maintaining the depth and complexity of the module. The nodes are divided into two groups. Node 1 of each path is defined as the starting node, and nodes 2-6 are defined as other nodes. The input of the starting node is the two image features of the current path, and the input of other nodes is the two image features of the current path and the output of the previous node. Step 2.3, Available Operations: For the starting node, there are 5 available operations: add1, sub1, abs1, cat, and g_cat. add1 adds two image features and then performs a convolution; sub1 subtracts two image features and then performs a convolution; abs1 subtracts two image features, takes the absolute value, and then performs a convolution; cat concatenates two image features and then performs a convolution; g_cat concatenates two image features along their corresponding channels and then performs a grouped convolution. For other nodes, there are also 5 available operations: add2, sub2, abs2, cat, and g_cat. add2 adds two image features to the output of the preceding node, concatenates them along their corresponding channels, and then performs a grouped convolution. The operations can be performed in groups: sub2 means subtracting the output of the previous node from each of the two image features, then concatenating the corresponding channels and performing group convolution; abs2 means subtracting the absolute values of the output of the previous node from each of the two image features, then concatenating the corresponding channels and performing group convolution; cat means concatenating the output of the previous node from each of the two image features and then performing convolution; g_cat means concatenating the output of the previous node from each of the two image features along the corresponding channels and then performing group convolution. Therefore, there are a total of 8 operations that can be performed between nodes: add1, sub1, abs1, add2, sub2, abs2, cat, and g_cat. Step 2.4, Hybrid Model Construction: Multiply the five operations between every two nodes by different structural weights. , where i represents a node, and the sum of the five weights is 1, construct a hybrid model; Step 3: Perform structure search on the existing dataset. Divide the existing dataset into two parts and use them to train the structure weights α and model weights w respectively, to obtain several different hybrid models. Select the structure with the largest F1 score on the validation set during the search process as the independent model. Step 4: Repeat Step 3 to obtain several independent models. Then, substitute the independent models into the overall framework of the remote sensing image change detection model and evaluate them on the change detection task. On the existing dataset, perform a short-time evaluation on the several independent models obtained in the search phase to obtain the optimal independent model. Step 5: Substitute the optimal independent model into the overall framework of the remote sensing image change detection model to achieve remote sensing image change detection.
2. The remote sensing image change detection method based on neural network structure search according to claim 1, characterized in that: The specific steps of Step 1 are as follows: Step 1.1: The image feature extraction module is used to extract image features at different scales; Step 1.2: The change feature extraction module is constructed based on the neural network structure search. This module is used to extract change features from the same scale of two images to obtain change features Di at different scales, where the value of i is n1. Step 1.3, Feature Fusion Module: Input the variation feature Di, and obtain the fused feature Pi through the bidirectional pyramid network BiFPN. BiFPN contains bidirectional fusion paths from bottom to top and from top to bottom, which enables semantic information and receptive field information to flow between different scales. Step 1.4: The classifier is used to stitch the Pi images together and feed them into a classifier composed of a fully connected network to obtain the output image.
3. The remote sensing image change detection method based on neural network structure search according to claim 2, characterized in that: The specific implementation method of Step 1.1 is as follows; Step 1.1.1 Multi-scale image features based on ResNet50 residual network: Input two images, ResNet50 residual network extracts four sets of image features at different scales. Specifically, the features of the first image are {C1i|(i=1,2,3,4)} and the features of the second image are {C2i|(i=1,2,3,4)} obtained by ResNet50 Layer2, Layer3, Layer4, and Layer5. 1, 2, 3, and 4 represent 1 / 4, 1 / 8, 1 / 16, and 1 / 32 of the original image size, respectively. Step 1.1.2: Image features based on dilated convolution pooling pyramid (ASPP): Input C14 and C24 features at 1 / 32 of the original image size, and use dilated convolution to upsample their size to 1 / 16 of the original image size, thereby improving semantic information while maintaining the receptive field.
4. The remote sensing image change detection method based on neural network structure search according to claim 1, characterized in that: The specific steps of Step 3 are as follows: Step 3.1, Training set partitioning: Divide the CDD training set into train1 and train2, which are used for training the structural weights α and model weights w, respectively; Step 3.2, Structure Weight Optimization: Fix the model weights w, train the hybrid model on train1, and obtain the structure weights that minimize the loss. The optimized algorithm is Adam, with a learning rate of 3e-3; Step 3.3, Model Weight Optimization: Fixed Structure Weights The hybrid model is trained on train2 to obtain the model weights w that minimize the loss. The optimization algorithm is stochastic gradient descent (SGD) with an initial learning rate of 0.
025. Step 3.4, Independent Model Determination: The structure search continues for several epochs. After each epoch, based on... The largest value in the search determines the independent model structure, and the independent model is evaluated on the CDD validation set. The model with the highest F1 score is the independent model obtained in this structure search.
5. The remote sensing image change detection method based on neural network structure search according to claim 1, characterized in that: The F1 value is calculated in Step 3 as follows; Let TP represent the number of correctly predicted positive samples, FP represent the number of incorrectly predicted positive samples, TN represent the number of correctly predicted negative samples, and FN represent the number of incorrectly predicted negative samples. The evaluation metrics are calculated as follows: 。 6. The remote sensing image change detection method based on neural network structure search according to claim 1, characterized in that: In Step 4, on the CDD dataset, the independent models obtained in the search phase are trained for 35 epochs and their F1 scores are evaluated. The model with the highest F1 score is the best independent model. Then, the best independent model is trained for 100 epochs and its F1 score is evaluated. The specific implementation method is as follows. Step 4.1 Independent Model Evaluation: The parameters are set as follows: the optimization algorithm is selected as stochastic gradient descent (SGD), the learning rate is set to 0.05, each independent model is trained for 35 epochs, and evaluated on the CDD validation set every 5 epochs. The independent model with the largest F1 score is selected as the final model. Step 4.2 Final Model Evaluation: The parameters are set as follows: the optimization algorithm is selected as stochastic gradient descent (SGD), the learning rate is set to 0.05, the training time is 100 epochs, and the model is evaluated on the validation set every 5 epochs. Finally, the model with the highest F1 score is selected to validate the results on the CDD test set.