A method and apparatus for identifying road changes in remote sensing images
By employing self-supervised contrastive learning and transfer learning methods, road change identification is performed using unlabeled remote sensing images. This solves the problem of insufficient labeled samples, achieves efficient and rapid road change identification, improves identification accuracy, and supports urban management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUIZHOU TUZHI INFORMATION TECH CO LTD
- Filing Date
- 2021-11-24
- Publication Date
- 2026-05-26
AI Technical Summary
Existing technologies struggle to efficiently identify road changes, especially when labeled samples are insufficient and unlabeled remote sensing image data is abundant. Current methods cannot meet the long-term and efficient identification needs of road changes.
A self-supervised contrastive learning method is adopted. Unlabeled road remote sensing images are acquired, cropped, and contrasted. A deep learning model is used for self-supervised training, and a small amount of labeled data is combined for transfer learning to generate a pre-trained model for recognizing road changes.
It enables rapid and accurate identification of road changes without the need for manual data labeling, improving the identification accuracy to 96%, and also features faster training speed and higher efficiency, supporting dynamic management of urban road changes.
Smart Images
Figure CN114078237B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geographic information remote sensing data analysis technology, and in particular to a method and apparatus for identifying road changes. Background Technology
[0002] With the rapid development of industrialization and urbanization in my country, the number of roads has increased rapidly. Especially in recent years, the scale of Chinese cities has expanded dramatically, and there has been a phenomenon of disorderly expansion, resulting in the creation of many new roads and the disappearance of many old ones. The original roads have undergone significant changes in width, branching, and other aspects. Marking these road changes requires a large amount of manpower and hinders real-time monitoring of road conditions.
[0003] In recent years, thanks to the development of artificial intelligence, many fields have seen computers replacing human labor. Satisfactory results have been achieved in remote sensing image classification, and it has been widely applied in areas such as urban planning and identification. However, existing AI identification technologies have a high demand for highly accurate annotation, and the annotations must also change as roads change over time, making it difficult to meet the long-term and efficient requirements for identifying road changes.
[0004] To address the problem of insufficient labeled samples, a common approach is to employ various data augmentation techniques to generate more samples. This has indeed improved the model's performance and has been widely used, but the improvement is insufficient for practical applications. In addition, the development of transfer learning techniques has slightly alleviated these problems. This involves pre-training the model on a larger training set relevant to the current task and initializing the parameters of the existing task model with the parameters of the pre-trained model. However, there is currently no large-scale remote sensing image training set, which poses a challenge to the application of transfer learning.
[0005] Given the scarcity of labeled remote sensing images, but the abundance of unlabeled remote sensing imagery globally, the key lies in utilizing this vast amount of unprocessed data—a crucial aspect that self-supervised learning can address. Self-supervised learning does not rely on any labeled image data; it directly uses the images themselves to design auxiliary tasks, thereby learning higher-order and more complex features of the images and applying them to downstream tasks to improve performance. Currently, there is considerable research on the processing and application of unlabeled remote sensing imagery data.
[0006] Patent CN112766089A discloses a cross-domain road extraction method based on a global-local adversarial learning framework, comprising the following steps: Step 1, acquiring labeled source domain data and unlabeled target domain images, forming a training set for network training, and normalizing and data augmenting the training set; Step 2, constructing a global-local adversarial learning framework; Step 3, training on the training set based on the constructed global-local adversarial learning framework, optimizing model parameters until convergence; Step 4, predicting from the test set of the target domain based on the converged global-local adversarial learning framework, and obtaining the road extraction result using the output road segmentation probability map. The implementation of step 3 is as follows: Step 3.1, initialize the network parameters of the global-local adversarial learning framework, fix the cropping size of the training set, the number of input images per batch, and the initial learning rate for segmentation and adversarial learning; Step 3.2, train the generator using the SGD optimizer and the discriminator using the Adam optimizer. The segmentation training and adversarial training are performed alternately in steps. The network parameters of the global-local adversarial learning framework are continuously updated through forward propagation and backward feedback until convergence.
[0007] This patent discloses a method that uses both labeled and unlabeled remote sensing images to construct a road recognition model. However, simply employing an adversarial learning framework requires a large number of positive and negative samples for adversarial learning, consuming significant manual effort to collect labeled data for training, and can only extract and process the road framework. In contrast, this invention researches self-supervised learning methods, using a self-supervised contrastive learning approach to train a large number of unlabeled images to construct a recognition model. This requires no manual intervention and utilizes momentum methods to update parameters, resulting in faster and more efficient training.
[0008] Patent CN113486827A discloses a multi-source remote sensing image transfer learning method based on domain adversarial and self-supervised training. The deep learning model is pre-trained in the source domain and then trained within a transfer learning framework based on domain adversarial and self-supervised training. Remote sensing image data of the target domain is input into the trained deep learning model, which outputs prediction results for the target domain. The transfer learning framework includes a domain adversarial training module and a self-supervised training module. The domain adversarial training module includes one generator and one or more discriminators. The self-supervised training module includes an information entropy calculation module and a pseudo-label selection module. The information entropy calculation module calculates the information entropy of the prediction result for each sample, and the pseudo-label selection module selects pseudo-labels based on the information entropy of the prediction result for each sample.
[0009] While this patent discloses the use of self-supervised training methods to optimize remote sensing image transfer learning models, it focuses more on the fine-tuning of transfer learning and neglects the powerful pre-training capabilities of self-supervised learning itself. This invention, through research on self-supervised learning, employs MoCo's self-supervised contrastive learning method to perform contrastive learning training on unlabeled data, thereby obtaining a pre-trained model applicable to downstream tasks in various fields. It exhibits strong transfer capabilities in the remote sensing image domain and has practical significance for the specific task of road change recognition in remote sensing images. Summary of the Invention
[0010] To address the shortcomings of existing technologies, this invention provides a method and apparatus for identifying road changes in remote sensing images using self-supervised comparative learning artificial intelligence technology; specifically, it is achieved through the following technical solution:
[0011] The purpose of this invention is to provide a method for identifying road changes in remote sensing images, comprising:
[0012] Acquire a large number of new and old remote sensing images of unmarked roads;
[0013] The new and old remote sensing images are cropped based on the land parcel vector range corresponding to the road to be identified, resulting in new and old cropped remote sensing images that include the land parcel vector range.
[0014] By comparing the old and new cropped remote sensing images, a remote sensing image of the changes in the old and new roads is obtained;
[0015] Remote sensing images of changes in old and new roads are input into a deep learning model for self-supervised training to obtain a pre-trained model;
[0016] A classification and recognition model was trained using a small number of labeled remote sensing images of road changes to obtain a self-supervised model parameter transfer model.
[0017] The remote sensing image of the road change to be identified is input into the classification and recognition model with parameters transferred from the self-supervised model, and the road change results corresponding to the road to be identified are obtained from the output of the classification and recognition model.
[0018] Preferably, the comparison processing involves performing a special subtraction between the new road remote sensing image and the old road remote sensing image during the data processing process. The value is 1 for road areas present in the new image but absent in the old image, 2 for road areas absent in the new image but present in the old image, and 0 for all other cases.
[0019] Preferably, the self-supervised training involves using one or more of the following deep learning models: VGG16, ResNet18, ResNet50, or DenseNet121. A contrastive self-supervised learning method is employed for training, meaning that an encoder is trained during the training process to ensure that the encoder's output vector is similar to the corresponding key in a large dictionary and dissimilar to other keys. The loss function used is InfoNCE, which achieves the effect that the loss decreases when the query vector q is similar to its positive key value k and dissimilar to its negative key value k. The specific formula is shown below:
[0020]
[0021] Where K is the number of negative key values. It's a hyperparameter.
[0022] Preferably, the contrast-based self-supervised learning method uses the momentum contrast (MoCo) method to store the dictionary, that is, a queue is used to store the dictionary; that is, each new batch of encoding is entered into the queue after it is completed, and then the oldest encoding is dequeued.
[0023] Preferably, the remote sensing images of changes between old and new roads are subjected to self-supervised training, and the specific steps are as follows:
[0024] Data augmentation is performed on each remote sensing image to obtain a query image and a template image;
[0025] The inputs are fed into a deep learning model to extract query features and template features, respectively.
[0026] Assuming the template features are stable, we set the template features to not use gradient update parameters;
[0027] The matching degree of features is calculated using matrix multiplication;
[0028] Set the NCE loss function to ensure the maximum matching degree between the image and the image derived from it;
[0029] The parameters of the query network are updated using gradient descent based on the loss, and the parameters of the template network are updated using momentum.
[0030] Repeat the above steps to complete the training of self-supervised contrastive learning and obtain a pre-trained model.
[0031] Preferably, the training method for the classification and recognition model that transfers self-supervised model parameters specifically includes:
[0032] Obtain labeled remote sensing image samples of road changes and the types of road changes contained in the samples;
[0033] The parameters of the pre-trained model are retained, and a linear classification network layer is connected after the encoder to identify road changes.
[0034] The model was trained using labeled remote sensing images of road changes, and the parameters were fine-tuned to obtain the trained model.
[0035] Labeled remote sensing image samples of road changes are input into the training model to obtain the predicted road change type output by the training model;
[0036] Based on the road change type and the predicted road change type, the training loss value is calculated, and the loss function is as follows:
[0037]
[0038] Where y is the desired output and a is the output of the neuron.
[0039] When the loss value is less than 0.1, the trained model is used as the final road change recognition model, that is, the classification and recognition model that transfers the parameters of the self-supervised model.
[0040] Another objective of this invention is to provide a remote sensing image road change recognition device, comprising a road change acquisition module to be identified, a cropped remote sensing image acquisition module, a self-supervised contrastive learning feature processing module, and a target recognition result acquisition module;
[0041] The road change acquisition module is used to acquire new and old remote sensing images of the road to be identified.
[0042] The cropped remote sensing image acquisition module is used to crop the old and new remote sensing images according to the vector range of the land parcels corresponding to the road to be identified, so as to obtain new and old cropped remote sensing images that only contain the vector range of the land parcels corresponding to the road to be identified.
[0043] The self-supervised contrastive learning feature processing module is used for self-supervised training of unlabeled cropped remote sensing images to obtain a feature extractor containing specific parameters.
[0044] The target recognition result acquisition module is used to input the remote sensing image of the road change to be identified into the road change recognition model, and obtain the road change type corresponding to the road change to be identified output by the road change recognition model.
[0045] Preferably, the remote sensing image road change recognition device further includes a model training sample acquisition module, a land parcel remote sensing image generation module, a road change remote sensing image generation module, a self-supervised learning model parameter generation module, a predicted road change type acquisition module, a loss value calculation module, and a road change type recognition module acquisition module.
[0046] The model training sample acquisition module is used to acquire model training samples; the model training samples include remote sensing images of old and new roads and the types of road changes contained in the remote sensing images;
[0047] The land parcel remote sensing image generation module is used to crop the old and new remote sensing images based on the road vector range of the multiple old and new land parcels, and generate cropped old and new road remote sensing images corresponding to the multiple land parcels.
[0048] The road change remote sensing image generation module is used to compare and process cropped remote sensing images of old and new roads to generate new and old road change images corresponding to the multiple plots of land.
[0049] The self-supervised learning model parameter generation module is used to perform self-supervised training and update model parameters on images of changes between old and new roads, generate parameters for the first few layers of the network for feature extraction, and obtain the trained model.
[0050] The predicted road change type acquisition module is used to input the remote sensing image of the road change to be identified into the training model and obtain the predicted road change type of the multiple plots output by the training model.
[0051] The loss value calculation module is used to calculate the loss value of the trained model based on the road change type and the predicted road change type.
[0052] The road change type identification module is used to obtain the training model as the final road change identification model when the loss value is within a preset range.
[0053] The advantages of this invention compared to the prior art are:
[0054] This invention uses benchmark models in multiple image classification fields, such as VGG16, ResNet18, ResNet50, or DenseNet121, as the backbone network for self-supervised learning, and the results in image classification can reach the top level in the industry.
[0055] This invention uses an image comparison processing method to generate comparative images of road changes. It uses the image processing library of MATLAB to process the noise generated during the image subtraction process and fills the holes that appear in the road change images. Finally, it sets thresholds to assign values of 0, 1, and 2 to different blocks. The final result has an overlap rate of up to 99% with the manually labeled data.
[0056] This invention employs MoCo's self-supervised contrastive learning method, which can fully train the information contained in the image itself, allowing images to learn against each other and extract higher-order information from the image, thereby generating a better image classification model. In the task of road change recognition, it improves the accuracy by 20% compared with traditional techniques, reaching 96%. Moreover, most of the data in the training process is initially unlabeled, requiring no manual intervention, resulting in faster training speed and higher efficiency.
[0057] This invention collects and organizes road change data in a timely manner, combines it with regularly updated satellite remote sensing imagery, and utilizes techniques such as self-supervised comparative learning, transfer learning, and remote sensing image visual analysis for analysis and processing. This allows for rapid and accurate acquisition of the current status of road changes, improving the model's automation and intelligence without requiring manual data labeling, and achieving rapid and accurate automated identification of road change types. Furthermore, by timely updating road change information across the city, it comprehensively enhances the dynamic management level of transportation planning and urban construction, providing strong data support and automation technology for effective urban planning management and scientific decision-making. Attached Figure Description
[0058] Figure 1 A flowchart of the remote sensing image road change recognition method provided by the present invention;
[0059] Figure 2 A schematic diagram of the road change recognition device provided by the present invention;
[0060] Figure 3 This is a schematic diagram of the application system of the road change recognition method provided in Embodiment 3 of the present invention. Detailed Implementation
[0061] The technical solution of the present invention will be further defined below with reference to specific embodiments, but the scope of protection is not limited to the description made.
[0062] Example 1
[0063] A method for identifying road changes in remote sensing images, flowchart as follows: Figure 1 As shown, it includes the following steps:
[0064] Step 1: Acquire a large number of new and old remote sensing images of unmarked roads;
[0065] Remote sensing images of roads refer to remote sensing images of land parcels containing roads, obtained through satellite photography. Since remote sensing images are updated over time, the latest remote sensing image is compared with a remote sensing image from one year ago.
[0066] Step 2: Crop the new and old remote sensing images based on the vector range of the land parcels corresponding to the road to be identified, and obtain new and old cropped remote sensing images that include the vector range of the land parcels;
[0067] The land parcel vector range refers to the area of the land parcel where the road is located.
[0068] Step 3: Compare the old and new cropped road remote sensing images to obtain remote sensing images showing the changes in the old and new roads;
[0069] The comparison process involves a special subtraction operation between the new and old road remote sensing images during data processing. For road areas present in the new image but absent in the old image, the result is 1; for road areas absent in the new image but present in the old image, the result is 2; and for all other cases, the result is 0.
[0070] Step 4: Input the remote sensing images of the changes between the old and new roads into the deep learning model for self-supervised training to obtain a pre-trained model. The specific steps are as follows:
[0071] Step (4.1): Perform data augmentation on each remote sensing image to obtain the query image and template image.
[0072] In this embodiment, considering the large amount of remote sensing image data used, a road change database can be pre-constructed; the road change database includes information such as road traffic conditions, road change conditions, road change time, and old and new remote sensing images of the road.
[0073] Step (4.2): Input the data into the deep learning model to extract query features and template features respectively;
[0074] During model construction, two networks, f_k and f_q, are used to map input information to a feature space. This feature space is represented by a vector of length c. The f_k network extracts features from the template graph, while the f_q network extracts features from the query graph. Both networks use the ResNet50 architecture, but can be modified as needed to use one or more of VGG16, ResNet18, or DenseNet121.
[0075] Step (4.3): Assuming the template features are stable, set the template features to not use gradient update parameters;
[0076] Although f_k and f_q use the same network structure, they need to be distinguished in terms of parameter updates. During training, similarity matching is performed between query features and template features, so it is assumed that the template features are stable and there is no need to use gradient algorithms to update parameters quickly.
[0077] Step (4.4): Calculate the feature matching degree using matrix multiplication;
[0078] Step (4.5): Set the NCE loss function to ensure the maximum matching degree between the image and its derived images;
[0079] Step (4.6): Update the parameters of the query network using gradient Lie method based on the loss, and update the parameters of the template network using momentum method;
[0080] Repeat steps (4.1) to (4.6) until the loss value of InfoNCE reaches the required value or the number of training iterations reaches the required value to obtain the pre-trained model.
[0081] Step 5: Train the pre-trained model using a small number of labeled remote sensing image samples of road changes to obtain a classification and recognition model that transfers the parameters of the self-supervised model. The specific training method is as follows:
[0082] Step (5.1): Obtain labeled remote sensing image samples of road changes and the types of road changes contained in the samples;
[0083] Step (5.2): Retain the parameters of the pre-trained deep learning model, and connect a linear classification network layer after the encoder to identify road changes;
[0084] Step (5.3): Use remote sensing image samples of road changes for training, fine-tune the parameters of the model, and obtain the trained model;
[0085] Step (5.4): Input the labeled remote sensing image samples of road changes into the training model, and obtain the predicted road change type of the multiple plots output by the training model;
[0086] Step (5.5): Calculate the loss value based on the road change type and the predicted road change type. The loss function is shown below:
[0087]
[0088] Where y is the desired output and a is the output of the neuron.
[0089] Step (5.6): When the loss value is less than 0.1, the trained road change recognition model is used as the final road change recognition model, that is, the classification recognition model that transfers the parameters of the supervised model.
[0090] Step 6: Input the remote sensing image of the road changes to be identified into the classification and recognition model that transfers the parameters of the self-supervised model, and obtain the road change results corresponding to the road to be identified output by the classification and recognition model.
[0091] Example 2
[0092] A remote sensing image road change recognition device, such as Figure 2 As shown, it includes the following modules:
[0093] The road change acquisition module 410 is used to acquire new and old remote sensing images of the road to be identified.
[0094] The remote sensing image cropping module 420 is used to crop the remote sensing image according to the land parcel vector range corresponding to the road to be identified, so as to obtain a cropped remote sensing image that only contains the land parcel vector range corresponding to the road to be identified.
[0095] The self-supervised contrastive learning feature processing module 430 is used for self-supervised training of unlabeled road change remote sensing images to obtain a feature extractor containing specific parameters.
[0096] The target recognition result acquisition module 440 is used to input the remote sensing image of the road change to be identified into a pre-fine-tuned road change recognition model, and to obtain the road change type corresponding to the road change to be identified output by the road change recognition model.
[0097] Furthermore, the remote sensing image road change recognition device also includes:
[0098] The model training sample acquisition module is used to acquire model training samples; the samples include remote sensing images of old and new roads and the types of road changes contained in the remote sensing images;
[0099] The land parcel remote sensing image generation module is used to crop the new and old remote sensing images according to the road vector range of the multiple new and old land parcels, and generate new and old road cropped remote sensing images corresponding to the multiple land parcels.
[0100] The road change remote sensing image generation module is used to compare and process the cropped remote sensing images of new and old roads to generate road change images corresponding to the multiple plots of land.
[0101] The self-supervised learning model parameter generation module is used to perform self-supervised training and update model parameters based on road change images, generate parameters for feature extraction in the first few layers of the network, and obtain the trained model.
[0102] The predicted road change type acquisition module is used to input the road change image into the training model and acquire the predicted road change type of the multiple plots output by the training model.
[0103] The loss value calculation module is used to calculate the loss value of the trained model based on the road change type and the predicted road change type.
[0104] The road change type identification module is used to obtain the training model as the final road change identification model when the loss value is within a preset range.
[0105] Example 3
[0106] This embodiment applies the identification method of Embodiment 1 and the identification device of Embodiment 2 to the actual road identification process, forming a road identification change identification system, such as... Figure 3 As shown, it includes a road change database, a sample training management module, a data preprocessing module, an intelligent recognition module, and a human-computer interaction module.
[0107] I. Road Change Database
[0108] The road change database contains information such as road traffic conditions, road changes, the time of road changes, and old and new remote sensing images of roads.
[0109] II. Data Preprocessing Module
[0110] (1) Labelless remote sensing image data acquisition module
[0111] Based on the existing large amount of unlabeled remote sensing image data and update time, the data can be imported into the road change database in batches without manual processing and labeling, becoming a dataset for self-supervised training.
[0112] (2) Data extraction submodule
[0113] The data extraction submodule retrieves remote sensing images of land parcels with roads from the road change database and stores them in a temporary storage directory according to the parcel name. Then it retrieves the image of the next parcel until all parcel images have been retrieved.
[0114] (3) Image cropping submodule
[0115] The data cropping submodule retrieves remote sensing images and corresponding road vector data from the temporary storage directory for cropping, and the cropped images are saved in the cropped image directory.
[0116] (4) Road change image generation submodule
[0117] The road change image generation submodule extracts and processes data from the cropped image catalog. During data processing, a special subtraction is performed between the new and old road remote sensing images: a value of 1 is assigned to road areas present in the new image but absent in the old image; a value of 2 is assigned to road areas absent in the new image but present in the old image; and a value of 0 is assigned to all other cases. The processed image is then saved in the image to be identified catalog.
[0118] III. Sample Training Management Module
[0119] (1) Sample Management Submodule
[0120] The sample management submodule mainly enables operations such as browsing and deleting various types of samples, and selecting various types of samples to participate in training.
[0121] (2) Sample editing tools
[0122] The sample editing tool primarily handles labeled images to be identified, as these images are few in number and easy to process. The system provides a plot sample editing tool, whose main functions include image selection, sample modification, saving, and uploading. The functions are described below:
[0123] Image selection: Images can be selected from the database or uploaded. This can be done individually or in batches using code.
[0124] Sample modification: The selected sample can be manually modified.
[0125] Save and Upload: After the sample area is manually determined, it needs to be saved and uploaded to the server for updates.
[0126] (3) Training Task Management Submodule
[0127] After the samples are preprocessed, a self-supervised contrastive learning component is used to train the pre-trained model. Then, transfer learning technology and remote sensing image visual analysis components are used to identify and train the sample features, and extract the sample features to form a sample library.
[0128] Training task management mainly implements functions such as training sample selection, task start and stop, task monitoring, and recognition rate check. The functions are described below:
[0129] Training sample selection: Used to select samples to participate in training from all preprocessed samples.
[0130] Task Start and Stop: After selecting training samples, start self-supervised training; after selecting labeled samples, select the corresponding road change category to start deep learning training; if training needs to be stopped for any reason during the training process, the training can be terminated.
[0131] Task monitoring: Used to monitor the training process, displaying training progress in the form of a progress bar, including start time, estimated end time, and other information.
[0132] Recognition rate check: After fine-tuning the pre-trained model, one or more images can be manually selected for recognition. The manual check results are in line with expectations. If the results do not meet the requirements, samples can be added or replaced for retraining.
[0133] IV. Intelligent Recognition Module
[0134] (1) Self-supervised contrastive learning component
[0135] It mainly includes the basic framework for machine learning, components for batch data acquisition and storage, GPU and CPU scheduling components, and self-supervised structure components.
[0136] (2) Transfer learning components
[0137] It mainly includes the basic framework of machine learning, empirical and fine-tuning reference components for hyperparameters, and components for real-time storage and modification of trained models.
[0138] (3) Remote sensing image visual analysis component
[0139] A library of neural network visual analysis algorithm models for remote sensing big data.
[0140] V. Human-Computer Interaction Module
[0141] (1) Recognition Result Display Submodule
[0142] The recognition results display mainly includes the display and query of recognition results. The recognition result information includes road changes and handling status, etc. Map positioning is supported; clicking on an image can directly locate the corresponding map information.
[0143] (2) Task control submodule
[0144] Business personnel can stop, start, delete, and set monitoring intervals for each project and plot in the image catalog. If the system cannot find the required image at the monitoring time, it will prompt that there is no image and prompt the business personnel to conduct on-site verification.
[0145] (3) Manually determine the sub-module
[0146] Business personnel can identify the types of road changes automatically recognized by the system. After confirming the type of change for certain roads, they can report it according to regulatory requirements. For road change types that are difficult for the system to accurately identify through remote sensing images, business personnel can verify them through on-site confirmation. Photos of on-site verification can be uploaded to the system as supporting materials for reporting.
[0147] The specific embodiments described in this application are intended to enable those skilled in the art to gain a more comprehensive understanding of this application, but do not limit this application in any way. Therefore, those skilled in the art should understand that modifications or equivalent substitutions can still be made to this application; and all technical solutions and improvements that do not depart from the spirit and technical essence of this application should be covered within the scope of protection of this patent application.
Claims
1. A method for identifying road changes in remote sensing images, characterized in that, include: Acquire a large number of new and old remote sensing images of unmarked roads; The new and old remote sensing images are cropped based on the land parcel vector range corresponding to the road to be identified, resulting in new and old cropped remote sensing images that include the land parcel vector range. By comparing the old and new cropped remote sensing images, a remote sensing image of the changes in the old and new roads is obtained; Remote sensing images of changes in old and new roads are input into a deep learning model for self-supervised training to obtain a pre-trained model; A classification and recognition model was trained using a small number of labeled remote sensing images of road changes to obtain a self-supervised model parameter transfer model. The remote sensing image of the road change to be identified is input into the classification and recognition model with parameters transferred from the self-supervised model, and the road change results corresponding to the road to be identified are obtained from the output of the classification and recognition model.
2. The method for identifying road changes in remote sensing images according to claim 1, characterized in that, The comparison processing involves performing a special subtraction between the new and old road remote sensing images during the data processing process. Specifically, a road area present in the new image but absent in the old image is assigned a value of 1, a road area absent in the new image but present in the old image is assigned a value of 2, and all other cases are assigned a value of 0.
3. The method for identifying road changes in remote sensing images according to claim 1, characterized in that, The self-supervised training includes: The deep learning model used for self-supervised training is one or more of VGG16, ResNet18, ResNet50, or DenseNet121. A contrast-based self-supervised learning method is used for training, that is, the encoder is trained during the training process to ensure that the output vector of the encoder is similar to the corresponding key in a large dictionary and is not similar to other keys; The loss function used is InfoNCE, which achieves the effect that the loss decreases when the query vector q is similar to its positive key value k and dissimilar to its negative key value. The specific formula is shown below: Where K is the number of negative key values. It's a hyperparameter.
4. The method for identifying road changes in remote sensing images according to claim 3, characterized in that, The contrast-based self-supervised learning method uses a momentum contrast approach to store the dictionary, i.e., a queue is used; that is, each new batch of encoding is entered into the queue after completion, and then the oldest encoding is dequeued.
5. The remote sensing image road change recognition method according to claim 1, characterized in that, The method for self-supervised training using remote sensing images of changes between old and new roads involves the following steps: Data augmentation is performed on each remote sensing image to obtain a query image and a template image; The inputs are fed into a deep learning model to extract query features and template features, respectively. Assuming the template features are stable, we set the template features to not use gradient update parameters; The matching degree of features is calculated using matrix multiplication; Set the NCE loss function to ensure the maximum matching degree between the image and the image derived from it; The parameters of the query network are updated using gradient descent based on the loss, and the parameters of the template network are updated using momentum. Repeat the above steps to complete the training of self-supervised contrastive learning and obtain a pre-trained model.
6. The method for identifying road changes in remote sensing images according to claim 1, characterized in that, The training method for the classification and recognition model that transfers self-supervised model parameters specifically includes: Acquire labeled remote sensing image samples of road changes and the types of road changes they contain; The parameters of the pre-trained model are retained, and a linear classification network layer is connected after the encoder to identify road changes. The model was trained using labeled remote sensing images of road changes, and the parameters were fine-tuned to obtain the trained model. Labeled remote sensing image samples of road changes are input into the training model to obtain the predicted road change type output by the training model; The training loss value is calculated based on the road change type and the predicted road change type. If the loss value is within a preset range, the trained model will be used as the final road change recognition model, that is, a classification and recognition model that transfers the parameters of the self-supervised model.
7. A remote sensing image road change recognition device, characterized in that, The remote sensing image road change recognition method for performing any one of claims 1-6 includes a road change acquisition module, a cropped remote sensing image acquisition module, a self-supervised contrastive learning feature processing module, and a target recognition result acquisition module. The road change acquisition module is used to acquire new and old remote sensing images of the road to be identified; The cropped remote sensing image acquisition module is used to crop the new remote sensing image and the old remote sensing image according to the land parcel vector range corresponding to the road to be identified, so as to obtain a new and old cropped remote sensing image that only contains the land parcel vector range corresponding to the road to be identified. The self-supervised contrastive learning feature processing module is used for self-supervised training of unlabeled road change remote sensing images to obtain a feature extractor containing specific parameters. The target recognition result acquisition module is used to input the remote sensing image of the road change to be identified into the road change recognition model, and to obtain the road change type corresponding to the road change to be identified output by the road change recognition model.
8. The remote sensing image road change recognition device according to claim 7, characterized in that, The remote sensing image road change recognition device also includes a model training sample acquisition module, a land parcel remote sensing image generation module, a road change remote sensing image generation module, a self-supervised learning model parameter generation module, a predicted road change type acquisition module, a loss value calculation module, and a road change type recognition module acquisition module. The model training sample acquisition module is used to acquire model training samples; the model training samples include remote sensing images of old and new roads and the types of road changes contained in the remote sensing images; The land parcel remote sensing image generation module is used to crop the new remote sensing image and the old remote sensing image based on the road vector range of multiple new and old land parcels corresponding to the model training samples, and generate new and old road-cropped remote sensing images corresponding to the multiple land parcels. The road change remote sensing image generation module is used to compare and process cropped remote sensing images of old and new roads to generate new and old road change images corresponding to the multiple plots of land. The self-supervised learning model parameter generation module is used to compare the processed images of changes in the old and new roads, perform self-supervised training to update the model parameters, generate the parameters for feature extraction in the first few layers of the network, and obtain the trained model. The predicted road change type acquisition module is used to input the remote sensing image of the road change to be identified into the training model and acquire the predicted road change type of the multiple plots output by the training model. The loss value calculation module is used to calculate the loss value of the trained model based on the road change type and the predicted road change type. The road change type identification module is used to obtain the training model as the final road change identification model when the loss value is within a preset range.