High-resolution remote sensing image change detection method based on reinforcement learning

Abstract

The invention relates to a reinforcement learning-based high-resolution remote sensing image change detection method, and relates to the technical field of remote sensing image detection, and the method comprises the steps: collecting a first remote sensing image and a second remote sensing image of a target region at a first moment and a second moment, and obtaining a first imaging condition parameter and a second imaging condition parameter; acquiring a remote sensing content category set in the target region; performing remote sensing identification error test on the remote sensing content identifier by combining the first and second imaging condition parameters to obtain first and second remote sensing identification error rate sets; configuring reinforcement learning data and a reinforcement learning remote sensing content recognizer according to the first remote sensing recognition error rate set and the second remote sensing recognition error rate set, recognizing remote sensing contents of the first remote sensing image and the second remote sensing image, obtaining first remote sensing content distribution and second remote sensing content distribution, and calculating remote sensing content change information. According to the method, the problems of misrecognition of remote sensing content and insufficient change detection precision caused by the fact that imaging condition differences of data acquired at different times are not considered in traditional remote sensing image change detection are solved.

Description

Technical Field

[0001] This application relates to the field of remote sensing image detection, and in particular to a method for detecting changes in high-resolution remote sensing images based on reinforcement learning. Background Technology

[0002] With the increasing demands for spatial information accuracy in fields such as environmental monitoring and land use assessment, the accuracy of remote sensing image change detection results has become a key technical requirement to support decision-making.

[0003] Currently, traditional remote sensing image change detection methods do not fully consider the differences in imaging conditions of data collected at different times, and cannot specifically eliminate interference caused by changes in lighting and atmospheric conditions. This not only easily leads to misidentification of remote sensing content, but also reduces the reliability of change detection results and increases the decision-making risks of subsequent applications based on the detection data. Summary of the Invention

[0004] This application provides a high-resolution remote sensing image change detection method based on reinforcement learning, which improves the problem of misidentification caused by different imaging conditions when remote sensing data are acquired at different times, and enhances the accuracy of high-resolution remote sensing image change detection.

[0005] The embodiments of this application disclose the following technical solutions: This application provides a high-resolution remote sensing image change detection method based on reinforcement learning, the method comprising: Acquire first and second remote sensing images of the target area at the first and second time points, and obtain first and second imaging condition parameters at the first and second time points. Obtain all remote sensing content categories within the target area to obtain a set of remote sensing content categories; Based on the first imaging condition parameters and the second imaging condition parameters, and in conjunction with the remote sensing content category set, a remote sensing recognition error test is performed on the pre-constructed remote sensing content recognizer to obtain a first remote sensing recognition error rate set and a second remote sensing recognition error rate set. Based on the first remote sensing identification error rate set and the second remote sensing identification error rate set, reinforcement learning data is configured to perform reinforcement learning on the remote sensing content recognizer, and remote sensing content recognition is performed on the first remote sensing image and the second remote sensing image to obtain the first remote sensing content distribution and the second remote sensing content distribution, and remote sensing content change information is calculated.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application proposes a high-resolution remote sensing image change detection method based on reinforcement learning. Through a step-by-step process of acquiring remote sensing data, constructing a remote sensing content recognizer, testing remote sensing recognition errors, optimizing with reinforcement learning, and calculating remote sensing content change information, it achieves accurate detection of land cover changes in high-resolution remote sensing images of the target area. First, high-resolution remote sensing images of the target area at the first and second time points are acquired, simultaneously obtaining imaging condition parameters including illumination and atmospheric parameters for the corresponding time points. Then, all land cover categories are extracted from the historical remote sensing data of the target area to form a remote sensing content category set. Subsequently, based on the imaging condition parameters and the remote sensing content category set, error testing is conducted on the pre-constructed remote sensing content recognizer to obtain a set of remote sensing recognition error rates for the two time points. Next, differentiated reinforcement learning data is configured according to the error rates to train and optimize the remote sensing content recognizer until the verification accuracy converges. Finally, the optimized remote sensing content recognizer is used for content recognition in the images at the two time points, outputting the land cover distribution results and calculating the area change amplitude of each category to obtain remote sensing content change information.

[0007] The technical solution of this application solves the problems of low recognition accuracy due to differences in imaging conditions at different times, difficulty in adapting fixed recognition models to complex environments, and lack of targeted data support for the calculation of ground feature changes in traditional remote sensing image change detection. It avoids misjudgment and missed judgment of changes due to interference from imaging conditions or insufficient model accuracy, and improves the accuracy and reliability of change detection in high-resolution remote sensing images. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A flowchart illustrating the high-resolution remote sensing image change detection method based on reinforcement learning provided in this application embodiment; Figure 2 This is a schematic diagram illustrating the process of obtaining the remote sensing identification error rate set provided in the embodiments of this application. Detailed Implementation

[0010] This application provides a high-resolution remote sensing image change detection method based on reinforcement learning, which addresses the technical problem in the prior art where differences in imaging conditions when remote sensing data are acquired at different times can easily lead to misidentification of remote sensing content, thereby affecting the accuracy of remote sensing image change detection.

[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0013] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0014] Example 1, as shown in the appendix Figure 1 As shown, this application provides a high-resolution remote sensing image change detection method based on reinforcement learning, the method comprising the following steps: S110: Acquire the first and second remote sensing images of the target area at the first and second time points, and obtain the first and second imaging condition parameters at the first and second time points. In this embodiment of the application, in the scenario of high-resolution remote sensing image change detection, in order to eliminate the interference of different imaging conditions at different times on the accuracy of change detection, it is necessary to first acquire high-resolution remote sensing images for two key observation times of the target area, and simultaneously obtain the imaging condition parameters at the corresponding times, so as to ensure that subsequent content recognition can be carried out based on image details, and at the same time, rely on the imaging condition parameters to quantify the impact of environmental differences on the recognition results.

[0015] Specifically, the first step is to focus on the target area and determine the first and second time points for comparative analysis. Remote sensing images of the target area are acquired at these two time points, and the acquired images must be of high resolution to ensure that the detailed information of the remote sensing content in the images can be accurately identified.

[0016] Furthermore, while acquiring remote sensing images, the first imaging condition parameters corresponding to the first moment and the second imaging condition parameters corresponding to the second moment are recorded simultaneously.

[0017] Each imaging condition parameter includes both illumination and atmospheric conditions. By collecting both types of condition parameters, a key basis is provided for subsequent analysis of the impact of imaging conditions on remote sensing content recognition errors.

[0018] This step first acquires high-resolution remote sensing images at different times to ensure the detailed foundation for content recognition, and then simultaneously acquires imaging condition data including illumination and atmospheric parameters to form a basic dataset corresponding to remote sensing images and imaging conditions. This provides data support for subsequent remote sensing recognition error testing and the construction of reinforcement learning models adapted to different imaging conditions.

[0019] Step S110 in the method provided in this application embodiment includes: The first and second remote sensing images of the target area were acquired at the first and second time points, respectively, with high-resolution remote sensing images being acquired. Acquire the first imaging condition parameters and the second imaging condition parameters at the first and second time points, wherein each imaging condition parameter includes illumination condition parameters and atmospheric condition parameters.

[0020] In this embodiment of the application, in order to obtain remote sensing image data with temporal comparability and sufficient detailed information, and at the same time capture the environmental differences that affect the recognition of remote sensing content at different times, it is necessary to collect high-resolution remote sensing images for key observation times of the target area, and simultaneously record the illumination and atmospheric parameters at the corresponding times, so as to ensure that the interference of imaging condition differences on the detection accuracy of remote sensing image changes can be accurately analyzed in the future.

[0021] Specifically, high-resolution remote sensing images of the target area at the first and second time points were acquired first.

[0022] Among them, the high-resolution images can clearly present the boundary outline of water bodies, structural details of buildings, and paving features of roads in the target area, thereby avoiding misidentification problems such as confusion between vegetation and bare land or omission of small buildings when remote sensing content is identified due to insufficient image resolution.

[0023] During the acquisition process, it is necessary to ensure that the spatial resolution of the images at the two time points is consistent, for example, using a spatial resolution specification of 0.5 meters to 2 meters. At the same time, control the basic parameters such as the shooting angle and coverage of the images to enable direct comparison between the first and second remote sensing images in spatial dimension, and avoid misalignment of the target area due to deviation in the shooting range.

[0024] While acquiring images, the system simultaneously obtains the first imaging condition parameters corresponding to the first moment and the second imaging condition parameters corresponding to the second moment. Each imaging condition parameter includes illumination condition parameters and atmospheric condition parameters.

[0025] Specifically, illumination condition parameters mainly include indicators such as illumination intensity, solar altitude angle, and illumination direction. These indicators directly affect the brightness, contrast, and shadow characteristics of ground objects in remote sensing images.

[0026] In addition, atmospheric condition parameters include indicators such as atmospheric transparency, cloud cover, and atmospheric aerosol concentration. When atmospheric transparency is low or cloud cover is high, the image will appear dark overall and the details will be blurred. Differences in atmospheric aerosol concentration will lead to errors in the calculation of ground reflectance.

[0027] Furthermore, during the acquisition of imaging condition parameters, it is necessary to ensure the timeliness and accuracy of the data.

[0028] For illumination condition parameters, they can be recorded in real time by illumination sensors mounted on remote sensing equipment, and supplemented and corrected by combining them with the observation data of the local meteorological station at the same time. For atmospheric condition parameters, atmospheric remote sensing technology can be used to obtain atmospheric vertical profile data, and combined with the real-time monitoring results of the regional atmospheric environment monitoring network to avoid deviations in subsequent error analysis caused by the acquisition error of imaging condition parameters.

[0029] At the same time, it is necessary to associate and label the first and second remote sensing images and their corresponding imaging condition parameters. For example, add attribute tags such as "acquisition time - illumination intensity - solar elevation angle - atmospheric transparency - cloud cover" to each remote sensing image to ensure that the remote sensing image data under the corresponding imaging conditions can be quickly indexed when conducting remote sensing identification error tests, thereby improving the efficiency of data retrieval and analysis.

[0030] For example, if the target area is a suburban area, the first time point is selected as a clear noon in June 2024, and the second time point is selected as a cloudy morning in January 2025.

[0031] During the initial acquisition, a high-resolution remote sensing image with a spatial resolution of 1 meter was obtained. The first imaging condition parameters recorded simultaneously were: illumination intensity of 80,000 lux, solar altitude angle of 75°, atmospheric transparency of 0.9, and cloud cover of 5%.

[0032] During the second acquisition, a high-resolution remote sensing image with a spatial resolution of 1 meter was also obtained. Among the recorded second imaging condition parameters, the illumination intensity was 35,000 lux, the solar altitude angle was 30°, the atmospheric transparency was 0.7, and the cloud cover was 60%.

[0033] The above acquisition method ensures the consistency of spatial resolution between the high-resolution remote sensing images at two different times, allowing them to be directly used for comparison of ground features. It also fully captures the differences in illumination and atmospheric conditions, laying a data foundation for the effective implementation of high-resolution remote sensing image change detection methods.

[0034] S120: Obtain all remote sensing content categories within the target area to obtain a set of remote sensing content categories; In this embodiment of the application, in order to clarify the specific objects of remote sensing content identification and change analysis and avoid identification omissions or invalid calculations caused by fuzzy detection range, it is necessary to first extract all existing remote sensing content categories from the historical remote sensing data of the target area to improve the efficiency and accuracy of change detection of high-resolution remote sensing images.

[0035] Specifically, the first step is to acquire historical remote sensing data of the target area within a preset time range. The preset time range must be determined based on the actual conditions of the target area to ensure that the extracted remote sensing content categories comprehensively reflect the actual land cover composition of the target area.

[0036] Furthermore, based on the acquired historical remote sensing data, all remote sensing content categories within the target area are systematically extracted.

[0037] During the extraction process, it is necessary to identify and classify the land cover information in the historical remote sensing data one by one. That is, to distinguish specific categories such as vegetation, water bodies, buildings, and roads from historical images of different time periods, to ensure that all remote sensing content that has appeared in the target area is covered without duplication or omission, and finally integrated to form a complete set of remote sensing content categories.

[0038] This step extracts all remote sensing content categories for the target area from historical remote sensing data, providing specific testing dimensions for subsequent remote sensing identification error testing. This allows error testing to be conducted for each remote sensing content category in combination with different imaging conditions, ensuring that the test results are more targeted.

[0039] Step S120 in the method provided in this application embodiment includes: Acquire historical remote sensing data of the target area within a preset time range; Extract all remote sensing content categories within the target area from the historical remote sensing data to obtain a set of remote sensing content categories.

[0040] In this embodiment of the application, in order to clarify the specific analysis object of high-resolution remote sensing image change detection and avoid content omissions or invalid analysis in the subsequent identification and change calculation process, it is necessary to sort out all the remote sensing content categories that have existed in the past observation data of the target area to improve the targeting of the detection and the accuracy of the results.

[0041] Specifically, the first step is to acquire historical remote sensing data of the target area within a preset time frame.

[0042] The preset time range is determined by combining the land cover change characteristics of the target area. For example, historical remote sensing data from the past 5-10 years are selected to cover both possible seasonal land cover changes in the target area and long-term land cover types, so as to ensure that the acquired historical remote sensing data can comprehensively reflect the overall land cover composition of the target area.

[0043] Meanwhile, during the data acquisition process, it is necessary to ensure that the resolution of historical remote sensing data is not lower than that of the first and second remote sensing images used for subsequent change detection. For example, if the resolution of the subsequent remote sensing images used for detection is 1 meter, then the resolution of the historical data must be ≥1 meter to ensure that the details of ground features extracted from the historical data can accurately match the content of the subsequent detection images and avoid deviations in the judgment of ground feature categories due to resolution differences.

[0044] After acquiring historical remote sensing data, the next step is to extract all remote sensing content categories within the target area. During this extraction process, each image in the historical remote sensing data needs to be identified and classified on a region-by-region and feature-by-feature basis.

[0045] For example, from historical images from different years and seasons, vegetation types including forests, farmland, grasslands, and shrubs can be identified; water body types including rivers, lakes, reservoirs, and ponds can be identified; building types including residential buildings, industrial plants, commercial buildings, and public facilities can be identified; road types including highways, urban arterial roads, rural roads, and park roads can be identified; and land features such as bare land, gardens, and wetlands can be identified.

[0046] At the same time, it is necessary to deduplicatize and integrate the identified remote sensing content categories. For example, deciduous forests and evergreen forests should be classified into the vegetation category, and asphalt roads and cement roads should be classified into the road category. In the end, a set of remote sensing content categories covering all actual land cover types in the target area will be formed.

[0047] Furthermore, when extracting remote sensing content categories, it is also necessary to cross-validate the historical imagery by combining its temporal attributes with the features of the ground objects.

[0048] For example, if a certain area shows green coverage in spring and yellowish-brown in autumn in historical images, the area can be identified as farmland; if a certain area is gradually built from scratch and shows architectural spectrum and structural characteristics, it can be identified as an architectural area.

[0049] Ultimately, by combining temporal attributes with ground feature characteristics, the extraction method described above avoids misclassification caused by relying solely on a single historical image, ensuring the accuracy of the remote sensing content category set. Simultaneously, each extracted category of remote sensing content needs to be labeled with the time of its first and last appearance in historical images, providing a reference for subsequent analysis of long-term trends in ground features.

[0050] S130: Based on the first imaging condition parameters and the second imaging condition parameters, and in conjunction with the remote sensing content category set, perform remote sensing recognition error testing on the pre-constructed remote sensing content recognizer to obtain a first remote sensing recognition error rate set and a second remote sensing recognition error rate set. In this embodiment of the application, in order to quantify the recognition deviation of the recognizer under different imaging conditions and different combinations of remote sensing content categories, it is necessary to first conduct a systematic error test on the pre-constructed remote sensing content recognizer based on the acquired imaging condition parameters and remote sensing content category set, so as to identify the accuracy shortcomings of the recognizer in different scenarios.

[0051] Specifically, the first step is to traverse and combine the first imaging condition parameters, the second imaging condition parameters, and the remote sensing content category set to generate multiple first remote sensing features and multiple second remote sensing features.

[0052] Each first remote sensing feature contains a first imaging condition parameter and any one category from the remote sensing content category set, and each second remote sensing feature contains a second imaging condition parameter and any one category from the remote sensing content category set. This combination method can cover the combination scenarios of imaging conditions and all land cover categories at two time points, ensuring that no error tests are missed.

[0053] Furthermore, based on the generated multiple first and second remote sensing features, data indexing is performed on the remote sensing data within the historical time period to filter out remote sensing images that match each feature, forming multiple first test remote sensing image sets and multiple second test remote sensing image sets. The indexing and matching here must ensure that the imaging conditions, land cover categories, and corresponding features of the historical remote sensing data are completely consistent.

[0054] Furthermore, the test remote sensing images in multiple first test remote sensing image sets and multiple second test remote sensing image sets are respectively input into a pre-constructed remote sensing content recognizer. The recognizer identifies the content of each test image and outputs multiple first test remote sensing content sets and multiple second test remote sensing content sets.

[0055] At the same time, the proportion of test remote sensing content in each test remote sensing content set that matches the remote sensing content category in the corresponding remote sensing feature is calculated. This proportion is the remote sensing recognition accuracy in the corresponding scenario.

[0056] Furthermore, by calculating "1-remote sensing recognition accuracy", multiple first remote sensing recognition error rates and multiple second remote sensing recognition error rates are obtained respectively, and then integrated to form a set of first remote sensing recognition error rates and a set of second remote sensing recognition error rates.

[0057] This step, through systematic error testing, clearly quantifies the error of the remote sensing content recognizer under different imaging conditions and combinations of ground cover categories, laying the foundation for error analysis to improve the final change detection accuracy.

[0058] As attached Figure 2 As shown, step S130 in the method provided in this application embodiment includes: Based on the first imaging condition parameters, the second imaging condition parameters, and the remote sensing content category set, multiple first remote sensing features and multiple second remote sensing features are obtained by traversing and combining them. Each first remote sensing feature includes the first imaging condition parameters and any one remote sensing content category. Indexing is performed within the remote sensing data over a historical period based on multiple first remote sensing features and multiple second remote sensing features to obtain multiple first test remote sensing image sets and multiple second test remote sensing image sets. The test remote sensing images in multiple first test remote sensing image sets and multiple second test remote sensing image sets are respectively input into the remote sensing content recognizer to obtain multiple first test remote sensing content sets and multiple second test remote sensing content sets as recognition outputs; Calculate the proportion of test remote sensing content in each test remote sensing content set that matches the remote sensing content category in the corresponding remote sensing feature, obtain multiple first remote sensing recognition accuracies and multiple second remote sensing recognition accuracies, and calculate the first remote sensing recognition error rate set and the second remote sensing recognition error rate set.

[0059] In this embodiment of the application, in order to quantify the recognition deviation of the pre-built remote sensing content recognizer under different imaging conditions and different combinations of remote sensing content categories, it is necessary to obtain a set of recognition error rates through error testing based on the acquired first and second imaging condition parameters and remote sensing content category sets, so as to improve the accuracy of the final high-resolution remote sensing image change detection.

[0060] Specifically, the first imaging condition parameters and the second imaging condition parameters are first combined with the remote sensing content category set. That is, the first imaging condition parameters are paired with each category in the remote sensing content category set to form multiple first remote sensing features. The second imaging condition parameters are also paired with each category in the remote sensing content category set to form multiple second remote sensing features.

[0061] For example, if the first imaging condition parameter is "sunny noon in June 2024 (illuminance 80000 lux, atmospheric transparency 0.9)", and the remote sensing content category set includes "vegetation, water, buildings, roads", then the generated first remote sensing features include "sunny noon in June 2024 + vegetation", "sunny noon in June 2024 + water", "sunny noon in June 2024 + buildings", and "sunny noon in June 2024 + roads".

[0062] Similarly, if the second imaging condition parameter is "January 2025, cloudy morning (illuminance 35000 lux, atmospheric transparency 0.7)," the generated second remote sensing features include "January 2025, cloudy morning + vegetation" and "January 2025, cloudy morning + water body," etc. By traversing and combining them, the scene of the imaging conditions and all land cover categories at the two time points can be fully covered to avoid missing test scenes.

[0063] Furthermore, after generating the first and second remote sensing features, the remote sensing data within the historical time period are indexed according to the features to filter out remote sensing images that match each feature, forming multiple first test remote sensing image sets and multiple second test remote sensing image sets.

[0064] During the indexing process, it is essential to ensure that the imaging conditions, land cover categories, and corresponding features of historical remote sensing data are completely consistent.

[0065] For example, for the first remote sensing feature of "sunny noon in June 2024 + vegetation", it is necessary to select remote sensing images containing vegetation categories that were taken under the same light intensity and atmospheric transparency conditions from historical remote sensing data. The resolution of the images must be consistent with the resolution of the first and second remote sensing images used for change detection in order to ensure that the test data can truly reflect the recognition situation in the corresponding scene and avoid the distortion of test results due to data matching deviation.

[0066] Furthermore, the test remote sensing images obtained from multiple first test remote sensing image sets and multiple second test remote sensing image sets are respectively input into a pre-built remote sensing content recognizer.

[0067] In the method provided in this application embodiment, the step of "pre-constructing the remote sensing content recognizer" includes: Based on remote sensing data, a set of sample remote sensing images is collected, and the content distribution within each remote sensing image is labeled to obtain a set of sample remote sensing content distributions. Each sample remote sensing content distribution includes remote sensing entity content with all coordinates within the remote sensing image. A network architecture for a remote sensing content recognizer is constructed based on deep learning. The remote sensing content recognizer is trained under supervised supervision using the sample remote sensing image set and the sample remote sensing content distribution set until the verification accuracy converges, thus completing the pre-construction.

[0068] In this embodiment of the application, in order to accurately identify various remote sensing contents in high-resolution remote sensing images, it is necessary to construct a remote sensing content recognizer through the steps of sample collection and annotation, network architecture construction and supervised training, so as to lay the foundation for subsequent error quantification and reinforcement learning optimization.

[0069] Specifically, firstly, based on remote sensing data covering different imaging conditions and different types of land cover, a sufficient amount of sample remote sensing images are collected and a sample remote sensing image set is formed to ensure that the sample images include images under different lighting and atmospheric conditions, such as sunny noon and cloudy early morning, as well as covering a variety of remote sensing content categories such as vegetation, water bodies, buildings, and roads.

[0070] Meanwhile, the image resolution must be consistent with the resolution of the first and second remote sensing images used for change detection (both are 1 meter) to ensure the compatibility between the sample data and the actual detection data.

[0071] Furthermore, the content distribution within each sample remote sensing image is finely annotated, and the annotation range must cover all coordinates within the remote sensing image to clearly identify the remote sensing entity content corresponding to each coordinate.

[0072] For example, if a coordinate corresponds to a farmland area, it is labeled as "vegetation-farmland"; if another coordinate corresponds to a river area, it is labeled as "water body-river". By labeling each coordinate, a sample remote sensing content distribution set containing complete land cover distribution information is formed, providing accurate label data for the supervised training of the recognizer.

[0073] Furthermore, after completing sample collection, a network architecture for a remote sensing content recognizer was constructed based on deep learning. Considering the needs of remote sensing image land cover recognition, a deep learning network based on image semantic segmentation (ResNet) was selected as the basic architecture to effectively capture the texture, spectral, and structural features of land covers in remote sensing images, adapting to the recognition needs of different types of land covers.

[0074] Specifically, during the architecture construction process, the number of network layers, convolution kernel size, pooling method and other parameters need to be set reasonably according to the scale of the sample data and the complexity of the land cover categories, so as to ensure that the network can fully learn the land cover features in the sample, and avoid overfitting or slow convergence due to overly complex architecture.

[0075] For example, if the sample data size is 20,000 to 40,000 images and the land cover categories include 5 types: vegetation, water bodies, buildings, roads, and bare land, a 12-layer network architecture can be constructed. The encoding end has 5 layers (including the input layer). The input layer receives 512×512×3 remote sensing images. The 2nd to 5th convolutional layers use 3×3 convolution kernels (stride 1, same padding). After the 2nd to 4th convolutional layers, a 2×2 max pooling layer (stride 2, no padding) is set.

[0076] Meanwhile, the decoding end has 7 layers: layers 6-9 are 2×2 deconvolutional layers (stride 2), layers 10-11 are 3×3 convolutional layers, and layer 12 is the output layer. At the same time, skip connections are set between the corresponding layers of the encoding and decoding ends.

[0077] In addition, if the sample data size is less than 10,000 images and there are only 3 types of land features, it can be simplified to a 10-layer architecture. One convolutional and pooling layer is reduced at the encoding end, and some convolutional layers are replaced with 1×1 convolutional kernels to reduce network complexity.

[0078] Furthermore, the constructed remote sensing content recognizer is trained under supervision using a set of sample remote sensing images and a set of sample remote sensing content distributions.

[0079] During training, sample remote sensing images are input into the remote sensing content recognizer, which outputs the ground feature content recognition results. The recognition results are then compared with the real annotations in the sample remote sensing content distribution set. The recognition error is quantified by calculating loss functions such as cross-entropy loss, and the network parameters are continuously adjusted based on the gradient descent algorithm in existing technology to optimize the recognition accuracy of the recognizer.

[0080] Meanwhile, during the training process, it is necessary to divide the training set, validation set, and test set. The training set is used for iterative parameter updates, and the validation set is used to monitor the model performance during the training process to avoid overfitting. When the recognition accuracy on the validation set no longer improves for several consecutive rounds and reaches a convergence state, training is stopped. At this point, the remote sensing content recognizer has a stable and reliable basic recognition capability and has completed the pre-construction.

[0081] Furthermore, during supervised training, the training effect needs to be verified periodically. For example, after every 10 rounds of training, the accuracy of the remote sensing content recognizer should be tested using validation set data. If a downward trend in accuracy is observed, training parameters such as the learning rate and batch size should be adjusted in a timely manner, or more diverse sample data should be added to avoid overfitting and ensure that the remote sensing content recognizer maintains a stable learning state, thereby improving its ability to generalize the recognition of different land cover categories.

[0082] For example, if the accuracy of the validation set drops from 92% to 88% after a certain round of training, the learning rate can be adjusted from 0.001 to 0.0005, or the batch size can be increased from 32 images / batch to 64 images / batch. If it is found that the recognition accuracy of the "water body-pond" category is consistently below 80%, 500 sample remote sensing images containing ponds are added and re-integrated into the sample set for training until the recognition accuracy of each category on the validation set is stable above 85% and there is no significant decrease for 5 consecutive rounds, so as to ensure that the remote sensing content recognizer has a balanced and reliable basic recognition capability.

[0083] Ultimately, the remote sensing content recognizer constructed through the above steps can accurately identify different types of remote sensing ground features and adapt to different combinations of imaging conditions and ground feature categories in subsequent error tests, providing a stable testing tool for calculating the first and second remote sensing recognition error rate sets.

[0084] Furthermore, the test remote sensing images in multiple first test remote sensing image sets and multiple second test remote sensing image sets corresponding to the first and second times of the target area are respectively input into the pre-constructed remote sensing content recognizer to identify and obtain multiple first test remote sensing content sets and multiple second test remote sensing content sets corresponding to each test remote sensing image.

[0085] Furthermore, the percentage of test remote sensing content within each test remote sensing content set that matches the remote sensing content category within the corresponding remote sensing feature is calculated.

[0086] Meanwhile, the average accuracy of all images within the same test image set is taken as the first remote sensing recognition accuracy corresponding to the first remote sensing feature.

[0087] For example, if the first test remote sensing image set corresponding to "June 2024 sunny noon + vegetation" contains 15 images, and the recognition accuracy of each image is 93%, 91%, 92%, 89%, 94%, 90%, 92%, 91%, 88%, 93%, 90%, 92%, 89%, 91%, and 93%, respectively, the above accuracy rates are added together and divided by 15 to obtain an average accuracy rate of 91.2%. This value is the first remote sensing recognition accuracy rate corresponding to the first remote sensing feature of "June 2024 sunny noon + vegetation".

[0088] Similarly, the accuracy rates of first remote sensing identification corresponding to all first remote sensing features and the accuracy rates of second remote sensing identification corresponding to all second remote sensing features can be calculated.

[0089] Finally, by calculating "1 - recognition accuracy", each accuracy rate is converted into a corresponding error rate, and the results are integrated to form the first remote sensing recognition error rate set and the second remote sensing recognition error rate set.

[0090] For example, if the accuracy rate of the first remote sensing identification of "sunny noon in June 2024 + vegetation" is 91.2%, then the corresponding error rate of the first remote sensing identification is 1-91.2%=8.8%; if the accuracy rate of the second remote sensing identification of "cloudy morning in January 2025 + water body" is 87.5%, then the corresponding error rate of the second remote sensing identification is 1-87.5%=12.5%.

[0091] Simultaneously, the error rates corresponding to all the calculated first remote sensing features are summarized to form a first remote sensing identification error rate set, and the error rates corresponding to all the second remote sensing features are summarized to form a second remote sensing identification error rate set.

[0092] Finally, the first and second remote sensing recognition error rate sets obtained through the above steps can clearly show the accuracy performance of the remote sensing content recognizer under different imaging conditions and combinations of ground cover categories, providing accurate data basis for subsequent configuration of reinforcement learning data based on error rates and targeted optimization of the recognizer accuracy.

[0093] S140: Based on the first remote sensing identification error rate set and the second remote sensing identification error rate set, configure reinforcement learning data, perform reinforcement learning on the remote sensing content recognizer, and perform remote sensing content recognition on the first remote sensing image and the second remote sensing image to obtain the first remote sensing content distribution and the second remote sensing content distribution, and calculate the remote sensing content change information.

[0094] In this embodiment of the application, in order to improve the accuracy of the remote sensing content recognizer in recognizing the content of remote sensing images at the first and second time points, it is necessary to first determine the average error level based on the error rate set to configure reinforcement learning data, then optimize the recognizer through reinforcement learning, and finally use the optimized recognizer to complete image recognition and change calculation, so as to ensure that the change detection results can truly reflect the changes in ground features in the target area.

[0095] Specifically, firstly, based on the first set of remote sensing identification error rates and the second set of remote sensing identification error rates, the first average remote sensing identification error rate and the second average remote sensing identification error rate are calculated. That is, by taking the arithmetic mean of all error rate data in the first set of remote sensing identification error rates, the first average remote sensing identification error rate, which can represent the overall identification error level under the first imaging condition, is obtained.

[0096] Similarly, the arithmetic mean of all error rate data in the second remote sensing identification error rate set is taken to obtain the second average remote sensing identification error rate, thereby clarifying the overall accuracy difference of the identifier under the imaging conditions at the two time points.

[0097] Furthermore, reinforcement learning data is configured based on the calculated first average remote sensing recognition error rate and the second average remote sensing recognition error rate.

[0098] Furthermore, the configured reinforcement learning data is used to perform reinforcement learning on the remote sensing content recognizer. After the reinforcement learning is completed, remote sensing content recognition is performed on the first and second remote sensing images.

[0099] Specifically, the first and second remote sensing images are input into the optimized remote sensing content recognizer, and the recognizer will output the distribution of the first and second remote sensing content to ensure that the distribution information of ground features at the two times can be accurately matched.

[0100] Finally, based on the distribution of the first and second remote sensing contents, information on changes in remote sensing contents is calculated.

[0101] This step, through the technical solution of "error quantification - data configuration - model optimization - recognition calculation", not only solves the accuracy deviation problem of remote sensing content recognizer under different imaging conditions, but also realizes accurate detection of changes in remote sensing content in the target area.

[0102] Step S140 in the method provided in this application embodiment includes: Based on the first set of remote sensing identification error rates and the second set of remote sensing identification error rates, the first average remote sensing identification error rate and the second average remote sensing identification error rate are calculated. Based on the first average remote sensing identification error rate and the second average remote sensing identification error rate, configure reinforcement learning data; The reinforcement learning data is used to perform reinforcement learning on the remote sensing content recognizer, and the reinforcement learning is completed after the verification accuracy converges.

[0103] The first and second remote sensing images are input into the remote sensing content recognizer that has completed reinforcement learning, and the first and second remote sensing content distributions are output respectively. Based on the first and second remote sensing content distributions, the area change rate of each remote sensing content category within the target area is calculated as remote sensing content change information.

[0104] In this embodiment of the application, in order to optimize the remote sensing content recognizer for the accuracy difference under the first and second imaging conditions, it is necessary to first calculate the average error level based on the error rate set to configure differentiated reinforcement learning data, then optimize the remote sensing content recognizer through reinforcement learning, and finally use the optimized remote sensing content recognizer to complete image recognition and change calculation, so as to ensure that the change detection results can truly reflect the dynamic changes of ground features in the target area.

[0105] Specifically, the first average remote sensing identification error rate and the second average remote sensing identification error rate are first calculated based on the first remote sensing identification error rate set and the second remote sensing identification error rate set.

[0106] Specifically, the arithmetic mean of all error rate data in the first remote sensing recognition error rate set is taken to obtain the first average remote sensing recognition error rate that can represent the overall error level of the remote sensing content recognizer under the first imaging condition.

[0107] For example, suppose the first set of remote sensing identification error rates includes "June 2024 sunny noon + vegetation" 8.8%, "June 2024 sunny noon + water body" 9.5%, "June 2024 sunny noon + building" 7.9%, and "June 2024 sunny noon + road" 8.2%. Add these four error rate data together (8.8% + 9.5% + 7.9% + 8.2% = 34.4%), and then divide by the number of data points, 4, to get the first average remote sensing identification error rate of 8.6% (34.4% / 4 = 8.6%).

[0108] Similarly, the arithmetic mean of all error rate data in the second remote sensing identification error rate set is taken to obtain the second average remote sensing identification error rate.

[0109] For example, suppose the second set of remote sensing identification error rates includes "January 2025, cloudy morning + vegetation" 10.8%, "January 2025, cloudy morning + water body" 12.5%, "January 2025, cloudy morning + buildings" 11.2%, and "January 2025, cloudy morning + roads" 9.9%. Add these four error rate data together (10.8% + 12.5% ​​+ 11.2% + 9.9% = 44.4%), and then divide by the number of data points, 4, to get the second average remote sensing identification error rate of 11.1% (44.4% / 4 = 11.1%).

[0110] At this point, by comparing the two average error rates, it is possible to intuitively determine under which imaging condition the overall accuracy of the remote sensing content recognizer is weaker, providing a quantitative standard for subsequent data configuration.

[0111] Furthermore, reinforcement learning data is configured based on the calculated first average remote sensing recognition error rate and the second average remote sensing recognition error rate.

[0112] In the method provided in this application embodiment, "configuring reinforcement learning data based on the first average remote sensing identification error rate and the second average remote sensing identification error rate" includes: Obtain reinforcement learning data; Based on the first average remote sensing identification error rate and the second average remote sensing identification error rate, the reinforcement learning data volume is divided to obtain the first data volume and the second data volume. According to the first data volume and the second data volume, collect the first enhanced remote sensing image set and the second enhanced remote sensing image set under the first imaging condition parameters and the second imaging condition parameters, and collect the corresponding first enhanced remote sensing content distribution set and the second enhanced remote sensing content distribution set. The first enhanced remote sensing image set, the second enhanced remote sensing image set, the first enhanced remote sensing content distribution set, and the second enhanced remote sensing content distribution set are integrated into reinforcement learning data.

[0113] In this embodiment of the application, in order to ensure that reinforcement learning can accurately improve the performance of the remote sensing content recognizer in scenarios with high error rates, it is necessary to configure the reinforcement learning data through a process of determining the total amount of data, dividing the data according to the error rate, collecting matching data and integrating it, so as to ensure that subsequent reinforcement learning can efficiently make up for the accuracy defects of the recognizer.

[0114] Specifically, the first step is to obtain the amount of reinforcement learning data. The determination of the amount of data needs to be based on a comprehensive judgment of the network complexity of the remote sensing content recognizer, training hardware resources, and expected optimization results.

[0115] For example, if the remote sensing content recognizer is a 12-layer ResNet architecture and the training device has 24GB of video memory, the amount of reinforcement learning data can be set to 2000 sets. This can meet the sample size required for network parameter optimization and avoid excessively long training cycles or memory overflow due to excessive data volume.

[0116] Meanwhile, it is necessary to ensure that each set of data includes one remote sensing image and corresponding content distribution labels, and that the image resolution is consistent with the resolution of the first and second remote sensing images (both are 1 meter) to ensure the data's adaptability to subsequent identification tasks.

[0117] Furthermore, based on the first average remote sensing recognition error rate and the second average remote sensing recognition error rate, the acquired reinforcement learning data volume is divided to obtain the first data volume and the second data volume.

[0118] Specifically, the division follows the principle of "the larger the error rate, the larger the amount of training data". The amount of data is allocated by calculating the ratio of the two average error rates to ensure that the imaging conditions with weaker remote sensing content recognition accuracy can obtain more training samples to support them.

[0119] For example, if the first average remote sensing recognition error rate is 8.6% and the second average remote sensing recognition error rate is 11.1%, the ratio of their recognition error rates is approximately 4:5. If the total amount of reinforcement learning data is 2000 sets, then the first data set is divided into 889 sets (2000×4 / 9≈889) and the second data set is divided into 1111 sets (2000×5 / 9≈1111) according to this ratio. This allows the second imaging condition (with a higher error rate) to obtain more training data, thereby specifically improving the recognition accuracy under it.

[0120] Furthermore, after determining the first data volume and the second data volume, enhancement data under the first imaging condition parameters and the second imaging condition parameters are collected according to these two data volumes respectively.

[0121] Specifically, when collecting the first enhanced remote sensing image set, it is necessary to select remote sensing images that are completely consistent with the first imaging condition parameters, such as "sunny noon in June 2024, light intensity 80,000 lux, atmospheric transparency 0.9". In addition, the images must cover all categories in the remote sensing content category set to ensure that the training samples can cover all land feature recognition scenarios under the imaging conditions.

[0122] At the same time, all remote sensing entities at all coordinates within each first enhanced remote sensing image are labeled to form a first enhanced remote sensing content distribution set. For example, coordinates (150, 250) are labeled as "vegetation-grassland" and coordinates (400, 500) are labeled as "building-industrial plant" to ensure that the labeling information completely matches the real ground features in the image.

[0123] Similarly, when collecting the second enhanced remote sensing image set, it is necessary to match the second imaging condition parameters, such as "January 2025, cloudy morning, light intensity 35000 lux, atmospheric transparency 0.7", and label them to form the second enhanced remote sensing content distribution set. The labeling standard must be consistent with the first enhanced remote sensing content distribution set to avoid affecting the training effect due to labeling differences.

[0124] Finally, the first enhanced remote sensing image set, the second enhanced remote sensing image set, the first enhanced remote sensing content distribution set, and the second enhanced remote sensing content distribution set are integrated to form complete reinforcement learning data.

[0125] Specifically, during the integration process, it is necessary to establish a one-to-one mapping relationship between images and corresponding labels. For example, each first enhanced remote sensing image is assigned a unique identifier and associated with the labeled data with the same identifier in the first enhanced remote sensing content distribution set, so as to ensure that the remote sensing content recognizer can accurately call the images and corresponding labels for learning during subsequent training.

[0126] At the same time, the integrated reinforcement learning data needs to be formatted uniformly, converting the image format to PNG and the labeled data to JSON format, so that the training framework can read and calculate them, ultimately forming reinforcement learning data with a clear structure and standardized format.

[0127] Furthermore, the integrated reinforcement learning data is used to train the remote sensing content recognizer through reinforcement learning.

[0128] During specific training, the first enhanced remote sensing image set and the second enhanced remote sensing image set are first input into the recognizer in batches. The number of images in each batch is set according to the performance of the training equipment. For example, based on a 24GB video memory device, 32 images are input in each batch. The remote sensing content recognizer identifies the image content and outputs the predicted distribution results of land cover categories.

[0129] Furthermore, the prediction results are compared with the real annotations in the corresponding first enhanced remote sensing content distribution set and the second enhanced remote sensing content distribution set. The recognition error of each image is calculated by the cross-entropy loss function. Then, the error is backpropagated based on the gradient descent algorithm in the existing technology to adjust the parameters such as the convolutional layer weights and deconvolutional layer biases in the remote sensing content recognizer network to reduce the recognition bias.

[0130] During training, it is necessary to divide the data into a reinforcement training set (accounting for 80% of the reinforcement learning data) and a reinforcement validation set (accounting for 20%). The reinforcement training set is used for iterative parameter updates, and the reinforcement validation set is used for real-time monitoring of the generalization performance of the recognizer.

[0131] After each round of training, the validation accuracy of the recognizer is tested using a reinforced validation set. If the validation accuracy does not improve for 10 consecutive rounds, for example, if it stabilizes above 95% and fluctuates by less than 0.5%, the validation accuracy is considered to have converged, and reinforcement learning is stopped. At this point, the remote sensing content recognizer has optimized the shortcomings under the first and second imaging conditions and has a more stable ability to recognize ground features.

[0132] After reinforcement learning is completed, the first and second remote sensing images of the target area are input into the optimized remote sensing content recognizer. Based on the land feature characteristics learned during training under different imaging conditions, the remote sensing content recognizer performs coordinate-by-coordinate land feature classification on the first remote sensing image and outputs the first remote sensing content distribution containing the spatial distribution of all land features in the target area at the first time point. For example, coordinates (200, 300) are "water body - river" and coordinates (500, 600) are "vegetation - farmland".

[0133] Similarly, the second remote sensing image is identified, and the distribution of the second remote sensing content at the second time point is output. Both images use the same coordinate system to ensure that the distribution of ground features can be accurately mapped to the same spatial location.

[0134] Finally, based on the first and second remote sensing content distributions, the area change of each remote sensing content category within the target area is calculated as remote sensing content change information.

[0135] In the specific calculation, firstly, for each land feature category, the number of its coordinates in the first remote sensing content distribution is counted, and combined with the image resolution (1 meter), the actual area at the first moment is calculated (1 coordinate corresponds to 1 square meter, and the number of coordinates is the area in square meters).

[0136] Similarly, the actual area of ​​the land cover category at the second time point in the second remote sensing content distribution is calculated in the same way.

[0137] Furthermore, the rate of change is calculated using the formula: "Area change range = (Area at second time point - Area at first time point) / Area at first time point × 100%". For example, if the vegetation area is 50,000 square meters at the first time point and 45,000 square meters at the second time point, the area change range is (45,000 - 50,000) / 50,000 × 100% = -10%, indicating that the vegetation area of ​​the target area decreased by 10% between the two time points.

[0138] In addition, the building category has an area of ​​10,000 square meters at the first moment and 15,000 square meters at the second moment, with a change of (15,000-10,000) / 10,000=50%, which means that the building area increases by 50%.

[0139] Finally, by summarizing the area changes of all land cover categories, a complete set of remote sensing content change information is formed. This information can be directly used in scenarios such as land use change assessment and ecological environment monitoring in the target area. It can accurately reflect the dynamic evolution of land cover types in the target area from the first time point to the second time point, and provide data support for the subsequent formulation of land planning, ecological protection and other strategies.

[0140] The embodiments of this application, through the specific implementation methods described above, achieve the following technical effects: This application proposes a high-resolution remote sensing image change detection method based on reinforcement learning. First, high-resolution remote sensing images of the target area at the first and second time points are acquired, simultaneously obtaining imaging condition parameters including illumination and atmospheric parameters to ensure sufficient image detail and quantifiable environmental differences. Next, all land cover categories are extracted from historical remote sensing data of the target area to form a remote sensing content category set, clarifying the specific objects for subsequent identification and change analysis. Then, an error test is conducted on a pre-constructed remote sensing content recognizer. Imaging condition parameters and the remote sensing content category set are combined to generate remote sensing features. Historical images are indexed and matched to construct a test set, which is then input into the remote sensing content recognizer to calculate the accuracy and error rate, forming error rate sets for the two time points. Next, the average error rate is calculated based on the error rate sets. Reinforcement learning data is configured according to the principle of "the larger the error rate, the more data," and the remote sensing content recognizer is trained and optimized until the verification accuracy converges. The optimized remote sensing content recognizer is then used for content recognition in the two time point images, outputting the land cover distribution results and calculating the area change amplitude of each category, ultimately obtaining remote sensing content change information.

[0141] The method provided in this application, through the technical solution of "data acquisition - category classification - error testing - reinforcement learning optimization - calculation of change amplitude", solves the problems of low recognition accuracy and difficulty in adapting fixed models to complex scenes caused by differences in imaging conditions in traditional remote sensing image change detection. It improves the accuracy and reliability of high-resolution remote sensing image change detection and provides technical support for land planning, ecological assessment and other fields.

[0142] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0143] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0144] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. A high-resolution remote sensing image change detection method based on reinforcement learning, characterized in that, The method includes: Acquire first and second remote sensing images of the target area at the first and second time points, and obtain first and second imaging condition parameters at the first and second time points. Obtain all remote sensing content categories within the target area to obtain a set of remote sensing content categories; Based on the first imaging condition parameters and the second imaging condition parameters, and in conjunction with the remote sensing content category set, a remote sensing recognition error test is performed on the pre-constructed remote sensing content recognizer to obtain a first remote sensing recognition error rate set and a second remote sensing recognition error rate set. Based on the first remote sensing identification error rate set and the second remote sensing identification error rate set, reinforcement learning data is configured to perform reinforcement learning on the remote sensing content recognizer, and remote sensing content recognition is performed on the first remote sensing image and the second remote sensing image to obtain the first remote sensing content distribution and the second remote sensing content distribution, and remote sensing content change information is calculated.

2. The high-resolution remote sensing image change detection method based on reinforcement learning according to claim 1, characterized in that, Acquire first and second remote sensing images of the target area at the first and second time points, and obtain first and second imaging condition parameters at the first and second time points, including: The first and second remote sensing images of the target area were acquired at the first and second time points, respectively, with high-resolution remote sensing images being acquired. Acquire the first imaging condition parameters and the second imaging condition parameters at the first and second time points, wherein each imaging condition parameter includes illumination condition parameters and atmospheric condition parameters.

3. The high-resolution remote sensing image change detection method based on reinforcement learning according to claim 1, characterized in that, Obtain all remote sensing content categories within the target area to obtain a set of remote sensing content categories, including: Acquire historical remote sensing data of the target area within a preset time range; Extract all remote sensing content categories within the target area from the historical remote sensing data to obtain a set of remote sensing content categories.

4. The high-resolution remote sensing image change detection method based on reinforcement learning according to claim 1, characterized in that, The steps for pre-building the remote sensing content recognizer include: Based on remote sensing data, a set of sample remote sensing images is collected, and the content distribution within each remote sensing image is labeled to obtain a set of sample remote sensing content distributions. Each sample remote sensing content distribution includes remote sensing entity content with all coordinates within the remote sensing image. A network architecture for a remote sensing content recognizer is constructed based on deep learning. The remote sensing content recognizer is trained under supervised supervision using the sample remote sensing image set and the sample remote sensing content distribution set until the verification accuracy converges, thus completing the pre-construction.

5. The high-resolution remote sensing image change detection method based on reinforcement learning according to claim 1, characterized in that, Based on the first and second imaging condition parameters, and in conjunction with the remote sensing content category set, a remote sensing recognition error test is performed on the pre-constructed remote sensing content recognizer to obtain a first remote sensing recognition error rate set and a second remote sensing recognition error rate set, including: Based on the first imaging condition parameters, the second imaging condition parameters, and the remote sensing content category set, multiple first remote sensing features and multiple second remote sensing features are obtained by traversing and combining them. Each first remote sensing feature includes the first imaging condition parameters and any one remote sensing content category. Indexing is performed within the remote sensing data over a historical period based on multiple first remote sensing features and multiple second remote sensing features to obtain multiple first test remote sensing image sets and multiple second test remote sensing image sets. The test remote sensing images in multiple first test remote sensing image sets and multiple second test remote sensing image sets are respectively input into the remote sensing content recognizer to obtain multiple first test remote sensing content sets and multiple second test remote sensing content sets as recognition outputs; Calculate the proportion of test remote sensing content in each test remote sensing content set that matches the remote sensing content category in the corresponding remote sensing feature, obtain multiple first remote sensing recognition accuracies and multiple second remote sensing recognition accuracies, and calculate the first remote sensing recognition error rate set and the second remote sensing recognition error rate set.

6. The high-resolution remote sensing image change detection method based on reinforcement learning according to claim 1, characterized in that, Based on the first remote sensing identification error rate set and the second remote sensing identification error rate set, reinforcement learning data is configured, and reinforcement learning is performed on the remote sensing content recognizer, including: Based on the first set of remote sensing identification error rates and the second set of remote sensing identification error rates, the first average remote sensing identification error rate and the second average remote sensing identification error rate are calculated. Based on the first average remote sensing identification error rate and the second average remote sensing identification error rate, configure reinforcement learning data; The reinforcement learning data is used to perform reinforcement learning on the remote sensing content recognizer, and the reinforcement learning is completed after the verification accuracy converges.

7. The high-resolution remote sensing image change detection method based on reinforcement learning according to claim 6, characterized in that, Based on the first average remote sensing identification error rate and the second average remote sensing identification error rate, reinforcement learning data is configured, including: Obtain reinforcement learning data; Based on the first average remote sensing identification error rate and the second average remote sensing identification error rate, the reinforcement learning data volume is divided to obtain the first data volume and the second data volume. According to the first data volume and the second data volume, collect the first enhanced remote sensing image set and the second enhanced remote sensing image set under the first imaging condition parameters and the second imaging condition parameters, and collect the corresponding first enhanced remote sensing content distribution set and the second enhanced remote sensing content distribution set. The first enhanced remote sensing image set, the second enhanced remote sensing image set, the first enhanced remote sensing content distribution set, and the second enhanced remote sensing content distribution set are integrated into reinforcement learning data.

8. The high-resolution remote sensing image change detection method based on reinforcement learning according to claim 1, characterized in that, Remote sensing content identification is performed on the first and second remote sensing images to obtain the distribution of the first and second remote sensing content, and remote sensing content change information is calculated, including: The first and second remote sensing images are input into the remote sensing content recognizer that has completed reinforcement learning, and the first and second remote sensing content distributions are output respectively. Based on the first and second remote sensing content distributions, the area change rate of each remote sensing content category within the target area is calculated as remote sensing content change information.

Images