A multi-temporal super-resolution reconstruction method based on change prior guidance

Abstract

The application discloses a multi-temporal super-resolution reconstruction method based on change prior guidance, and belongs to the technical field of remote sensing image processing. The method first pre-processes multi-temporal remote sensing images to construct a standardized time series dataset; then extracts multi-temporal change information to generate a pixel-level change prior map; further, low-resolution image features and the change prior map are jointly modeled, a change perception attention mechanism is used to adaptively assign temporal weights, and a change region enhancement branch and a non-change region fusion branch are used for differential reconstruction; meanwhile, physical constraints such as spectral consistency and spatial autocorrelation are introduced to establish a joint optimization framework for reconstruction and change detection; and finally, high-resolution reconstruction results and a change detection map are output. The application effectively solves the change region reconstruction artifact problem and realizes high-precision and high-consistency multi-temporal remote sensing image super-resolution reconstruction.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image processing, specifically relating to a multi-temporal super-resolution reconstruction method based on change prior guidance. Background Technology

[0002] With the rapid development of remote sensing technology, high-resolution remote sensing images are playing an increasingly important role in resource surveys, environmental monitoring, disaster assessment, urban planning, and land change analysis. However, limited by the imaging mechanism of satellite remote sensing platforms and sensor hardware conditions, remote sensing observations often cannot simultaneously achieve high spatial resolution and high temporal resolution; there is an inherent trade-off between the two. To obtain observational data that combines rich spatial details and continuous temporal information, super-resolution reconstruction using multi-temporal low-resolution remote sensing image sequences has become an important research direction in the field of remote sensing image processing.

[0003] While existing multi-temporal super-resolution methods can extract complementary information between temporal images to some extent, most assume that images acquired at different times are essentially identical, lacking explicit modeling of truly changing regions. When the monitored scene contains changes in land cover, disaster disturbances, seasonal succession, or human activities, the information utilization methods of changed and unchanged regions differ significantly. If a uniform fusion strategy is still adopted, it is easy to cause artifacts, edge blurring, and texture distortion in changed regions, while also failing to fully utilize redundant complementary information in unchanged regions. In addition, some existing methods mainly focus on pixel-level reconstruction accuracy, ignoring the constraints of remote sensing images in terms of spectral consistency, spatial structure preservation, and accuracy of change representation. As a result, although the reconstruction results are visually improved, their application in subsequent change detection and land cover interpretation tasks remains limited. Therefore, there is an urgent need to propose a multi-temporal remote sensing image super-resolution reconstruction method that can incorporate prior change information, balance the enhancement of changed regions with stable fusion of unchanged regions, and perform joint optimization in conjunction with physical prior constraints. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-temporal super-resolution reconstruction method based on change prior guidance. This method can effectively solve the problems of obvious artifacts in the reconstruction of changed regions, insufficient utilization of complementary information in unchanged regions, and weak ability to restore the structure and texture details of the reconstructed image in the prior art.

[0005] To achieve the above objectives, this invention provides a multi-temporal super-resolution reconstruction method based on changing priors, comprising the following steps:

[0006] S1. Acquire multi-temporal low-resolution remote sensing image sequences, perform radiometric calibration, atmospheric correction, geometric correction and spatial registration on the image sequences, and construct a standardized time-series dataset;

[0007] S2. Construct a change detection module to perform temporal difference analysis on the standardized time series dataset, extract multi-temporal change information, and generate a pixel-level change prior map.

[0008] S3. Construct a change-guided super-resolution reconstruction module, jointly model the features of multi-temporal low-resolution images with the change prior map, adaptively allocate the feature weights of each temporal phase using a change-aware attention mechanism, and perform differentiated reconstruction of different regions using a change region enhancement branch and an unchanged region fusion branch.

[0009] S4. Construct a physical prior constraint and multi-task collaborative optimization module. In the reconstruction process, introduce constraints such as spectral consistency, spatial autocorrelation, change consistency and texture authenticity to establish a joint optimization framework for the super-resolution reconstruction task and change detection task. Use paired low-resolution-high-resolution image data to supervise the training of the model and optimize the network parameters.

[0010] S5. Input the multi-temporal low-resolution remote sensing images to be reconstructed into the trained model, and output the high-resolution reconstruction results and change detection map.

[0011] Furthermore, the specific steps in S1 for acquiring multi-temporal low-resolution image sequences, performing radiometric calibration, atmospheric correction, and spatial alignment to construct a standardized time-series dataset include:

[0012] S1.1 Acquire multiple low-resolution remote sensing images of the same area at different time phases, and use sensor calibration parameters to convert the original observation values ​​into radiance values ​​or apparent reflectance.

[0013] S1.2 Perform atmospheric correction, denoising, and geometric fine correction on multiple frames of low-resolution remote sensing images to eliminate the effects of atmospheric disturbance, sensor noise, and imaging geometric differences during the imaging process.

[0014] S1.3. Using geographic reference information, control point information, or resampling methods, images from different time phases are unified to the same geographic coordinate system and spatial resolution to achieve pixel-level or sub-pixel-level spatial alignment and construct a standardized time-series dataset.

[0015] Furthermore, in S2, a change detection module is constructed to perform temporal difference analysis on the standardized time-series dataset, extract multi-temporal change information, and generate a pixel-level change prior map, specifically including the following:

[0016] S2.1 Multi-temporal feature extraction: Representing multi-temporal low-resolution remote sensing image sequences as... The low-resolution images from each time phase are then input into a feature encoder with shared parameters to obtain the corresponding deep feature representations.

[0017]

[0018]

[0019] in, This represents the low-resolution input image at time phase t; T represents the total number of time phases. For feature encoder; For the set of trainable parameters of the feature encoder; This represents the deep feature map extracted at time phase t; This is a feature mapping operator used to standardize and map deep features; This represents the t-th phase feature after standardization;

[0020] S2.2, Calculation of phase difference response, based on the characteristics of reference phase r. Using the reference time phase as a baseline, calculate the difference response between the remaining time phases and the reference time phase, and generate a single-phase change response diagram:

[0021]

[0022]

[0023] in, Indicates the reference phase index; Indicates the first The absolute difference characteristics of each time phase relative to the reference time phase r; This indicates element-wise absolute value operation; Concat indicates concatenation and stacking along the channel dimension; A is a change response prediction network; Use the Sigmoid activation function; This represents the probability diagram of the change response at time t.

[0024] S2.3 Generation of Change Prior Map: Aggregate the change response maps of all non-reference time phases to obtain a continuous change prior map, and further generate a binary change prior map:

[0025]

[0026] }

[0027] Where P represents a continuous change prior map, used to characterize the intensity of change at each pixel position; Represents the binary transformation prior map; (x,y) represents the spatial coordinates of the image. Indicates the threshold for judging changes; when When, it indicates that the position (x, y) belongs to the region of change; when When the position (x, y) is within the unchanged region, it indicates that the position is within the unchanged region.

[0028] Furthermore, in S3, a change-guided super-resolution reconstruction module is constructed, which jointly models the features of multi-temporal low-resolution images with the change prior map, adaptively allocates the feature weights of each temporal phase using a change-aware attention mechanism, and performs differentiated reconstruction of different regions using a change region enhancement branch and an unchanged region fusion branch. Specifically, this includes the following:

[0029] S3.1 Feature recalibration guided by change prior: Multi-temporal features are modulated using a continuous change prior map P to obtain change-guided features, and global representations are extracted.

[0030]

[0031] in, This represents the t-th phase characteristic after being guided by the changed prior; This represents element-wise multiplication; Indicates the change in prior modulation coefficients;

[0032] S3.2 Calculation of change-aware attention weights: Map global features to temporal importance scores and normalize them to obtain change-aware attention weights for each temporal phase.

[0033]

[0034] in, This represents the original score for the t-th time phase; This represents a trainable weight matrix; Indicates the bias term;

[0035] S3.3 Construction of the Change Region Enhancement Branch: Weighted aggregation of the change region portion is performed to obtain the change region aggregated features, which are then input into the change region enhancement branch.

[0036]

[0037]

[0038] in, Indicates the aggregation characteristics of the change region; Indicates the enhanced branching in the region of change; This represents the reconstructed features of the changed region; the changed region enhancement branch is used to strengthen the information expression ability of edge, texture and structurally abrupt regions.

[0039] S3.4 Construction of the Unchanged Region Fusion Branch: Weighted aggregation is performed on the unchanged region portions to obtain the aggregated features of the unchanged regions, which are then input into the unchanged region fusion branch.

[0040]

[0041] in, This indicates the aggregation characteristics of unchanged regions; A complement to a priori diagram representing continuous changes;

[0042] S3.5, Dual-branch fusion and initial super-resolution reconstruction: This method fuses the reconstruction features of changed regions with those of unchanged regions, and obtains the initial high-resolution reconstruction results of the reference time phase through the upsampling reconstruction module.

[0043]

[0044]

[0045] in, Indicates the dual-branch fusion feature; This indicates the upsampling reconstruction module; This represents the set of trainable parameters for the upsampling reconstruction module; Indicates reference phase The initial high-resolution reconstruction results.

[0046] Furthermore, in S4, a physical prior constraint and multi-task collaborative optimization module is constructed. During the reconstruction process, constraints such as spectral consistency, spatial autocorrelation, change consistency, and texture realism are introduced to establish a joint optimization framework for the super-resolution reconstruction task and the change detection task. The model is then trained under supervision using paired low-resolution and high-resolution image data to optimize network parameters. Specifically, the following steps are included:

[0047] S4.1 Spectral Consistency Constraint Construction: By calculating the pixel-level differences between the reconstructed result and the high-resolution real image, a spectral consistency loss is constructed.

[0048]

[0049] in, This indicates a loss of spectral uniformity. The corresponding high-resolution real image when referring to a reference; This represents the corresponding high-resolution reconstruction result when referring to the reference. Denotes the L1 norm; Represents the total number of elements in the high-resolution image; spectral consistency loss is used to constrain the reconstruction result to maintain consistency with the radiometric response of the real image in each band.

[0050] S4.2 Construction of Change Consistency Constraints: The predicted change map is output by the change detection head and compared with the actual change annotation map to construct the change consistency loss.

[0051]

[0052] in, This represents a probability diagram of predicted changes. Indicates a change detection head; This represents the set of trainable parameters for the change detection head.

[0053] S4.3, Texture Authenticity Constraint Construction: By comparing the spatial gradients of the reconstructed image and the real image, a texture authenticity loss is constructed.

[0054]

[0055] in, This indicates a loss of texture realism; Represents the spatial gradient operator; A gradient map representing the reconstructed image; The gradient map represents the real image; texture realism loss is used to enhance the high-frequency texture details and edge realism of the reconstructed image. Then, the spectral consistency loss, variation consistency loss and texture realism loss are weighted and summed to construct the total loss function, and the model parameters are updated using gradient descent.

[0056] Furthermore, in step S5, the multi-temporal low-resolution remote sensing images to be reconstructed are input into the trained model, and high-resolution reconstruction results and change detection maps are output. Specific steps include:

[0057] S5.1 Model Inference Output: Input the multi-temporal low-resolution image sequence to be reconstructed into the trained overall model to obtain the high-resolution reconstruction result and the change probability map:

[0058]

[0059] in, This represents the overall model mapping after training is complete; This indicates the output of the high-resolution reconstruction result of the reference time phase; This represents a probability plot of the output changes.

[0060] S5.2 Change detection map generation: The change probability map is binarized according to a preset threshold to obtain the final change detection map.

[0061]

[0062] in, This represents the final change detection map; Indicates the threshold for change detection; when When, it indicates that the position (x, y) is determined to be a region of change; when When the position (x, y) is determined to be an unchanged region, the final output is a high-resolution reconstruction result and a change detection map.

[0063] Beneficial Effects: This invention effectively improves the overall performance of multi-temporal remote sensing image super-resolution reconstruction by introducing a change prior guidance mechanism, a multi-temporal feature adaptive fusion strategy, and a physical prior constraint joint optimization method. By constructing a change detection module and generating a pixel-level change prior map, explicit distinction between changed and unchanged regions is achieved, significantly improving the problems of change region artifacts, edge blurring, and detail distortion caused by insufficient modeling of temporal differences in existing methods. By setting up a change region enhancement branch and an unchanged region fusion branch, differentiated reconstruction strategies are adopted for different regions, which not only enhances the structural expression and texture restoration capabilities of changed regions but also fully explores the redundant and complementary information of unchanged regions in multi-temporal observations, improving the reconstruction consistency between background and steady-state regions. By introducing constraints such as spectral consistency, spatial autocorrelation, change consistency, and texture realism, the reliability of reconstruction results in terms of spectral response, spatial structure, and change expression is strengthened, avoiding the over-smoothing problem caused by relying solely on pixel regression. By establishing a multi-task collaborative optimization framework for super-resolution reconstruction and change detection, mutual promotion between the two types of tasks is achieved, further improving the model's reconstruction accuracy, detail fidelity, and change recognition capabilities. Compared with existing technologies, this invention can obtain clearer, more realistic and more suitable high-resolution reconstruction results for remote sensing applications in complex scenarios, and has good application value in fields such as resource surveys, environmental monitoring, disaster assessment, urban planning and land change analysis. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0065] Figure 2 This is a flowchart for generating the a priori diagram of the present invention.

[0066] Figure 3 This is a flowchart of the super-resolution reconstruction process guided by the changes of this invention;

[0067] Figure 4 This is a flowchart of the physical prior constraints and multi-task joint optimization process of the present invention. Detailed Implementation

[0068] The invention will now be further described with reference to the accompanying drawings.

[0069] Example

[0070] Furthermore, such as Figure 1 As shown, a multi-temporal super-resolution reconstruction method based on change prior guidance includes the following steps:

[0071] S1. Acquire multi-temporal low-resolution remote sensing image sequences, perform radiometric calibration, atmospheric correction, geometric correction and spatial registration on the image sequences, and construct a standardized time-series dataset;

[0072] S2. Construct a change detection module to perform temporal difference analysis on the standardized time series dataset, extract multi-temporal change information, and generate a pixel-level change prior map.

[0073] S3. Construct a change-guided super-resolution reconstruction module, jointly model the features of multi-temporal low-resolution images with the change prior map, adaptively allocate the feature weights of each temporal phase using a change-aware attention mechanism, and perform differentiated reconstruction of different regions using a change region enhancement branch and an unchanged region fusion branch.

[0074] S4. Construct a physical prior constraint and multi-task collaborative optimization module. In the reconstruction process, introduce constraints such as spectral consistency, spatial autocorrelation, change consistency and texture authenticity to establish a joint optimization framework for the super-resolution reconstruction task and change detection task. Use paired low-resolution-high-resolution image data to supervise the training of the model and optimize the network parameters.

[0075] S5. Input the multi-temporal low-resolution remote sensing images to be reconstructed into the trained model, and output the high-resolution reconstruction results and change detection map.

[0076] Furthermore, such as Figure 2 As shown, the specific steps in S1 for acquiring multi-temporal low-resolution image sequences, performing radiometric calibration, atmospheric correction, and spatial alignment, and constructing a standardized time-series dataset include:

[0077] S1.1 Acquire multiple low-resolution remote sensing images of the same area at different time phases, and use sensor calibration parameters to convert the original observation values ​​into radiance values ​​or apparent reflectance.

[0078] S1.2 Perform atmospheric correction, denoising, and geometric fine correction on multiple frames of low-resolution remote sensing images to eliminate the effects of atmospheric disturbance, sensor noise, and imaging geometric differences during the imaging process.

[0079] S1.3. Using geographic reference information, control point information, or resampling methods, images from different time phases are unified to the same geographic coordinate system and spatial resolution to achieve pixel-level or sub-pixel-level spatial alignment and construct a standardized time-series dataset.

[0080] Furthermore, such as Figure 2 As shown, the change detection module constructed in S2 performs temporal difference analysis on the standardized time-series dataset, extracts multi-temporal change information, and generates a pixel-level change prior map, specifically including the following:

[0081] S2.1 Multi-temporal feature extraction: Representing multi-temporal low-resolution remote sensing image sequences as... The low-resolution images from each time phase are then input into a feature encoder with shared parameters to obtain the corresponding deep feature representations.

[0082]

[0083]

[0084] in, This represents the low-resolution input image at time phase t; T represents the total number of time phases. For feature encoder; For the set of trainable parameters of the feature encoder; This represents the deep feature map extracted at time phase t; This is a feature mapping operator used to standardize and map deep features; This represents the t-th phase feature after standardization;

[0085] S2.2, Calculation of phase difference response, based on the characteristics of reference phase r. Using the reference time phase as a baseline, calculate the difference response between the remaining time phases and the reference time phase, and generate a single-phase change response diagram:

[0086]

[0087]

[0088] Where r represents the reference phase index; This represents the absolute difference characteristic of the t-th time phase relative to the reference time phase r; This indicates element-wise absolute value operation; Concat indicates concatenation and stacking along the channel dimension; A is a change response prediction network; Use the Sigmoid activation function; This represents the probability diagram of the change response at time t.

[0089] S2.3 Generation of Change Prior Map: Aggregate the change response maps of all non-reference time phases to obtain a continuous change prior map, and further generate a binary change prior map:

[0090]

[0091]

[0092] Where P represents a continuous change prior map, used to characterize the intensity of change at each pixel position; Represents the binary transformation prior map; (x,y) represents the spatial coordinates of the image. Indicates the threshold for judging changes; when When, it indicates that the position (x, y) belongs to the region of change; when When the position (x, y) is within the unchanged region, it indicates that the position is within the unchanged region.

[0093] Furthermore, such as Figure 3 As shown, the change-guided super-resolution reconstruction module in S3 jointly models multi-temporal low-resolution image features with change prior maps, adaptively allocates the feature weights of each temporal phase using a change-aware attention mechanism, and performs differentiated reconstruction of different regions using a change region enhancement branch and an unchanged region fusion branch. Specifically, it includes the following:

[0094] S3.1 Feature recalibration guided by change prior: Multi-temporal features are modulated using a continuous change prior map P to obtain change-guided features, and global representations are extracted.

[0095]

[0096] in, This represents the t-th phase characteristic after being guided by the changed prior; This represents element-wise multiplication; Indicates the change in prior modulation coefficients;

[0097] S3.2 Calculation of change-aware attention weights: Map global features to temporal importance scores and normalize them to obtain change-aware attention weights for each temporal phase.

[0098]

[0099] in, This represents the original score for the t-th time phase; This represents a trainable weight matrix; Indicates the bias term;

[0100] S3.3 Construction of the Change Region Enhancement Branch: Weighted aggregation of the change region portion is performed to obtain the change region aggregated features, which are then input into the change region enhancement branch.

[0101]

[0102]

[0103] in, Indicates the aggregation characteristics of the change region; Indicates the enhanced branching in the region of change; This represents the reconstructed features of the changed region; the changed region enhancement branch is used to strengthen the information expression ability of edge, texture and structurally abrupt regions.

[0104] S3.4 Construction of the Unchanged Region Fusion Branch: Weighted aggregation is performed on the unchanged region portions to obtain the aggregated features of the unchanged regions, which are then input into the unchanged region fusion branch.

[0105]

[0106] in, This indicates the aggregation characteristics of unchanged regions; A complement to a priori diagram representing continuous changes; Indicates the merging of unchanged regions into branches;

[0107] S3.5, Dual-branch fusion and initial super-resolution reconstruction: This method fuses the reconstruction features of changed regions with those of unchanged regions, and obtains the initial high-resolution reconstruction results of the reference time phase through the upsampling reconstruction module.

[0108]

[0109]

[0110] in, Indicates the dual-branch fusion feature; This indicates the upsampling reconstruction module; This represents the set of trainable parameters for the upsampling reconstruction module; This represents the initial high-resolution reconstruction result for the reference phase r.

[0111] Furthermore, such as Figure 4 As shown, in step S4, a physical prior constraint and multi-task collaborative optimization module is constructed. During the reconstruction process, constraints such as spectral consistency, spatial autocorrelation, change consistency, and texture realism are introduced to establish a joint optimization framework for the super-resolution reconstruction task and the change detection task. Paired low-resolution and high-resolution image data are used to supervise the training of the model and optimize the network parameters. Specifically, the steps include:

[0112] S4.1 Spectral Consistency Constraint Construction: By calculating the pixel-level differences between the reconstructed result and the high-resolution real image, a spectral consistency loss is constructed.

[0113]

[0114] in, This indicates a loss of spectral uniformity. The corresponding high-resolution real image when referring to a reference; This represents the corresponding high-resolution reconstruction result when referring to the reference. Denotes the L1 norm; Represents the total number of elements in the high-resolution image; spectral consistency loss is used to constrain the reconstruction result to maintain consistency with the radiometric response of the real image in each band.

[0115] S4.2 Construction of Change Consistency Constraints: The predicted change map is output by the change detection head and compared with the actual change annotation map to construct the change consistency loss.

[0116]

[0117] in, This represents a probability diagram of predicted changes. Indicates a change detection head; This represents the set of trainable parameters for the change detection head.

[0118] S4.3, Texture Authenticity Constraint Construction: By comparing the spatial gradients of the reconstructed image and the real image, a texture authenticity loss is constructed.

[0119]

[0120] in, This indicates a loss of texture realism; Represents the spatial gradient operator; A gradient map representing the reconstructed image; The gradient map represents the real image; texture realism loss is used to enhance the high-frequency texture details and edge realism of the reconstructed image. Then, the spectral consistency loss, variation consistency loss and texture realism loss are weighted and summed to construct the total loss function, and the model parameters are updated using gradient descent.

[0121] Furthermore, such as Figure 4 As shown, in step S5, the multi-temporal low-resolution remote sensing images to be reconstructed are input into the trained model, and the high-resolution reconstruction results and change detection maps are output. Specific steps include:

[0122] S5.1 Model Inference Output: Input the multi-temporal low-resolution image sequence to be reconstructed into the trained overall model to obtain the high-resolution reconstruction result and the change probability map:

[0123]

[0124] in, This represents the overall model mapping after training is complete; This indicates the output of the high-resolution reconstruction result of the reference time phase; This represents a probability plot of the output changes.

[0125] S5.2 Change detection map generation: The change probability map is binarized according to a preset threshold to obtain the final change detection map.

[0126]

[0127] in, This represents the final change detection map; Indicates the threshold for change detection; when When, it indicates that the position (x, y) is determined to be a region of change; when When the position (x, y) is determined to be an unchanged region, the final output is a high-resolution reconstruction result and a change detection map.

[0128] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. The scope of protection of the present invention should be determined by the scope of protection of the appended claims.

Claims

1. A multi-temporal super-resolution reconstruction method based on changing priors, characterized in that, Includes the following steps: S1. Acquire multi-temporal low-resolution remote sensing image sequences, perform radiometric calibration, atmospheric correction, geometric correction and spatial registration on the image sequences, and construct a standardized time-series dataset; S2. Construct a change detection module to perform temporal difference analysis on the standardized time series dataset, extract multi-temporal change information, and generate a pixel-level change prior map. S3. Construct a change-guided super-resolution reconstruction module, jointly model the features of multi-temporal low-resolution images with the change prior map, adaptively allocate the feature weights of each temporal phase using a change-aware attention mechanism, and perform differentiated reconstruction of different regions using a change region enhancement branch and an unchanged region fusion branch. S4. Construct a physical prior constraint and multi-task collaborative optimization module. In the reconstruction process, introduce constraints such as spectral consistency, spatial autocorrelation, change consistency and texture authenticity to establish a joint optimization framework for the super-resolution reconstruction task and change detection task. Use paired low-resolution-high-resolution image data to supervise the training of the model and optimize the network parameters. S5. Input the multi-temporal low-resolution remote sensing images to be reconstructed into the trained model, and output the high-resolution reconstruction results and change detection map.

2. The multi-temporal super-resolution reconstruction method based on changing priors as described in claim 1, characterized in that, S1 includes the following steps: S1.1 Acquire multiple low-resolution remote sensing images of the same area at different time phases, and use sensor calibration parameters to convert the original observation values ​​into radiance values ​​or apparent reflectance. S1.2 Perform atmospheric correction, denoising, and geometric fine correction on multiple frames of low-resolution remote sensing images to eliminate the effects of atmospheric disturbance, sensor noise, and imaging geometric differences during the imaging process. S1.

3. Using geographic reference information, control point information, or resampling methods, images from different time phases are unified to the same geographic coordinate system and spatial resolution to achieve pixel-level or sub-pixel-level spatial alignment and construct a standardized time-series dataset.

3. The multi-temporal super-resolution reconstruction method based on changing priors as described in claim 1, characterized in that, S2 includes the following steps: S2.1 Multi-temporal feature extraction: Representing multi-temporal low-resolution remote sensing image sequences as... The low-resolution images from each time phase are then input into a feature encoder with shared parameters to obtain the corresponding deep feature representations. in, This represents the low-resolution input image at time phase t; T represents the total number of time phases. For feature encoder; For the set of trainable parameters of the feature encoder; This represents the deep feature map extracted at time phase t; This is a feature mapping operator used to standardize and map deep features; This represents the t-th phase feature after standardization; S2.2, Calculation of phase difference response, based on the characteristics of reference phase r. Using the reference time phase as a baseline, calculate the difference response between the remaining time phases and the reference time phase, and generate a single-phase change response diagram: Where r represents the reference phase index; This represents the absolute difference characteristic of the t-th time phase relative to the reference time phase r; This represents element-wise absolute value operation; This indicates that the network is spliced ​​and stacked along the channel dimension; A represents the change response prediction network. for Activation function; This represents the probability diagram of the change response at time t. S2.3 Generation of Change Prior Map: Aggregate the change response maps of all non-reference time phases to obtain a continuous change prior map, and further generate a binary change prior map: Where P represents a continuous change prior map, used to characterize the intensity of change at each pixel position; Represents the binary transformation prior map; (x,y) represents the spatial coordinates of the image. Indicates the threshold for judging changes; when When, it indicates that the position (x, y) belongs to the region of change; when When the position (x, y) is within the unchanged region, it indicates that the position is within the unchanged region.

4. The multi-temporal super-resolution reconstruction method based on changing priors as described in claim 1, characterized in that, S3 includes the following steps: S3.1 Feature recalibration guided by change prior: Multi-temporal features are modulated using a continuous change prior map P to obtain change-guided features, and global representations are extracted. in, This represents the t-th phase characteristic after being guided by the changed prior; This represents element-wise multiplication; Indicates the change in prior modulation coefficients; S3.2 Calculation of change-aware attention weights: Map global features to temporal importance scores and normalize them to obtain change-aware attention weights for each temporal phase. in, This represents the original score for the t-th time phase; This represents a trainable weight matrix; Indicates the bias term; S3.3 Construction of the Change Region Enhancement Branch: Weighted aggregation of the change region portion is performed to obtain the change region aggregated features, which are then input into the change region enhancement branch. in, Indicates the aggregation characteristics of the change region; Indicates the enhanced branching in the region of change; This represents the reconstructed features of the changed region; the changed region enhancement branch is used to strengthen the information expression ability of edge, texture and structurally abrupt regions. S3.4 Construction of the Unchanged Region Fusion Branch: Weighted aggregation is performed on the unchanged region portions to obtain the aggregated features of the unchanged regions, which are then input into the unchanged region fusion branch. in, This indicates the aggregation characteristics of unchanged regions; A complement to a priori diagram representing continuous changes; Indicates the merging of unchanged regions into branches; S3.5, Dual-branch fusion and initial super-resolution reconstruction: This method fuses the reconstruction features of changed regions with those of unchanged regions, and obtains the initial high-resolution reconstruction results of the reference time phase through the upsampling reconstruction module. in, Indicates the dual-branch fusion feature; This indicates the upsampling reconstruction module; This represents the set of trainable parameters for the upsampling reconstruction module; Indicates reference phase The initial high-resolution reconstruction results.

5. The multi-temporal super-resolution reconstruction method based on changing priors as described in claim 1, characterized in that, S4 includes the following steps: S4.1 Spectral Consistency Constraint Construction: By calculating the pixel-level differences between the reconstructed result and the high-resolution real image, a spectral consistency loss is constructed. in, This indicates a loss of spectral uniformity. The corresponding high-resolution real image when referring to a reference; This represents the corresponding high-resolution reconstruction result when referring to the reference. Denotes the L1 norm; Represents the total number of elements in the high-resolution image; spectral consistency loss is used to constrain the reconstruction result to maintain consistency with the radiometric response of the real image in each band. S4.2 Construction of Change Consistency Constraints: The predicted change map is output by the change detection head and compared with the actual change annotation map to construct the change consistency loss. in, This represents a probability diagram of predicted changes. Indicates a change detection head; This represents the set of trainable parameters for the change detection head. S4.3, Texture Authenticity Constraint Construction: By comparing the spatial gradients of the reconstructed image and the real image, a texture authenticity loss is constructed. in, This indicates a loss of texture realism; Represents the spatial gradient operator; A gradient map representing the reconstructed image; The gradient map represents the real image; texture realism loss is used to enhance the high-frequency texture details and edge realism of the reconstructed image. Then, the spectral consistency loss, variation consistency loss and texture realism loss are weighted and summed to construct the total loss function, and the model parameters are updated using gradient descent.

6. The multi-temporal super-resolution reconstruction method based on changing priors as described in claim 1, characterized in that, S5 includes the following steps: S5.1 Model Inference Output: Input the multi-temporal low-resolution image sequence to be reconstructed into the trained overall model to obtain the high-resolution reconstruction result and the change probability map: in, This represents the overall model mapping after training is complete; This indicates the output of the high-resolution reconstruction result of the reference time phase; This represents a probability plot of the output changes. S5.2 Change detection map generation: The change probability map is binarized according to a preset threshold to obtain the final change detection map. in, This represents the final change detection map; Indicates the threshold for change detection; when When, it indicates that the position (x, y) is determined to be a region of change; when When the position (x, y) is determined to be an unchanged region, the final output is a high-resolution reconstruction result and a change detection map.

Images