A method and system for creating a multi-temporal remote sensing image database

Abstract

The application discloses a kind of multi-temporal remote sensing image database creation method and system, and the application relates to remote sensing image data processing technical field, comprising the following steps: S1: receiving multi-temporal remote sensing image data as input;S2: each image is carried out preliminary quality assessment, and this assessment is by extracting the radiation statistical characteristics of image, cloud amount index and the spectral correlation with reference image to construct multidimensional quality feature vector.The multi-temporal remote sensing image database creation method and system solve the core technical problems that traditional multi-temporal remote sensing image database creation process relies on manual intervention, processing efficiency is low and radiation consistency is difficult to guarantee;Through the complete quality guarantee system established from preliminary assessment to final decision, the standardization, automation control of mass image data radiation quality is realized, and the artificial cost is reduced and the database construction efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image data processing technology, specifically to a method and system for creating a multi-temporal remote sensing image database. Background Technology

[0002] With the rapid development of Earth observation technology, the demand for multi-temporal remote sensing image data is increasing across various industries, making the construction of large-scale, high-quality multi-temporal remote sensing image databases an inevitable trend. However, existing database creation methods generally focus on basic functions such as efficient image storage, rapid indexing, and query management, neglecting a crucial aspect affecting the ultimate value of the data: how to automatically ensure the core quality of radiometric consistency among the images imported into the database. Current practices typically rely on cumbersome manual preprocessing before database construction or the default use of calibrated data sources. This not only makes the entire process inefficient and heavily reliant on individual expert experience but also results in inherent radiometric differences at the underlying layer of the constructed database. This noise, which is not a reflection of actual changes in ground features, is introduced into the database and severely misleads subsequent advanced applications such as change detection and time series analysis. These problems are often only discovered after the database is put into use, causing huge rework costs and analytical biases. Therefore, the current technical system lacks an embedded, intelligent process to automatically identify and handle radiometric inconsistencies, which has become a major obstacle to the automated construction of high-quality multi-temporal remote sensing image databases. The present invention aims to solve the practical problem that existing methods cannot intelligently ensure the radiometric consistency of images entering the database during the automated creation of multi-temporal remote sensing image databases. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for creating a multi-temporal remote sensing image database, so as to solve the problems mentioned in the background art.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for creating a multi-temporal remote sensing image database, comprising the following steps:

[0005] S1: Receive multi-temporal remote sensing image data as input;

[0006] S2: Perform a preliminary quality assessment on each image. This assessment constructs a multidimensional quality feature vector by extracting the radiometric statistics, cloud cover index and spectral correlation with the baseline image of the image, and uses a pre-trained lightweight classifier to analyze the vector to output an initial consistency score.

[0007] S3: Based on the initial consistency score, intelligent routing is performed to divide the images into high-confidence pass stream, pending stream, and high-confidence failure stream. The images in the high-confidence pass stream directly enter the secondary verification queue, the images in the pending stream flow into the policy execution engine, and the images in the high-confidence failure stream are marked as unqualified and logged.

[0008] S4: For the images in the processing stream, the policy execution engine selects a radiometric normalization algorithm for initial processing to generate the first generation of processed data. This selection is based on the specific composition of the multidimensional quality feature vector and adapts one algorithm from a variety of radiometric normalization algorithms.

[0009] S5: Input the high-confidence through-stream imagery and the first-generation processed data into the secondary verification queue for secondary verification to obtain the secondary verification result. The secondary verification uses a deep verification model to calculate a new round of consistency score. This deep verification model is different from the primary quality assessment model and includes local verification on a preset stable ground feature sample area to ensure that the processed imagery and the reference imagery maintain spectral stability on typical ground features.

[0010] S6: Perform secondary routing based on the secondary verification results: Data that passes the secondary verification is approved for entry into the database, while data that fails the secondary verification is returned to the strategy execution engine along with a diagnostic report;

[0011] S7: For the data returned to the strategy execution engine, analyze the reasons for the failure based on the diagnostic report, and switch or combine another radiation normalization algorithm for secondary processing to generate the second-generation processed data;

[0012] S8: Submit the second-generation processed data to the final arbiter, which uses the physical spectral model to perform the final evaluation and outputs the final evaluation result. The final evaluation result is either pass or fail. Passed data is stored in the multi-temporal remote sensing image database, while failed data is marked as difficult data and the automated process is terminated.

[0013] The entire process utilizes a closed-loop mechanism of primary quality assessment, intelligent triage, strategy execution, secondary verification, and final decision-making to automate the creation of a multi-temporal remote sensing image database and ensure the radiometric consistency of the images entering the database.

[0014] Furthermore, the primary quality assessment extracts the radiometric statistical features of the images, including calculating the mean, variance, and histogram distribution of the images; extracts the cloud cover index of the images, including identifying cloud-covered areas and calculating the cloud coverage ratio through image segmentation technology; and extracts the spectral correlation between the images and the reference images, including calculating the correlation coefficient between the two images in the same band.

[0015] The lightweight classifier is a support vector machine or decision tree model, and its pre-training process uses historical multi-temporal remote sensing image data and its corresponding radiometric consistency labels for supervised learning; the initial consistency score is a numerical score used to quantify the degree of radiometric difference between the image and the reference image.

[0016] In the intelligent traffic splitting process, the criteria for a high-confidence flow to pass is that the initial consistency score is higher than the first threshold. The criteria for a flow to be processed is that the initial consistency score is between the first and second thresholds. The criteria for a high-confidence failed flow is that the initial consistency score is lower than the second threshold. The first and second thresholds are preset through historical data statistical analysis.

[0017] Furthermore, the process of the strategy execution engine selecting a radiation normalization algorithm includes: matching a predefined algorithm selection rule based on the radiation statistical features and cloud cover index in the multidimensional quality feature vector, which maps the feature vector to a specific algorithm; the various radiation normalization algorithms include histogram matching algorithm, relative radiation normalization algorithm and normalization algorithm based on pseudo-invariant features;

[0018] When selecting algorithms, the radiation normalization algorithm with low computational complexity is given priority for initial processing. After the initial processing generates the first generation of processed data, the system records the algorithm used and its parameters for reference during secondary processing. The strategy execution engine also has a built-in algorithm performance evaluation module, which updates the algorithm selection rules based on historical processing results to optimize subsequent processing efficiency.

[0019] Furthermore, the deep verification model used in the secondary verification is a convolutional neural network model or a random forest model, and the training data includes multi-temporal remote sensing images and their spectral stability labels on stable ground cover sample areas; the calculation of the new round of consistency score includes global radiometric statistics, as well as the calculation of local spectral deviations on stable ground cover sample areas.

[0020] The stable land cover sample area is a pre-selected area with stable land cover types, including water bodies, vegetation, and bare land. Local verification is performed by comparing the spectral reflectance values ​​of the processed image and the reference image in the sample area. If the deviation exceeds the preset tolerance value, it is determined that the second verification has failed. In the second diversion, the data that passes the second verification must meet the following requirements: the consistency score of the new round is higher than the third threshold, and the local spectral deviation is lower than the preset tolerance value. The diagnostic report attached to the data that fails the second verification includes the failure reason analysis, recommended processing algorithm, and parameter adjustment suggestions.

[0021] Furthermore, during the secondary processing, the strategy execution engine, based on the failure cause analysis in the diagnostic report, selects a radiation normalization algorithm with high computational complexity or combines multiple radiation normalization algorithms for processing; combining multiple radiation normalization algorithms includes first using a histogram matching algorithm for preprocessing, and then using a normalization algorithm based on pseudo-invariant features for adjustment; after the secondary processing generates the second-generation processed data, the processing effects of the first and second generations are compared, and the algorithm performance evaluation module is updated;

[0022] The physical spectral model used by the final arbiter is a model based on atmospheric radiative transfer theory. This model simulates the atmospheric conditions at the time of image acquisition and calculates the residual between the theoretical spectral value and the actual value. During the final evaluation, if the residual is lower than a preset threshold, it is judged as passing; otherwise, it is judged as failing. The evaluation process of the final arbiter is independent of other verification steps and does not rely on historical data.

[0023] Furthermore, the closed-loop mechanism also includes a feedback loop, in which the evaluation results of the final arbiter are used to optimize the lightweight classifier for primary quality assessment and the deep verification model for secondary verification; the optimization process is implemented through machine learning, using the data passed by the final arbiter as positive samples to update the model parameters; the automation of the entire process is coordinated by a workflow engine, which monitors the execution status of each step and triggers retries or alarms when timeouts or errors occur; the workflow engine also records the processing history of each image, including evaluation scores, algorithms used, and verification results, for subsequent auditing and analysis.

[0024] Furthermore, the storage structure of the multi-temporal remote sensing image database adopts a layered design, including a raw data layer, a processing intermediate data layer, and a finished data layer. The raw data layer stores the input multi-temporal remote sensing images; the processing intermediate data layer stores the initial quality assessment results, the first-generation processed data, and the second-generation processed data; the finished data layer stores the data passed by the final arbiter; each layer of data is associated with metadata, including acquisition time, sensor type, processing history, and radiometric consistency score; the database also provides a query interface to support retrieval based on time range, geographical region, and radiometric consistency level.

[0025] A system for creating a multi-temporal remote sensing image database, used to implement a method for creating a multi-temporal remote sensing image database, includes:

[0026] The data receiving module is used to receive multi-temporal remote sensing image data;

[0027] The primary quality assessment module is used to perform a primary quality assessment on each image. It constructs a multidimensional quality feature vector by extracting the radiometric statistical features, cloud cover index, and spectral correlation with the baseline image. A pre-trained lightweight classifier is then used to analyze the vector to output an initial consistency score.

[0028] The intelligent routing module is used to perform intelligent routing based on the initial consistency score, dividing the images into high-confidence pass streams, pending streams, and high-confidence failure streams. Images in the high-confidence pass stream directly enter the secondary verification queue, images in the pending stream flow into the strategy execution engine, and images in the high-confidence failure stream are marked as unqualified and logged.

[0029] The strategy execution engine module is used to select a radiation normalization algorithm for initial processing and generate the first generation of processed data. This selection is based on the specific composition of the multidimensional quality feature vector to adapt one algorithm from a variety of radiation normalization algorithms.

[0030] The secondary verification module inputs the high-confidence through-stream imagery and the first-generation processed data into the secondary verification queue for secondary verification to obtain the secondary verification result. The secondary verification uses a deep verification model to calculate a new round of consistency score. This deep verification model is different from the primary quality assessment model and includes local verification on a preset stable ground feature sample area to ensure that the processed imagery and the reference imagery maintain spectral stability on typical ground features.

[0031] The secondary routing module is used to perform secondary routing based on the secondary verification results: data that passes the secondary verification is approved and stored in the database, while data that fails the secondary verification is returned to the strategy execution engine along with a diagnostic report;

[0032] The secondary processing module is used to analyze the reasons for processing failures based on the diagnostic report of the data from the returned strategy execution engine, and switch or combine another radiation normalization algorithm for secondary processing to generate second-generation processed data.

[0033] The final arbiter module submits the second-generation processed data to the final arbiter, which uses a physical spectral model to perform a final evaluation and outputs the final evaluation result. The final evaluation result is either pass or fail. Data that fails is marked as difficult data and the automated process is terminated.

[0034] The storage module is used to store the data that has passed the final evaluation into a multi-temporal remote sensing image database;

[0035] The various modules of the system are connected via a bus or network to form a closed-loop processing pipeline.

[0036] Furthermore, the primary quality assessment module also includes a feature extraction unit and a classifier unit; the feature extraction unit is configured to compute radiation statistical features, cloud cover indicators, and spectral correlations in parallel; the classifier unit is periodically retrained using new data to maintain accuracy; the policy execution engine module includes an algorithm library unit and a selection unit; the algorithm library unit stores various radiation normalization algorithms; the selection unit selects an algorithm in real time based on multidimensional quality feature vectors; the secondary verification module includes a deep model unit and a local verification unit; the deep model unit loads a pre-trained deep verification model; the local verification unit accesses the geographic information system to obtain the coordinates of stable ground feature sample areas; the final arbiter module integrates physical spectral simulation software, which calculates theoretical spectral values ​​based on atmospheric parameters.

[0037] Furthermore, the system also includes a monitoring and management module, which displays the status of the processing pipeline in real time, including the input and output data volume, processing time, and error logs of each module. The monitoring and management module provides a graphical user interface for manually adjusting the diversion threshold, algorithm selection rules, and verification parameters. The system records all operations through a log database, supporting fault recovery and performance analysis. The storage module adopts a distributed file system, supporting multi-node parallel access to improve processing efficiency under large data volumes. The system integrates with external remote sensing data processing platforms through an application programming interface to achieve data sharing and collaborative processing.

[0038] This invention provides a method and system for creating a multi-temporal remote sensing image database. It has the following beneficial effects:

[0039] This method and system for creating a multi-temporal remote sensing image database effectively solves the core technical problems of traditional multi-temporal remote sensing image database creation, such as reliance on manual intervention, low processing efficiency, and difficulty in ensuring radiometric consistency, by constructing an automated processing workflow that includes multiple quality assessments, intelligent triage decisions, and closed-loop feedback optimization. Its established complete quality assurance system, from initial assessment to final decision-making, enables standardized and automated control of the radiometric quality of massive image data, reducing labor costs and improving database construction efficiency.

[0040] The method and system for creating this multi-temporal remote sensing image database ensures reliable radiometric consistency of the imported image data through dynamic algorithm selection, multi-layer verification mechanisms, and a continuously self-optimizing system design. This intelligent processing mode not only significantly improves the robustness and adaptability of the data processing workflow but also provides a high-quality data foundation for subsequent advanced applications such as remote sensing time-series analysis and change detection, demonstrating significant practical application value. Attached Figure Description

[0041] Figure 1This is a flowchart illustrating a method for creating a multi-temporal remote sensing image database according to the present invention.

[0042] Figure 2 This is a schematic diagram of the module interaction of a system for creating a multi-temporal remote sensing image database according to the present invention;

[0043] Figure 3 This is a quality assessment state transition diagram for a multi-temporal remote sensing image database creation system according to the present invention.

[0044] Figure 4 The algorithm selection decision tree is used for the creation system of a multi-temporal remote sensing image database according to the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Please see Figures 1 to 4 This invention provides a technical solution: a method for creating a multi-temporal remote sensing image database, comprising the following steps:

[0047] S1: Receive multi-temporal remote sensing image data as input;

[0048] S2: Perform a preliminary quality assessment on each image. This assessment constructs a multidimensional quality feature vector by extracting the radiometric statistics, cloud cover index and spectral correlation with the baseline image of the image, and uses a pre-trained lightweight classifier to analyze the vector to output an initial consistency score.

[0049] S3: Based on the initial consistency score, intelligent routing is performed to divide the images into high-confidence pass stream, pending stream, and high-confidence failure stream. The images in the high-confidence pass stream directly enter the secondary verification queue, the images in the pending stream flow into the policy execution engine, and the images in the high-confidence failure stream are marked as unqualified and logged.

[0050] S4: For the images in the processing stream, the policy execution engine selects a radiometric normalization algorithm for initial processing to generate the first generation of processed data. This selection is based on the specific composition of the multidimensional quality feature vector and adapts one algorithm from a variety of radiometric normalization algorithms.

[0051] S5: Input the high-confidence through-stream imagery and the first-generation processed data into the secondary verification queue for secondary verification to obtain the secondary verification result. The secondary verification uses a deep verification model to calculate a new round of consistency score. This deep verification model is different from the primary quality assessment model and includes local verification on a preset stable ground feature sample area to ensure that the processed imagery and the reference imagery maintain spectral stability on typical ground features.

[0052] S6: Perform secondary routing based on the secondary verification results: Data that passes the secondary verification is approved for entry into the database, while data that fails the secondary verification is returned to the strategy execution engine along with a diagnostic report;

[0053] S7: For the data returned to the strategy execution engine, analyze the reasons for the failure based on the diagnostic report, and switch or combine another radiation normalization algorithm for secondary processing to generate the second-generation processed data;

[0054] S8: Submit the second-generation processed data to the final arbiter, which uses the physical spectral model to perform the final evaluation and outputs the final evaluation result. The final evaluation result is either pass or fail. Passed data is stored in the multi-temporal remote sensing image database, while failed data is marked as difficult data and the automated process is terminated.

[0055] It should be further explained that, in the specific implementation process, the system first receives multi-temporal remote sensing image data from different times and different sensors as input. For each input image, the system performs a preliminary quality assessment. This step constructs a multi-dimensional quality feature vector by extracting the image's radiometric statistics, cloud cover index, and its spectral correlation with the reference image. This feature vector is then fed into a pre-trained lightweight classifier, which analyzes these features and outputs an initial consistency score. This score quantifies the degree of radiometric difference between the current image and the reference image.

[0056] Based on the initial consistency score, the system performs intelligent routing. When the score is higher than the preset first threshold, the image is classified into the high-confidence pass stream and directly enters the subsequent secondary verification queue; when the score is between the first and second thresholds, the image is marked as the pending stream and sent to the policy execution engine for further processing; and when the score is lower than the second threshold, the image is classified into the high-confidence failure stream, and the system will mark it as unqualified data and record detailed logs for subsequent review.

[0057] For the streaming imagery entering the strategy execution engine, the system dynamically selects the most suitable radiometric normalization algorithm from a variety of available algorithms for initial processing, based on the specific composition of its multidimensional quality feature vector. This selection process is based on predefined rules; for example, for radiometric differences caused by variations in atmospheric conditions, an atmospheric correction-based normalization algorithm might be selected; for problems caused by sensor differences, a sensor-specific correction algorithm might be chosen. After the initial processing, the first-generation processed data is generated.

[0058] Next, the system sends the original high-confidence throughstream imagery along with the first-generation processed data into a secondary verification queue. Here, the system employs a deep verification model for secondary verification. This model differs from the one used in the initial assessment, featuring a more complex structure and stronger discriminative capabilities. The secondary verification not only calculates a new round of consistency scores but also performs local checks on pre-selected stable feature sample areas. By comparing the spectral characteristics of the processed imagery with the baseline imagery in these sample areas, it ensures that the processed imagery maintains spectral stability for typical features.

[0059] Based on the results of the secondary verification, the system performs a secondary data splitting operation. Data that passes the secondary verification is approved for entry into the final database, while data that fails is sent back to the policy execution engine along with a detailed diagnostic report explaining the specific reasons for the failure and improvement suggestions.

[0060] For images that are returned, the strategy execution engine will adjust the processing strategy based on the analysis results of the diagnostic report. It may choose to switch to another radiometric normalization algorithm or combine multiple algorithms for secondary processing to generate second-generation processed data.

[0061] Finally, the system submits the second-generation processed data to the final arbiter. The final arbiter uses an evaluation method based on a physical spectral model, calculating the residual between the theoretical and actual spectral values ​​by simulating the physical conditions at the time of image acquisition. Only when the residual is below a preset strict threshold will the data be finally approved for storage in the multi-temporal remote sensing image database; otherwise, it will be marked as problematic data, the automated processing flow will be terminated, and the data will be reported to the system administrator.

[0062] The entire processing flow forms a complete closed-loop system through primary quality assessment, intelligent triage, strategy execution, secondary verification, and final decision-making. The system also establishes a feedback mechanism, using the final decision-making results to optimize the classifier for primary assessment and the deep model for secondary verification, continuously improving processing accuracy and efficiency through continuous learning. This multi-layered, multi-stage processing mechanism with feedback optimization ensures reliable radiometric consistency of the final image data, providing a high-quality data foundation for subsequent remote sensing data analysis applications.

[0063] The primary quality assessment extracts the radiometric statistical characteristics of images, including calculating the mean, variance, and histogram distribution of the images; extracts cloud cover indicators from images, including identifying cloud-covered areas and calculating the cloud coverage ratio through image segmentation techniques; and extracts the spectral correlation between images and reference images, including calculating the correlation coefficient between the two images in the same band.

[0064] The lightweight classifier is a support vector machine or decision tree model. Its pre-training process uses historical multi-temporal remote sensing image data and its corresponding radiometric consistency labels for supervised learning. The initial consistency score is a numerical score used to quantify the degree of radiometric difference between the image and the reference image.

[0065] In the intelligent traffic splitting process, the criteria for a high-confidence flow to pass is that the initial consistency score is higher than the first threshold. The criteria for a flow to be processed is that the initial consistency score is between the first and second thresholds. The criteria for a high-confidence failed flow is that the initial consistency score is lower than the second threshold. The first and second thresholds are preset through historical data statistical analysis.

[0066] It should be further explained that for each input remote sensing image, the system first performs radiometric statistical feature extraction. This involves calculating the mean pixel value for each band to reflect the overall brightness level, calculating the variance of pixel values ​​to characterize the image's contrast, and generating histograms of each band to analyze the concentration and dispersion trends of pixel values. Regarding cloud cover extraction, the system uses image segmentation technology to divide the image into cloud and non-cloud areas, determining the cloud coverage ratio by calculating the ratio of cloud area pixels to the total number of pixels. In the spectral correlation analysis with the reference image, the system performs pixel-level matching between the image to be evaluated and a rigorously selected reference image in corresponding bands, and then calculates the correlation coefficient of their spectral response values ​​within the same geographical area.

[0067] The three dimensions of features mentioned above together constitute a complete multidimensional quality feature vector. The lightweight classifier used in this system can be either a support vector machine model or a decision tree model. Its pre-training process involves collecting a large amount of historical multi-temporal remote sensing image data, supplemented by manually labeled radiometric consistency tags as supervision signals, and then iteratively training. The trained classifier can analyze the input multidimensional quality feature vector and output a quantified initial consistency score, which objectively reflects the degree of radiometric difference between the current image and the reference image.

[0068] It should be further explained that, in the initial quality assessment stage, the system extracts three dimensions of features for each image, including the following:

[0069] Radiation statistics include calculating the mean μ and variance σ² for each band:

[0070] Where μ represents the arithmetic mean of all pixel values ​​in a certain band of the image, reflecting the overall brightness level of that band; x i : Represents the value of the i-th pixel in the image in this band; N: Represents the total number of pixels in the image in this band.

[0071] ;

[0072] Where σ² represents the variance of all pixel values ​​in a certain band of the image, reflecting the degree of dispersion of pixel values ​​relative to the mean, i.e., the contrast of the image; x i : Represents the value of the i-th pixel in the image in this band; μ: Represents the arithmetic mean of all pixel values ​​in a certain band of the image, reflecting the overall brightness level of that band; N: Represents the total number of pixels in the image in that band.

[0073] The cloud cover index identifies cloud-covered areas and calculates the cloud coverage ratio using image segmentation technology. , where P cloud N represents the cloud coverage ratio. cloud N represents the number of pixels in the cloud area. total This represents the total number of pixels.

[0074] The spectral correlation is calculated using the Pearson correlation coefficient formula:

[0075] ;

[0076] Where r represents the Pearson correlation coefficient, ranging from -1 to 1, used to measure the spectral similarity between the image to be evaluated and the reference image within the same geographic area, and x i : Represents the value of the i-th pixel in the image within this band, i.e., the value of the image to be evaluated at the i-th pixel position; y i : Represents the value of the reference image at the corresponding i-th pixel position; μ x : Represents the mean value of all pixels involved in the calculation of the image to be evaluated; μ y : Represents the average value of all pixels involved in the calculation of the reference image.

[0077] The multidimensional feature vector formed by these features is input into a pre-trained lightweight classifier, which employs a support vector machine model and has the following decision function:

[0078] ;

[0079] Where, f(x): represents the decision function of the classifier, whose output determines the category of the input sample x; sign(...): the sign function, which returns +1 or -1 depending on the sign of the result calculated within the parentheses, representing different categories; α i: represents the Lagrange multipliers, which are parameters obtained during model training; y i : Represents the true class label of training sample i; K(x) i ,x): Represents the kernel function used to compute the kernel function of the training sample x. i b represents the similarity between the input sample x and the input sample x in the feature space; b is the bias term. The classifier outputs the initial consistency score S. initial The value ranges from 0 to 1.

[0080] Based on the initial consistency score, the system performs intelligent triage. This process sets two key thresholds: a first threshold and a second threshold, which are pre-defined after statistical analysis of a large amount of historical data. In setting the intelligent triage threshold corresponding to the initial consistency score, the sample size of the historical data must include at least 1000 multi-temporal remote sensing images covering different sensor types, shooting times, geographical regions, and atmospheric conditions to ensure the reliability of the statistical results. During the statistical analysis, it is assumed that the historical consistency score follows a normal distribution, and the KS test (Kolmogorov-Smirnov test) is used. (test) Verify the distribution fit. When the p-value corresponding to the test statistic is greater than 0.05, it is confirmed that the normal distribution hypothesis is met. The adjustment coefficients k1 and k2 are in the range of 0.8-1.2 and 0.6-1.0, respectively. The specific values ​​can be adjusted according to the strictness of the actual application scenario. For example, in scientific research scenarios with high data quality requirements, k1=1.2 and k2=0.6 can be used. In engineering scenarios with priority to processing efficiency, k1=0.8 and k2=1.0 can be used. The optimization of k1 and k2 is achieved through 5-fold cross-validation. That is, by dividing the historical data into 5 groups, the threshold model is trained with 4 groups of data and validated with 1 group of data in turn. Finally, the coefficient combination that makes the diversion accuracy the highest is selected. When the initial consistency score of an image exceeds the first threshold, the system classifies it as a high-quality image, places it in the high-confidence pass stream, and directly enters the subsequent processing stage. When the score is between the first and second thresholds, the system marks it as a pending stream and sends it to the policy execution engine for further processing. When the score is below the second threshold, it is placed in the high-confidence failure stream, and the system marks it as unqualified data and records detailed log information, including image metadata, evaluation score, and reason for failure, for subsequent manual review or system optimization.

[0081] It should be further explained that, based on the initial consistency score, the system performs intelligent traffic routing. Two thresholds T are set. high and T low When S initial >T high When T, the image enters the high-confidence pass stream; when T low ≤S initial ≤T highWhen S enters the pending stream; initial <T low At that time, it enters the high-confidence failure stream. These thresholds are determined through statistical analysis of historical data, and the specific calculation method is as follows:

[0082] ;

[0083] ;

[0084] Among them, T high : Indicates the high threshold used for triage; images with an initial consistency score higher than this value are considered high-quality images; T low : Indicates the low threshold used for triage; images with an initial consistency score below this value are considered unqualified; μ s σ represents the average initial consistency score of all images in the historical data; s : Represents the standard deviation of the initial consistency score of all images in the historical data; k1, k2: are adjustment coefficients used to fine-tune the strictness of the threshold according to actual needs.

[0085] The entire initial quality assessment and intelligent triage process constitutes a complete automated quality inspection link. Through comprehensive analysis and intelligent classification of image features, it provides a reliable data foundation for subsequent processing. The design of this link fully considers various image quality issues that may be encountered in practical applications. Through multi-dimensional feature extraction and intelligent decision-making, it ensures the accuracy and reliability of the triage results, providing an important guarantee for the stable operation of the entire system.

[0086] The process of the strategy execution engine selecting a radiation normalization algorithm includes: matching predefined algorithm selection rules based on radiation statistical features and cloud cover indicators in a multidimensional quality feature vector, which maps the feature vector to a specific algorithm; various radiation normalization algorithms include histogram matching algorithm, relative radiation normalization algorithm, and normalization algorithm based on pseudo-invariant features;

[0087] When selecting algorithms, the radiation normalization algorithm with low computational complexity is given priority for initial processing. After the initial processing generates the first generation of processed data, the system records the algorithm used and its parameters for reference during secondary processing. The strategy execution engine also has a built-in algorithm performance evaluation module, which updates the algorithm selection rules based on historical processing results to optimize subsequent processing efficiency.

[0088] It should be further explained that for the image stream to be processed, the policy execution engine selects the radiation normalization algorithm based on the feature vector. The selection process is based on rule-based reasoning, with the rule form being: IF (Condition) THEN (Action); for example: IF (P cloud>0.3) AND (r<0.8) THEN Select Algorithm(A) The algorithm library includes various algorithms such as histogram matching and relative radiation normalization.

[0089] It should be further explained that the policy execution engine internally maintains an algorithm library containing various radiometric normalization algorithms, including at least histogram matching, relative radiometric normalization, and normalization algorithms based on pseudo-invariant features. When an image stream enters the policy execution engine, the engine parses the corresponding multi-dimensional quality feature vector, particularly the radiometric statistics and cloud cover index. The system matches the specific values ​​of the feature vectors with conditions in the rule library according to predefined algorithm selection rules, thereby determining the most suitable radiometric normalization algorithm for processing the current image. The algorithm selection rules of the policy execution engine need to cover various complex feature combination scenarios, specifically including: when the cloud cover index P... cloud When P > 0.3 and the spectral correlation r < 0.8, and the radiation statistics show large differences in radiation distribution (variance ratio > 2.0), a combination of the relative radiation normalization algorithm and the normalization algorithm based on pseudo-invariant features is selected; when P cloud When P ≤ 0.3 and r < 0.8, choose a combination of histogram matching algorithm and relative radiation normalization algorithm; when P cloud When P ≤ 0.3 and r ≥ 0.8, but the radiation statistics show large distribution differences, the histogram matching algorithm is selected; when P cloud When P > 0.3 and r ≥ 0.8, but the land cover type is mainly complex (such as urban-rural fringe areas), the normalization algorithm based on pseudo-invariant features is selected; when P cloud When the variance ratio of radiation statistical features is greater than 0.4 and less than 1.5, and the land cover type is simple (such as large areas of bare land), the relative radiation normalization algorithm is selected. The matching priority of all rules is executed in the order of "cloud cover index → ​​radiation distribution difference → spectral correlation → land cover complexity" to ensure the relevance and effectiveness of algorithm selection. For example, when the feature vector shows that the image has a high cloud coverage ratio, the system may prioritize the relative radiation normalization algorithm, which has less impact on cloud areas; when the feature vector indicates that there is a large difference in radiation distribution pattern between the image and the reference image, the histogram matching algorithm may be selected for overall distribution adjustment.

[0090] During algorithm selection, the system comprehensively considers processing efficiency and effectiveness, prioritizing algorithms with relatively low computational complexity for initial processing to ensure overall system processing speed. After each algorithm execution, the system automatically records detailed information such as the specific algorithm type used, parameter settings, and processing time. These records provide a reference for subsequent possible secondary processing. The strategy execution engine also includes a built-in algorithm performance evaluation module. This module collects historical processing data, including feedback on the processing effects of different algorithms under various characteristic conditions, to build an algorithm performance knowledge base. When the processing data accumulates to a certain scale, the system dynamically adjusts and optimizes the algorithm selection rules based on this actual processing effect data, making subsequent algorithm selection decisions more accurate and efficient.

[0091] This dynamic algorithm selection mechanism enables the system to adopt the most suitable processing strategy for images with different characteristics, ensuring both processing effectiveness and efficiency. Through continuous learning and optimization of algorithm selection rules, the system can adapt to various complex real-world application scenarios, improving the overall intelligence level of processing. The entire strategy execution process forms a self-improving closed-loop system, providing reliable technical support for radiometric consistency processing of multi-temporal remote sensing images.

[0092] The deep validation model used in the second verification is a convolutional neural network model or a random forest model. The training data includes multi-temporal remote sensing images and their spectral stability labels in stable ground cover sample areas. The calculation of the new round of consistency scores includes global radiometric statistics and local spectral bias calculations in stable ground cover sample areas.

[0093] The stable land cover sample area is a pre-selected area with stable land cover types, including water bodies, vegetation, and bare land. Local verification is performed by comparing the spectral reflectance values ​​of the processed image and the reference image in the sample area. If the deviation exceeds the preset tolerance value, it is judged as failing the second verification. In the second diversion, the data that passes the second verification must meet the following requirements: the consistency score of the new round is higher than the third threshold, and the local spectral deviation is lower than the preset tolerance value. The diagnostic report attached to the data that fails the second verification includes the failure reason analysis, recommended processing algorithm, and parameter adjustment suggestions.

[0094] It should be further explained that the system employs a specially trained deep validation model for secondary verification. This model can be either a convolutional neural network or a random forest model. The model's training process uses a large amount of multi-temporal remote sensing imagery as training samples. These training samples are all precisely labeled manually and have clear spectral stability labels in stable land cover areas. When secondary verification is required, the system first inputs the image to be verified into the deep validation model. This model comprehensively analyzes the global radiometric characteristics and local texture information of the image, outputting a new consistency score. Unlike the initial assessment, this score not only considers the overall radiometric statistical characteristics of the image but also pays special attention to the spectral performance of the image in specific land cover type areas.

[0095] While completing the model evaluation, the system performs local verification operations in parallel. The system accesses a pre-defined database of stable land cover sample areas. These sample areas are regions with stable land cover types, determined through long-term observation, primarily including typical land cover types such as water bodies, vegetation, and bare land. For each sample area, the system extracts the pixel spectral values ​​of the image to be verified and the reference image within that area, calculating the spectral reflectance difference for each band. The system compares these differences band by band with pre-defined tolerance values. Only when the spectral deviations of all bands are below the corresponding tolerance values ​​is the local verification of that sample area considered passed.

[0096] In the final triage decision, the system needs to consider both the new round of consistency score and the local verification results. Only when the new round of consistency score of the image is higher than the set third threshold and passes the local verification of all preset stable ground feature sample areas will the image be approved to enter the next processing stage. The third threshold for secondary verification is clearly related to the first threshold for initial diversion. The value of the third threshold is 0.1-0.2 higher than the first threshold. For example, when the first threshold is set to 0.8, the third threshold can be set to 0.9-1.0 to reflect the higher requirements of secondary verification for radiation consistency. The determination of the local spectral deviation tolerance value needs to be combined with the land cover type, sensor accuracy, and application scenario requirements. The local spectral deviation tolerance value for water bodies is set to 0.01-0.03, for vegetation to 0.02-0.04, and for bare land to 0.03-0.05. For high-resolution sensors with a spatial resolution higher than 10 meters, the tolerance value is reduced by 20% overall. For databases used in high-precision application scenarios such as change detection, the tolerance value is further reduced by 30% to ensure that the spectral stability of typical land cover meets the requirements of high-level applications. For images that fail to meet any of the conditions, the system will generate a detailed diagnostic report that clearly indicates the specific reasons for the failure, including which sample areas have spectral deviations, the specific values ​​of the deviations and their band distributions, and will also recommend more suitable processing algorithms and parameter adjustment suggestions based on the analysis results.

[0097] It should be further explained that in the second verification phase, the deep verification model adopts a convolutional neural network structure, including convolutional layers, pooling layers, and fully connected layers. Its output is the new round of consistency score S. deep Simultaneously, local tests were conducted to calculate spectral bias in stable ground cover sample areas:

[0098] ;

[0099] Where Δλ represents the absolute value of the spectral reflectance deviation at a specific wavelength band λ; R λ : Represents the average spectral reflectance value of band λ in the stable ground feature sample area of ​​the processed image; R λref : Represents the average spectral reflectance value of band λ in the reference image over the same stable ground feature sample area. When Δλ for all bands is less than the tolerance value Δ... max At that time, a partial inspection was conducted.

[0100] This secondary verification mechanism, which combines deep learning models and physical feature verification, leverages the powerful feature learning capabilities of modern machine learning methods while maintaining the physical interpretability based on the spectral characteristics of ground objects, thus ensuring the accuracy and reliability of the verification results. By setting multi-dimensional evaluation criteria, the system can effectively identify radiometric inconsistencies that may have been missed in the initial assessment, providing a more accurate basis for decision-making in subsequent processing.

[0101] During the secondary processing, the strategy execution engine analyzes the failure reasons in the diagnostic report and selects a radiation normalization algorithm with high computational complexity or a combination of multiple radiation normalization algorithms for processing. The combination of multiple radiation normalization algorithms includes first using a histogram matching algorithm for preprocessing, and then using a normalization algorithm based on pseudo-invariant features for adjustment. After the secondary processing generates the second-generation processed data, the effects of the first and second generation processing are compared, and the algorithm performance evaluation module is updated.

[0102] The physical spectral model used by the final arbiter is based on the atmospheric radiative transfer theory. This model simulates the atmospheric conditions at the time of image acquisition and calculates the residual between the theoretical spectral values ​​and the actual values. During the final evaluation, if the residual is lower than a preset threshold, the image is deemed to have passed; otherwise, it is deemed to have failed. The evaluation process of the final arbiter is independent of other verification steps and does not rely on historical data.

[0103] It should be further explained that when image data is returned from the secondary verification stage, the strategy execution engine receives the accompanying diagnostic report and analyzes its contents to determine the root cause of the initial processing failure. Based on the specific problem type indicated in the diagnostic report, the system initiates the corresponding secondary processing strategy. For cases with systematic deviations in radiation distribution patterns, the system may employ a combination of algorithms. For example, it might first use a histogram matching algorithm to coarsely adjust the overall radiation distribution of the image, and then use a normalization algorithm based on pseudo-invariant features to finely correct local areas. If the diagnostic report indicates that the initial processing algorithm does not respond well to a specific land cover type, the system will switch to another radiation normalization algorithm more suitable for the current land cover characteristics.

[0104] During the secondary processing, the system establishes a processing effect tracking mechanism, automatically recording and comparing the improvements in key indicators between the data after the first-generation processing and the data after the second-generation processing. This comparative data is fed back to the algorithm performance evaluation module in real time to update the performance records of each algorithm in different processing scenarios, thereby continuously optimizing the decision-making capabilities of the entire system.

[0105] The second-generation processed data, generated after secondary processing, is directly submitted to the final arbiter. The final arbiter uses a physical spectral model based on atmospheric radiative transfer theory. This model simulates atmospheric conditions during image acquisition, including aerosol optical thickness and water vapor content, and calculates the theoretical spectral reflectance based on these parameters. Subsequently, the system compares the spectral measurements of the actual image with the theoretical calculations point-by-point, generating a residual distribution map. The final arbiter analyzes the statistical characteristics of this residual map. When both the mean and dispersion of the residuals are below a preset stringent threshold, the image is deemed to have passed the final quality inspection; if either indicator exceeds the threshold range, it is deemed to have failed. The residual threshold setting for the final arbitration is directly related to the sensor noise level and the range of atmospheric condition fluctuations, with the mean residual threshold being the most important factor. The calculation method is to take 1.5 times the standard deviation of the sensor noise in the corresponding band, where the sensor noise standard deviation can be obtained from the sensor's factory parameters or measured through dark-field imaging experiments; the residual standard deviation threshold... The value ranges from 0.02 to 0.05. When atmospheric conditions fluctuate significantly during image acquisition, such as when the coefficient of variation of aerosol optical thickness (AOT) is greater than 0.3, This can be relaxed to 0.05 when atmospheric conditions are stable, such as when the AOT coefficient of variation is less than 0.1. Tightened to 0.02, the aforementioned quantitative correlation ensures that the residual threshold setting is both scientific and practically feasible. This final evaluation process is completely independent of all previous verification steps of the system and does not rely on any historical statistical data, thus ensuring the objectivity and accuracy of the evaluation results.

[0106] It should be further explained that the final arbiter uses a physical spectral model based on atmospheric radiative transfer theory to calculate the theoretical reflectivity:

[0107] ;

[0108] Among them, R theoretical θ is the theoretical reflectivity. s θ is the solar zenith angle. v To observe the zenith angle, φ is the relative azimuth angle, AOT is the aerosol optical thickness, and WV is the water vapor content. The residual between the actual reflectance and the theoretical value is calculated as follows:

[0109] ;

[0110] Where ε represents the residual, i.e., the difference between the measured value and the theoretical value; R measured : Represents the spectral reflectance value actually measured from the image to be adjudicated; R theoretical This represents the theoretical spectral reflectance value calculated using an atmospheric radiative transfer model based on atmospheric condition parameters at the time of image acquisition; the statistical characteristics of the residuals, including the mean μ, are analyzed. ε and standard deviation σ ε When |μ ε |<ε mean max And σ ε <ε std max At that time, the image is deemed to have passed the final inspection.

[0111] Through this progressive processing and verification mechanism, the system can effectively solve complex radiometric consistency problems in automated processes, while continuously improving processing efficiency through self-optimization, providing a reliable technical guarantee for building a high-quality multi-temporal remote sensing image database.

[0112] The closed-loop mechanism also includes a feedback loop, where the evaluation results of the final arbiter are used to optimize the lightweight classifier for primary quality assessment and the deep validation model for secondary verification. The optimization process is implemented through machine learning, using the data passed by the final arbiter as positive samples to update the model parameters. The automation of the entire process is coordinated by a workflow engine that monitors the execution status of each step and triggers retries or alarms when timeouts or errors occur. The workflow engine also records the processing history of each image, including the evaluation score, the algorithm used, and the verification results, for subsequent auditing and analysis.

[0113] It should be further explained that the system establishes a complete feedback loop, using the evaluation results of the final arbiter as an optimization signal to be transmitted back to the front-end quality assessment module. Specifically, all image data that passes the final arbiter's verification and their corresponding processing history, including primary evaluation feature vectors, consistency scores at each stage, and the final arbitration results, are automatically collected and labeled as high-quality samples by the system. After screening and processing, these sample data serve as new training datasets, used to periodically update the lightweight classifier in the primary quality assessment module and the deep validation model in the secondary verification module. During the model update process, the system employs incremental learning techniques, incorporating the feature patterns of new samples while maintaining existing knowledge. This allows the classifier and validation model to gradually adapt to the data characteristics of new remote sensing images, continuously improving their evaluation accuracy.

[0114] The entire processing flow is coordinated and monitored by a workflow engine, which tracks the status of each image data point in real time. The workflow engine maintains a status log table, recording the complete processing path of each image from data reception, through initial assessment, intelligent triage, policy execution, secondary verification, to final decision. When an anomaly occurs in any processing stage, such as processing time exceeding a preset limit or a module returning an error code, the workflow engine automatically initiates a retry operation according to a preset retry policy, while simultaneously recording the anomaly information. If the retry still fails to complete the process normally, the system triggers an alarm mechanism to notify the system administrator for manual intervention.

[0115] The workflow engine also maintains a complete processing profile for each image, detailing its evaluation score at each stage, the processing algorithms used, verification results, and status transition timestamps. This historical processing data is not only used for subsequent auditing and analysis but also provides valuable data support for system optimization. By analyzing this historical data, system administrators can identify bottlenecks in the processing flow and adjust system parameter configurations, thereby continuously improving the overall system's processing efficiency and reliability.

[0116] It should be further explained that the system establishes a feedback loop mechanism, using the data that passes the final verification to update the primary classifier and the deep validation model. The update process employs incremental learning, and the objective function is:

[0117] ;

[0118] Among them, L new : Represents the total loss function value after model update; L old : Represents the total loss function value before model update; η: Represents the learning rate, used to control the step size of each model update; Σl(y i ,f(x i): This represents summing the loss over all newly added training samples, where l is the value for calculating the predicted value f(x) for a single sample. i ) and the real label y i The loss function for the difference between them.

[0119] It should be further explained that the coordination of the entire system is achieved through a workflow engine, whose state transition function is:

[0120] ;

[0121] Here, S represents state and E represents event. The monitoring module collects system metrics in real time, including throughput Q and resource utilization U, to ensure stable system operation.

[0122] This closed-loop feedback and coordination mechanism ensures the system's ability to continuously self-optimize. By constantly learning new, high-quality sample features, the system's evaluation accuracy and processing precision are steadily improved. Meanwhile, the workflow engine's end-to-end monitoring guarantees the system's operational stability, making the entire creation process of the multi-temporal remote sensing image database both efficient and reliable.

[0123] The multi-temporal remote sensing image database adopts a layered storage structure, including a raw data layer, a processing intermediate data layer, and a finished data layer. The raw data layer stores the input multi-temporal remote sensing images; the processing intermediate data layer stores the initial quality assessment results, first-generation processed data, and second-generation processed data; the finished data layer stores the data that has passed the final arbiter. Each layer of data is associated with metadata, including acquisition time, sensor type, processing history, and radiometric consistency score. The database also provides a query interface to support searches based on time range, geographic region, and radiometric consistency level.

[0124] It should be further explained that the system employs a layered storage architecture to manage data at different stages. The raw data layer stores all input raw multi-temporal remote sensing images, organized according to acquisition time, sensor type, and geographical extent. Each image file includes complete metadata information. The intermediate data layer stores various intermediate results generated throughout the quality assessment and processing process, including multi-dimensional quality feature vectors and initial consistency scores generated in the initial quality assessment stage, first-generation processed data generated by the policy execution engine, and second-generation processed data generated after secondary processing. These intermediate data are associated with their corresponding raw images through unique identifiers. The finished data layer specifically stores qualified image data that has passed the final arbiter's verification. This data has undergone the system's complete processing flow and possesses reliable radiometric consistency quality assurance.

[0125] Each data layer is associated with a complete metadata description. This metadata records detailed attributes such as data acquisition time, sensor model, spatial resolution, and band information. It also includes a complete processing history, including processing algorithms used at each stage, parameter settings, consistency score changes, and the final quality assessment results. By establishing a unified metadata management mechanism, the system ensures complete traceability of all data.

[0126] The database system provides a flexible query interface, allowing users to retrieve image data within a specific period based on time range criteria, filter data by geographical area, and support quality filtering based on radiometric consistency scores. This hierarchical storage structure, combined with a rich metadata management mechanism, ensures the traceability of the data processing process while providing users with convenient and efficient data retrieval and access capabilities, thus constructing a well-structured, well-managed, and user-friendly multi-temporal remote sensing image database system.

[0127] A system for creating a multi-temporal remote sensing image database, used to implement a method for creating a multi-temporal remote sensing image database, includes:

[0128] The data receiving module is used to receive multi-temporal remote sensing image data;

[0129] The primary quality assessment module is used to perform a primary quality assessment on each image. It constructs a multidimensional quality feature vector by extracting the radiometric statistical features, cloud cover index, and spectral correlation with the baseline image. A pre-trained lightweight classifier is then used to analyze the vector to output an initial consistency score.

[0130] The intelligent routing module is used to perform intelligent routing based on the initial consistency score, dividing the images into high-confidence pass streams, pending streams, and high-confidence failure streams. Images in the high-confidence pass stream directly enter the secondary verification queue, images in the pending stream flow into the strategy execution engine, and images in the high-confidence failure stream are marked as unqualified and logged.

[0131] The strategy execution engine module is used to select a radiation normalization algorithm for initial processing and generate the first generation of processed data. This selection is based on the specific composition of the multidimensional quality feature vector to adapt one algorithm from a variety of radiation normalization algorithms.

[0132] The secondary verification module inputs the high-confidence through-stream imagery and the first-generation processed data into the secondary verification queue for secondary verification to obtain the secondary verification result. The secondary verification uses a deep verification model to calculate a new round of consistency score. This deep verification model is different from the primary quality assessment model and includes local verification on a preset stable ground feature sample area to ensure that the processed imagery and the reference imagery maintain spectral stability on typical ground features.

[0133] The secondary routing module is used to perform secondary routing based on the secondary verification results: data that passes the secondary verification is approved and stored in the database, while data that fails the secondary verification is returned to the strategy execution engine along with a diagnostic report;

[0134] The secondary processing module is used to analyze the reasons for processing failures based on the diagnostic report of the data from the returned strategy execution engine, and switch or combine another radiation normalization algorithm for secondary processing to generate second-generation processed data.

[0135] The final arbiter module submits the second-generation processed data to the final arbiter, which uses a physical spectral model to perform a final evaluation and outputs the final evaluation result. The final evaluation result is either pass or fail. Data that fails is marked as difficult data and the automated process is terminated.

[0136] The storage module is used to store the data that has passed the final evaluation into a multi-temporal remote sensing image database;

[0137] The various modules of the system are connected via bus or network to form a closed-loop processing pipeline.

[0138] It should be further explained that the system includes a data receiving module, which is equipped with a network interface and a data buffer to continuously receive remote sensing image data from different data sources and perform format parsing and basic information verification on the received data. The primary quality assessment module includes a feature extraction unit and a classifier unit. The feature extraction unit calculates the radiometric statistical features, cloud cover index, and spectral correlation of the image by calling an image processing library. The classifier unit loads a pre-trained lightweight classification model to analyze the extracted feature vectors and output an initial consistency score.

[0139] The intelligent triage module performs classification operations based on the scoring results. This module includes a threshold comparator and a data routing controller, which can guide image data to different processing paths according to preset threshold conditions. The strategy execution engine module has a built-in algorithm library management unit and a selection decision unit. The algorithm library stores executable code for various radiometric normalization algorithms, and the selection decision unit determines the applicable processing algorithm based on the feature vector content of the input image through rule matching. The secondary verification module integrates a deep verification model and a local verification unit. The deep verification model is implemented by calling a deep learning framework, and the local verification unit obtains the spatial coordinate information of stable ground feature sample areas by accessing a geographic information database and performs spectral feature comparison in a specified area.

[0140] The secondary data distribution module includes a result analyzer and a data scheduler, capable of determining the next step of data flow based on the verification results. The secondary processing module is equipped with an algorithm combination controller and a parameter adjuster, capable of selecting new processing strategies based on the content of the diagnostic report. The final adjudicator module runs a physical spectral simulation program, built upon the atmospheric radiative transfer equation, capable of simulating spectral response characteristics under different atmospheric conditions. The storage module employs a hierarchical database design, containing raw data tables, intermediate data tables, and final data tables, with each table linked through image identifiers.

[0141] The aforementioned modules exchange data and transmit commands via a system bus. In a network deployment environment, cross-node communication is achieved through message queues and remote calls. The system ensures data format compatibility between modules by defining a unified data interface specification. The entire processing pipeline adopts an event-driven architecture. When a module completes processing, it automatically triggers the execution flow of the next module and updates the processing status to the central status manager. This modular design and closed-loop processing mechanism enable the system to automatically and continuously complete the entire processing process from data reception to the storage of qualified images, effectively ensuring the radiometric consistency quality of images across different time phases in the final database.

[0142] The primary quality assessment module also includes a feature extraction unit and a classifier unit. The feature extraction unit is configured to compute radiation statistics, cloud cover indices, and spectral correlations in parallel. The classifier unit is periodically retrained with new data to maintain accuracy. The policy execution engine module includes an algorithm library unit and a selection unit. The algorithm library unit stores various radiation normalization algorithms. The selection unit selects an algorithm in real time based on multidimensional quality feature vectors. The secondary verification module includes a deep model unit and a local verification unit. The deep model unit loads a pre-trained deep verification model. The local verification unit accesses the geographic information system to obtain the coordinates of stable ground feature sample areas. The final arbiter module integrates physical spectral simulation software, which calculates theoretical spectral values ​​based on atmospheric parameters.

[0143] It should be further explained that the feature extraction unit of the primary quality assessment module adopts a parallel computing architecture. The radiometric statistical feature calculation subunit calculates the statistical moments of each band by traversing the image pixel values. The cloud cover index extraction subunit uses an image segmentation algorithm based on spectral features to identify cloud-covered areas. The spectral correlation analysis subunit completes the similarity calculation with the reference image through spatial registration and pixel-level computation. The classifier unit adopts a periodic update mechanism. When the system accumulates a certain number of new samples, it automatically triggers the model retraining process, maintaining continuous optimization of classification accuracy through incremental learning.

[0144] The algorithm library unit of the strategy execution engine module organizes various radiation normalization algorithms in a pluggable manner. Each algorithm is encapsulated as an independent execution unit and equipped with a complete parameter configuration interface. The selection decision unit has a built-in rule inference engine, which performs pattern matching between the numerical features of multi-dimensional quality feature vectors and preset rule conditions. The rule base supports dynamic updates, and system administrators can adjust rule weights and matching conditions based on actual processing results. The deep model unit of the secondary verification module loads a pre-trained neural network model file containing a complete network structure and training parameters, enabling it to quickly complete forward inference calculations after receiving the image to be verified. The local verification unit establishes a stable connection with the geographic information system database, automatically retrieving the spatial range of corresponding stable ground feature sample areas based on the geographic coordinate information of the image, and performing refined spectral stability analysis within these specific areas.

[0145] The atmospheric radiative transfer simulation software integrated into the final adjudicator module is developed based on a mature atmospheric physics model. This software takes atmospheric parameter observation data acquired during image acquisition, including aerosol concentration, water vapor content, and ozone concentration, and obtains theoretical spectral values ​​by solving the radiative transfer equation. The comparison and analysis of the software output with actual image data is completed in a dedicated quality assessment subunit. This subunit generates a detailed residual analysis report, providing quantitative evidence for the final adjudication.

[0146] Data interaction between modules employs a unified message protocol to ensure data integrity and consistency across different processing stages. The system utilizes a message middleware for asynchronous communication between modules; this loosely coupled architecture ensures that maintenance and upgrades of individual modules do not affect the overall system's normal operation. Furthermore, a health monitoring mechanism is in place, with each module periodically sending heartbeat signals to the central monitor, reporting its operational status and resource usage in real time, thus guaranteeing stable system operation.

[0147] The system also includes a monitoring and management module, which displays the status of the processing pipeline in real time, including the input and output data volume, processing time, and error logs for each module. The monitoring and management module provides a graphical user interface for manually adjusting the diversion threshold, algorithm selection rules, and verification parameters. The system records all operations through a log database, supporting fault recovery and performance analysis. The storage module adopts a distributed file system, supporting multi-node parallel access to improve processing efficiency under large data volumes. The system integrates with external remote sensing data processing platforms through an application programming interface to achieve data sharing and collaborative processing.

[0148] It should be further explained that the system's monitoring and management module collects real-time operational status information from each processing module through a data acquisition agent. This includes the input queue length of the data receiving module, the number of pending tasks in each processing stage, the resource utilization of the algorithm execution unit, and the disk space usage of the storage module. This status data is reported to the central status manager via a message bus. After aggregation and analysis, it is visualized through a graphical user interface. The interface uses different colors to distinguish the operational status levels of different parts of the system, allowing administrators to grasp the overall operational status of the system in real time through the interface elements.

[0149] The monitoring and management module provides a parameter configuration interface, allowing administrators with appropriate permissions to adjust system operating parameters according to actual business needs. Administrators can modify threshold parameters used in the intelligent traffic triage process, adjust the algorithm selection rule weights in the policy execution engine, or optimize the evaluation criteria in the secondary verification process. All parameter modification operations ensure data consistency through a transaction mechanism; modified parameters take effect in real time and are persistently stored in the system configuration database. The system also provides parameter verification functionality; when an administrator attempts to set illogical parameter combinations, the system will issue a warning and refuse to execute the illegal configuration.

[0150] The system comprehensively records all operational events and processing activities through a logging component. Log data is stored in a dedicated log database in chronological order, and each log record includes key information such as timestamp, operation type, execution module, and operation result. The log system supports combined queries based on multiple conditions, facilitating troubleshooting and performance analysis. The system also provides a log archiving mechanism, periodically transferring historical log data to long-term storage devices while maintaining its searchability.

[0151] The storage module employs a distributed file system architecture, distributing image data and metadata across multiple storage nodes and ensuring data reliability through a redundant replication mechanism. The system uses a data sharding strategy to divide large image files into fixed-size data blocks, which can be read and written in parallel, thereby improving processing throughput under large data volumes. The storage module also implements a load balancing mechanism, automatically distributing access requests to less loaded storage nodes to prevent any single node from becoming a performance bottleneck.

[0152] The system integrates with external remote sensing data processing platforms through defined standard application programming interfaces (APIs). These interfaces are based on common network protocols and data format specifications, supporting functions such as authentication, data querying, and status checks. External systems can submit data processing requests, obtain processing status, download qualified data, or upload auxiliary data by calling these interfaces. The system also provides a data exchange queue to buffer data transmission between systems, ensuring stable processing capabilities even under high concurrency scenarios. This open architecture design allows the system to flexibly integrate into the existing remote sensing data processing ecosystem, enabling collaborative work and data sharing with different platforms.

[0153] It should be further explained that the storage module adopts a distributed architecture, and the sharding strategy formula for the storage module is as follows:

[0154] ;

[0155] Among them, Shard ID : Represents the calculated data shard identifier, used to determine which shard the data is stored on; hash(Image ID `()` indicates that the unique identifier ImageID of the image is hashed to obtain an integer value; `mod` indicates the modulo operation; `N` represents the total number of data shards in the system. Load balancing is achieved through a consistent hashing algorithm to ensure even data distribution.

[0156] In the specific implementation of the multi-temporal remote sensing image database creation method and system involved in this invention, the technical solution is realized through a complete closed-loop processing flow. The system first receives remote sensing image data from different sensors at different times, and then initiates a processing mechanism that includes multiple verifications and adaptive optimization.

[0157] Once the image data is input into the system, the primary quality assessment module begins operation. This module evaluates the overall radiometric level by calculating the statistical characteristics of pixel values ​​in each band of the image. Simultaneously, it uses image segmentation technology to identify cloud-covered areas and calculate the cloud cover ratio. Furthermore, it performs pixel-level registration and spectral correlation analysis between the image to be evaluated and a rigorously selected benchmark image. These analytical results collectively form a multidimensional feature vector, which is input into a pre-trained lightweight classification model. This model analyzes these features and outputs a quantified initial consistency score, which objectively reflects the degree of radiometric difference between the current image and the benchmark image.

[0158] Based on the initial consistency score, the system performs intelligent data triage. This process sets two key thresholds, which are pre-defined after statistical analysis of a large amount of historical data. When the initial consistency score of an image is higher than the first threshold, the system classifies it as a high-quality image and directly proceeds to the subsequent processing stage; when the score is between the first and second thresholds, the system sends it to the strategy execution engine for further processing; and when the score is lower than the second threshold, it is marked as unqualified data, and detailed logs are recorded for subsequent review.

[0159] The strategy execution engine incorporates various radiation normalization algorithms, including histogram matching, relative radiation normalization, and normalization based on pseudo-invariant features. The engine dynamically selects the most suitable algorithm for initial processing based on the specific composition of the multidimensional feature vectors of the input image. During the selection process, the system comprehensively considers processing efficiency and effectiveness, prioritizing algorithms with relatively low computational complexity for initial processing. After each algorithm execution, the system automatically records the algorithm type and parameter settings used; these records provide a reference for subsequent processing.

[0160] The data processed initially, along with the data that passed the preliminary assessment, enters the second verification stage. This stage employs a deep validation model for more rigorous quality checks. This model differs from the one used in the preliminary assessment, featuring a more complex structure and stronger discriminative capabilities. The second verification not only calculates a new round of consistency scores but also conducts local verification on pre-selected stable land cover sample areas. These sample areas are stable land cover types identified through long-term observation, including typical land cover types such as water bodies, vegetation, and bare land. The system extracts the pixel spectral values ​​of the image to be verified and the reference image within these sample areas, calculating the spectral reflectance differences for each band. Only when the spectral deviation of all bands is below the corresponding tolerance value is the local verification of that sample area considered passed.

[0161] Based on the results of the second verification, the system performs a second data splitting operation. Data that simultaneously meets the requirements of the new round of consistency scoring and local verification is approved for entry into the final database, while data that fails is sent back to the strategy execution engine along with a detailed diagnostic report. The diagnostic report clearly indicates the specific reasons for failure, including which sample areas exhibited spectral bias, the specific values ​​of the bias, and the distribution of the bias bands. It also recommends more suitable processing algorithms and parameter adjustment suggestions based on the analysis results.

[0162] For images that are rejected, the strategy execution engine adjusts the processing strategy based on the analysis results of the diagnostic report. The system may choose to switch to another radiometric normalization algorithm or combine multiple algorithms for secondary processing. During secondary processing, the system establishes a processing effect tracking mechanism, automatically recording and comparing the improvements in key indicators between the initial and secondary processing. This comparative data is fed back to the algorithm performance evaluation module in real time to update the performance records of each algorithm in different processing scenarios.

[0163] The data after secondary processing is directly submitted to the final arbiter. The final arbiter uses a physical spectral model based on atmospheric radiative transfer theory for evaluation. This model simulates the atmospheric conditions at the time of image acquisition and calculates the theoretical spectral reflectance value based on these parameters. Subsequently, the system compares the spectral measurements of the actual image with the theoretical calculations, generating a residual distribution map. The final arbiter analyzes the statistical characteristics of this residual map; when both the mean and dispersion of the residuals are below a preset stringent threshold, the image is deemed to have passed the final quality inspection.

[0164] The system also establishes a complete feedback loop, transmitting the final arbiter's evaluation results as optimization signals back to the front-end quality assessment module. All image data that passes the final arbiter's verification, along with their corresponding processing history, are automatically collected and marked as high-quality samples by the system. These samples are used to periodically update the analysis models in the primary quality assessment module and the secondary verification module. During the model update process, the system employs incremental learning techniques, incorporating the feature patterns of new samples while maintaining existing knowledge.

[0165] The entire processing workflow is coordinated and monitored by a unified workflow engine, which tracks the status of each image data point in real time. When an anomaly occurs in a processing stage, the workflow engine automatically initiates a retry operation based on a preset retry strategy. The system also maintains a complete processing archive for each image. This historical processing data is not only used for subsequent auditing and analysis but also provides valuable data support for system optimization.

[0166] In terms of system architecture, each processing module exchanges data and transmits commands via a system bus or network connection. The system adopts a hierarchical storage architecture to manage data at different stages. The raw data layer stores all input raw images, the intermediate processing data layer stores various intermediate results generated throughout the quality assessment and processing process, and the finished product data layer specifically stores qualified image data that has passed final inspection. Each type of data in each data layer is associated with a complete metadata description, ensuring that all data has complete traceability.

[0167] The system also includes a monitoring and management module, which displays the system's operating status through a graphical user interface and allows administrators to adjust system parameters. The storage module employs a distributed file system architecture, improving processing efficiency under large data volumes through data sharding and load balancing mechanisms. The system integrates with external remote sensing data processing platforms via a standard application programming interface, supporting data sharing and collaborative processing.

[0168] This multi-level, multi-stage quality inspection and processing mechanism, combined with closed-loop feedback and adaptive optimization capabilities, ensures reliable radiometric consistency of the final imagery, providing a high-quality data foundation for subsequent remote sensing data analysis applications. The entire technical solution effectively solves the technical problems of traditional methods—relying on manual intervention, low processing efficiency, and difficulty in guaranteeing quality—through automated processing flows and intelligent decision-making mechanisms.

[0169] The beneficial effects of this invention lie in its construction of an automated processing workflow that includes multiple quality assessments, intelligent triage decisions, and closed-loop feedback optimization. This effectively solves the core technical problems of traditional multi-temporal remote sensing image database creation, such as reliance on manual intervention, low processing efficiency, and difficulty in ensuring radiometric consistency. The complete quality assurance system it establishes, from initial assessment to final decision, enables standardized and automated control of the radiometric quality of massive image data, reducing labor costs and improving database construction efficiency.

[0170] Through dynamic algorithm selection, multi-layered verification mechanisms, and a continuously self-optimizing system design, reliable radiometric consistency quality of the imported image data is ensured. This intelligent processing mode not only significantly improves the robustness and adaptability of the data processing workflow but also provides a high-quality data foundation for subsequent advanced applications such as remote sensing time-series analysis and change detection, demonstrating significant practical application value.

[0171] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0172] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for creating a multi-temporal remote sensing image database, characterized in that, Includes the following steps: S1: Receive multi-temporal remote sensing image data as input; S2: Perform a preliminary quality assessment on each image. This assessment constructs a multidimensional quality feature vector by extracting the radiometric statistics, cloud cover index and spectral correlation with the baseline image of the image, and uses a pre-trained lightweight classifier to analyze the vector to output an initial consistency score. S3: Based on the initial consistency score, intelligent routing is performed to divide the images into high-confidence pass stream, pending stream, and high-confidence failure stream. The images in the high-confidence pass stream directly enter the secondary verification queue, the images in the pending stream flow into the policy execution engine, and the images in the high-confidence failure stream are marked as unqualified and logged. S4: For the images in the processing stream, the policy execution engine selects a radiometric normalization algorithm for initial processing to generate the first generation of processed data. This selection is based on the specific composition of the multidimensional quality feature vector and adapts one of the various radiometric normalization algorithms. When selecting, the radiometric normalization algorithm with low computational complexity is given priority for initial processing. S5: Input the high-confidence through-stream imagery and the first-generation processed data into the secondary verification queue for secondary verification to obtain the secondary verification result. The secondary verification uses a deep verification model to calculate a new round of consistency score. This deep verification model is different from the primary quality assessment model and includes local verification on a preset stable ground feature sample area to ensure that the processed imagery and the reference imagery maintain spectral stability on typical ground features. S6: Perform secondary routing based on the secondary verification results: Data that passes the secondary verification is approved for entry into the database, while data that fails the secondary verification is returned to the strategy execution engine along with a diagnostic report; S7: For the data returned to the strategy execution engine, analyze the reasons for the failure based on the diagnostic report, and switch or combine another radiation normalization algorithm for secondary processing to generate the second-generation processed data; During the secondary processing, the strategy execution engine selects a radiation normalization algorithm with high computational complexity or combines multiple radiation normalization algorithms for processing based on the failure reasons analysis in the diagnostic report. S8: Submit the second-generation processed data to the final arbiter, which uses a physical spectral model for final evaluation and outputs the final evaluation result. The final evaluation result is either pass or fail. Passed data is stored in the multi-temporal remote sensing image database, while failed data is marked as problematic data and the automated process is terminated.

2. The method for creating a multi-temporal remote sensing image database according to claim 1, characterized in that: The primary quality assessment extracts the radiometric statistical features of the images, including calculating the mean, variance, and histogram distribution of the images; extracts cloud cover indicators of the images, including identifying cloud-covered areas and calculating the cloud coverage ratio through image segmentation technology; and extracts the spectral correlation between the images and the reference images, including calculating the correlation coefficient between the two images in the same band. The lightweight classifier is a support vector machine or decision tree model, and its pre-training process uses historical multi-temporal remote sensing image data and its corresponding radiometric consistency labels for supervised learning. The initial consistency score is a numerical score used to quantify the degree of radiometric difference between the image and the reference image; In the intelligent traffic splitting process, the criteria for a high-confidence flow to pass is that the initial consistency score is higher than the first threshold. The criteria for a flow to be processed is that the initial consistency score is between the first and second thresholds. The criteria for a high-confidence failed flow is that the initial consistency score is lower than the second threshold. The first and second thresholds are preset through historical data statistical analysis.

3. The method for creating a multi-temporal remote sensing image database according to claim 1, characterized in that: The process of the strategy execution engine selecting a radiation normalization algorithm includes: matching a predefined algorithm selection rule with the radiation statistical features and cloud cover index in the multidimensional quality feature vector, which maps the feature vector to a specific algorithm; the various radiation normalization algorithms include histogram matching algorithm, relative radiation normalization algorithm and normalization algorithm based on pseudo-invariant features; After the initial processing generates the first generation of processed data, the algorithm used and its parameters are recorded. The strategy execution engine also has a built-in algorithm performance evaluation module, which updates the algorithm selection rules based on historical processing results to optimize subsequent processing efficiency.

4. The method for creating a multi-temporal remote sensing image database according to claim 1, characterized in that: The deep verification model used in the secondary verification is a convolutional neural network model or a random forest model. The training data includes multi-temporal remote sensing images and their spectral stability labels in stable ground cover sample areas. The calculation of the new round of consistency score includes global radiometric statistics and local spectral deviation calculations in stable ground cover sample areas. The stable land cover sample area is a pre-selected area with stable land cover types, including water bodies, vegetation, and bare land. Local verification is performed by comparing the spectral reflectance values ​​of the processed image and the reference image in the sample area. If the deviation exceeds the preset tolerance value, it is determined that the second verification has failed. In the second diversion, the data that passes the second verification must meet the following requirements: the consistency score of the new round is higher than the third threshold, and the local spectral deviation is lower than the preset tolerance value. The diagnostic report attached to the data that fails the second verification includes the failure reason analysis, recommended processing algorithm, and parameter adjustment suggestions.

5. The method for creating a multi-temporal remote sensing image database according to claim 1, characterized in that: Combining multiple radiation normalization algorithms involves first using a histogram matching algorithm for preprocessing, then using a normalization algorithm based on pseudo-invariant features for adjustment; after secondary processing to generate second-generation processed data, the effects of the first and second generation processing are compared, and the algorithm performance evaluation module is updated. The physical spectral model used by the final arbiter is a model based on atmospheric radiative transfer theory. This model simulates the atmospheric conditions at the time of image acquisition and calculates the residual between the theoretical spectral values ​​and the actual values. In the final evaluation, if the residual is lower than a preset threshold, it is judged as passing; otherwise, it is judged as failing.

6. The method for creating a multi-temporal remote sensing image database according to claim 1, characterized in that: The evaluation results of the final arbiter are used to optimize the lightweight classifier for primary quality assessment and the deep validation model for secondary verification; the optimization process is implemented through machine learning, using the data passed by the final arbiter as positive samples to update the model parameters.

7. The method for creating a multi-temporal remote sensing image database according to claim 1, characterized in that: The storage structure of the multi-temporal remote sensing image database adopts a hierarchical design, including a raw data layer, a processing intermediate data layer, and a finished data layer. The raw data layer stores the input multi-temporal remote sensing images; the processing intermediate data layer stores the initial quality assessment results, the first-generation processed data, and the second-generation processed data; the finished data layer stores the data that has passed the final arbiter. Each layer of data is associated with metadata, including acquisition time, sensor type, processing history, and radiometric consistency score. The database also provides a query interface to support retrieval based on time range, geographic region, and radiometric consistency level.

8. A system for creating a multi-temporal remote sensing image database, used to implement the method for creating a multi-temporal remote sensing image database as described in any one of claims 1 to 7, characterized in that, include: The data receiving module is used to receive multi-temporal remote sensing image data; The primary quality assessment module is used to perform a primary quality assessment on each image. It constructs a multidimensional quality feature vector by extracting the radiometric statistical features, cloud cover index, and spectral correlation with the baseline image. A pre-trained lightweight classifier is then used to analyze the vector to output an initial consistency score. The intelligent routing module is used to perform intelligent routing based on the initial consistency score, dividing the images into high-confidence pass streams, pending streams, and high-confidence failure streams. Images in the high-confidence pass stream directly enter the secondary verification queue, images in the pending stream flow into the strategy execution engine, and images in the high-confidence failure stream are marked as unqualified and logged. The strategy execution engine module is used to select a radiation normalization algorithm for initial processing and generate the first generation of processed data. This selection is based on the specific composition of the multidimensional quality feature vector to adapt one algorithm from a variety of radiation normalization algorithms. The secondary verification module inputs the high-confidence through-stream imagery and the first-generation processed data into the secondary verification queue for secondary verification to obtain the secondary verification result. The secondary verification uses a deep verification model to calculate a new round of consistency score. This deep verification model is different from the primary quality assessment model and includes local verification on a preset stable ground feature sample area to ensure that the processed imagery and the reference imagery maintain spectral stability on typical ground features. The secondary routing module is used to perform secondary routing based on the secondary verification results: data that passes the secondary verification is approved and stored in the database, while data that fails the secondary verification is returned to the strategy execution engine along with a diagnostic report; The secondary processing module is used to analyze the reasons for processing failures based on the diagnostic report of the data from the returned strategy execution engine, and switch or combine another radiation normalization algorithm for secondary processing to generate second-generation processed data. The final arbiter module submits the second-generation processed data to the final arbiter, which uses a physical spectral model to perform a final evaluation and outputs the final evaluation result. The final evaluation result is either pass or fail. Data that fails is marked as difficult data and the automated process is terminated. The storage module is used to store the data that has passed the final evaluation into a multi-temporal remote sensing image database.

9. The system for creating a multi-temporal remote sensing image database according to claim 8, characterized in that: The primary quality assessment module also includes a feature extraction unit and a classifier unit; the feature extraction unit is configured to perform parallel computation of radiation statistical features, cloud cover indicators, and spectral correlation. The classifier unit is periodically retrained with new data to maintain accuracy; The strategy execution engine module includes an algorithm library unit and a selection unit; the algorithm library unit stores various radiation normalization algorithms; the selection unit selects an algorithm in real time based on a multidimensional quality feature vector; the secondary verification module includes a deep model unit and a local verification unit; the deep model unit loads a pre-trained deep verification model; the local verification unit accesses the geographic information system to obtain the coordinates of stable ground feature sample areas; the final arbiter module integrates physical spectral simulation software, which calculates theoretical spectral values ​​based on atmospheric parameters.

10. The system for creating a multi-temporal remote sensing image database according to claim 8, characterized in that: The system also includes a monitoring and management module, which displays the status of the processing pipeline in real time, including the input and output data volume, processing time, and error logs of each module. The monitoring and management module provides a graphical user interface for manually adjusting the diversion threshold, algorithm selection rules, and verification parameters. The system records all operations through a log database, supporting fault recovery and performance analysis. The storage module uses a distributed file system, supporting parallel access across multiple nodes to improve processing efficiency under large data volumes. The system integrates with external remote sensing data processing platforms through an application programming interface (API) to achieve data sharing and collaborative processing.

Images