An AI-based method for target change detection by fusing temporal remote sensing imagery

Abstract

This invention relates to the field of image data processing technology and discloses an AI target change detection method that integrates time-series remote sensing imagery. The method includes: calculating the differential information entropy value of the remote sensing imagery on a spaceborne edge computing unit and comparing it with a trigger threshold; only when the information entropy value is higher than the threshold is the changed area extracted to form an incremental information packet and transmitted to the ground; after analyzing the incremental information packet, the ground station also generates instructions based on the identified environmental interference characteristics and regional stability, and feeds them back to the spaceborne unit to bidirectionally adjust the trigger threshold. This invention avoids the invalid transmission of massive amounts of unchanging background data by placing the information filtering decision at the source of data generation, thus significantly reducing the delay in acquiring key change information; the established space-ground closed-loop feedback mechanism also enables the entire monitoring method to continuously optimize its judgment criteria during use, improving the applicability and reliability of the method.

Description

Technical Field

[0001] This invention relates to an AI target change detection method that integrates time-series remote sensing imagery, belonging to the field of image data processing technology. Background Technology

[0002] In the field of image data processing technology, a common technical process for detecting changes in time-series remote sensing images is as follows: the original images are acquired by an on-orbit satellite, all data of the original images are transmitted to the ground center via a communication link, and then the computing equipment at the ground center is used to analyze and process the data to identify the areas of change on the Earth's surface.

[0003] With the deployment of large-scale satellite constellations, the total amount of data generated in orbit has increased significantly. However, the bandwidth capacity of satellite-to-ground communication has not increased in tandem due to physical and cost factors. This has led to a problem in the practical application of the above-mentioned technical process: a large amount of raw image data needs to be queued for downlink transmission, which means that images containing changing information cannot be acquired and analyzed in a timely manner in time-sensitive applications such as emergency response. At the same time, since the unchanging background areas in the images occupy the majority, a large amount of communication and storage resources are consumed in order to transmit this data that does not contain new information.

[0004] To address the aforementioned issues, simply adding ground stations or enhancing ground computing capabilities cannot alter the information processing timeline. The limitations of this process are: 1. The data value identification step occurs after data transmission, which reduces the value of images containing information with time-sensitive changes due to transmission delays; 2. At the satellite-based data generation end, there is a lack of an on-orbit operational technology for preliminary screening of image information before transmission, resulting in all data, regardless of whether it contains changes, being included in the transmission process. Therefore, the technical problem this invention aims to solve is how to construct an image data processing method capable of performing pre-diagnosis of image changes at the satellite edge and determining whether to transmit high-resolution image data based on the diagnostic results. Furthermore, this method can utilize the analysis results from ground stations to provide feedback and adjust the diagnostic criteria at the satellite end. Summary of the Invention

[0005] This invention provides an AI target change detection method that integrates time-series remote sensing images. Its main purpose is to solve the problems of large data transmission volume and untimely information acquisition caused by the data value identification step being placed after data transmission and the lack of a satellite-ground closed-loop adjustment mechanism in the existing remote sensing image processing workflow.

[0006] To achieve the above objectives, the present invention provides an AI target change detection method based on fused temporal remote sensing imagery, comprising the following steps:

[0007] Step a: On the spaceborne edge computing unit, acquire the current period remote sensing image and generate a reference benchmark based on the historical remote sensing image;

[0008] Step b: The onboard edge computing unit calculates the information entropy value of the differential information between the current period remote sensing image and the reference benchmark;

[0009] Step c: Compare the information entropy value with a trigger threshold. When the information entropy value is higher than the trigger threshold, extract the target block corresponding to the changed area from the current period remote sensing image, form an incremental information packet and transmit it to the ground station. If the information entropy value is not higher than the trigger threshold, abandon the original data transmission of the current period remote sensing image on the spaceborne edge computing unit side.

[0010] Step d: The ground station receives and analyzes incremental information packets to confirm target changes. Based on the analysis of at least one incremental information packet received in history, the ground station identifies and extracts the characteristic signature of environmental interference and quantifies the continuous unupdated duration of the reference benchmark for each area.

[0011] Step e: The ground station generates an adjustment command based on the feature signature of environmental interference and the duration of continuous non-update, and sends it to the onboard edge computing unit to adjust the trigger threshold of the corresponding area: when the current change feature is detected to match the feature signature in the environmental log, the trigger threshold is increased; when the duration of continuous non-update in a certain area reaches a stable duration threshold, the trigger threshold is decreased.

[0012] Preferably, the reference benchmark in step a is a low-resolution thumbnail of historical remote sensing imagery; the object for calculating the information entropy value in step b is the difference image between the corresponding low-resolution thumbnail generated from the current period's remote sensing imagery and the reference benchmark.

[0013] Preferably, the incremental information packet in step c is a high-resolution differential image between the target block of the current period remote sensing image and its corresponding historical remote sensing image block.

[0014] Preferably, step d, before analyzing the incremental information packet, further includes: calculating the texture entropy value of each sub-region within the incremental information packet based on the differential information within it; and in the subsequent change confirmation analysis, assigning weights to the motion vector features from each sub-region, the weight values ​​being determined based on the texture entropy value of the sub-region, wherein the motion vector features are derived from the optical flow field calculation performed on the incremental information packet.

[0015] Preferably, the adjustment of the trigger threshold in step e follows the following rules: ,in, It is the adjusted trigger threshold. This is the base trigger threshold for this area. It is the count value of the silent period corresponding to the continuous period without updates. It is an adjustment function that decreases as the silent cycle count increases and has a value less than 1.

[0016] Preferably, the method further includes: before step c is executed, the spaceborne edge computing unit obtains the Doppler frequency shift value measured by the communication payload of the carrier platform, and generates pre-correction data representing the platform's own motion based on the value; the pre-correction data is transmitted to the ground station together with the incremental information packet; and the ground station, before analyzing the incremental information packet in step d, first uses the pre-correction data to compensate for the motion-related features in the information packet.

[0017] Preferably, after confirming the target change, step d further includes a self-verification step for deciding whether to update the reference benchmark: injecting standardized noise into the target block in memory to generate adversarial test samples; calculating the temporal residual fluctuation trends of the original target block and the adversarial test samples relative to the historical reference benchmark; and updating the reference benchmark with the current change only when the consistency of the two fluctuation trends is higher than a confidence threshold.

[0018] Preferably, the method further includes: a ground station extracting differential information from multiple incremental information packets received from the same geographical location and arranged in a time series to form a residual time series; calculating the variance of the residual time series; and generating an early warning signal indicating that there is a risk of physical instability at that geographical location when the variance value is continuously lower than a risk threshold.

[0019] Preferably, the optical flow field calculation includes tracking motion vectors of local blocks in the incremental information packet; the weights are used to adjust the contribution of each motion vector to the change confirmation result in the subsequent statistical clustering analysis, wherein the contribution of motion vectors originating from sub-regions with high texture entropy values ​​is higher than the contribution of motion vectors originating from sub-regions with low texture entropy values.

[0020] Preferably, the feature signature of environmental interference in step d is extracted by performing optical flow field clustering analysis on incremental information packets that have been historically identified as being triggered by cloud shadows, and by statistically modeling the motion vectors that are separated and have large-scale and single motion patterns.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. This invention constructs a novel time-series remote sensing image processing method. By pre-calculating the information entropy value of differential information on the onboard edge computing unit and comparing this information entropy value with a trigger threshold adjusted by feedback from the ground station, it serves as the decision-making basis for whether to transmit high-resolution data to the ground. This avoids the ineffective transmission of massive amounts of unchanging background image data back to the ground, allowing limited space-to-ground transmission bandwidth resources to be prioritized for transmitting data on changing areas with high information value. This not only alleviates the contradiction between the onboard data generation speed and the space-to-ground transmission capacity from a mechanism perspective, but also provides a feasible data acquisition path for time-sensitive applications such as emergency response.

[0023] 2. This invention establishes a closed-loop adjustment mechanism spanning satellite and ground, enabling the system to cope with complex monitoring environments. By analyzing historically received incremental information packets, the ground station can not only identify and extract characteristic signatures of environmental interference such as cloud shadows, but also quantify the long-term stability of the monitoring area. Subsequently, these analysis results are transformed into bidirectional adjustment commands for the onboard trigger threshold. This method of transforming the fine analysis capabilities of the ground end into continuous optimization of the initial diagnostic strategy on the satellite end allows the entire monitoring system to gradually adapt to the environmental characteristics of specific areas during use, suppressing invalid triggers while remaining alert to subtle changes that may occur in long-term silent areas.

[0024] 3. In the analysis process of the ground station, this invention introduces the secondary utilization of the texture features of the incremental information packet itself. Before performing optical flow field calculation to confirm the target change, the texture entropy value of each sub-region within the differential information is calculated first, and weights are assigned to the subsequent motion vector analysis based on the texture entropy value. This approach enables the ground-side analysis model to reduce the interference of unreliable motion vectors caused by missing feature points on the final judgment when dealing with weak texture areas such as water surfaces or deserts, thereby improving the applicability of the method under various surface types and the accuracy of change confirmation.

[0025] 4. This invention also provides a self-verification step for updating the reference benchmark. Before the ground station confirms a change and plans to update the reference benchmark with it, the physical authenticity of the change is determined by injecting standardized noise into the target block in memory and comparing the consistency of the temporal residual fluctuation trend with and without noise injection. The update is only performed when the change shows stability to weak disturbances. This method can effectively identify and filter out false changes caused by accidental factors such as sensor transient noise, avoid contaminated images from entering the reference benchmark library, and ensure the accuracy of subsequent change detection and the reliability of long-term operation.

[0026] 5. The method of this invention further expands the application dimensions of remote sensing data. By calculating the variance of the residual time series of multiple incremental information packets received from the same geographical location, the differential information, which was originally an intermediate product of change detection, is transformed into a new indicator that can characterize the physical stability of ground objects. When the residual time series variance is continuously lower than a risk threshold, a physical instability risk warning can be generated. This makes this method not only able to capture sudden events, but also to monitor slow changes such as long-term cumulative changes such as ground subsidence, providing a new technical approach for infrastructure health monitoring and geological disaster early warning. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the interaction process of the method of the present invention;

[0028] Figure 2 This is a schematic diagram illustrating the change of the differential information entropy value with the observation period in this invention;

[0029] Figure 3 This is a diagram of the system's functional modules and data flow architecture. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below. It should be understood that the specific embodiments described herein are intended to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0031] This application provides an AI target change detection method based on time-series remote sensing imagery. The system architecture consists of a satellite-borne edge computing unit deployed on an in-orbit satellite and a ground station located on a ground facility. The two units interact bidirectionally via a satellite-to-ground communication link. The satellite-borne edge computing unit performs real-time change pre-diagnosis on the acquired remote sensing imagery and determines whether to generate and transmit an incremental information packet containing only information about the changed area based on the diagnosis results. The ground station confirms and analyzes the received incremental information packet and generates feedback control commands based on the analysis results to continuously adjust the diagnostic criteria of the satellite-borne edge computing unit. For clarity, the implementation process of this invention is illustrated using a wide-area forest fire monitoring application scenario. In this scenario, high timeliness of early small fire point information is required, while limited satellite-to-ground transmission bandwidth is a constraint. This method involves periodically revisiting the same wide-area forest using an in-orbit satellite. The specific steps are as follows: On the satellite-borne edge computing unit, the current observation data is acquired... The system generates remote sensing images for a given period and creates a reference benchmark based on these historical images. Considering the limitations of onboard storage and computing resources, a feasible implementation is to downsample a historical remote sensing image covering the same geographical area captured by the satellite in a previous observation period. This downsampling is performed in orbit, for example, by reducing the resolution by a factor of 16 horizontally and vertically using bilinear interpolation, resulting in a low-resolution thumbnail with a smaller data volume. This low-resolution thumbnail is then used as the reference benchmark and stored in the onboard memory. The onboard edge computing unit calculates the information entropy value of the difference between the current period's remote sensing image and the reference benchmark. To adapt to the computing power limitations at the onboard edge, the calculation process is designed as follows: First, a corresponding low-resolution thumbnail is generated in real-time from the newly acquired current period's remote sensing image using the same algorithm as the reference benchmark. Then, the grayscale difference between corresponding pixels in the current thumbnail and the reference benchmark thumbnail is calculated to obtain a difference image. This difference image is then traversed once to statistically analyze its grayscale histogram. Finally, based on this histogram and using formulas or other methods, the information entropy of the difference image is calculated. Or it can generate the normalized information entropy of the difference image, where, It can be generated as the probability of a specific gray value appearing in a differential image. The entropy value calculated in this way can characterize the degree of texture and structure change between two images.

[0032] The information entropy value is compared with a preset trigger threshold. When the information entropy value is higher than the trigger threshold, the target block corresponding to the changed area is extracted from the current period's remote sensing image, forming an incremental information packet and transmitting it to the ground station. If the information entropy value is not higher than the trigger threshold, the original data transmission of the current period's remote sensing image is abandoned at the spaceborne edge computing unit. The trigger threshold here is a dynamically settable parameter. Its value is set based on suppressing entropy fluctuations caused by sensor background noise or uniform illumination changes while ensuring the required sensitivity to target changes. When the entropy value exceeds the threshold, the system determines that a change with high information content has occurred and only extracts data from the original high-resolution current period's remote sensing image. In remote sensing imagery, one or more high-resolution image blocks, i.e., target blocks, are extracted corresponding to the thumbnail regions that triggered entropy changes. Furthermore, to further compress the data volume, differential images of the high-resolution target block and its corresponding historical high-resolution blocks can be calculated. This high-resolution differential image block is then used as the core payload, encapsulated into an incremental information packet, assigned a high transmission priority, and sent via the downlink. Before execution, a platform motion feedforward compensation step can be introduced to handle potential interference from minor attitude jitters during satellite platform flight on subsequent ground-based motion analysis. Given that the communication payload continuously measures Doppler shift to maintain link stability, this step can be repeated... Using this signal, the Doppler frequency shift value measured by the communication payload of the carrier platform is obtained on the onboard edge computing unit. Based on this value, the equivalent motion vector introduced by the platform's own motion is calculated, generating pre-correction data. This pre-correction data is embedded in the metadata of the incremental information packet and transmitted together, thus providing prior calibration information for ground-based analysis. The ground station receives and analyzes the incremental information packet to confirm target changes. The ground station processes the incremental information packets that have been initially screened and are lightweight. During analysis, to address the issue of insufficient reliability of optical flow field calculation results in weakly textured regions (such as calm water surfaces or dense vegetation canopies), an adaptive weight correction mechanism based on texture entropy can be introduced. That is, when using optical flow algorithms... Before tracking motion vectors, the texture entropy value of each sub-region within the received differential image patch is calculated. At this point, the texture entropy value is used as a quantitative indicator of the texture richness within that region. Subsequently, when performing optical flow field calculations and statistical cluster analysis on the output motion vectors, each motion vector is assigned a weight value. The value of this weight value is positively correlated with the texture entropy value of the sub-region from which it originates. For example, a linear mapping function from 0.1 to 1.0 can be set. In this way, vectors originating from high texture entropy regions will have a higher contribution in subsequent cluster analysis, while vectors originating from low texture entropy regions with lower reliability will have their influence effectively suppressed, thereby improving the applicability of change confirmation under different land surface types.

[0033] After confirming target changes, to prevent spurious changes caused by accidental factors such as sensor transient noise from contaminating the reference base library, a self-verification step can be executed. This step is initiated before the system plans to update the reference base using the current changes. It injects a normalized, weak Gaussian noise into the confirmed changed target block in memory, for example, setting the standard deviation to 1 to 2 gray levels to generate an adversarial test sample. Then, it calculates the temporal residual fluctuation trends of the original target block and the adversarial test sample relative to the historical reference base, for example, composed of the image sequences of the previous N periods. Finally, it compares the time series correlation coefficients of the two fluctuation trends. Due to physical changes... The generated signal strength is typically higher than the injected weak noise, and its two trends will exhibit a high degree of consistency. Therefore, the system only confirms the change as physically real and performs a reference baseline update when the consistency metric exceeds a preset confidence threshold (e.g., 0.95). Otherwise, the change is marked as suspected transient interference and no update is performed. Based on the analysis of one or more incremental packets received historically, the ground station generates adjustment commands and sends them to the onboard edge computing unit via the uplink to dynamically adjust the trigger threshold for a specific area. This step establishes a closed-loop adjustment mechanism between the ground station and the onboard edge computing unit. Interaction can be achieved through the satellite's conventional uplink and downlink communication links. To ensure the reliability of command transmission, two communication channels can be configured, one primary and one backup, for example, using different frequency bands or passing through different ground gateway stations. When the primary channel is blocked, the command can be switched to the backup channel for transmission. The generation logic of the adjustment command includes two aspects. First, to deal with false triggers caused by insignificant environmental interference such as cloud shadows, the ground station performs optical flow field clustering analysis on incremental information packets that have been historically confirmed to be triggered by interference such as cloud shadows, extracts motion vectors with characteristics such as large scale and single motion mode, and statistically models them to form the feature signature of environmental interference. When the subsequent changes monitored in orbit are consistent with the changes sent by the ground station, the ground station will then use these features to generate the characteristic signature of the environmental interference. When the signature matches, the onboard edge computing unit locally increases the trigger threshold for that region to suppress the re-triggering of such interference. Secondly, to address the issue of long-term, slow changes such as glacial melting or land subsidence being difficult to detect, the ground station maintains a silent period count for each monitoring region's reference baseline. This count records the number of consecutive observation periods for which the baseline has not been successfully triggered for updating. When the count for a region exceeds a preset stability duration threshold (e.g., 100 consecutive observation periods), the ground station determines that the region has entered a high-sensitivity monitoring state and generates instructions to cause the onboard unit to process the region according to a preset function. To lower its trigger threshold, where, The adjusted trigger threshold, This is the base trigger threshold for this area. This is the count value for the silent period, while Then it is a random An adjustment function that increases and monotonically decreases, with values ​​less than 1, can be set as follows: ,in This is the preset maximum silent period threshold.

[0034] Where there is no conflict, the method of this invention can also expand the application of remote sensing data by performing secondary analysis on multiple incremental information packets stored at a ground station from the same geographical location and arranged in a time series. Specifically, the ground station extracts the differential information from these incremental information packets to form a residual time series and calculates the time series variance of the series. This variance can be used as an indicator to characterize the physical stability of ground features. The technical principle is that in an area experiencing creep-type landslides, the continuous small displacements of the ground surface will cause the residual value of its image to remain continuously non-zero but with small fluctuations over a long period of time, and its time series variance will therefore show an abnormally low value. In contrast, the residual value of a physically stable area is normally zero. Occasional sensor or atmospheric noise can produce instantaneous high amplitude values, which in turn can lead to higher time-series variance values. Therefore, when the residual time-series variance value of a certain area is consistently lower than a risk threshold calibrated based on historical data or a physical model of that area, the system can generate an early warning signal indicating a risk of physical instability in that geographical location, thus providing a technical approach for monitoring long-term cumulative slow changes. In the deployment or recalibration phase of this invention, the initial trigger threshold and dynamic adjustment function of each monitoring area are generated by a standardized set of parameter deterministic procedures. The starting point of these procedures is to obtain a standardized time-series image sample library that has been cross-labeled and includes confirmed areas of no change, confirmed areas of interference, and confirmed areas of real change. First, by traversing candidate confidence coefficients... And calculate each The true positive rate (TPR) and false positive rate (FPR) corresponding to the values ​​in the sample database are used to plot a receiver operating characteristic (ROC) curve, which characterizes the relationship between detection sensitivity and false alarm rate under a given sensor noise baseline. This curve is a chart that visualizes the classifier performance, with FPR on the horizontal axis and TPR on the vertical axis. The final result is... The value is the utility function on the ROC curve that makes a specific task possible. The coefficient value corresponding to the point that reaches the maximum value, where the weight... and It is a constant preset based on the tolerance for missed and false alarms according to the monitoring task; secondly, the threshold decay adjustment function. Its specific form was determined to be morphological parameters and This directly stems from the statistical analysis results of the time interval sequence of historical change events in the region. The value of is inversely proportional to the coefficient of variation of the time interval sequence, while The value of the threshold decay is inversely proportional to the mean of the sequence, thus the more regular the change, the more drastic the threshold decay response, while the more random the change, the smoother the response.

[0035] The adjustment command sent from the ground station to the onboard edge computing unit is a structured command packet containing an asynchronous execution timestamp and a multimodal interference signature. The asynchronous execution timestamp specifies the exact future moment the command will be activated onboard, compensating for the inherent latency in space-to-ground communication and command processing. The multimodal interference signature, based on the original kinematic features derived from optical flow field analysis, further integrates spectral residual features and thermal infrared inertia features. The spectral residual feature refers to a normalized exponent defined by the stable relationship of reflectance of different surface cover types across different spectral bands, which exhibits specific anomalous fluctuations during cloud cover events. The thermal infrared... External inertia refers to the physical property of ground objects exhibiting significantly different rates of heating and cooling under solar radiation due to differences in heat capacity compared to clouds. When a spaceborne unit matches a suspected interference event, it executes a hierarchical verification logic. That is, only when the kinematic, spectral residual, and thermal infrared inertia characteristics of an event all match the signature in the command packet will it be finally determined as environmental interference and a threshold increase operation be performed. At the same time, any geographical location that is determined to be the same type of interference for three or more consecutive cycles will have its suppression logic automatically suspended, and the incremental information packet of the event will be forcibly transmitted to the ground station for manual review with high priority.

[0036] Example 1: In an application of high-frequency fire monitoring in a wide-area forest area, the technical solution of the present invention operates as follows: When a remote sensing satellite equipped with the method of the present invention performs an imaging observation of a vast and remote forest area every 90 minutes according to its predetermined orbit, within the interval between two observation cycles, a small fire spot with an initial area of ​​less than 5x5 pixels, still in the smoldering stage, is triggered by lightning in a local area of ​​the forest area. During the next flyby of the area, the satellite's imaging payload acquires a remote sensing image of the current cycle containing this small fire spot and several gigabytes (GB) of data. At this time, the onboard edge computing unit deployed on the satellite does not directly store the original image data into the downlink transmission queue. Instead, it first generates a low-resolution thumbnail of the image in orbit according to the above steps, and then performs a difference calculation between it and a reference thumbnail representing the previous cycle stored on the satellite. This difference is then used to calculate the normalized information entropy of the differentiated image, which is then used for general image data processing or generation. Due to the presence of this smoldering fire point, even though it occupies a very small proportion in the entire image, the change in its physical state still produces non-random texture changes in the corresponding differential image local block, causing the information entropy value calculated for this block to exceed the trigger threshold set for this forest area. This entropy value exceeding the limit triggers subsequent steps. The onboard edge computing unit does not initiate the transmission of the entire gigabyte-level original image, but only extracts the high-resolution target block corresponding to the entropy value change area from the original high-resolution image, and performs high-resolution differential operation on it with the corresponding block in the historical image to generate an incremental information packet with a data size of several kilobytes (KB). This information packet is given high transmission priority and is immediately transmitted to the ground station through the downlink. Given its data volume, this transmission process does not occupy the broadband data downlink channel that usually requires queuing for several hours, and is completely received by the ground station within a few minutes.

[0037] Upon receiving the incremental information packet, the ground station immediately initiated the analysis process. The change detection model was able to concentrate its computational resources on the pre-screened target block, thereby quickly segmenting, identifying, and confirming fire points within the block and generating early warning information. During this process, the onboard entropy change triggering mechanism and the ground station incremental calibration mechanism formed a seamless workflow. The onboard mechanism eliminated the need for downlink resources to be used for transmitting static background data, which accounted for over 99% of the data. Meanwhile, the ground station, by shifting its processing focus from entire images to a small target block, significantly reduced the processing time required for change confirmation. This information processing workflow effectively captured change information. The bottleneck has shifted from the satellite-to-ground transmission link to the edge computing on the satellite and the analysis of extremely small data packets by the ground station. This improves the timeliness of acquiring high-value information without changing the physical bandwidth of the satellite-to-ground link. Ultimately, the emergency response department received early warning information containing the geographical coordinates of the smoldering fire point within the first satellite observation window after the fire broke out, thus creating conditions for emergency forces to intervene before the fire spread. At the same time, the ground station generated an adjustment command based on the confirmed fire point information and sent it to the satellite. This command dynamically and specifically lowered the entropy change trigger threshold in the area surrounding the fire point, enabling the system to have higher monitoring sensitivity for any subtle changes in the fire in the area during subsequent observations.

[0038] Example 2: To objectively and quantitatively evaluate the technical effects of the method of the present invention in terms of data transmission efficiency and information acquisition latency, an experimental platform based on semi-physical simulation was built. The platform consists of an embedded processing module simulating a spaceborne edge computing unit and a server simulating a ground station. The two are connected by a bandwidth-limited and latency-configurable communication link to simulate a space-to-ground communication environment. The experimental data source is a set of publicly available time-series remote sensing image sequences covering various surface types and including various scenarios from unchanging environmental interference to minor and significant target changes. As a control, a control group simulating the existing full-data transmission method was run simultaneously. The control group used the same hardware and dataset as the experimental group, but its onboard module sent complete raw image data to the downlink in each cycle.

[0039] In the experiment, the trigger threshold of the on-board entropy change triggering mechanism is a key setting parameter. The determination of its value aims to balance the sensitivity of change detection with the economy of data transmission. The setting process is as follows: First, a series of image sequences without any surface changes are selected from the dataset, and the average information entropy of their continuous difference images is calculated. with standard deviation Then, a confidence coefficient is set according to the false alarm rate requirements of the target application scenario. The trigger threshold can then be determined by In this experiment, based on the sensor noise characteristics of the dataset used, the following was measured: for , for To operate at a higher confidence level, set The trigger threshold is then determined to be After the experiment was launched, the experimental group and the control group processed the same time-series image sequence simultaneously. This sequence simulated four test scenarios in turn: Scenario 1, no identifiable changes on the ground surface; Scenario 2, large sparse cloud shadows moving, constituting environmental interference; Scenario 3, two new vehicles added to a parking lot, constituting a minor target change; Scenario 4, a temporary building in a construction site being demolished, constituting a significant target change. During the test cycle of each scenario, the system recorded the total amount of downlink data transmitted, the end-to-end latency from the occurrence of the change to the ground station confirming the change information, and the average processor utilization rate of the onboard edge computing unit. The core data records are shown in Table 1.

[0040] Table 1 shows the performance data of the method of the present invention and existing methods under different test scenarios.

[0041]

[0042] As shown in Table 1, in scenarios one and two, the downlink data transmission volume of the method of the present invention is zero. This is because the differential information entropy value calculated by the onboard edge computing unit does not exceed the preset trigger threshold of 0.37, thus the downlink transmission is not initiated. In contrast, the existing method transmits all 1024MB of original image data. In scenarios three and four, the method of the present invention is triggered, and the downlink data transmission volumes are 0.82MB and 5.15MB, respectively, which are only 0.08% and 0.5% of the original image data volume. Correspondingly, the latency of obtaining change information is also shortened from about 3600 seconds in the existing method to less than 20 seconds. It should be noted that because the method of the present invention performs entropy calculation on the satellite, its onboard CPU utilization rate is slightly higher than that of the existing method, but its value of 2.8% is still within the normal normal workload range of the low-power embedded processor.

[0043] Example 3: This example combines Figures 1 to 3 This paper describes the implementation of an AI-based target change detection method that integrates temporal remote sensing imagery. Figure 1 As shown, the spaceborne payload first acquires the current periodic remote sensing image and hands it over to the spaceborne edge computing unit. This unit performs pre-diagnosis of the data by generating a low-resolution thumbnail, comparing it with a historical reference benchmark, and calculating the differential information entropy value. It then determines the subsequent actions based on a judgment condition alt, namely whether the information entropy value is higher than a trigger threshold. If the information entropy value > the trigger threshold, the changed area is extracted from the high-resolution image, and incremental information packets are generated and transmitted to the ground station via the space-to-ground link. After receiving and analyzing the target changes, the ground station can send a change warning to the user terminal. If the information entropy value ≤ the trigger threshold, the spaceborne edge computing unit directly abandons the transmission of the original data, thereby saving transmission bandwidth.

[0044] like Figure 2 As shown in the figure, the horizontal axis represents the observation period, and the vertical axis represents the information entropy value. The solid line represents the differential information entropy value calculated for each period, while the dashed line represents the preset trigger threshold. As can be seen from the figure, before approximately the 35th observation period and after the 60th observation period, the differential information entropy value is lower than the trigger threshold, indicating that the system's judgment has not changed significantly. However, between the 35th and 60th observation periods, the differential information entropy value is consistently higher than the trigger threshold, indicating that the system has continuously detected incremental information that needs to be transmitted downlink during this period.

[0045] like Figure 3As shown in the figure, the system comprises two core components: a spaceborne edge computing unit and a ground station. The spaceborne edge computing unit is responsible for acquiring the current period's remote sensing imagery and calculating the information entropy value of the differential information based on historical remote sensing imagery references. When the information entropy value is determined to be greater than the trigger threshold, the target block of the changed area is extracted and an incremental information packet is formed. Simultaneously, this unit can also acquire Doppler frequency shift values ​​to generate pre-correction data. Conversely, if the entropy value is not greater than the threshold, the original data transmission is abandoned. The ground station is responsible for receiving and analyzing the incremental information packet. By confirming the target change, it completes the monitoring task. Its internal analysis process also includes calculating the texture entropy value of the sub-region to assist in optical flow field calculation and motion vector tracking, performing self-verification through noise injection testing to determine whether to update the reference benchmark, and conducting physical instability risk warning through residual temporal variance analysis. Crucially, the ground station also generates trigger threshold adjustment instructions by identifying environmental interference feature signatures and quantifying the continuous unupdated duration of the region, and sends these feedback adjustment instructions back to the spaceborne edge computing unit, thus forming a closed-loop optimization adjustment mechanism.

[0046] Example 4: In an application for dynamic monitoring of a near-shore port encompassing a large area of ​​water and land, the method needs to address a complex working condition resulting from the superposition of multiple factors, such as weak water surface texture, cloud shadow movement, and slow movement of the target being measured. In this scenario, some core parameters and internal processing logic of the method of this invention are configured and implemented as follows: In the ground station analysis process, to address the problem of unreliable motion vectors in the optical flow field calculation output due to the lack of stably trackable feature points in areas with weak texture, such as the water surface, the weight assignment mechanism is defined by a deterministic mapping function; specifically, the normalized texture entropy value is calculated for each sub-region within the received incremental information packet. After this value ranges from 0 to 1, it represents the weight value of the motion vector corresponding to this sub-region. follow The quadratic function relationship is used to suppress the computational impact of low texture entropy regions to a greater extent. According to this relationship, the weight value corresponding to a sub-region with a texture entropy value of 0.3 is 0.09, while the weight value corresponding to a sub-region with a texture entropy value of 0.8 is 0.64. Thus, when performing weighted statistical clustering on all motion vectors in the subsequent process, the contribution of motion vectors originating from low texture entropy regions such as the water surface to the final judgment result is effectively reduced.

[0047] Meanwhile, to enable the spaceborne edge computing unit to distinguish between entropy fluctuations caused by target changes and environmental factors such as cloud shadows, the extraction and matching logic at the ground station and the spaceborne end is defined by a closed-loop process. At the ground station, the system collects incremental information packets triggered by entropy changes but subsequently identified by the analysis module as large-scale, single-mode motion packets through optical flow field clustering and ultimately excluded. These motion vector fields, confirmed as being caused by cloud shadow interference, are statistically analyzed, and the mean and standard deviation of their motion vectors in direction and velocity are extracted to form a vector field containing the interference type: cloud shadow, and the vector mean: [...]. , Vector standard deviation: [ , Structured data such as [fields] are used as feature signatures, which are sent to the satellite via the uplink; on the on-board edge computing unit side, when a new entropy limit violation event occurs, the on-orbit system estimates a motion vector through a block matching algorithm. If the direction value of the motion vector is within [fields], the system will determine the motion vector. , The velocity value is within the range of [ ] and its velocity value is within [ ]. , Within the range of ], the system will classify the triggered event as an interference event that matches the feature signature, and automatically increase the trigger threshold used for this judgment by a fixed coefficient, which can be 1.5.

[0048] Furthermore, to enable the system to effectively monitor long-term changes such as slow coastline erosion that may occur within the port area, a stability duration threshold—the threshold for the continuous period without updates used to activate the high-sensitivity monitoring mode—is determined by a mechanism that adaptively adjusts based on historical data. For a newly included monitoring area, the system first adopts a universal initial stability duration threshold, which can be 100 observation periods. During continuous operation, the ground station records the time point at which each successful triggering and confirmation of actual change in the area occurs, and uses this to calculate the average time interval of change in the area. After the system has run for a preset initial learning period, which can be 1000 observation cycles, the corresponding stability duration threshold is determined. It will be automatically updated to a multiple of the average change interval, which can be set to... In this way, the monitoring sensitivity adjustment rhythm of each area is configured according to its own level of activity.

[0049] Example 5: In the application of long-term health monitoring and risk warning for a large bridge structure, in order to ensure the reliability of change detection results and effectively identify minor deformations caused by structural fatigue or ground settlement, some internal verification and judgment procedures of the present invention are implemented as follows: When an analysis system deployed on a ground station analyzes an incremental information packet and confirms a change, before planning to use the change to update the reference benchmark, an internal self-verification step is initiated. This step first injects a normalized Gaussian noise with a standard deviation of 1 gray level into the target block that caused the change in memory to generate an adversarial test sample, and calculates the time-series residual sequences of the original target block and the adversarial test sample relative to the historical reference benchmark. Then, the system calculates the Pearson correlation coefficient between these two residual sequences. Only when the correlation coefficient is higher than a preset confidence threshold of 0.98 is the change considered to be a real change originating from the physical structure and used to update the reference benchmark.

[0050] To provide early warning of potential physical instability risks to the bridge structure, the risk threshold is set according to an engineering procedure based on offline calibration. After a comprehensive structural overhaul of the bridge, the system first collects a series of remote sensing images of the structure in a stable state over a period of time as a calibration dataset. The ground station processes the differential information in the calibration dataset to form one or more residual time series representing the impact of environmental noise on the bridge in a healthy state, and calculates the mean and standard deviation of the variance of the residual time series in the healthy state. Finally, the risk threshold used for online real-time monitoring is set as the value obtained by subtracting three times the standard deviation from the mean variance of the healthy state. When the variance of the residual time series formed by multiple incremental information packets received from the same geographical location during online monitoring is consistently lower than this calibrated risk threshold, the system generates an early warning signal.

[0051] Example 6: In an application where the method of the present invention is first deployed to monitor a previously uncovered geographical area, a standardized pre-deployment procedure is required to ensure system initialization and module parameter calibration. In this procedure, the first and second consecutive transit observations of the satellite over the new area are used for system initialization. The complete data of the original remote sensing image acquired during the first transit is downlinked to the ground station and processed by the ground station to be established as the initial reference benchmark for the entire monitoring mission. The change pre-diagnosis function of the onboard edge computing unit is activated from the second transit observation and uses the image of the first transit as the first historical remote sensing image required for its on-orbit calculation.

[0052] In the ground station configuration section of this procedure, the data processing flow for incremental information packets generated on-board as differential images is set as follows: Upon receiving an incremental information packet, the ground station first retrieves the historical remote sensing image block corresponding to the spatial location of the incremental information packet from its database. Then, through matrix addition, the differential image block and the historical remote sensing image block are overlaid pixel-by-pixel, thereby reconstructing the target block corresponding to the current period's remote sensing image at the ground end. Subsequently, the reconstructed current target block and the original historical remote sensing image block are used as input data pairs and fed into a change detection model based on a Siamese network architecture or a dual-channel encoder for analysis to complete change confirmation and classification. This is for the calibration of the pre-correction data generation mechanism based on Doppler frequency shift. To compensate for systematic errors that may be introduced by platform motion, the procedure also includes an offline calibration step. During the initial deployment phase of the system, the ground station calculates the theoretical Doppler frequency shift caused by the theoretical relative radial velocity between the satellite and the ground station under different imaging geometry angles based on high-precision satellite orbital elements. On the other hand, it records the Doppler frequency shift values ​​actually measured and transmitted from the satellite communication payload. By comparing these two sets of data, the system can calculate a correction coefficient or lookup table to compensate for platform attitude deviations or internal clock drift. This correction information is then injected uplink into the onboard edge computing unit to optimize the accuracy of the generated pre-corrected data in subsequent missions. At this point, the monitoring system has completed its deployment and calibration in the new area and entered normal operation.

[0053] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0054] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An AI-based target change detection method integrating temporal remote sensing imagery, characterized in that, Includes the following steps: Step a: On the spaceborne edge computing unit, acquire the current period remote sensing image and generate a reference benchmark based on the historical remote sensing image; Step b: The onboard edge computing unit calculates the information entropy value of the differential information between the current period remote sensing image and the reference benchmark; Step c: Compare the information entropy value with a trigger threshold. When the information entropy value is higher than the trigger threshold, extract the target block corresponding to the changed area from the current period remote sensing image, form an incremental information packet and transmit it to the ground station. If the information entropy value is not higher than the trigger threshold, abandon the original data transmission of the current period remote sensing image on the spaceborne edge computing unit side. Step d: The ground station receives and analyzes incremental information packets to confirm target changes. Based on the analysis of at least one incremental information packet received in history, the ground station identifies and extracts the characteristic signature of environmental interference and quantifies the continuous unupdated duration of the reference benchmark for each area. In step d, after confirming the target change, a self-verification step is included to determine whether to update the reference benchmark: injecting normalized noise into the target block in memory to generate adversarial test samples; calculating the temporal residual fluctuation trends of the original target block and the adversarial test samples relative to the historical reference benchmark; and updating the reference benchmark with the current change only when the consistency of the two fluctuation trends is higher than a confidence threshold. Step e: Based on the feature signature of environmental interference and the duration of continuous non-update, the ground station generates an adjustment command and sends it to the onboard edge computing unit to adjust the trigger threshold of the corresponding area: when the current change feature is detected to match the feature signature in the environmental log, the trigger threshold is increased; when the duration of continuous non-update in a certain area reaches a stable duration threshold, the trigger threshold is decreased.

2. The AI ​​target change detection method based on fused temporal remote sensing imagery according to claim 1, characterized in that, The reference benchmark in step a is a low-resolution thumbnail of historical remote sensing imagery; the object for calculating the information entropy value in step b is the difference image between the corresponding low-resolution thumbnail generated from the current period's remote sensing imagery and the reference benchmark.

3. The AI ​​target change detection method based on fused temporal remote sensing imagery according to claim 1, characterized in that, The incremental information packet in step c is a high-resolution differential image between the target block of the current period remote sensing image and its corresponding historical remote sensing image block.

4. The AI ​​target change detection method based on fused temporal remote sensing imagery according to claim 3, characterized in that, Step d, before analyzing the incremental information packet, also includes: calculating the texture entropy value of each sub-region within the incremental information packet based on the differential information within it; and in the subsequent change confirmation analysis, assigning weights to the motion vector features from each sub-region, the weight values ​​being determined based on the texture entropy value of the sub-region, wherein the motion vector features are derived from the optical flow field calculation performed on the incremental information packet.

5. The AI ​​target change detection method based on fused temporal remote sensing imagery according to claim 1, characterized in that, The adjustment of lowering the trigger threshold in step e follows these rules: ,in, It is the adjusted trigger threshold. This is the base trigger threshold for this area. It is the count value of the silent period corresponding to the continuous period without updates. It is an adjustment function that decreases as the silent cycle count increases and has a value less than 1.

6. The AI ​​target change detection method based on fused temporal remote sensing imagery according to claim 1, characterized in that, The method also includes: before step c is executed, the onboard edge computing unit obtains the Doppler frequency shift value measured by the communication payload of the carrier platform, and generates pre-correction data representing the platform's own motion based on the value; the pre-correction data is transmitted to the ground station along with the incremental information packet; and the ground station, before analyzing the incremental information packet in step d, first uses the pre-correction data to compensate for the motion-related features in the information packet.

7. The AI ​​target change detection method based on fused temporal remote sensing imagery according to claim 1, characterized in that, The method also includes: a ground station extracting differential information from multiple incremental information packets received from the same geographical location and arranged in a time series to form a residual time series; calculating the variance of the residual time series; and generating an early warning signal indicating that there is a risk of physical instability at that geographical location when the variance value is consistently below a risk threshold.

8. The AI ​​target change detection method based on fused temporal remote sensing imagery according to claim 4, characterized in that, Optical flow field calculation includes tracking motion vectors of local blocks in incremental information packets; weights are used to adjust the contribution of each motion vector to the change confirmation results in subsequent statistical cluster analysis, wherein the contribution of motion vectors originating from sub-regions with high texture entropy values ​​is higher than that of motion vectors originating from sub-regions with low texture entropy values.

9. The AI ​​target change detection method based on fused temporal remote sensing imagery according to claim 1, characterized in that, The feature signature of environmental interference in step d is extracted by performing optical flow field clustering analysis on incremental information packets that have been historically identified as being triggered by cloud shadows, and by statistically modeling the motion vectors that are separated and have large-scale and single motion patterns.

Images