AI target change detection method fused with time sequence remote sensing image

Abstract

The invention relates to the technical field of image data processing, and discloses an AI target change detection method fusing a time sequence remote sensing image, which comprises the following steps of: calculating a difference information entropy value of a remote sensing image on a satellite-borne edge calculation unit, and comparing the difference information entropy value with a trigger threshold value; only when the information entropy value is higher than the threshold value, the change area is extracted to form an incremental information packet to be transmitted to the ground; and after analyzing the incremental information packet, the ground station generates an instruction according to the identified environmental interference characteristics and regional stability, feeds back the instruction to the satellite-borne unit, and bidirectionally adjusts a trigger threshold value. According to the invention, an information screening decision is preposed to a data generation source, so that invalid transmission of massive invariable background data is avoided; therefore, the acquisition delay of the key change information is obviously reduced; the established satellite-ground closed-loop feedback mechanism enables the whole monitoring method to continuously optimize the judgment standard during use, and the applicability and reliability of the method are improved.

Description

TECHNICAL FIELD

[0001] The application relates to an AI target change detection method for fusing time-series remote sensing images and belongs to the technical field of image data processing. BACKGROUND

[0002] In the field of image data processing, for change detection of time-series remote sensing images, a commonly used technical process is as follows: original images are acquired by an on-orbit satellite, all data of the original images are transmitted to a ground center through a communication link, and then the data are analyzed and processed by a computing device of the ground center to identify changed areas on the ground.

[0003] With the deployment of large-scale satellite constellations, the total amount of on-orbit data has increased significantly, while the bandwidth capacity of satellite-ground communication is limited by physical and cost factors and has not increased synchronously, which leads to a problem in the actual application of the above technical process: a large amount of original image data needs to be queued for downlink transmission, so that images containing change information cannot be acquired and analyzed in time in time-sensitive applications such as emergency response; at the same time, since the unchanged background area in the image accounts for a large proportion, a large amount of communication and storage resources are occupied for transmitting these data that do not contain new information.

[0004] For the above problem, simply increasing the ground station or improving the ground computing capacity cannot change the time sequence of information processing, and the limitations of the process are as follows: 1. The data value identification step is after data transmission, which reduces the value of images containing time-sensitive change information due to the delay in the transmission process; 2. On the on-orbit end of data generation, there is a lack of a technical link that can run on-orbit and be used for preliminary screening of image information before transmission, so that all data, whether containing changes or not, are included in the transmission process. Therefore, how to construct an image data processing method that can perform change pre-diagnosis on images on the edge of the satellite and decide whether to transmit high-resolution data of the image according to the diagnosis result, and at the same time, can use the analysis result of the ground station to feedback and adjust the diagnosis standard on the on-orbit end, has become a technical problem to be solved by the present application. SUMMARY

[0005] The application provides an AI target change detection method for fusing time-series remote sensing images, which mainly aims to solve the problems of large data transmission and untimely information acquisition caused by the fact that the data value identification step is after data transmission and lacks a star-ground closed-loop adjustment mechanism in the existing remote sensing image processing process.

[0006] To achieve the above purpose, the AI target change detection method for fusing time-series remote sensing images provided by the application comprises the following steps: Step a, obtaining current period remote sensing image on the satellite-borne edge computing unit, and generating reference benchmark based on historical remote sensing image; Step b, the satellite-borne edge computing unit calculates the information entropy value of the difference information between the current period remote sensing image and the reference benchmark; Step c, comparing the information entropy value with a trigger threshold value, when the information entropy value is higher than the trigger threshold value, extracting the target block corresponding to the change area from the current period remote sensing image, forming an incremental information package and transmitting to the ground station, if the information entropy value is not higher than the trigger threshold value, abandoning the original data transmission of the current period remote sensing image on the satellite-borne edge computing unit side; Step d, the ground station receives and analyzes the incremental information package to confirm the target change, and the ground station identifies the feature signature of the environmental interference and the continuous non-updating time length of the reference benchmark of each region based on the analysis of at least one incremental information package received in the past; Step e, the ground station generates adjustment instructions based on the feature signature of the environmental interference and the continuous non-updating time length, and sends them to the satellite-borne edge computing unit to adjust the trigger threshold value of the corresponding region: when the current change feature matches the feature signature in the environmental log, the trigger threshold value is increased; when the continuous non-updating time length of a region reaches a stable time length threshold, the trigger threshold value is reduced.

[0007] Preferably, the reference benchmark in step a is a low-resolution thumbnail of the historical remote sensing image; the object for calculating the information entropy value in step b is the difference image between the corresponding low-resolution thumbnail generated by the current period remote sensing image and the reference benchmark.

[0008] Preferably, the incremental information package in step c is a high-resolution difference image between the target block of the current period remote sensing image and the corresponding historical remote sensing image block.

[0009] Preferably, before analyzing the incremental information package in step d, it further includes: calculating the texture entropy value of each sub-region inside the incremental information package based on the difference information inside the incremental information package; and in the subsequent change confirmation analysis, assigning a weight to the motion vector feature derived from each sub-region, the value of the weight being determined according to the texture entropy value of the sub-region, wherein the motion vector feature is derived from the optical flow field calculation on the incremental information package.

[0010] Preferably, the adjustment of reducing the trigger threshold value in step e follows the following rules: , wherein, is the adjusted trigger threshold value, is the basic trigger threshold value of the region, is the silent period count value corresponding to the continuous non-updating time length, is an adjustment function that decreases as the silent period count value increases and the value is less than 1.

[0011] Preferably, the method further comprises: before step c, the on-board edge computing unit acquires the Doppler shift value measured by the communication payload of the platform, and generates pre-correction data representing the motion of the platform itself based on the value; the pre-correction data is transmitted to the ground station together with the incremental information package; and the ground station, before analyzing the incremental information package in step d, first compensates the motion-related features in the information package using the pre-correction data.

[0012] Preferably, after confirming the target change in step d, the method further comprises a self-verification step for deciding whether to update the reference baseline: injecting standardized noise into the target block in memory to generate an adversarial test sample; calculating the time-series residual fluctuation trend of the original target block and the adversarial test sample with respect to the historical reference baseline, respectively; only when the consistency of the two fluctuation trends is higher than a confidence threshold, the reference baseline is updated using the current change.

[0013] Preferably, the method further comprises: the ground station, for a plurality of incremental information packages received from the same geographical location and arranged in time sequence, extracts the differential information inside to form a residual time series; calculates the variance value of the residual time series; when the variance value is continuously lower than a risk threshold, generates an early warning signal indicating that the geographical location is at risk of physical instability.

[0014] Preferably, the optical flow field calculation includes motion vector tracking of the incremental information package in local blocks; the weight is used to adjust the contribution of each motion vector to the change confirmation result in subsequent statistical clustering analysis, wherein the contribution of the motion vector from the sub-region with high texture entropy value is higher than the contribution of the motion vector from the sub-region with low texture entropy value.

[0015] Preferably, the feature signature of the environmental interference in step d is extracted by performing optical flow field clustering analysis on the incremental information packages historically confirmed to be triggered by cloud shadows, and statistically modeling the motion vectors with large-scale and single motion pattern characteristics separated out.

[0016] Compared with the prior art, the present application has the following advantages: 1. The present application constructs a new time-series remote sensing image processing method, which pre-calculates the information entropy value of the differential information on the on-board edge computing unit, compares the information entropy value with the trigger threshold adjusted by the ground station, and uses the comparison as the basis for deciding whether to transmit high-resolution data to the ground, thereby avoiding the transmission of a large amount of unchanged background image data to the ground, so that the limited satellite-ground transmission bandwidth resources can be preferentially used for transmitting data of the changed area with high information value, which not only alleviates the contradiction between the data generation speed on the satellite and the satellite-ground transmission capacity, but also provides a feasible data acquisition path for time-sensitive applications such as emergency response.

[0017] 2、The present application establishes a closed-loop adjustment mechanism across the star-ground, enabling the system to cope with complex monitoring environments. The ground station analyzes the historical incremental information packets received, not only identifying and extracting the characteristic signatures of environmental interference such as cloud shadows, but also quantifying the long-term stability of the monitoring area. Subsequently, these analysis results are converted into bidirectional adjustment instructions for the trigger threshold on the satellite. This way of converting the ground station's detailed analysis capabilities into continuous optimization of the preliminary diagnostic strategy on the satellite enables the entire monitoring system to gradually adapt to the environmental characteristics of a specific area during use, suppressing invalid triggers while remaining vigilant for possible weak changes in long-term silent areas.

[0018] 3、In the analysis process of the ground station, the present application introduces the secondary use of the texture characteristics of the incremental information packets. Before calculating the optical flow field to confirm target changes, the texture entropy values of each sub-region within the differential information are calculated, and the texture entropy values are used to assign weights to subsequent motion vector analysis. This approach enables the ground station's analysis model to reduce the interference of unreliable motion vectors caused by the absence of feature points in weak texture areas such as water surfaces or deserts, thereby improving the applicability and accuracy of change confirmation of the method under various types of ground surfaces.

[0019] 4、The present application also provides a self-verification step for reference benchmark updates. Before the ground station confirms a change and plans to update the reference benchmark, it injects standardized noise into the target block in memory and compares the consistency of the timing residual fluctuation trend with and without noise injection to determine the physical authenticity of the change. Only when the change exhibits stability to weak disturbances will the update be performed. This approach can effectively identify and filter out false changes caused by accidental factors such as sensor transient noise, preventing contaminated images from entering the reference benchmark library and ensuring the accuracy of subsequent change detection and the reliability of long-term operation.

[0020] 5、The method of the present application further expands the application dimension of remote sensing data. By calculating the variance of the residual timing sequence of multiple incremental information packets received from the same geographic location, the differential information, which is originally an intermediate product of change detection, is converted into a new indicator that can represent the physical stability of ground objects. When the residual timing variance consistently falls below a risk threshold, a physical instability risk warning can be generated. This enables the method not only to capture sudden events but also to monitor slow changes such as land subsidence, providing a new technical approach for infrastructure health monitoring and geological disaster warning. BRIEF DESCRIPTION OF DRAWINGS

[0021] Fig. 1 An interaction flowchart for the method of the present application; Fig. 2 A schematic diagram of the differential information entropy value of the present application changing with the observation period; Fig. 3 A system function module and data flow architecture diagram of the present application. DETAILED DESCRIPTION

[0022] To make the purpose, technical scheme and advantages of the present application clearer, the technical scheme of the present application will be described in detail below, and it should be understood that the specific embodiments described herein are intended to explain the present application, rather than to limit the protection scope of the present application.

[0023] The embodiment of the present application provides an AI target change detection method for fusing time sequence remote sensing images. The system architecture is composed of a satellite-borne edge computing unit deployed on an on-orbit satellite and a ground station arranged in a ground facility. The two communicate with each other through a space-ground communication link to realize bidirectional interaction of data and instructions. The satellite-borne edge computing unit is used to perform real-time change pre-diagnosis on the acquired remote sensing images, and to determine whether to generate and downlink an incremental information package containing only change area information according to the diagnosis result. The ground station is used to perform change confirmation and in-depth analysis on the received incremental information package, and to generate feedback control instructions according to the analysis result to continuously adjust the diagnosis standard of the satellite-borne edge computing unit. For the sake of clarity, the execution process of the present application is described in the application scenario of wide-area forest fire monitoring. In this scenario, both high timeliness of early small fire point information and limited space-ground transmission bandwidth are required. The method performs periodic revisit observation on the same wide-area forest by the on-orbit satellite, and the specific steps are as follows: on the satellite-borne edge computing unit, the remote sensing image of the current observation period is acquired, and the reference benchmark is generated based on the remote sensing image of the historical period. Considering the limitation of on-board storage and computing resources, a feasible implementation is to perform on-orbit downsampling processing on a historical remote sensing image covering the same geographical area taken by the satellite in the previous observation period, for example, to reduce the resolution of the image by 16 times in the horizontal and vertical directions by using a bilinear interpolation algorithm, to generate a low-resolution thumbnail with smaller data volume, and to store the low-resolution thumbnail as the reference benchmark in the on-board storage. The satellite-borne edge computing unit calculates the information entropy value of the differential information between the current period remote sensing image and the reference benchmark. To adapt to the computing power condition of the edge terminal on the satellite, the calculation process is designed as follows: first, the newly acquired current period remote sensing image is used to generate a corresponding low-resolution thumbnail in real time by using the same algorithm as that for generating the reference benchmark. Then, the gray value difference between the corresponding pixel points of the current thumbnail and the reference benchmark thumbnail is calculated to obtain a differential image. Finally, the normalized information entropy of the differential image is calculated based on the histogram of the differential image and by using the formula or or generate Or it can be generated as the probability of a specific grayscale value appearing in the differential image. The entropy value calculated in this way can represent the degree of texture and structure change between the two images.

[0024] The information entropy value is compared with a preset triggering threshold, when the information entropy value is higher than the triggering threshold, the target block corresponding to the changed area is extracted from the current period remote sensing image, the incremental information package is formed and transmitted to the ground station, if the information entropy value is not higher than the triggering threshold, the original data transmission of the current period remote sensing image is abandoned on the side of the on-board edge computing unit; The triggering threshold here is a dynamically set parameter, and the setting of the value of the triggering threshold is based on the premise of ensuring the required sensitivity to the target change, suppressing the entropy value fluctuation caused by the background noise of the sensor or uniform illumination change, when the entropy value exceeds the threshold, the system determines that a change with high information content has occurred, and only the high-resolution image block corresponding to the thumbnail area triggering the entropy value change is extracted from the original high-resolution current period remote sensing image, that is, the target block, in other words, to further compress the data volume, the difference image of the high-resolution target block and the corresponding historical high-resolution block can be calculated, and the high-resolution difference image block is taken as the core load and encapsulated into an incremental information package, and then transmitted via the downlink after being given a higher transmission priority; Before execution, a platform motion feedforward compensation step can also be introduced to process the interference that the slight attitude jitter of the satellite platform in orbit flight may cause to the subsequent ground end motion analysis; Since the communication load is used to measure the Doppler shift continuously and stably to maintain the stability of the link, this step can reuse the signal to obtain the Doppler shift value measured by the communication load of the bearing platform on the on-board edge computing unit, and based on the value, an equivalent motion vector introduced by the platform motion is calculated to generate a pre-correction data, and the pre-correction data is embedded in the metadata of the incremental information package and transmitted together, thereby providing prior calibration information for the analysis of the ground end; The ground station receives and analyzes the incremental information package to confirm the target change, and the processing object of the ground station is the lightweight incremental information package screened on the satellite, when analyzing, to cope with the problem of insufficient reliability of the optical flow field calculation result in the weak texture area (such as calm water or dense vegetation canopy), an adaptive weight correction mechanism based on texture entropy can be introduced, that is, before the motion vector is tracked by using the optical flow algorithm, the texture entropy values of each sub-region in the received difference image block are calculated, at this time, the texture entropy value is used as a quantitative indicator of the texture richness in the region, then, when the optical flow field is calculated and the motion vector is statistically clustered and analyzed, each motion vector is given a weight value, the value of the weight value is positively correlated with the texture entropy value of the sub-region from which the weight value is derived, for example, a linear mapping function from 0.1 to 1.0 can be set, in this way, the vector derived from the high texture entropy region will have a higher contribution in the subsequent clustering analysis, and the vector derived from the low texture entropy region with low reliability will be effectively suppressed, thereby improving the applicability of change confirmation in different ground types.

[0025] After confirming the target change, in order to prevent the reference base library from being contaminated by pseudo-changes caused by accidental factors such as sensor transient noise, a self-verification step can also be performed; this step is started before the system plans to use the current change to update the reference base, and it injects a standardized weak Gaussian noise into the target block that has been confirmed to have changed in the memory, for example, setting the standard deviation to 1 to 2 gray levels to generate an adversarial test sample, and then respectively calculates the original target block and the adversarial test sample relative to the historical reference base, for example, the time series residual fluctuation trend composed of the image sequence of the previous N cycles, and finally compares the time series correlation coefficient of the two fluctuation trends. The signal strength generated is usually higher than the weak noise injected, and the two trends will show a high degree of consistency. Therefore, only when the consistency metric is higher than a preset confidence threshold (for example, 0.95) will the system finally confirm that the change is physically real and perform an update of the reference benchmark. Otherwise, the change will be marked as a suspected transient interference and no update will be performed. The ground station generates an adjustment instruction based on the analysis results of one or more incremental information packets received in the past and sends it to the satellite-borne edge computing unit via the uplink to dynamically adjust the trigger threshold of a specific area. This step constructs a closed-loop adjustment mechanism between the ground station and the satellite-borne edge computing unit, and its instructions Interaction can be achieved through the satellite's conventional uplink and downlink communication links. In order to ensure the reliability of command transmission, two main and backup communication channels can be configured. For example, different frequency bands or different ground gateway stations are used. When the main 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, in order to deal with false triggering caused by worthless environmental interference such as cloud shadows, the ground station performs optical flow field clustering analysis on incremental information packets that have been confirmed to be triggered by interference such as cloud shadows in the past, extracts motion vectors with characteristics such as large scale and single motion pattern, and performs statistical modeling on them to form characteristic signatures of environmental interference. When the subsequent on-orbit monitoring changes with the characteristics sent by the ground station, When the characteristic signature matches, the onboard edge computing unit will accordingly and locally increase the trigger threshold of the area to suppress the re-triggering of such interference; secondly, in order to deal with the problem that long-period slow changes such as glacier melting or ground subsidence are difficult to detect, the ground station maintains a silent cycle count value for the reference benchmark of each monitoring area to record the number of observation cycles in which the benchmark has not been successfully triggered and updated. When the count value of a certain area exceeds a preset stable time threshold (for example, 100 consecutive observation cycles), the ground station determines that the area has entered a high-sensitivity monitoring state and generates an instruction to enable the onboard unit to process the area according to a preset function To lower its trigger threshold, is the adjusted trigger threshold, is the basic trigger threshold of the area, is the silent period count value, and is a monotonically decreasing adjustment function with a value less than 1, which can be set as , where is a preset maximum silence period threshold.

[0026] In the case of no conflict, the method of the present application can also expand the application of remote sensing data by performing secondary analysis on the stored multiple incremental information packets from the same geographical location and arranged in time sequence at the ground station; Specifically, the ground station extracts the differential information in these incremental information packets to form a residual time sequence, and calculates the time sequence variance value of the sequence, which can be used as an indicator of the physical stability of the ground object. The technical principle is that in an area where a creep landslide is occurring, the continuous small displacement of the ground surface will cause the residual value of the image to be continuously non-zero but with small fluctuation amplitude for a long time, and the time sequence variance will therefore exhibit an abnormally low value. On the contrary, in an area with stable physical state, the residual value is normally zero, and occasional sensor or atmospheric noise will produce transient high amplitude, which will make the time sequence variance value higher. Therefore, when the residual time sequence variance value of a certain area is continuously lower than a risk threshold value calibrated based on the historical data or physical model of the area, the system can generate a warning signal indicating that there is a physical instability risk at the geographical location, thereby providing a technical approach for monitoring long-period cumulative slow changes. In the deployment or re-calibration stage of the present application, the initial trigger threshold and dynamic adjustment function of each monitoring area are determined by a standardized parameter deterministic procedure. The starting point of the procedure is to obtain a standardized time sequence image sample library containing confirmed unchanged areas, confirmed interference areas and confirmed real changed areas through cross-labeling. First, by traversing the candidate confidence coefficients and calculating the true positive rate TPR and false positive rate FPR corresponding to each value in the sample library, a receiver operating characteristic curve (ROC curve) is drawn to represent the relationship between detection sensitivity and false positive rate under a given sensor noise baseline. The ROC curve is a chart that visualizes the performance of a classifier, with the FPR on the horizontal axis and the TPR on the vertical axis. The final value is the coefficient value corresponding to the point on the ROC curve that maximizes the utility function of the specific task, where the weights and are constants preset according to the tolerance of false negatives and false positives in the monitoring task; secondly, the threshold decay adjustment function is determined as , where the shape parameters and are directly derived from the statistical analysis results of the time interval sequence of historical real change events in the area, ​the value of which is inversely proportional to the coefficient of variation of the sequence of time intervals, while the value of which is inversely proportional to the mean of the sequence, so that the more regular the variation, the more intense the threshold decay response, and the more random the variation, the more gradual the response.

[0027] The adjustment instruction sent by the ground station to the on-board edge computing unit is a structured instruction package containing an asynchronous execution timestamp and a multi-modal interference feature signature, wherein the asynchronous execution timestamp specifies a specific future time when the instruction is activated on the satellite, so as to compensate for the inherent delay of satellite-ground communication and instruction processing, and the multi-modal interference feature signature further integrates spectral residual features and thermal infrared inertial features on the basis of the original kinematic features based on optical flow field analysis. The spectral residual feature refers to a normalized index that will produce a specific abnormal fluctuation when cloud shadow shielding occurs, which is defined by the stable relationship of reflectivity of surface cover types at different spectral bands. The thermal infrared inertial feature refers to the physical property that ground objects show different temperature rising and falling rates under the change of sunshine and clouds due to different heat capacities. When matching a suspected interference event, the on-board unit will execute a hierarchical verification logic, that is, only when the features of an event in the three dimensions of kinematics, spectral residual and thermal infrared inertia are matched with the signature in the instruction package, it is finally determined as environmental interference and executes the threshold raising operation. At the same time, the suppression logic of any geographical location that is continuously determined as the same type of interference for three or more times will be automatically suspended, and the incremental information package of the event will be forcibly transmitted to the ground station at high priority for manual review.

[0028] Embodiment 1: The technical solution of the present application in an application of high-frequency fire monitoring in a wide forest area is as follows: when a remote sensing satellite configured with the method of the present application performs imaging observation on a vast and remote forest area every 90 minutes according to its predetermined orbit, a small fire point with an initial area less than 5x5 pixels and in the smoldering stage is caused by lightning in the local area within the interval of two observation periods; in the next flyover of the area, the imaging load of the satellite obtains the current period remote sensing image containing the small fire point and with a data volume of several gigabytes (GB); at this time, the on-board edge computing unit deployed on the satellite does not directly store the original image data into the downlink transmission queue, but first generates a low-resolution thumbnail of the image on orbit according to the above steps, and differentiates it with the reference thumbnail stored on the satellite representing the reference of the last period, and then calculates the normalized information entropy of the differential image. In general, image data processing or Due to the presence of this smoldering fire, even though it accounts for a very small proportion of the entire image, the change in its physical state still produces non-random texture changes in the corresponding local block of the differential image, causing the information entropy value calculated for this block to exceed the trigger threshold set for the forest area; this entropy value exceeding the limit triggers the subsequent steps. The onboard edge computing unit does not initiate the transmission of the entire gigabyte-level original image. Instead, it only extracts the high-resolution target block corresponding to the area of ​​entropy change from the original high-resolution image, and performs a high-resolution differential operation on it with the corresponding block in the historical image, generating an incremental information packet with a data volume of several kilobytes (KB); this information packet is given high transmission priority and immediately transmitted to the ground station via the downlink. Due to its data volume, this transmission process does not occupy the conventional broadband data downlink channel that requires hours of queuing, and is completely received by the ground station within minutes.

[0029] After the ground station receives the incremental information packet, it immediately starts the analysis process, and the change detection model is able to concentrate its computing resources on this pre-screened target block, thereby completing the segmentation, identification and confirmation of the fire points in the block in a short time, and generating early warning information; in this process, the on-board entropy change trigger mechanism and the ground station incremental precision mechanism form a front-to-back workflow. The operation of the on-board mechanism makes it unnecessary for the downlink to occupy transmission resources for the transmission of static background data that accounts for more than 99%, and the ground station greatly shortens the processing time required for change confirmation because its processing object is changed from the entire image to the target block with extremely small data volume. This way of constructing the information processing process obtains change information The bottleneck of acquisition has shifted from the satellite-to-ground transmission link to the on-board edge computing and ground station analysis of extremely small data packets, thereby improving the timeliness of obtaining 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 during the first satellite observation window after the fire broke out, thus creating conditions for emergency forces to intervene before the fire spreads; at the same time, the ground station generated an adjustment instruction based on the confirmed fire point information and sent it to the satellite. This instruction will dynamically and targetedly lower the entropy change trigger threshold of the area around the fire point, so that the system will have higher monitoring sensitivity for any subtle changes in the fire in the area in subsequent observations.

[0030] Example 2: To objectively quantify the technical effects of the method of the present application in terms of data transmission efficiency and information acquisition latency, a semi-physical simulation test platform was built, which consisted of an embedded processing module simulating a satellite-borne edge computing unit and a server simulating a ground station, connected by a bandwidth-limited and latency-configurable communication link to simulate the satellite-ground communication environment. The test data source was a set of public time-series remote sensing image sequences covering various types of ground surfaces, including various scenarios from no environmental interference to minor and significant target changes. As a control, a control group simulating the existing full-volume transmission method was run synchronously. The control group used the same hardware and data set as the test group, but the on-board module sent the complete raw image data to the downlink in each cycle.

[0031] In the test, the trigger threshold of the on-board entropy change trigger mechanism was a key setting parameter, and its value was determined to balance the sensitivity of change detection and the economy of data transmission. The setting process was as follows: first, a series of image sequences with no ground surface changes in the data set were selected, and the mean information entropy of their consecutive difference images was calculated and the standard deviation Then, a confidence coefficient was set according to the false positive rate requirement of the target application scenario Then the trigger threshold can be derived from In this test, for the sensor noise characteristics of the data set used, it was measured that , To run at a higher confidence level, set Then the trigger threshold is determined as After the test started, the test group and the control group processed the same time-series image sequence synchronously. The sequence simulated four test scenarios in turn: scenario one, no identifiable changes on the ground surface; scenario two, large and sparse cloud shadow movement, constituting environmental interference; scenario three, two new vehicles in a parking lot, constituting minor target changes; scenario four, a temporary building in a construction site was demolished, constituting significant target changes. In each test cycle of each scenario, the system recorded the total amount of data transmitted, the end-to-end latency from the occurrence of the change to the ground station's confirmation of the change information, and the average processor occupancy rate of the satellite-borne edge computing unit. The core data records are shown in Table 1.

[0032] Table 1 shows the performance data of the method of the present application and the existing method in different test scenarios ​​From the data in Table 1, in scene one and scene two, the downlink transmission data amount of the method of the application is zero, because the difference information entropy value calculated by the spaceborne edge computing unit does not exceed the preset trigger threshold of 0.37, so that the downlink transmission is not started, while the existing mode transmits all 1024MB of the original image data; in scene three and scene four, the method of the application is triggered, and the downlink transmission data amount is 0.82MB and 5.15MB respectively, which is only 0.08% and 0.5% of the original image data amount, correspondingly, the change information acquisition delay is also shortened from about 3600 seconds of the existing mode to less than 20 seconds, it should be noted that the CPU occupancy rate of the application method is slightly higher than that of the existing mode due to the execution of entropy value calculation on the star, but the value 2.8% is still within the normal working load range of low-power embedded processors.

[0033] Embodiment 3: This embodiment combines Figs. 1 to 3 , an AI target change detection method for fusing time-series remote sensing images is realized and described, as shown in Fig. 1 , the spaceborne load first acquires the current period remote sensing image and delivers it to the spaceborne edge computing unit, the unit compares the low-resolution thumbnail with the historical reference benchmark and calculates the difference information entropy value to make a preliminary diagnosis on the data, and determines the subsequent action according to a judgment condition alt, i.e. whether the information entropy value is higher than a trigger threshold, if the information entropy value > trigger threshold, the change area is extracted from the high-resolution image, and the incremental information package is generated and transmitted to the ground station via the space-ground link, after receiving and analyzing and confirming the target change, the ground station can send the change warning to the user terminal, if the information entropy value ≤ trigger threshold, the spaceborne edge computing unit directly gives up the transmission of the original data, so as to achieve the purpose of saving transmission bandwidth.

[0034] As shown in Fig. 2 , the horizontal coordinate in the figure is the observation period, and the vertical coordinate is the information entropy value, the solid line represents the difference information entropy value calculated period by period, and the dashed line represents the preset trigger threshold, from the figure, it can be seen that before about the 35th observation period and after the 60th observation period, the difference information entropy value is lower than the trigger threshold, indicating that the system determines that there is no significant change, and between the 35th and 60th observation periods, the difference information entropy value is continuously higher than the trigger threshold, indicating that the system continuously detects the incremental information that needs to be transmitted downward during this period.

[0035] As shown in Fig. 3As shown, the figure contains two core parts of the satellite-borne edge computing unit and the ground station. The satellite-borne edge computing unit is responsible for obtaining the current period remote sensing image, and calculating the information entropy value of the differential information based on the historical remote sensing image reference benchmark. When it is determined that the information entropy value is greater than the trigger threshold value, the changed area target block is extracted and the incremental information package is formed. At the same time, the unit can also obtain the Doppler shift value to generate pre-correction data. On the contrary, if the entropy value is not greater than the threshold value, the original data transmission is abandoned. The ground station is responsible for receiving and analyzing the incremental information package, and completing the monitoring task by confirming the target change. The internal analysis process also includes calculating the texture entropy value of the sub-area to assist the calculation of the optical flow field and the motion vector tracking, self-verifying by injecting noise test to determine whether to update the reference benchmark, and warning the physical instability risk through residual time series variance analysis. It is crucial that the ground station also generates trigger threshold adjustment instructions by identifying environmental interference characteristic signatures and quantifying the continuous non-updating time length of the region, and returns the feedback adjustment instructions to the satellite-borne edge computing unit, thereby forming a closed-loop optimization adjustment mechanism.

[0036] In an application of dynamically monitoring a near-sea port containing a large area of water and land, the method needs to cope with a composite working condition formed by the superposition of multiple factors such as weak texture on the water surface, moving cloud shadows, and slow movement of the target to be detected. In this scenario, part of the core parameters and internal processing logic in the method of the application are configured and implemented as follows. In the ground station analysis process, to solve the problem that the calculation of the optical flow field output unreliable motion vectors due to the lack of stable feature points in weak texture areas such as water surface, the weight assignment mechanism is defined by a deterministic mapping function. Specifically, after calculating the normalized texture entropy value of each sub-area in the received incremental information package , the value range is 0 to 1, the weight value of the corresponding motion vector of the sub-area follows the relationship . This quadratic function relationship aims to more effectively suppress the calculation impact of low texture entropy areas. According to this relationship, the weight value corresponding to a sub-area with a texture entropy value of 0.3 is 0.09, and the weight value corresponding to a sub-area with a texture entropy value of 0.8 is 0.64. In this way, when all the motion vectors are weighted and clustered, the contribution of the motion vectors from low texture entropy areas such as water surface to the final judgment result is effectively reduced.

[0037] Meanwhile, in order to enable the on-board edge computing unit to distinguish the entropy fluctuation caused by target changes from that caused by environmental factors such as cloud shadows, the environmental interference feature signature, the extraction of which at the ground station and the matching logic at the on-board end are specified by a closed-loop process; on the ground station side, the system collects those incremental information packets that are triggered by entropy changes but are finally excluded by the analysis module through optical flow field clustering recognition of large-scale and single motion patterns, and performs statistics on the motion vector fields confirmed to be generated by cloud shadow interference, extracts the mean and standard deviation of the motion vector in direction and speed, and forms a structured data containing the fields of interference type: cloud shadow, vector mean: [ , ], vector standard deviation: [ , ] and other fields as a feature signature, which is sent to the satellite through the uplink; on the on-board edge computing unit side, when a new entropy out-of-limit event occurs, the on-orbit system estimates a motion vector through block matching algorithm, if the direction value of the motion vector is within the interval of [ , ] and its speed value is within the interval of [ , ], the system classifies this triggering event as a disturbance event matching the feature signature, and automatically temporarily increases the triggering threshold used for this judgment by a fixed coefficient, which can be 1.5.

[0038] In addition, in order to enable the system to effectively monitor the long-period changes such as slow coastal erosion that may occur in the port area, the stable time length threshold, i.e. the threshold of continuous un-updated time length for starting high-sensitivity monitoring mode, is determined by a mechanism based on adaptive adjustment of historical data; for a newly included monitoring area, the system first uses a universal initial stable time length threshold, which can be 100 observation periods, in continuous operation, the ground station records the time points of each successful triggering and confirmation of real changes of the area, and calculates the average time interval of changes of the area , after the system runs for a preset initial learning period, which can be 1000 observation periods, the corresponding stable time length threshold 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 change activity.

[0039] In the application of long-term health monitoring and risk early warning of a large bridge structure, to ensure the reliability of the change detection results and effectively identify the small deformation caused by structural fatigue or ground subsidence, part of the internal verification and judgment procedures of the method of the present application is implemented as follows: when a set of analysis systems deployed in the ground station analyzes an incremental information package and confirms a change, an internal self-verification step is started before the planned use of the change to update the reference benchmark. The step first injects a standardized Gaussian noise with a standard deviation of 1 gray level into the target block causing the change in memory to generate an adversarial test sample, and calculates the time series residual sequence of the original target block and the adversarial test sample with respect to the historical reference benchmark, respectively. Then, the system calculates the Pearson correlation coefficient between the two residual sequences. Only when the correlation coefficient is higher than a preset confidence threshold of 0.98, the change is identified as a real change of the physical structure and used to update the reference benchmark.

[0040] To warn of the possible physical instability risk of the bridge structure, the risk threshold is set according to an offline calibration engineering procedure. After the bridge completes a comprehensive structural inspection, the system first collects a series of remote sensing images within a period of time under the stable state of the structure as a calibration data set. The ground station processes the differential information in the calibration data set to form one or more residual time series representing the influence of environmental noise on the bridge in the 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 for online real-time monitoring is set as the value obtained by subtracting three times the standard deviation from the mean variance in the healthy state. When the residual time series formed by multiple incremental information packages received from the same geographical location in online monitoring continuously has a variance value lower than the calibrated risk threshold, the system generates a warning signal.

[0041] In an application of deploying the method of the present application in a previously uncovered geographical area for the first time, to ensure the initialization of the system and the calibration of the module parameters, a standardized pre-deployment procedure needs to be performed. In the procedure, the first and second consecutive overpass observations of the satellite in the new area are used for system initialization. The complete data of the original remote sensing image obtained in the first overpass is downlinked to the ground station and processed by the ground station to establish the initial reference benchmark for the entire monitoring task. The change pre-diagnosis function of the on-board edge computing unit is activated from the second overpass observation and uses the image of the first overpass as the first historical remote sensing image required for its on-orbit calculation.

[0042] In the ground station configuration part of the procedure, for the increment information package generated on the satellite as a differential image, the data processing flow is set as follows: when receiving an increment information package, the ground station first retrieves the historical remote sensing image block corresponding to the spatial position of the increment information package from its database, then performs pixel-level superposition between the differential image block and the historical remote sensing image block through matrix addition operation, thereby reconstructing the target block corresponding to the current period remote sensing image at the ground end, and then, the reconstructed current target block and the original historical remote sensing image block are sent as input data pairs into a change detection model based on a twin network architecture or a double-channel encoder for analysis to complete change confirmation and classification; a calibration mechanism for pre-correction data based on Doppler frequency shift is generated to compensate for systematic errors that may be introduced by platform motion, and the procedure also includes an offline calibration step; during the initial deployment stage of the system, the ground station calculates the theoretical Doppler frequency shift that should be generated by the theoretical relative radial velocity between the satellite and the ground station under different imaging geometric angles based on high-precision satellite orbital elements, and records the Doppler frequency shift values actually measured and downloaded from the satellite communication payload, by comparing the two sets of data, the system can calculate a correction coefficient or lookup table for compensating for platform attitude deviation or internal clock drift, and the correction information is then uplinked to the on-board edge computing unit to optimize the accuracy of the generated pre-correction data in subsequent tasks, thus the monitoring system completes the deployment and calibration in a new area and enters the normal operation state.

[0043] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not limiting, and although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. An AI target change detection method integrating time-series remote sensing images, characterized in that: The following steps are involved: Step a: On the onboard edge computing unit, obtain 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 remote sensing image of the current period 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 remote sensing image of the current period, form an incremental information packet, and transmit it to the ground station. If the information entropy value is not higher than the trigger threshold, the original data transmission of the remote sensing image of the current period is abandoned on the satellite-borne edge computing unit side. In step d, the ground station receives and analyzes the incremental information packet to confirm the target change, and the ground station identifies and extracts the characteristic signature of the environmental interference based on the analysis of at least one incremental information packet received historically, and quantifies the continuous non-updated duration of the reference benchmark of each area; In step e, the ground station generates an adjustment instruction based on the characteristic signature of the environmental interference and the continuous non-update duration and sends it to the onboard edge computing unit to adjust the trigger threshold of the corresponding area: when it is detected that the current change characteristics match the characteristic signature in the environmental log, the trigger threshold is increased; when the continuous non-update duration of a certain area reaches a stable duration threshold, the trigger threshold is lowered.

2. The AI ​​target change detection method for fusion of time-series remote sensing images according to claim 1 is characterized in that: The reference benchmark in step a is a low-resolution thumbnail of the historical remote sensing image; the object for calculating the information entropy value in step b is the difference image between the corresponding low-resolution thumbnail generated by the remote sensing image of the current period and the reference benchmark.

3. The AI ​​target change detection method for fusion of time-series remote sensing images according to claim 1 is characterized in that: The incremental information package 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 for fusion of time-series remote sensing images according to claim 3 is characterized in that: Before analyzing the incremental information packet, step d also includes: calculating the texture entropy value of each sub-region within the incremental information packet based on the differential information within the incremental information packet; and in the subsequent change confirmation analysis, assigning weights to the motion vector features derived from each sub-region, and the weight values ​​are determined according to the texture entropy values ​​of the sub-regions, wherein the motion vector features are derived from the optical flow field calculation of the incremental information packet.

5. The AI ​​target change detection method for fusing time-series remote sensing images according to claim 1 is characterized in that: The adjustment to lower the trigger threshold in step e follows the following rules: ,in, is the adjusted trigger threshold, is the basic trigger threshold of the area, Is the silent period count value corresponding to the continuous non-updated time. It is a regulation function that decreases as the silent period count value increases and its value is less than 1.

6. The AI ​​target change detection method for fusing time-series remote sensing images according to claim 1 is characterized in that: The method also includes: before executing step c, 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; transmits the pre-correction data together with the incremental information packet to the ground station; 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 for fusing time-series remote sensing images according to claim 1 is characterized in that: After confirming the target change in step d, a self-verification step is also included to decide whether to update the reference benchmark: standardized noise is injected into the target block in memory to generate adversarial test samples; Calculate the fluctuation trend of the time series residuals of the original target block and the adversarial test sample relative to the historical reference benchmark respectively; The reference benchmark is updated with the current change only if the consistency of the two fluctuation trends is above a confidence threshold.

8. The AI ​​target change detection method for fusing time-series remote sensing images according to claim 1 is characterized in that: The method also includes: a ground station extracts 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 sequence; calculates the variance value of the residual time series sequence; and generates a warning signal indicating that there is a risk of physical instability at the geographical location when the variance value is continuously lower than a risk threshold.

9. The AI ​​target change detection method for fusing time-series remote sensing images according to claim 4 is characterized in that: The optical flow field calculation includes tracking the motion vectors of local blocks of incremental information packets; the weights are used to adjust the contribution of each motion vector to the change confirmation result in the subsequent statistical clustering analysis, where the contribution of the motion vector originating from the sub-region with high texture entropy value is higher than that of the motion vector originating from the sub-region with low texture entropy value.

10. The AI ​​target change detection method for fusing time-series remote sensing images according to claim 1 is characterized in that: The characteristic signature of the environmental interference in step d is extracted by performing optical flow field clustering analysis on the incremental information packets that have been historically confirmed to be triggered by cloud shadows, and statistically modeling the separated motion vectors with large-scale and single motion pattern characteristics.

Images