A dynamic algorithm autonomous selection method and system for a spaceborne remote sensing task

Abstract

The embodiment of the application provides a kind of dynamic algorithm autonomous selection method and system for spaceborne remote sensing task, it is related to spaceborne remote sensing intelligent processing technical field.The method can be after obtaining observation image data, for the image slice in observation image data, execute complexity probe, obtain invalid pixel ratio and interest area attribution mark etc.complexity characteristics.Again, according to complexity characteristics, generate slice processing level, and according to task type and slice processing level, multiple image slices are organized slice scheduling.Then based on slice scheduling result, call model image library matching processing model, image processing is executed to image slice.The method can be for spaceborne remote sensing task, pre-establish multiple different complexity processing model, and with image slice as basic processing unit, the complexity characteristics of each slice are obtained by complexity probe, to match the processing model corresponding to different slices, to make full and reasonable use of limited computing power resources on satellite.

Description

Technical Field

[0001] This application relates to the field of spaceborne remote sensing intelligent processing technology, and in particular to a dynamic algorithm autonomous selection method and system for spaceborne remote sensing missions. Background Technology

[0002] With the development of high-resolution Earth observation satellites, wide-swath survey satellites, and multi-satellite networked remote sensing systems, the amount of observational data acquired by spaceborne platforms continues to increase. For payloads such as high-resolution optics and multispectral optics, a single imaging mission can cover a large area, generating a large volume of data. However, the time window available for intelligent processing onboard is short, and onboard computing resources, storage resources, and parallel processing capabilities are relatively limited. Under high-resolution imaging conditions, the entire image contains a large number of different types of local regions, such as regions within the region of interest, regions outside the region of interest, cloud-occluded regions, clear and effective regions, and regions with varying degrees of texture complexity. These local regions have different processing values ​​in remote sensing tasks such as target detection, scene recognition, anomaly screening, change recognition, or image enhancement. However, in the onboard processing chain, it is difficult to fully distinguish them before they enter the main task processing.

[0003] Unlike ground-based high-performance computing environments, spaceborne platforms are subject to multiple constraints during on-orbit operation, including computing power, storage capacity, power consumption budget, task timing, and parallel processing capabilities. For on-orbit intelligent processing tasks requiring rapid response, uniformly performing complex model processing on the entire landscape survey data can easily lead to excessively long processing times, task backlogs, or subsequent tasks waiting; conversely, uniformly adopting the most lightweight processing method may fail to guarantee the processing quality of key or high-value areas.

[0004] As the level of intelligence on spacecraft increases, the goals of on-orbit processing of observation data have gradually expanded from simple preprocessing, compression, and transmission to more complex task modes such as on-orbit detection, rapid screening, local fine-grained analysis, and result pre-generation. In this process, if the entire field of observation data is still treated as a whole with approximately uniform processing value, and a single processing model is consistently applied to the same type of task, it will be difficult to fully adapt to the differences in data quality, spatial attribution, and task value among different local areas within the field. Therefore, how to prioritize limited computing power for data units that are more worthy of processing and more suitable for fine-grained processing has become a pressing technical problem to be solved in the spacecraft intelligent processing chain. Summary of the Invention

[0005] In view of this, embodiments of this application provide a dynamic algorithm autonomous selection method and system for spaceborne remote sensing missions, in order to solve the technical problem that limited onboard computing power is difficult to prioritize for remote sensing missions that are more worthwhile to process and more suitable for fine processing.

[0006] According to a first aspect of this application, a dynamic algorithm autonomous selection method for spaceborne remote sensing missions is provided, the method comprising: Acquire observation image data and the task type corresponding to the observation image data, wherein the observation image data includes multiple image slices obtained by segmenting the whole landscape observation image; A complexity probe is performed on the image slice to obtain the complexity features of the image slice, which include the proportion of invalid pixels and the region of interest attribution markers; A slice processing level is generated based on the complexity feature, and the slice processing level is calculated using the proportion of invalid pixels and the region of interest attribution label. Organize slice scheduling for multiple image slices according to the task type and the slice processing level to obtain slice scheduling results; Based on the slice scheduling result, the model profile library is invoked to match the processing model. The model profile library configures multiple candidate models with different complexity levels for the same task type and stores the model capabilities and resource profiles corresponding to each candidate model. The processing model is used to perform image processing on the image slice.

[0007] In some embodiments, obtaining the observed image data and the task type corresponding to the observed image data includes: Receive mission planning data and overall landscape survey images acquired by onboard image acquisition equipment, wherein the mission planning data includes at least one of ground mission instructions and onboard autonomous planning information; Read the task type from the task planning data; Obtain slice configuration parameters, including payload space resolution, cache organization method, and task type; Slicing rules are set according to the slice configuration parameters. The slice rules include at least the slice size; the slice rules also include one or more of the following: slice step size, overlap width, edge padding method, and row and column numbering; the slice size is obtained according to the payload space resolution, the cache organization method, and the task type. The overall landscape observation image is divided into multiple image slices according to the slicing rules to generate the observation image data.

[0008] In some embodiments, a complexity probe is performed on the image slice to obtain the complexity features of the image slice, including: Invalid pixel detection is performed on the image slice to determine the invalid pixel ratio, which is the ratio of cloud pixels, sensor bad pixels, missing pixels and fill values ​​calculated by the invalid pixel detection method. Perform region of interest (ROI) attribution detection on the image slices to obtain ROI attribution labels, which are classification labels obtained by overlaying slice spatial index and ROI range. The complexity features of the image slice are generated using the proportion of invalid pixels and the region of interest attribution markers as core features.

[0009] In some embodiments, performing a complexity probe on the image slice to obtain the complexity features of the image slice further includes: Extract complex characterization parameters from the image slices, wherein the complex characterization parameters include at least one of texture statistics, edge density, and feature encoding results; The complexity of the scenario is calculated based on the aforementioned complex representation quantity; Extract quality characteristics from the image slices, the quality characteristics including at least one of blur level, saturation ratio and basic quality marker; Calculate the data quality level based on the quality characterization parameters; Prior information on key events is obtained. This prior information is pre-generated by the ground mission planning system and the ground application system based on the prior mission configuration table. The prior information on key events includes at least one of the following: geographical scope of the event, event type, and priority identifier. The complexity features of the image slices are generated using the scene complexity, the data quality level, and the prior information of the key events as extended factor features.

[0010] In some embodiments, generating a slice processing level based on the complexity feature includes: According to the task type, a preset level generation parameter corresponding to the task type is obtained. The preset level generation parameter includes a first weight, a second weight, and at least one preset processing level threshold. The first weight is used to perform a weighted calculation on the proportion of invalid pixels. The second weight is used to perform a weighted calculation on the degree of overlap corresponding to the region of interest attribution marker. The effective pixel ratio of the image slice is determined based on the invalid pixel ratio, wherein the effective pixel ratio is the difference between 1 and the invalid pixel ratio; The degree of overlap between the image slice and the region of interest is obtained, and the degree of overlap is determined based on the ratio of the overlapping area of ​​the image slice and the region of interest to the area of ​​the image slice. The processing level index is calculated based on the complementary value of the effective pixel ratio, the degree of overlap of the region of interest, the first weight, and the second weight; The slice processing level is determined according to the index range to which the processing level index belongs; the index range is a numerical range divided based on at least one preset processing level threshold.

[0011] In some embodiments, slice scheduling is organized for multiple image slices according to the task type and the slice processing level to obtain slice scheduling results, including: Iterate through the task type and slice processing level corresponding to the image slice; Obtain the model input specifications corresponding to each of the image slices, wherein the model input specifications include at least one of the following: input size, band combination, data type, numerical precision, and preprocessing configuration; Image slices with the same task type, the same slice processing level, and compatible model input specifications are grouped into the same batch execution group to obtain the slice scheduling result; the slice scheduling result includes at least one batch execution group and the execution order corresponding to the batch execution group; or, Write the image slices into the batch execution queue according to the task type and the slice processing level; The triggering conditions for the batch execution queue are obtained, and the triggering conditions include at least one of the following: the number of slices in the queue reaches the batch processing quantity threshold, the current cumulative waiting time of the queue reaches the waiting time threshold, and the whole scene end flag, task end flag, or data stream termination flag is received. The system monitors queue parameters and generates slice scheduling results according to the batch execution queue when the queue parameters meet the triggering condition. The queue parameters include at least one of the number of slices in the queue, the current cumulative waiting time of the queue, and the queue execution priority. The slice scheduling results include at least one batch execution queue and the execution order corresponding to the batch execution queue.

[0012] In some embodiments, calling the model profile library matching processing model based on the slice scheduling result includes: Read the task type and slice processing level corresponding to the slice scheduling result; Obtain model profile information of candidate models in the model profile library. The model profile information includes at least one of the following: task type identifier, model complexity level, model performance index, model capability boundary, inference time, resource consumption, applicable data conditions, and output result type. Obtain the current resource status of the spaceborne platform, wherein the current resource status includes at least one of available computing resources, available memory, remaining processing time, power budget, current queue load, and current model dwell status; Based on the task type, the slice processing level, and the model profile information, the processing method information is determined, which includes skip processing and model inference processing. For the image slice for which the processing method information is the model inference processing, a processing model is matched from the model profile library according to the task type, the slice processing level, and the model profile information; For the image slice whose processing method information is "skip processing", the slice parameters of the image slice are retained. The slice parameters include the slice index, processing mark, and basic quality mark.

[0013] In some embodiments, the method further includes: The image slices in the slice scheduling result are input into the processing model slice by slice or batch, so that the processing model performs image processing on the image slices in the on-board processing resource environment corresponding to the task type and the slice processing level; Read the task type, which is either a first type or a second type; The image processing results output by the processing model are obtained, and the data structure type of the image processing results is determined according to the task type; the image processing results include at least one of raster processing results and structured processing results; the raster processing results include at least one of segmentation mask, image enhancement, variable raster, and cell-level anomaly results; the structured processing results include at least one of category results, target bounding boxes, target attributes, confidence scores, and anomaly event records; When the task type is the first type, multiple raster-type processing results are spliced ​​or fused according to the slice index and slice spatial position corresponding to the image slice to generate a whole-scene raster task result corresponding to the whole-scene survey image. When the task type is the second type, the association between the structured processing result and the corresponding position in the whole landscape survey image is established according to the slice index; wherein, when the structured processing result contains slice local coordinates, the slice local coordinates are converted into whole landscape image coordinates or geographic coordinates, and the converted structured processing result is written into the task result data.

[0014] In some embodiments, based on the slice index and spatial location corresponding to the image slice, multiple raster-type processing results are stitched or fused to generate a whole-scene raster task result corresponding to the whole-scene survey image, including: Read the slice index corresponding to the image slice; The position of the image slice in the overall landscape assessment image is determined based on the slice index; Obtain the mission result structure and tile overlap information corresponding to the spaceborne remote sensing mission; According to the overall scene location, the task result structure, and the slice overlap information, the image processing result is written back to the overall scene survey image to generate task result data; If adjacent slices overlap, perform at least one of the following on multiple processing results in the overlapping region: confidence comparison, weighted fusion, voting fusion, center region priority, or duplicate target elimination.

[0015] According to a second aspect of this application, a dynamic algorithm autonomous selection system for spaceborne remote sensing missions is provided, the system comprising: The image slice management module is used to acquire observation image data and the task type corresponding to the observation image data. The observation image data includes multiple image slices obtained by segmenting the whole landscape observation image. The complexity probe module is used to perform a complexity probe on the image slice to obtain the complexity features of the image slice, the complexity features including the proportion of invalid pixels and the region of interest attribution marker; A processing level generation module is used to generate a slice processing level based on the complexity feature. The slice processing level is calculated using the proportion of invalid pixels and the region of interest attribution marker. The slice scheduling module is used to organize slice scheduling for multiple image slices according to the task type and the slice processing level, and obtain slice scheduling results; The model matching module is used to call the model profile library to match the processing model based on the slice scheduling result. The model profile library includes multiple candidate models with different complexity levels and the model capabilities and resource profiles corresponding to the candidate models. The processing model is used to perform image processing on the image slice.

[0016] By employing the above technical solutions, embodiments of this application provide a dynamic algorithm autonomous selection method and system for spaceborne remote sensing missions. The method, after acquiring observation image data, performs a complexity probe on image slices within the observation image data to obtain complexity features such as the proportion of invalid pixels and region of interest (ROI) attribution markers. Then, based on the complexity features, it generates slice processing levels and organizes slice scheduling for multiple image slices according to mission type and slice processing levels. Finally, based on the slice scheduling results, it calls a model image library to match processing models and perform image processing on the image slices. This method can pre-establish multiple processing models of different complexities for spaceborne remote sensing missions, using image slices as basic processing units. By obtaining the complexity features of each slice through complexity probes, it matches corresponding processing models to different slices, thereby fully and rationally utilizing the limited computing resources on the satellite.

[0017] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A schematic diagram of the dynamic algorithm autonomous selection method for spaceborne remote sensing missions provided in this application embodiment; Figure 2 This is a schematic diagram of the process for acquiring observation image data provided in an embodiment of this application; Figure 3 This is a schematic diagram illustrating the entire process of dynamic algorithm autonomous selection provided in the embodiments of this application; Figure 4 This is a schematic diagram illustrating the process of calling the model profile library matching and processing model based on slice scheduling results, as provided in an embodiment of this application. Figure 5 This is a schematic diagram of the slice-level image processing result organization and write-back process provided in the embodiments of this application; Figure 6 This is a schematic diagram of the structure of a dynamic algorithm autonomous selection system for spaceborne remote sensing missions provided in an embodiment of this application. Detailed Implementation

[0019] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0020] In this embodiment of the application, the dynamic algorithm autonomous selection method for spaceborne remote sensing missions can be applied to spaceborne platforms to perform remote sensing missions. The spaceborne platform refers to a remote sensing system that uses a spacecraft operating in space to carry remote sensing acquisition equipment and data processing equipment to observe the Earth or a specific region.

[0021] With the development of high-resolution Earth observation satellites, wide-swath survey satellites, and multi-satellite networked remote sensing systems, the amount of observational data acquired by spaceborne platforms continues to increase. For payloads such as high-resolution optics and multispectral optics, a single imaging mission can cover a large area, generating a large volume of data. However, the time window available for intelligent processing onboard is short, and onboard computing resources, storage resources, and parallel processing capabilities are relatively limited. Under high-resolution imaging conditions, the entire image contains a large number of different types of local regions, such as regions within the region of interest, regions outside the region of interest, cloud-occluded regions, clear and effective regions, and regions with varying degrees of texture complexity. These local regions have different processing values ​​in remote sensing tasks such as target detection, scene recognition, anomaly screening, change recognition, or image enhancement. However, in the onboard processing chain, it is difficult to fully distinguish them before they enter the main task processing.

[0022] Unlike ground-based high-performance computing environments, spaceborne platforms are subject to multiple constraints during on-orbit operation, including computing power, storage capacity, power consumption budget, task timing, and parallel processing capabilities. For on-orbit intelligent processing tasks requiring rapid response, uniformly performing complex model processing on the entire landscape survey data can easily lead to excessively long processing times, task backlogs, or subsequent tasks waiting; conversely, uniformly adopting the most lightweight processing method may fail to guarantee the processing quality of key or high-value areas.

[0023] As the level of intelligence on spacecraft increases, the goals of on-orbit processing of observation data have gradually expanded from simple preprocessing, compression, and transmission to more complex task modes such as on-orbit detection, rapid screening, local fine-grained analysis, and result pre-generation. In this process, if the entire landscape observation data is still treated as a whole with approximately uniform processing value, and a single processing model is consistently applied to the same type of task, it will be difficult to fully adapt to the differences in data quality, spatial attribution, and task value among different local areas within the landscape. This makes it difficult to prioritize limited computing power for data units that are more worthy of processing and more suitable for fine-grained processing.

[0024] In some embodiments, to reduce the resource consumption of the spaceborne platform, the lightest model can be uniformly applied to all observation data of the same type of remote sensing mission. Using the lightest model prioritizes ensuring the stable operation of the processing pipeline under limited computing power, avoiding processing timeouts or task backlogs due to overly heavy models.

[0025] However, the approach of uniformly applying the lightest model assumes that all data units have similar processing value, failing to differentiate between key and non-key areas, clear and cloud-obscured areas, or adjust processing intensity based on the varying validity of different data units. Therefore, while this solution saves resources, it is prone to underprocessing high-value areas.

[0026] In some embodiments, a model can be pre-fixed for each type of task, and it will not be adjusted during task execution. The fixed model approach, compared to the uniform lightweight model, allows different models to be selected for different tasks, but all data within the same task is still processed in the same way.

[0027] Because the observation data within the same task is not homogeneous, the processing method makes inefficient use of spaceborne platform resources. Taking target detection as an example, the same image may simultaneously contain sharp slices within the region of interest, slices outside the region of interest, and slices covered by clouds. If the same model is always used, it will lead to serious waste of computing power in some cases, while the processing capacity for some key areas is insufficient and lacks specificity.

[0028] In some embodiments, the model can also be switched based on the platform resource status. For example, a higher complexity model can be used when there is a large amount of computing power available, and a lightweight model can be switched when computing power is tight, cache usage is high, or power consumption is limited.

[0029] The decision to switch models based on platform resource status is primarily based on the platform's own condition, without fully considering the quality and value differences of the observed data itself. In other words, switching models based on platform resource status can only answer "which type of model is suitable for the current platform," but it cannot answer "whether the current data is worth running a more complex model." Therefore, it is still possible to allocate too much computing power to low-value data while under-processing high-value data.

[0030] In some embodiments, a quick overall assessment of the entire image can be performed first, and then a unified processing model can be assigned to the entire image. For example, when the cloud cover of the entire image is low, a higher-capacity model can be used uniformly, and when the cloud cover of the entire image is high, a lightweight model or skipping processing can be used uniformly, in order to reduce the difficulty of implementation and facilitate scheduling and control.

[0031] However, the method of quickly judging the entire scene and uniformly specifying the processing model results in a coarse-grained judgment of the entire scene. In real-world scenarios, only a local area within the same scene may be obscured by clouds, or only a portion of the scene may fall within the region of interest. If a unified decision is made on a scene-by-scene basis, the differences between local areas within the scene cannot be fully utilized, leading to effective and ineffective areas being treated equally.

[0032] Furthermore, while refining the decision-making granularity to row-by-row data or excessively fine units theoretically yields higher-resolution scheduling and control capabilities, it also introduces problems such as excessively frequent switching, complex index management, increased cache scheduling burden, and difficulties in result concatenation, which are detrimental to spaceborne engineering implementation. Therefore, in the spaceborne remote sensing mission processing methods shown in the above embodiments, some methods are too coarse-grained, making it difficult to reflect local differences; others are too fine-grained, making it difficult to control engineering complexity. Both methods struggle to achieve a reasonable balance between processing effectiveness and system implementation.

[0033] To address the technical challenge of prioritizing the use of limited onboard computing power for more valuable and refined remote sensing tasks, this application provides a dynamic algorithm selection method for spaceborne remote sensing tasks in some embodiments. Given the large volume of overall landscape survey data, limited onboard processing time, and constraints on computing and storage resources, different local regions within the same image often exhibit significant differences in cloud obstruction, region of interest (ROI) attribution, effective information density, and processing value. Therefore, this method can pre-establish multiple processing models of varying complexity for spaceborne remote sensing tasks, using image slices as basic processing units. Complexity probes are used to acquire the complexity characteristics of each slice, matching appropriate processing models to different slices. This method, under the condition of limited onboard resources, can form a more targeted processing organization based on the differences between different local regions in the same landscape survey data, prioritizing the use of limited onboard processing resources for observation data with greater processing value.

[0034] The method can be applied to a spaceborne platform or an electronic device that establishes a communication connection with the spaceborne platform and has data processing capabilities. The electronic device includes, but is not limited to, computers, servers, mobile terminals, smart wearable devices, and industrial control machines. For ease of description, this application embodiment uses a spaceborne platform as the execution subject of the method. It should be understood that the method can also be applied to other types of execution subjects, which are not shown one by one in this application embodiment. Figure 1 As shown, the method includes: S101. Obtain the observed image data and the corresponding task type.

[0035] When performing spaceborne remote sensing missions, the spaceborne platform can first acquire observation image data, which includes multiple image slices obtained by segmenting the overall landscape observation image. For example... Figure 2 As shown, in order to obtain observation image data, the spaceborne platform can first acquire the whole-view observation image through the spaceborne image acquisition equipment, and then divide the whole-view observation image into multiple image slices through image segmentation.

[0036] Each image slice can serve as the basic processing unit for spaceborne remote sensing missions. Image slices can alleviate the problem of excessively coarse granularity in the processing of overall landscape survey images, thereby distinguishing the differences in cloud cover, region of interest, and processing value among local areas within the scene.

[0037] Refining the process to line-by-line or excessively small units can lead to problems such as frequent switching, complex cache organization, and increased burden on result stitching. Therefore, in some embodiments, image slicing can use fixed-size slices or tiles as the decision granularity, thereby achieving a balance between processing specificity and engineering feasibility.

[0038] For example, if the overall landscape survey image is represented as X, dividing the overall landscape survey image X into N image slices can form observation image data. Then, the i-th image slice is denoted as: T i = (X i L i M i ), i = 1, 2, ..., N; Among them, X i L represents the image data corresponding to the i-th image slice. i M represents the spatial location or index information of the i-th image slice. i This represents the metadata associated with the i-th image slice. This metadata may retain only the few pieces of information most critical to model selection, such as imaging time, payload mode, observation area identifier, and basic quality markers, without exhaustively listing all remote sensing imaging parameters.

[0039] While acquiring observation image data, the system can also acquire the corresponding task type. This task type characterizes the specific onboard remote sensing task to which the onboard platform will apply the observation image data, representing the task scenario requiring onboard intelligent processing. Examples include onboard target detection, scene recognition, anomaly screening, change recognition, and image enhancement.

[0040] In some embodiments, when acquiring observation image data and the corresponding mission type, mission planning data and overall observation images acquired by the onboard image acquisition equipment can be received first, and the mission type can be read from the mission planning data. The mission planning data includes at least one of ground mission instructions and onboard autonomous planning information; that is, the mission type is determined by ground mission instructions or onboard autonomous planning information generated by onboard autonomous planning.

[0041] Then, the slicing configuration parameters are obtained, and slicing rules are set according to these parameters. The slicing configuration parameters include payload spatial resolution, cache organization method, and task type. The slicing rules include slice size, which is obtained based on the payload spatial resolution, cache organization method, and task type. The slicing rules also include slice step size, overlap width, edge padding method, and row and column numbering. The entire landscape observation image is then divided into multiple image slices according to the slicing rules to generate observation image data.

[0042] For example, the spaceborne platform can perform whole-view survey image reception and image slicing. That is, the spaceborne platform can receive a whole-view survey image X and, based on preset slice sizes or tile division rules, divide the whole-view survey image into multiple slices. The slice size can be configured according to the payload resolution, buffer organization method, and mission type.

[0043] S102. Perform a complexity probe on the image slice to obtain the complexity features of the image slice.

[0044] After acquiring the observed image data and task type, a complexity probe can be executed on the image slices to obtain their complexity features. This complexity probe refers to a complexity feature detection method implemented using a fast algorithm with lower overhead than the main task model. For example, in a satellite remote sensing data processing pipeline, the complexity probe can be deployed as an intelligent filter or task evaluation module through computational logic modules or algorithm flows. The complexity probe can quickly detect the complexity features of each image slice with extremely low computational cost before large-scale, time-consuming calculations such as deep learning inference and quantitative inversion are performed.

[0045] The complexity features include the invalid pixel ratio and region of interest (ROI) attribution markers. The invalid pixel ratio refers to the percentage of pixel area obscured by invalid pixels such as clouds in a single image slice of the overall landscape image. In other words, the invalid pixel ratio is the proportion of cloud pixels, sensor dead pixels, missing pixels, and filler values ​​calculated using invalid pixel detection methods. For example, the invalid pixel ratio can be obtained through rapid cloud detection or invalid pixel assessment.

[0046] The region of interest (ROI) attribution marker is a marker used to determine whether the spatial extent of the current image slice falls within the vector boundary of a user-specified region of interest (ROI). The ROI attribution marker can be obtained by overlaying the image slice spatial index with the ROI extent.

[0047] like Figure 3As shown, to obtain complexity features such as the proportion of invalid pixels and region of interest (ROI) attribution markers, in some embodiments, a complexity probe is performed on the image slice to obtain its complexity features. In this process, invalid pixel detection can be performed on the image slice to determine the proportion of invalid pixels. The invalid pixel proportion is the ratio of cloud pixels, sensor bad pixels, missing pixels, and fill values ​​calculated by an invalid pixel detection method. For example, the invalid pixel proportion can be the cloud occlusion ratio, which can be determined by fast cloud detection or invalid pixel detection. Fast cloud detection uses threshold-based band calculation, deep learning semantic segmentation, and other algorithms to quickly identify whether a pixel is a cloud, and then calculates the ratio of cloud pixels to the total number of pixels to determine the cloud occlusion ratio. Invalid pixel detection involves reading the image's built-in quality band (QA band) or using a lightweight image recognition model to identify image pixels, determining invalid pixels in the image slice such as clouds, cloud shadows, oversaturation, sensor bad pixels, and edge fill values, and calculating the proportion of invalid pixels in the entire image slice.

[0048] Next, region of interest (ROI) attribution detection is performed on the image slices to obtain ROI attribution labels. These labels are classification tags obtained by overlaying the slice's spatial index with the ROI's extent. Since each image slice carries geographic metadata, such as its latitude and longitude, and center point coordinates, spatial index overlay can be used to calculate spatial topological relationships. The outer rectangle or precise contour of the image slice is then overlaid with the user-uploaded ROI vector surface to obtain a Boolean or enumerated value representation of the determination result. For example, 1 indicates complete inclusion, 0 indicates exclusion, and values ​​between 0 and 1 indicate partial overlap. The task relevance level of the slice is determined based on the degree of ROI overlap, and the processing level is accordingly determined before it is sent to subsequent quantitative inversion or intelligent interpretation processes.

[0049] After obtaining the proportion of invalid pixels and the region of interest (ROI) attribution markers, the complexity features of the image slices are generated using these as core features. For example, before model matching, the spaceborne platform can first perform a complexity probe on each image slice to obtain key information in a low-cost manner, such as whether the slice is worth further processing and what processing level is suitable. The complexity probe can focus on the two most important types of features: the proportion of invalid pixels and the ROI attribution. Therefore, the complexity probe result for the i-th slice can be expressed as: C i =(u i a i ); Among them, C i This represents the result of the complexity probe, i.e., the complexity feature; u i Indicates the proportion of invalid pixels, a i Indicates the region of interest as a marker.

[0050] The aforementioned two complexity features, namely the proportion of invalid pixels and the region of interest attribution label, can support slice-level model selection. In order to improve the adaptability of model selection, in some embodiments, the complexity features may also include factors such as scene complexity, data quality level, and prior information of key events. These factors can be used as extended factor features to participate in the subsequent calculation process.

[0051] Therefore, when performing a complexity probe on an image slice to obtain its complexity features, complexity metrics can also be extracted from the image slice. These complexity metrics include at least one of texture statistics, edge density, and feature encoding results. The scene complexity is then calculated based on these complexity metrics. Simultaneously, quality metrics are extracted from the image slice. These quality metrics include at least one of blur level, saturation ratio, and basic quality markers. The data quality level is then calculated based on these quality metrics.

[0052] Next, prior information on key events is acquired. This prior information is pre-generated by the ground mission planning system and the ground application system based on a prior mission configuration table. For example, the prior information on key events includes at least one of the following: the geographical scope of the event, the event type, and the priority identifier. Then, using scene complexity, data quality level, and prior information on key events as extended factor features, complexity features of image tiles are generated.

[0053] For example, after determining the proportion of invalid pixels and the region of interest (ROI) attribution markers as the core inputs for generating processing levels, scene complexity can be obtained through texture statistics, edge density, or simple feature encoding. Texture statistics, for instance, can extract features such as contrast, correlation, and entropy from image slices using the Gray-Level Co-occurrence Matrix (GLCM), and quantify the roughness and regularity of the ground surface based on the extracted features, representing the scene complexity corresponding to the image slice. For example, a uniform desert texture has low complexity, while a fragmented urban texture has high complexity. Edge density can be obtained by extracting edge information from image slices using edge detection operators such as Canny and Sobel, and calculating the proportion of edge pixels per unit area. Higher edge density indicates more complex ground boundary and higher landscape fragmentation, thus increasing scene complexity. Simple feature encoding can generate efficient supplementary information for quickly calculating the image's brightness variance, information entropy, or histogram-based contrast, rapidly scoring the overall information content of the scene and determining whether it is a large homogeneous area or a complex heterogeneous area.

[0054] In addition to determining the complexity of a scene, data quality levels can be established using blur levels, saturation ratios, or basic quality markers. Data quality levels are quantitative indicators of data usability. They can be obtained through basic quality markers, such as by observing pixel quality bands and radiation saturation bands provided by image data products, using bit-level packaging to efficiently mark each pixel as filled data, cloud, cloud shadow, snow, water, and whether each band is saturated. Alternatively, by parsing cloud confidence or cloud marker bits in QA bands, the proportion of affected pixels within an image slice can be statistically determined, thus quantifying the proportion of invalid pixels in the image slice. Furthermore, the saturation ratio can be determined by reading the radiation saturation markers corresponding to the image slice. This leverages the characteristic that some sensors easily produce signal saturation for bright targets such as snow and sand; by statistically analyzing the proportion of pixels marked as saturated in the radiation saturation band, the authenticity and dynamic range of the slice's radiation information can be assessed.

[0055] Depending on the complexity of the scenario and the data quality level, key event prior information can be pre-generated by combining ground mission planning systems, ground application systems, or prior mission configuration tables. This key event prior information can be generated during the satellite mission planning phase to guide the satellite's mission execution process. For the ground mission planning system, the observable time window for each ground target can be pre-calculated by combining satellite orbit prediction, attitude maneuverability, ground target coordinates, and user requirements. Furthermore, by combining preset constraints such as user task priority, imaging mode, and side-swing angle limits stored in the prior mission configuration table, a global optimization arrangement of all candidate targets can be performed to form a long-term coarse planning scheme as prior information.

[0056] Dynamic, weather-related prior knowledge can also be generated through ground application systems. For example, ground weather forecast information can be injected into satellites to predict cloud cover within the observation window. This allows satellites to know before flying over a target that a certain path is likely to be obscured by clouds, thereby triggering short-cycle rolling replanning, abandoning the obscured task or replacing it with another task, achieving autonomous decision-making based on prior information. Prior information on key events is uploaded to the onboard platform before mission execution, and then, along with the scenario complexity and data quality level, participates in the generation of processing level as an optional extension factor.

[0057] S103. Generate slice processing level based on complexity characteristics.

[0058] After obtaining the complexity features of an image slice using a complexity probe, a slice processing level can be generated based on these features. The slice processing level is calculated using the proportion of invalid pixels and the region of interest (ROI) attribution marker.

[0059] To generate slice processing levels, in some embodiments, when generating slice processing levels based on complexity features, preset level generation parameters corresponding to the task type can be obtained first. These preset level generation parameters include a first weight, a second weight, and at least one preset processing level threshold. The first weight is used to perform a weighted calculation on the proportion of invalid pixels; the second weight is used to perform a weighted calculation on the degree of overlap corresponding to the region of interest (ROI) attribution markers. Then, the effective pixel proportion of the image slice is determined based on the proportion of invalid pixels, where the effective pixel proportion is the difference between 1 and the proportion of invalid pixels.

[0060] Then, the degree of overlap between the image slice and the region of interest (ROI) is obtained, wherein the degree of ROI overlap is determined based on the ratio of the overlapping area of ​​the image slice to the area of ​​the image slice. Next, a processing level index is calculated based on the complementary value of the effective pixel ratio, the degree of ROI overlap, a first weight, and a second weight. The slice processing level is then determined according to the index interval to which the processing level index belongs. The index interval is a numerical range divided based on at least one preset processing level threshold.

[0061] For example, when generating slice processing levels, the spaceborne platform can generate a slice processing level d for each slice based on the slice complexity probe results. i At the slice processing level d i In this case, the comprehensive processing indicator value, i.e., the processing level index, for each image slice can be calculated first based on the proportion of invalid pixels and the region of interest attribution markers. r i =ω1(1-u i )+ω2a i ; Where, r i This represents the processing level index of the i-th slice, i.e., the overall processing indicator value, where ω1 represents the first weight; u i Indicates the proportion of invalid pixels; ω2 represents the second weight; a i Indicates the region of interest as a marker.

[0062] Based on the processing level index obtained from the above formula, the index range to which the processing level index belongs can be determined, and the slice processing level can be determined according to the index range to which it belongs, that is: d i =0, r i <η1; d i =1, η1≤r i <η2; d i =2, η2≤r i <η3; d i =3, r i ≥η3; Where, r i The processing level index represents the i-th slice; η1, η2, and η3 represent the processing level thresholds; d i Indicates the slice processing level. When d i When d = 0, it means that the image slice can be skipped directly, or only the most basic markers can be retained; when .... i When d = 1, it indicates that the image slice is processed using the lightest possible model; when d i When d = 2, it indicates that the image slice is processed using the standard model; when d i When the value is 3, it indicates that the image slice is processed using a higher capability model. That is, when the slice meets both "high cloud cover" and "outside the region of interest", it can be directly marked as skipping processing; when the slice is within the region of interest and the proportion of invalid pixels is lower than the preset threshold, it will enter the standard processing or high capability processing mode; in other cases, it can enter the lightest processing mode or standard processing mode.

[0063] It should be noted that the slice processing level determination method described in the above embodiments is only used as an example to illustrate how to group a small amount of key probe information into processing levels, and does not limit the specific weighting form. In other embodiments, processing level indicators can also be output using rule tables, segmentation logic, or lookup tables. That is, the slice processing level can be generated by formulas or rule tables. If scenario complexity, data quality level, or prior knowledge of key events is introduced in other embodiments, these can be used as optional expansion factors to further refine the processing level.

[0064] S104. Organize slice scheduling for multiple image slices according to task type and slice processing level to obtain slice scheduling results.

[0065] After generating the slice processing level, multiple image slices can be organized and scheduled according to the task type and slice processing level to obtain the slice scheduling result. The slice scheduling result can include a set of image slices to be processed, which can be used in subsequent scheduling to process the image slices one by one.

[0066] In some embodiments, after the slice processing level is generated, the subsequent scheduling of image slices may not adopt an immediate slice-by-slice execution approach, but rather be organized according to the task type and processing level of the slice, in order to reduce the additional overhead caused by frequent model switching. Therefore, the slice scheduling result may also include batch execution groups or batch execution queues determined jointly according to task type and processing level.

[0067] There are several optional methods for organizing and scheduling multiple image slices. Specifically, in some embodiments, slices can be grouped by task type and processing level and then executed in batches. When organizing and scheduling multiple image slices according to task type and slice processing level to obtain the slice scheduling results, the task type and slice processing level corresponding to each image slice can be traversed, and the model input specifications corresponding to each image slice can be obtained. The model input specifications include at least one of the following: input size, band combination, data type, numerical precision, and preprocessing configuration.

[0068] Image slices with the same task type, the same slice processing level, and compatible model input specifications are then grouped into the same batch execution group to obtain the slice scheduling result. The slice scheduling result includes at least one batch execution group and the execution order corresponding to the batch execution group.

[0069] For example, the spaceborne platform can first jointly group image slices according to their task type and processing level, grouping slices with the same task type and processing level into the same slice group. For instance, in object detection tasks, lightweight processed slices are grouped into the lightweight object detection group, standard processed slices into the standard object detection group, and higher-capacity processed slices into the high-capacity object detection group. Other task types can also be grouped in the same way.

[0070] In some embodiments, when organizing slice scheduling, slices can be enqueued in a mixed manner according to task type and processing level, and the queue can be executed in batches after certain conditions are met. Therefore, when organizing slice scheduling for multiple image slices according to task type and slice processing level, and obtaining the slice scheduling result, the image slices can be written into the batch execution queue according to task type and slice processing level. Then, the triggering conditions for the batch execution queue are obtained. The triggering conditions include at least one of the following: the number of slices in the queue reaches a batch processing quantity threshold; the current cumulative waiting time of the queue reaches a waiting time threshold; or a scene end flag, task end flag, or data stream termination flag is received.

[0071] The system monitors queue parameters and, when the queue parameters meet the trigger conditions, generates slice scheduling results according to the batch execution queues. The queue parameters include at least one of the following: the number of slices in the queue, the current cumulative waiting time of the queue, and the queue execution priority; the slice scheduling results include at least one batch execution queue and the execution order corresponding to the batch execution queues.

[0072] For example, after generating slice processing levels, the onboard platform may not immediately process all slices, but instead write them into corresponding processing queues according to task type and processing level. For instance, lightweight object detection slices are written to the first processing queue, standard object detection slices to the second, and lightweight scene recognition slices to the third. Then, for any processing queue, a batch processing round can be initiated when a preset trigger condition is met. The trigger conditions for batch execution queues may include: the number of slices in the queue reaching a corresponding batch processing threshold, and / or the current cumulative waiting time of the queue reaching a waiting time threshold. By monitoring queue parameters such as the number of slices in the queue and the current cumulative waiting time, a batch processing round can be initiated when any of the above trigger conditions are met. When a scene end flag, task end flag, or data stream termination flag is received, the remaining image slices in the queue are generated as the final batch to be executed. Therefore, the batch execution queue method is suitable for scenarios with continuous data inflow, high task concurrency, or the need to reduce model switching frequency.

[0073] S105. Based on the slice scheduling results, call the model profile library to match and process the model.

[0074] After slice scheduling is completed, a matching processing model can be invoked from the model image library based on the slice scheduling results. This processing model is used to perform image processing on the image slices. In other words, the spaceborne platform can invoke the corresponding processing model from the model image library to perform processing based on the task type and processing level of the slice. By matching processing models from the model image library, the model selection process is no longer about uniformly specifying a single model for the entire image, but rather matching processing methods for different slices separately.

[0075] Therefore, multiple candidate models can be pre-configured in the spaceborne platform. Different candidate models can be used to perform different remote sensing tasks and consume different spaceborne platform resources. For example, for the same type of task, the system pre-configures K candidate models with different complexity levels, denoted as: F={F （1） F （2） F （K）}; Different candidate models can correspond to different model levels such as the lightest processing, standard processing, and high-capacity processing; skip processing corresponds to a candidate processing method or processing level, rather than the candidate model itself.

[0076] To support subsequent model matching, the onboard platform can also create a corresponding model profile for each candidate model. In some embodiments, model profile information directly related to onboard scheduling may include task type identifier, model complexity level, inference time, resource consumption, applicable data conditions, and output result type. Inference time and resource consumption are related to both model type and slice attributes (such as slice size, numerical precision, etc.). These should be considered comprehensively, for example, using total inference time for filtering.

[0077] The model profile information of the k-th model can be represented as: P （k） =(h （k） g （k） , t （k） r （k） q （k） o （k） ); Among them, P （k） This represents the model profile information for the k-th model; h （k） Indicates the task type corresponding to the model; g （k） Indicates the level of model complexity; t （k） r represents the inference time of the model on the target spaceborne platform. （k） Indicates resource usage, q （k） Indicates the applicable data conditions, o （k） This indicates the type of output result. Resource usage can be represented as computing power usage, storage usage, or a combination thereof, depending on the project implementation needs; it is not necessary to expand all hardware resource variables during the handover phase.

[0078] like Figure 4 As shown, in some embodiments, when calling the model profile library to match and process models based on the slice scheduling results, the task type and slice processing level corresponding to the slice scheduling results can be read first, and the model profile information of candidate models in the model profile library can be obtained. Then, the current resource status of the spaceborne platform is obtained, and the processing method information is determined based on the task type, slice processing level, and model profile information. The current resource status includes at least one of available computing resources, available memory, remaining processing time, power budget, current queue load, and current model resident status; the processing method information includes skip processing and model inference processing.

[0079] For image slices whose processing method is model inference, a processing model can be matched from the model profile library based on the task type, slice processing level, and model profile information. For example, for each slice group formed in the above steps, or each processing queue that has met the triggering conditions, the spaceborne platform matches the corresponding processing model or processing method from the model profile library according to the corresponding task type, processing level, and model profile information. Under the target detection task, if a slice group or processing queue corresponds to a lightweight processing file, the lightweight detection model corresponding to that task is matched; if it corresponds to a standard processing file, the standard detection model is matched; if it corresponds to a high-capacity processing file, the high-precision detection model is matched; if it corresponds to a skip processing file, the skip processing method or the processing method that only retains basic markers is matched. The matching is based on the pre-established capability and resource profiles of each candidate model or processing method, rather than an unconstrained selection made temporarily at runtime.

[0080] During the model matching process, if multiple candidate models meet the matching requirements, the processing model set can be determined from the model profile library based on the task type, the slice processing level, and the model profile information. Then, the resource requirement information of each candidate model in the candidate model set is compared with the current resource status to exclude candidate models that do not meet the current resource constraints. When multiple candidate models satisfy the current resource constraints exist, the candidate models are ranked according to at least one of the model performance indicators, estimated total processing time, resource consumption, and model loading or switching overhead, and the processing model is determined based on the ranking result. The estimated total processing time is calculated based on the inference time, the number of image slices to be processed, and the model loading time. For example, the candidate model with the highest accuracy in the ranking result is preferentially selected as the processing model; or, the candidate model with the lowest resource consumption in the ranking result is preferentially selected as the processing model. The selection principle for the processing model can be determined comprehensively by considering the scarcity of available spaceborne resources and the output requirements.

[0081] For image slices whose processing method is "skip processing," the slice parameters can be retained. These parameters include the slice index, processing markers, and basic quality markers. In other words, for skip processing groups, only the slice index and basic markers are retained, and the slices are not included in the subsequent main task processing flow. This method is suitable for scenes where the entire image has been segmented and the slice set is relatively complete, helping to reduce the number of model loading and switching operations.

[0082] By applying the technical solutions of the above embodiments, the dynamic algorithm autonomous selection method for spaceborne remote sensing missions described in the above embodiments can pre-set multiple models of different complexities for the same type of mission, and establish capability and resource profiles for each model. Then, using observed image slices as basic processing units, key features of the slices are extracted through a low-cost complexity probe, and corresponding processing models or methods are matched to different slices accordingly. Therefore, the method can alleviate the processing load when applying the same processing method to all observation data. That is, the method uses image slices as basic processing units, allowing different regions to adopt different processing levels based on complexity characteristics such as the proportion of invalid pixels, region of interest affiliation, and processing value, thereby avoiding treating low-value slices and high-value slices equally.

[0083] The method also prioritizes limited onboard computing power for data processing that is more valuable. Specifically, it uses complexity probes to identify whether each slice contains high-invalidity pixels and whether it is located within a region of interest. Combined with model capabilities and resource profiles, it matches suitable models or processing methods to slices worthy of further processing, while skipping or performing minimal processing on obviously low-value or irrelevant slices. The method can further refine the levels by using factors such as scene complexity and data quality level. This reduces the waste of limited computing power on areas covered by thick clouds, outside regions of interest, or other low-value areas, allowing onboard processing resources to be prioritized for clear, important slices that are more likely to produce effective results.

[0084] The method also improves the targeting and consistency of model selection, with rules or indicators simultaneously referencing slice status, task type, and the model profile built on the target hardware. The profile provides processing capacity boundaries, resource consumption characteristics, and applicable conditions, giving model invocation a clear basis. This facilitates configuration and reuse, as well as unified management and expansion across different tasks, loads, and engineering platforms. Compared to switching methods based solely on simple thresholds, this matching method is more stable and easier to implement in engineering projects.

[0085] The method also balances processing effectiveness with engineering feasibility. Specifically, it uses image slices as the basic processing unit, striking a balance between processing granularity and engineering complexity. This allows for differentiated model calls for different parts of the scene while maintaining a clear data indexing, caching organization, and result management approach.

[0086] The method also reduces invalid processing and wasted computing power. In actual observation scenarios, the entire image often contains cloud-covered areas, areas outside the region of interest, and ordinary areas that do not require fine processing. If a high-complexity model is uniformly applied to these areas, it is easy to cause invalid processing. Therefore, the method first identifies the validity and value of the slices through a complexity probe, and then determines the corresponding processing method. This reduces unnecessary complex reasoning for obviously invalid or low-value areas, thereby reducing invalid reasoning within a limited computing budget and improving overall resource utilization efficiency.

[0087] The method also enhances the processing capabilities of key and high-quality regions. By differentiating processing value by slice, slices within regions of interest that possess higher image quality or warrant more detailed analysis can be matched with more suitable model levels. This improves the processing quality and usability of results for key and high-value regions without significantly increasing the overall processing burden.

[0088] The method also facilitates subsequent expansion to different tasks and models. Specifically, it organizes the processing flow according to model capabilities and resource profiling, complexity probes, and slice-level matching, without relying on a specific neural network structure or limiting it to a particular task type. Therefore, the method can be applied to various tasks such as object detection, scene classification, anomaly screening, cloud analysis, and change recognition. Furthermore, it can be expanded to new model levels or new complexity metrics based on different payloads, satellite platforms, and application requirements.

[0089] In some embodiments, as a refinement and extension of the specific implementation of the above embodiments, and to fully illustrate the specific implementation process of this embodiment, some embodiments of this application also provide a dynamic algorithm autonomous selection method for spaceborne remote sensing missions. The difference between this method and the above embodiments is that it can perform slice-level image processing and organize and write back the results, such as... Figure 5 As shown, the method includes: S201. Input the image slices from the slice scheduling results into the processing model slice by slice or in batches; S202, Read the task type; S203. Obtain the image processing results output by the processing model, and determine the data structure type of the image processing results according to the task type; S204. When the task type is the first type, according to the slice index and slice spatial position corresponding to the image slice, multiple raster-type processing results are spliced ​​or fused to generate a whole-scene raster task result corresponding to the whole-scene survey image. S205. When the task type is the second type, establish the association between the structured processing results and the corresponding positions in the whole landscape survey image based on the slice index.

[0090] After calling the model image library to match the processing model, the spaceborne platform can input the image slices from the slice scheduling results into the processing model slice by slice or in batches, so that the processing model can perform image processing on the image slices in the spaceborne processing resource environment corresponding to the task type and slice processing level.

[0091] Next, the task type is read, where the task type is either a first type or a second type; the first type represents tasks such as segmentation and augmentation that support write-back. The second type represents tasks where classification results, bounding boxes, attribute information, confidence scores, etc., are not applicable to write-back observation images.

[0092] Then, image slices from the same batch in the slice scheduling result are input into the processing model, so that the processing model performs image processing on the image slices under the onboard processing resource environment corresponding to the task type and slice processing level, and outputs image processing results. The image processing results output by the processing model are obtained, and the data structure type of the image processing results is determined according to the task type; the image processing results include at least one of raster-type processing results and structured processing results; raster-type processing results include at least one of segmentation mask, image enhancement, variable raster, and cell-level anomaly results; structured processing results include at least one of category results, target bounding boxes, target attributes, confidence scores, and anomaly event records.

[0093] When the task type is the first type, multiple raster-type processing results can be spliced ​​or fused according to the slice index and slice spatial position corresponding to the image slice to generate a whole-scene raster task result corresponding to the whole-scene survey image.

[0094] When the task type is the second type, the association between the structured processing result and the corresponding position in the whole landscape survey image can be established according to the slice index; wherein, when the structured processing result contains slice local coordinates, the slice local coordinates are converted into whole landscape image coordinates or geographic coordinates, and the converted structured processing result is written into the task result data.

[0095] In some embodiments, when the task type is the first type, the spaceborne platform can first read the slice index corresponding to the image slice and determine the overall position of the image slice in the overall landscape survey image based on the slice index. Then, it obtains the task result structure and slice overlap information corresponding to the spaceborne remote sensing task. Then, according to the overall position, task result structure, and slice overlap information, the image processing results are written back to the overall landscape survey image to generate task result data. If adjacent slices overlap, at least one of the following is performed on multiple processing results in the overlapping area: confidence comparison, weighted fusion, voting fusion, center region priority, or duplicate target elimination.

[0096] To avoid boundary effects, where the target itself is segmented by different slices, slice overlap should be considered when generating slices. When slice overlap is implemented, overlapping information is generated, requiring consideration of result fusion when writing back to the whole scene. Methods such as mask fusion or intersection of overlapping regions can be used to write the image processing results back to the whole scene survey image based on the whole scene location, task result structure, and slice overlap information, thus generating task result data.

[0097] For example, after calling the model profile library to match and process the model, the spaceborne platform can perform slice-level processing and output the results. That is, the spaceborne platform performs corresponding processing on slice groups or processing queues that have completed model matching. For slice groups or processing queues that need to execute the main task model, the matched model can be called for batch processing and the corresponding results can be output. For slices that are determined to skip processing, only the slice index, processing mark, or basic quality mark can be retained, without executing the complete task model. The output results can be defined according to the model profile as category results, target boxes, segmentation masks, anomaly marks, or enhanced images, etc. For slices using the batch execution method, although they are processed in the same batch, each slice still maintains an independent index relationship to facilitate subsequent result organization, write-back, and download.

[0098] Then, the results are organized and written back. The onboard platform can organize the processing results of each slice and write them back to the corresponding overall scene location or mission result structure according to the slice index. Since this method can use fixed slices as the basic processing unit, the result index relationship is clear, facilitating subsequent onboard caching, downlink organization, and ground interpretation.

[0099] By applying the technical solutions of the above embodiments, the dynamic algorithm autonomous selection method for spaceborne remote sensing missions described in the above embodiments can write back the results to the corresponding whole scene position or mission result structure according to the slice index through result organization and write-back, forming mission result data corresponding to the whole scene survey image. Under the premise of making full and reasonable use of the limited computing power resources on the satellite, the adaptability of the output result structure is guaranteed, and subsequent caching, downlink organization and ground interpretation are supported.

[0100] In some embodiments, as a specific implementation of the dynamic algorithm autonomous selection method for spaceborne remote sensing missions described in the above embodiments, some embodiments of this application also provide a dynamic algorithm autonomous selection system for spaceborne remote sensing missions, such as... Figure 6 As shown, the system includes: The image slice management module is used to acquire observation image data and the task type corresponding to the observation image data. The observation image data includes multiple image slices obtained by segmenting the whole landscape observation image. The complexity probe module is used to perform a complexity probe on the image slice to obtain the complexity features of the image slice, the complexity features including the proportion of invalid pixels and the region of interest attribution marker; A processing level generation module is used to generate a slice processing level based on the complexity feature. The slice processing level is obtained by performing a weighted calculation on the proportion of invalid pixels and the region of interest attribution label. The batch processing scheduling module is used to organize slice scheduling for multiple image slices according to the task type and the slice processing level, and obtain slice scheduling results; The model matching module is used to call the model profile library to match the processing model based on the slice scheduling result. The model profile library includes multiple candidate models with different complexity levels and the model capabilities and resource profiles corresponding to the candidate models. The processing model is used to perform image processing on the image slice.

[0101] A dynamic algorithm autonomous selection system for spaceborne remote sensing missions can be deployed in a payload processor, a spaceborne computing platform, a spaceborne AI computing unit, or a combination thereof. The system can receive whole-scene survey images through an image slicing management module, perform slicing, and establish slice indexes, spatial locations, and basic metadata records. The image slicing management module can also maintain the correspondence between slices and the whole-scene image to support the writing back and organization of subsequent processing results.

[0102] The system can also perform fast complexity detection on slices through a complexity probe module, generating complexity features related to model selection. These complexity features include one or more of the following: invalid pixel ratio or region of interest attribution; they can also include one or more of the following: scene complexity, data quality level, and prior knowledge of key events. The complexity probe module can be implemented using a fast algorithm with lower overhead than the main task model, to avoid the complexity probe itself becoming the main computational burden.

[0103] The system may also deploy a model profiling management module to store and maintain profiling information for multiple candidate models and related processing methods corresponding to the same type of task. In some embodiments, the profiling information includes one or more of the following: task type, complexity level, inference time, resource consumption, applicable data conditions, and output result type. For non-model processing methods such as skipping processing, the profiling information is at least used to characterize their resource consumption and output constraints. The model profiling management module can also be used to provide predefined criteria for subsequent model matching, rather than performing unconstrained model selection at runtime.

[0104] The system can also determine the processing level corresponding to each slice based on the slice complexity characteristics output by the complexity probe module through the processing level generation module. The processing level can include at least one or more of skip processing, lightest processing, standard processing, and high-capacity processing. The processing level generation module can generate the processing level using rule tables, segmentation logic, lookup tables, or weighted indexes.

[0105] The model matching module and batch processing scheduling module can be used to match corresponding processing models or methods to different slices based on their task type, processing level, and model profile information, and to perform subsequent scheduling and organization of the slices. The batch processing scheduling module can handle the scheduling control functions of grouping by level, enqueuing, and triggering batch processing.

[0106] In some embodiments, the batch processing scheduling module can group slices according to processing level and perform batch scheduling on slices of the same level. In other embodiments, the batch processing scheduling module writes slices of different levels into corresponding processing queues and performs batch scheduling when preset trigger conditions are met. The trigger conditions include one or more of the following: queue length reaches a threshold, cumulative waiting time reaches a threshold. Thus, the model matching module can complete model selection.

[0107] The system may further include a task execution module and a result organization module, used to call the matched model, perform actual task processing on slice groups of the same level or processing queues that meet the triggering conditions, and output corresponding results. The results may include one or more of the following: category results, bounding boxes, segmentation masks, anomaly markers, and enhanced images. The task execution module and result organization module are also used to organize the processing results according to slice indices and write the results back to the corresponding scene location or task result structure to support subsequent caching, download organization, and ground interpretation.

[0108] The aforementioned modules can be physically deployed separately or integrated as logical functional units within the same onboard application. In some embodiments, the processing level generation module and the model matching and batch processing scheduling module can be implemented independently or merged into a unified scheduling control module; however, regardless of the implementation method, the system should at least support generating processing levels for slices first, then completing model matching based on the processing levels and model profiles, and organizing subsequent processing flows by batch execution in groups by level or by queuing by level and triggering batch execution.

[0109] By applying the technical solutions of the above embodiments, the dynamic algorithm autonomous selection system for spaceborne remote sensing missions described in the above embodiments can, after the image slice management module acquires the observation image data, the complexity probe module performs complexity probes on the image slices in the observation image data to obtain complexity features such as the proportion of invalid pixels and region of interest (ROI) attribution markers. The processing level generation module then generates slice processing levels based on the complexity features, and the batch processing scheduling module organizes slice scheduling for multiple image slices according to task type and slice processing level. Then, the model matching module calls the model image library to match processing models based on the slice scheduling results and performs image processing on the image slices. The system can pre-establish multiple processing models of different complexities for spaceborne remote sensing missions, using image slices as basic processing units, obtaining the complexity features of each slice through complexity probes, and matching corresponding processing models for different slices to fully and rationally utilize the limited computing resources on the satellite.

[0110] It should be noted that other corresponding descriptions of the functional units involved in the dynamic algorithm autonomous selection system for spaceborne remote sensing missions provided in the embodiments of this application can be found in the corresponding descriptions in the dynamic algorithm autonomous selection method for spaceborne remote sensing missions provided in the above embodiments, and will not be repeated here.

[0111] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0112] The embodiments described above are merely examples of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application.

Claims

1. A dynamic algorithm autonomous selection method for spaceborne remote sensing missions, characterized in that, The method includes: Acquire observation image data and the task type corresponding to the observation image data, wherein the observation image data includes multiple image slices obtained by segmenting the whole landscape observation image; A complexity probe is performed on the image slice to obtain the complexity features of the image slice, which include the proportion of invalid pixels and the region of interest attribution markers; A slice processing level is generated based on the complexity features. The slice processing level is calculated using the invalid pixel ratio and the region of interest (ROI) attribution marker, including: obtaining preset level generation parameters corresponding to the task type, the preset level generation parameters including a first weight, a second weight, and at least one preset processing level threshold; the first weight is used to perform a weighted calculation on the invalid pixel ratio; the second weight is used to perform a weighted calculation on the overlap degree corresponding to the ROI attribution marker; determining the effective pixel ratio of the image slice based on the invalid pixel ratio, the effective pixel ratio being the difference between 1 and the invalid pixel ratio; obtaining the ROI overlap degree between the image slice and the ROI, the ROI overlap degree being determined based on the ratio of the overlap area between the image slice and the ROI to the area of ​​the image slice; calculating a processing level index based on the effective pixel ratio, the ROI overlap degree, the first weight, and the second weight; and determining the slice processing level according to the index interval to which the processing level index belongs; the index interval being a numerical interval divided based on at least one preset processing level threshold. Organize slice scheduling for multiple image slices according to the task type and the slice processing level to obtain slice scheduling results; Based on the slice scheduling result, the model profile library is invoked to match the processing model. The model profile library configures multiple candidate models with different complexity levels for the same task type and stores the model capabilities and resource profiles corresponding to each candidate model. The processing model is used to perform image processing on the image slice.

2. The method according to claim 1, characterized in that, Acquiring the observed image data and the corresponding task type, including: Receive mission planning data and overall landscape survey images acquired by onboard image acquisition equipment, wherein the mission planning data includes at least one of ground mission instructions and onboard autonomous planning information; Read the task type from the task planning data; Obtain slice configuration parameters, including payload space resolution, cache organization method, and task type; Slicing rules are set according to the slice configuration parameters. The slice rules include at least the slice size; the slice rules also include one or more of the following: slice step size, overlap width, edge padding method, and row and column numbering; the slice size is obtained according to the payload space resolution, the cache organization method, and the task type. The overall landscape observation image is divided into multiple image slices according to the slicing rules to generate the observation image data.

3. The method according to claim 1, characterized in that, A complexity probe is performed on the image slice to obtain its complexity features, including: Invalid pixel detection is performed on the image slice to determine the invalid pixel ratio, which is the ratio of cloud pixels, sensor bad pixels, missing pixels and fill values ​​calculated by the invalid pixel detection method. Perform region of interest (ROI) attribution detection on the image slices to obtain ROI attribution labels, which are classification labels obtained by overlaying slice spatial index and ROI range. The complexity features of the image slice are generated using the proportion of invalid pixels and the region of interest attribution markers as core features.

4. The method according to claim 3, characterized in that, Performing a complexity probe on the image slice to obtain the complexity features of the image slice, further includes: Extract complex characterization parameters from the image slices, wherein the complex characterization parameters include at least one of texture statistics, edge density, and feature encoding results; The complexity of the scenario is calculated based on the aforementioned complex representation quantity; Extract quality characteristics from the image slices, the quality characteristics including at least one of blur level, saturation ratio and basic quality marker; Calculate the data quality level based on the quality characterization parameters; Prior information on key events is obtained. This prior information is pre-generated by the ground mission planning system and the ground application system based on the prior mission configuration table. The prior information on key events includes at least one of the following: geographical scope of the event, event type, and priority identifier. The complexity features of the image slices are generated using the scene complexity, the data quality level, and the prior information of the key events as extended factor features.

5. The method according to claim 1, characterized in that, Slice scheduling is organized for multiple image slices according to the task type and the slice processing level, resulting in slice scheduling results, including: Iterate through the task type and slice processing level corresponding to the image slice; Obtain the model input specifications corresponding to each of the image slices, wherein the model input specifications include at least one of the following: input size, band combination, data type, numerical precision, and preprocessing configuration; Image slices with the same task type, the same slice processing level, and compatible model input specifications are grouped into the same batch execution group to obtain the slice scheduling result; the slice scheduling result includes at least one batch execution group and the execution order corresponding to the batch execution group; or, Write the image slices into the batch execution queue according to the task type and the slice processing level; The triggering conditions for the batch execution queue are obtained, and the triggering conditions include at least one of the following: the number of slices in the queue reaches the batch processing quantity threshold, the current cumulative waiting time of the queue reaches the waiting time threshold, and the whole scene end flag, task end flag, or data stream termination flag is received. The system monitors queue parameters and generates slice scheduling results according to the batch execution queue when the queue parameters meet the triggering condition. The queue parameters include at least one of the following: the number of slices in the queue, the current cumulative waiting time of the queue, and the queue execution priority. The slice scheduling results include at least one batch execution queue and the execution order corresponding to the batch execution queue.

6. The method according to claim 1, characterized in that, Based on the slice scheduling results, the model profile library is invoked to match and process the model, including: Read the task type and slice processing level corresponding to the slice scheduling result; Obtain model profile information of candidate models in the model profile library. The model profile information includes at least one of the following: task type identifier, model complexity level, model performance index, model capability boundary, inference time, resource consumption, applicable data conditions, and output result type. Obtain the current resource status of the spaceborne platform, wherein the current resource status includes at least one of available computing resources, available memory, remaining processing time, power budget, current queue load, and current model dwell status; Based on the task type, the slice processing level, and the model profile information, the processing method information is determined, which includes skip processing and model inference processing. For the image slice for which the processing method information is the model inference processing, a processing model is matched from the model profile library according to the task type, the slice processing level, and the model profile information; For the image slice whose processing method information is "skip processing", the slice parameters of the image slice are retained. The slice parameters include the slice index, processing mark, and basic quality mark.

7. The method according to claim 6, characterized in that, The method further includes: The image slices in the slice scheduling result are input into the processing model slice by slice or batch, so that the processing model performs image processing on the image slices in the on-board processing resource environment corresponding to the task type and the slice processing level; Read the task type, which is either a first type or a second type; The image processing results output by the processing model are obtained, and the data structure type of the image processing results is determined according to the task type; the image processing results include at least one of raster processing results and structured processing results; the raster processing results include at least one of segmentation mask, image enhancement, variable raster, and cell-level anomaly results; the structured processing results include at least one of category results, target bounding boxes, target attributes, confidence scores, and anomaly event records; When the task type is the first type, multiple raster-type processing results are spliced ​​or fused according to the slice index and slice spatial position corresponding to the image slice to generate a whole-scene raster task result corresponding to the whole-scene survey image. When the task type is the second type, the association between the structured processing result and the corresponding position in the whole landscape survey image is established according to the slice index; wherein, when the structured processing result contains slice local coordinates, the slice local coordinates are converted into whole landscape image coordinates or geographic coordinates, and the converted structured processing result is written into the task result data.

8. The method according to claim 7, characterized in that, Based on the slice index and spatial location corresponding to the image slice, multiple raster-type processing results are stitched or fused to generate a whole-scene raster task result corresponding to the whole-scene survey image, including: Read the slice index corresponding to the image slice; The position of the image slice in the overall landscape assessment image is determined based on the slice index; Obtain the mission result structure and tile overlap information corresponding to the spaceborne remote sensing mission; According to the overall scene location, the task result structure, and the slice overlap information, the image processing result is written back to the overall scene survey image to generate task result data; If adjacent slices overlap, perform at least one of the following on multiple processing results in the overlapping region: confidence comparison, weighted fusion, voting fusion, center region priority, or duplicate target elimination.

9. A dynamic algorithm autonomous selection system for spaceborne remote sensing missions, characterized in that, The system includes: The image slice management module is used to acquire observation image data and the task type corresponding to the observation image data. The observation image data includes multiple image slices obtained by segmenting the whole landscape observation image. The complexity probe module is used to perform a complexity probe on the image slice to obtain the complexity features of the image slice, the complexity features including the proportion of invalid pixels and the region of interest attribution marker; A processing level generation module is used to generate a slice processing level based on the complexity features. The slice processing level is calculated using the invalid pixel ratio and the region of interest (ROI) attribution marker. The module includes: obtaining preset level generation parameters corresponding to the task type, the preset level generation parameters including a first weight, a second weight, and at least one preset processing level threshold; the first weight being used to perform a weighted calculation on the invalid pixel ratio; the second weight being used to perform a weighted calculation on the overlap degree corresponding to the ROI attribution marker; determining the effective pixel ratio of the image slice based on the invalid pixel ratio, the effective pixel ratio being the difference between 1 and the invalid pixel ratio; obtaining the ROI overlap degree between the image slice and the ROI, the ROI overlap degree being determined based on the ratio of the overlap area between the image slice and the ROI to the area of ​​the image slice; calculating a processing level index based on the effective pixel ratio, the ROI overlap degree, the first weight, and the second weight; and determining the slice processing level according to the index interval to which the processing level index belongs, the index interval being a numerical interval divided based on at least one preset processing level threshold. The slice scheduling module is used to organize slice scheduling for multiple image slices according to the task type and the slice processing level, and obtain slice scheduling results; The model matching module is used to call the model profile library to match the processing model based on the slice scheduling result. The model profile library configures multiple candidate models with different complexity levels for the same task type and stores the model capabilities and resource profiles corresponding to each candidate model. The processing model is used to perform image processing on the image slice.

Images