An unattended on-board automated processing method for satellite remote sensing image data

Abstract

This invention discloses an unattended onboard automated processing method for satellite remote sensing image data, belonging to the field of satellite remote sensing data processing technology. The method includes: performing adaptive dynamic task scheduling based on onboard resource status, data priority, and task deadlines; performing modality adaptive adaptation, core preprocessing operations, and preprocessing quality checks for remote sensing image data of different modalities; automatically initiating corresponding intelligent processing flows to form a complete processing archive; constructing a multi-level quality detection mechanism; and introducing a layered fault diagnosis and self-healing mechanism. This invention achieves unattended automated operation of the entire process of image data acquisition, preprocessing, intelligent processing, quality detection, and data distribution / storage. It possesses autonomous fault diagnosis and recovery capabilities, dynamic task scheduling, and multi-modal adaptation capabilities, improving the real-time performance and reliability of image processing, adapting to various remote sensing satellites and multi-modal image processing scenarios; and significantly reducing the amount of data transmitted between satellite and ground, thus reducing the pressure on satellite-to-ground transmission.

Description

Technical Field

[0001] This invention relates to the field of satellite remote sensing data processing technology, specifically, to an unattended on-board automated processing method for satellite remote sensing image data. Background Technology

[0002] With the rapid development of satellite remote sensing technology, the large-scale deployment of remote sensing satellite constellations has propelled remote sensing image data into an era of "massive volume, multimodal nature, and high timeliness." The amount of image data collected daily by various remote sensing satellites is growing exponentially, placing extremely high demands on the real-time and autonomous nature of data processing. Traditional remote sensing image processing adopts a "satellite acquisition-ground transmission-ground processing" model, which suffers from the following core technical pain points and cannot meet the application requirements of unattended operation and real-time response:

[0003] 1. Reliance on ground-based manual intervention and low level of automation: Traditional on-board image processing can only complete simple data acquisition and storage. Core processes such as preprocessing, feature extraction, and quality inspection all require data to be sent down to the ground and the processing flow to be started manually. It cannot achieve unattended operation of the entire process on the satellite, has high response latency, and is difficult to adapt to scenarios that require real-time processing, such as disaster emergencies.

[0004] 2. High pressure and poor real-time performance of satellite-to-ground transmission: All massive amounts of remote sensing image data are downlinked to the ground for processing. Due to the limitations of satellite-to-ground communication bandwidth and visible arc, the transmission efficiency is low and the latency is high. This not only increases the cost of satellite-to-ground communication, but also causes the image processing cycle to be as long as several hours or even days, which cannot meet the needs of real-time monitoring and rapid response.

[0005] 3. Weak fault response capability and insufficient reliability: The on-board processing system lacks an autonomous fault diagnosis and recovery mechanism. When problems such as abnormal data acquisition, processing module failure, or storage overflow occur, it cannot be identified and handled autonomously, requiring manual intervention from the ground for troubleshooting. This leads to interruption of the processing flow and affects mission continuity. This problem is particularly prominent in scenarios with long communication delays, such as deep space exploration.

[0006] 4. Rigid task scheduling and low resource utilization: Existing on-board processing systems mostly adopt fixed task scheduling strategies, which cannot dynamically adjust the processing order according to the on-board computing power, storage resource status, and image data priority (such as images of emergency disaster areas). This easily leads to wasted computing power and delayed processing of critical tasks, and it is difficult to adapt to the differentiated processing needs of multimodal images.

[0007] 5. Poor adaptability and difficulty in compatibility with multiple scenarios: Existing on-board processing workflows are mostly customized for specific modes and specific satellite models, lacking universality and unable to flexibly adapt to multi-modal images such as optical, SAR, and hyperspectral images. Furthermore, they are difficult to adapt to on-board platforms with different computing power and storage configurations, resulting in poor scalability.

[0008] Currently, while there are preliminary attempts at automating on-board processing in related technologies, these are mostly limited to automating single stages (such as automatic preprocessing), failing to form a closed-loop architecture with full-process unattended operation. Furthermore, they lack core capabilities such as autonomous fault recovery and dynamic task scheduling, thus failing to address the aforementioned pain points. Therefore, developing a fully unattended, highly reliable, real-time, and scalable automated on-board processing workflow for satellite remote sensing images has become an urgent technical problem to be solved in the field of satellite remote sensing technology. Summary of the Invention

[0009] The purpose of this invention is to provide an unattended on-board automated processing method for satellite remote sensing image data, which solves the problems of existing technologies where all massive remote sensing image data is downlinked to the ground for processing, resulting in high pressure on satellite-to-ground transmission and poor real-time performance, as well as the low degree of automation in on-board processing, where automation is limited to a single link and does not form a closed-loop architecture of unattended operation throughout the entire process, thus failing to automatically realize dynamic task scheduling and autonomous fault recovery.

[0010] The present invention solves the above problems through the following technical solution:

[0011] An unattended on-board automated processing method for satellite remote sensing image data includes:

[0012] Step S1: Automatically acquire, preprocess, and cache multimodal image data;

[0013] Step S2: Based on the on-board resource status, data priority, and task deadline, perform adaptive dynamic task scheduling;

[0014] Step S3: For remote sensing image data of different modalities, a lightweight algorithm that is more practical than the theoretical optimum is adopted to automatically perform modal adaptive adaptation, core preprocessing operations and preprocessing quality inspection, ensuring that no manual intervention is required throughout the entire process;

[0015] Step S4: Based on the on-board lightweight AI processing platform, the corresponding intelligent processing flow is automatically started according to the preset task requirements to realize image feature extraction, land cover classification and target detection. The processing results are automatically associated and stored with the original image and the pre-processed image to form a complete processing archive.

[0016] Step S5: Construct a multi-level quality inspection mechanism to achieve automatic quality inspection throughout the entire process of preprocessing and intelligent processing. After passing the inspection, the data will be automatically distributed or stored.

[0017] Step S6: Introduce a hierarchical fault diagnosis and self-healing mechanism to achieve autonomous fault identification, diagnosis and recovery throughout the entire onboard processing process, ensuring the continuity of the process in an unattended state.

[0018] This invention enables fully unattended automated operation of image data from acquisition, preprocessing, intelligent processing, quality testing to data distribution / storage. It features autonomous fault diagnosis and recovery, dynamic task scheduling, and multimodal adaptation capabilities, while also meeting the low computing power and low power consumption requirements of satellites. This improves the real-time performance and reliability of image processing and is compatible with various remote sensing satellites and multimodal image processing scenarios. The entire process is completed on-board, significantly reducing the amount of data transmitted between satellite and ground, and alleviating the pressure on satellite-to-ground transmission.

[0019] Further, step S1 specifically includes:

[0020] Based on the satellite's preset orbital parameters and mission planning instructions, the remote sensing payload of the corresponding mode is automatically activated;

[0021] The remote sensing payload acquires remote sensing image data in optical, SAR, infrared and hyperspectral modes in real time, and records metadata of the remote sensing image data simultaneously.

[0022] The system automatically parses the acquired raw image data, adapts the format of data of different modalities, stores the parsed data on the satellite, and adopts a cyclic overwrite strategy to prioritize the retention of high-priority data to avoid storage overflow, while also performing data encryption caching.

[0023] Further, step S2 specifically includes:

[0024] The on-board computing power is collected in real time, a resource status assessment model is established, and the on-board processing capacity is dynamically judged. The on-board computing power includes processor utilization, storage capacity, and energy consumption status. The resource status assessment model is used to perform weighted normalization, bottleneck factor identification, and dynamic weight adjustment processing on the on-board computing power data, and output a comprehensive status classification.

[0025] Based on the imaging area, imaging quality, and task type of the remote sensing image data, the remote sensing image data is divided into priority levels.

[0026] By combining comprehensive status classification and priority levels, processing resources are dynamically allocated and the processing order is adjusted.

[0027] Furthermore, the strategy of dynamically allocating processing resources and adjusting the processing order by combining comprehensive status classification and priority levels includes: prioritizing the allocation of currently available resources to the highest priority and fastest-progressing tasks, specifically:

[0028] Employing a dynamic scheduler kernel based on a greedy algorithm, the scheduler scans the ready queue at each heartbeat and processes each waiting task. Calculate an immediate utility value :

[0029] ;

[0030] in, Score the data priority of the task; The time a task waits in the queue; the longer it waits, the greater the starvation penalty, preventing low-priority tasks from starving indefinitely. The degree of matching between the resources required for the mission and the current state of resources on the satellite; This refers to the dynamic weighting coefficients for task priority. The dynamic weighting coefficient for the hunger penalty;

[0031] In each scheduling cycle, select The task with the highest value is put into operation.

[0032] Furthermore, the strategy of dynamically allocating processing resources and adjusting processing order by combining comprehensive status classification and priority levels also includes one or more of the following:

[0033] (1) Computing power adaptive parallel control strategy: When the on-board computing power is sufficient, multi-modal and multi-batch images are processed in parallel, and the parallelism is dynamically adjusted by real-time monitoring of resource fragments;

[0034] (2) Preemptive scheduling and resource reclamation strategy: When a sudden high-priority task arrives and the system resources are already full, a preemptive mechanism is triggered, and the system searches for the best performing task currently in operation. The lowest priority task is not terminated. Instead, its current register state, program counter, and half-processed data cache address are packaged into a checkpoint and stored in non-volatile memory. The computing power and memory resources occupied by the task are forcibly reclaimed, and the sudden high-priority task is immediately put into the newly freed resource block for execution. After execution, the process jumps to the checkpoint to continue execution.

[0035] (3) Resumable download mechanism:

[0036] Three-phase commit storage is adopted:

[0037] Prepare: Writes intermediate processing data to a temporary buffer;

[0038] Snapshot: Records metadata about the current task progress;

[0039] Commit: Updates the status bit, marking the task as "stage completed";

[0040] When the system restarts or recovers from an exception, the bootloader first checks the exception database. If it finds a task that is interrupted, it reads its last checkpoint, restores the state, and requests resources from the scheduler again to continue execution from where it was interrupted, instead of starting from the beginning.

[0041] Furthermore, when onboard computing power is sufficient, the parallel processing of multimodal and multi-batch images, and the dynamic adjustment of parallelism through real-time monitoring of resource fragmentation, specifically include:

[0042] The on-board computing power is divided into multiple resource blocks;

[0043] Check if the current highest priority task can be split; if it can be split, proceed to the next step.

[0044] Parallel distribution: if immediate utility value If the resource density is high and the system has excess resource blocks, the scheduler will distribute the blocks to idle CPU cores or GPUs for streaming parallel processing.

[0045] Synchronously monitor memory bandwidth usage. If multiple tasks running in parallel cause memory bandwidth saturation, automatically reduce the number of parallel tasks and switch to serial pipeline mode.

[0046] Furthermore, step S3 specifically includes:

[0047] A lightweight decision logic based on statistical features is used to quickly calculate the entropy, kurtosis and number of bands of an image, and to complete the modal adaptation of optical, SAR, infrared and hyperspectral data.

[0048] The modality-adapted data is then processed using the corresponding modality preprocessing algorithm, which includes:

[0049] (1) Automatically perform brightness normalization, Gaussian noise reduction and geometric correction on optical images;

[0050] (2) Automatically perform Lee filtering to remove speckles and normalize the backscattering coefficient for SAR images;

[0051] (3) Automatically perform radiometric correction and band alignment for infrared and hyperspectral images;

[0052] All preprocessing algorithms are optimized for lightweight processing and computational power reduction. They all adopt a tile-based streaming processing architecture, where the image to be processed is divided into small blocks of fixed size and flows into the processing pipeline one by one. At the same time, the preprocessing algorithms integrate a computational graph pruning mechanism, which automatically skips unnecessary processing steps when a decrease in the overall state level is detected.

[0053] After preprocessing, key image indicators are automatically extracted for initial inspection, including:

[0054] Sharpness detection: The Laplacian variance operator is used to achieve the best balance between computational cost and sensitivity; a 3×3 integer convolution kernel is applied to the image to calculate the variance of the response value, and if it is lower than the dynamic threshold, it is judged as blurry;

[0055] Contrast evaluation: Directly calculate the dynamic range or RMS contrast of pixel values. If overexposed or hazy data is found, trigger automatic retry logic immediately.

[0056] Registration accuracy: The downsampled image is used to perform fast cross-correlation calculation. Only the root mean square error of the feature point residuals is counted. Once it exceeds the set pixel-level threshold, it is judged as a registration failure.

[0057] Images that fail the initial inspection will be automatically reprocessed after adjusting the denoising parameters or redoing the registration. If the images still fail after multiple preprocessing attempts, they will be marked as abnormal data, stored in the abnormal database, and automatically reported to the ground later.

[0058] Furthermore, step S5 specifically includes:

[0059] The quality of image preprocessing, the accuracy of intelligent processing results, and the completeness and correlation of overall detection data are checked to ensure that the processing flow is free of abnormalities.

[0060] Data that fails the test is automatically marked as abnormal and stored in the abnormal database. At the same time, the abnormal early warning mechanism is activated. When the space-to-ground communication link is clear, the abnormal information is automatically reported to the ground without manual intervention.

[0061] Data that passes the test is automatically distributed or stored according to preset rules: high-priority data is automatically downlinked to the ground receiver via the satellite-to-ground communication link; medium and low-priority data are automatically stored in the satellite's large-capacity storage module and then downlinked to the ground receiver in batches when the satellite-to-ground communication is idle.

[0062] Further, step S6 specifically includes:

[0063] Real-time monitoring of the operational status of each component and capture of fault signals;

[0064] By adopting a hierarchical diagnostic approach at the module, unit, and system levels, faults can be accurately located, and minor, general, and serious faults can be distinguished.

[0065] The system enables autonomous fault healing, including:

[0066] Minor faults will automatically trigger resource adjustments and restore process operation.

[0067] For general faults, the system automatically starts to replace redundant modules and resumes interrupted data transmission.

[0068] In the event of a serious malfunction, the system will automatically restart and return to its pre-malfunction state to continue processing tasks. At the same time, the serious malfunction information will be recorded and reported to the ground.

[0069] A redundant backup mechanism is employed to ensure rapid switching in the event of a failure, thus avoiding interruption of the processing flow.

[0070] Furthermore, it also includes step S7: after the ground update algorithm and mission planning instructions are received, they are automatically uploaded to the satellite system through the satellite-ground communication link. The satellite system automatically verifies the legality of the updated content. After the verification is passed, the update is automatically updated. After the update is completed, there is no need to restart. It is directly applied to the subsequent processing flow to achieve continuous optimization of the system.

[0071] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0072] (1) This invention operates autonomously throughout the entire process from image acquisition, preprocessing, intelligent processing to quality inspection, data output, and fault recovery, without any ground-based human intervention. This solves the problem of traditional on-board processing relying on ground intervention, making it suitable for unattended on-board operation scenarios, and especially suitable for missions with long communication latency such as deep space exploration.

[0073] (2) In this invention, the entire process is completed on the satellite, and only qualified processing results or key data are sent down to the ground, which greatly reduces the amount of data transmitted between the satellite and the ground, and compresses the image processing cycle from several hours / days to minutes, meeting the timeliness requirements of disaster emergency response, real-time monitoring and other scenarios.

[0074] (3) The present invention adopts a layered fault diagnosis and self-healing mechanism, combined with a backup design, which can autonomously identify and handle various faults, avoid interruption of the processing flow, improve the reliability of the on-board processing system, and reduce ground intervention costs.

[0075] (4) The present invention adopts an adaptive task scheduling strategy to dynamically allocate on-board computing power and storage resources, and adjusts the processing order according to data priority to avoid wasting computing power; it supports multimodal image adaptation, is compatible with various remote sensing payloads such as optical, SAR, and infrared, and is adaptable to on-board platforms with different computing power and storage configurations, with strong scalability.

[0076] (5) All processing algorithms in this invention are optimized for lightweight design, taking into account both processing accuracy and low power consumption and low computing power requirements. They can be widely used in various remote sensing satellites such as low orbit and high orbit, and are highly practical.

[0077] (6) The present invention can realize the dynamic updating of algorithms and mission planning through the automatic update mechanism of satellite-ground collaboration, without the need to modify the hardware of the satellite system, thereby improving the flexibility and scalability of the system and extending the functional life cycle of the satellite in orbit. Attached Figure Description

[0078] Figure 1 This is an architecture diagram of an embodiment of the present invention;

[0079] Figure 2 This is an architecture diagram of the adaptive task scheduling module according to an embodiment of the present invention;

[0080] Figure 3 This is a flowchart illustrating the fault autonomous diagnosis and self-healing process in an embodiment of the present invention. Detailed Implementation

[0081] The present invention will be further described in detail below with reference to embodiments, but the implementation of the present invention is not limited thereto.

[0082] This invention provides an unattended on-board automated processing method for satellite remote sensing image data. The implementation architecture and processing logic are described below. Figure 1 As shown, a closed-loop process is formed, encompassing multimodal image acquisition and caching, adaptive task scheduling, automatic preprocessing, intelligent processing, quality detection and output, autonomous fault diagnosis and self-healing, and satellite-ground collaborative updates. The core onboard hardware components (hybrid heterogeneous computing core, storage module, and payload adaptation module) are also demonstrated. The methods include:

[0083] Step S1: Automatically acquire, preprocess, and cache multimodal image data;

[0084] Step S2: Based on the on-board resource status, data priority, and task deadline, perform adaptive dynamic task scheduling;

[0085] Step S3: For remote sensing image data of different modalities, a lightweight algorithm that is more practical than the theoretical optimum is adopted to automatically perform modal adaptive adaptation, core preprocessing operations and preprocessing quality inspection, ensuring that no manual intervention is required throughout the entire process;

[0086] Step S4: Based on the on-board lightweight AI processing platform, the corresponding intelligent processing flow is automatically started according to the preset task requirements to realize image feature extraction, land cover classification and target detection. The processing results are automatically associated and stored with the original image and the pre-processed image to form a complete processing archive.

[0087] Step S5: Construct a multi-level quality inspection mechanism to achieve automatic quality inspection throughout the entire process of preprocessing and intelligent processing. After passing the inspection, the data will be automatically distributed or stored.

[0088] Step S6: Introduce a hierarchical fault diagnosis and self-healing mechanism to achieve autonomous fault identification, diagnosis and recovery throughout the entire onboard processing process, ensuring the continuity of the process in an unattended state.

[0089] This invention enables fully unattended automated operation of image data from acquisition, preprocessing, intelligent processing, quality testing to data distribution / storage. It features autonomous fault diagnosis and recovery, dynamic task scheduling, and multimodal adaptation capabilities, while also meeting the low computing power and low power consumption requirements of satellites. This improves the real-time performance and reliability of image processing and is compatible with various remote sensing satellites and multimodal image processing scenarios. The entire process is completed on-board, significantly reducing the amount of data transmitted between satellite and ground, and alleviating the pressure on satellite-to-ground transmission.

[0090] Furthermore, step S1 employs a multimodal sensor adaptation module, compatible with various remote sensing payloads such as optical, SAR, infrared, and hyperspectral sensors, to achieve automatic acquisition, format parsing, and caching of image data, specifically including:

[0091] Based on the satellite's preset orbital parameters and mission planning instructions, the remote sensing payload of the corresponding mode is automatically awakened without the need for manual triggering on the ground, which is suitable for the periodic and random mission requirements of the satellite in orbit.

[0092] The remote sensing payload acquires remote sensing image data in optical, SAR, infrared, and hyperspectral modes in real time, and simultaneously records metadata such as geographic coordinates, imaging time, and payload parameters of the remote sensing image data to ensure the integrity of data acquisition.

[0093] The system automatically parses the acquired raw image data (adapting to different modal data formats), stores the parsed data in the on-board hardened cache module, and adopts a cyclic overwrite strategy to prioritize and retain high-priority data to avoid storage overflow. At the same time, it performs data encryption caching to ensure data security.

[0094] Furthermore, in step S2, an adaptive task scheduling module is constructed to achieve dynamic task scheduling in an unattended state based on onboard resource status, data priority, and task deadline, thus solving the problems of rigid scheduling and low resource utilization in traditional scheduling. The architecture of the adaptive task scheduling module is described in [reference needed]. Figure 2 As shown, it specifically includes:

[0095] 1. Real-time resource awareness: Real-time collection of onboard computing power (CPU / GPU / FPGA utilization), storage capacity, energy consumption status, temperature and other parameters to establish a resource status assessment model and dynamically determine onboard processing capacity;

[0096] (1) The input terminal of the resource status assessment model is connected to the onboard computer (OBC) and the power control board (PCDU), and the following data is collected according to the sampling frequency:

[0097] Computing power dimensions: CPU utilization (%), GPU memory utilization (%), FPGA logic unit utilization (%), on-chip cache hit rate (%).

[0098] Storage dimensions: NAND remaining capacity (GB), read / write IOPS, bad block ratio, remaining lifetime (PE Cycle).

[0099] Energy consumption dimensions: total power consumption (W), current fluctuation (A), battery state of charge (SOC%), and solar panel power supply efficiency (%).

[0100] Thermal control dimensions: chip junction temperature and cabin ambient temperature.

[0101] (2) The resource status assessment model adopts a three-stage architecture of weighted normalization + bottleneck factor identification + dynamic weight adjustment, and outputs a comprehensive status classification; as follows:

[0102] Step 1: Indicator Normalization

[0103] Since the physical quantities have different dimensions, they are first mapped to the [0,1] interval using the Sigmoid function, where a higher value indicates more abundant resources.

[0104] ;

[0105] in, Represents the normalization calculation formula and outputs it. Normalized value;

[0106] Step 2: Bottleneck Factor Analysis

[0107] A "weakest link" correction is introduced. The upper limit of onboard processing power often depends on the shortest plank.

[0108] ;

[0109] ;

[0110] These represent the normalized computing power, storage, and energy consumption margin, respectively. As an intermediate variable; Weights for computing power, storage, and energy consumption margin; Indicates on-board resources; :Pick The minimum value in the range; θ: safety threshold (e.g., 0.2). If any single resource falls below 20%, the CPI will be significantly reduced, forcing the system into energy-saving mode.

[0111] Step 3: Dynamic Weighting

[0112] Unlike static weights, the model dynamically adjusts weights based on the task type. :

[0113] AI inference tasks: Focus (The computing power weight has been increased to 0.6).

[0114] Big data stitching task: Focus (Storage weight increased to 0.5).

[0115] Sunlight area / Shadow area: Emphasis (Energy consumption weight is dynamically adjusted).

[0116] (3) State Classification

[0117] Based on the CPI value, the model outputs four levels of evaluation results, which are then used by the upper-level scheduling system (such as the preprocessing system in point 3):

[0118] L1 (Ample) [0.8, 1.0]: Full-speed processing with abundant resources: Enables high-precision mode, allowing complex preprocessing without compression.

[0119] L2 (normal) [0.5, 0.8) stable operation standard handling: perform routine preprocessing and close unnecessary background processes.

[0120] L3 (stress) [0.2, 0.5) near saturation degradation processing: skip the calculation of non-critical indicators (such as only registering without enhancement) and reduce the sampling rate.

[0121] L4 (Depletion) [0, 0.2) Critical Alarm Circuit Breaker Protection: Suspend all preprocessing, retain only telemetry data, and cache the remaining data for later download.

[0122] 2. Data Priority Classification: The system automatically classifies image data into three priorities—high, medium, and low—based on the image imaging area (such as emergency disaster areas or key monitoring areas), imaging quality, and task type. High-priority data (such as disaster emergency images) is prioritized for processing.

[0123] (1) Three-dimensional priority scoring model (PSM)

[0124] The system calculates a Priority Index (PI) for each newly generated image data frame, with a value ranging from [0, 100]. The PI is composed of weighted scores from three core dimensions:

[0125] PI=α⋅S_region+β⋅S_quality+γ⋅S_task;

[0126] Imaging region: S_region, α (highest)

[0127] Disaster Zones (100): Earthquakes, Floods; Key Zones (70): Cities, Borders, Specific Targets; Regular Zones (30): Oceans, Deserts

[0128] Image quality: ,β

[0129] Clarity / Contrast Excellent (100); Medium (60); Poor (10)

[0130] Task type: γ

[0131] Emergency Task (100): User places an urgent order; Routine Task (60): Periodic monitoring; Archive Task (20): Data backup.

[0132] Dynamic weight adjustment (α,β,γ):

[0133] Disaster response mode: The α weight is increased to 0.7 to ensure that even images of slightly lower quality from the disaster area can be downloaded first.

[0134] Routine operation and maintenance mode: β weight is increased, and high-quality images are prioritized for downloading for mapping, avoiding excessive bandwidth consumption for transmitting waste images.

[0135] (2) Priority classification

[0136] Based on the calculated PI value, the data is categorized into a three-level queue.

[0137] P0 (High) [75-100] Real-time response level: Real-time command data for flood disaster sites, fire monitoring, and satellite overhead.

[0138] P1 (Medium) [40–74] Routine business level: urban planning updates, crop growth monitoring, routine inspections.

[0139] P2 (Low) [0–39] Archive Fill Level: Historical data retakes, invalid data with severe cloud cover.

[0140] 3. Dynamic scheduling strategy: Based on resource status and data priority, a greedy algorithm is used to dynamically allocate processing resources and adjust the processing order. When the on-board computing power is sufficient, multi-modal and multi-batch images are processed in parallel. When the computing power is insufficient, high-priority data is processed first, and low-priority data processing is paused. The processing will be automatically resumed after the resources are released to ensure that critical tasks are completed on time. At the same time, it supports task breakpoint resume to avoid loss of processing progress due to temporary failures.

[0141] This invention is based on a dynamic scheduler kernel using a greedy algorithm. The algorithm's core idea is to achieve the highest global throughput through locally optimal decision-making. It can ensure the real-time performance of mission-critical tasks (P0) and zero system crashes even in extreme environments where on-board resources (CPI) change rapidly.

[0142] (1) Design the utility function under multiple constraints:

[0143] Employing a dynamic scheduler kernel based on a greedy algorithm, the scheduler scans the ready queue at each heartbeat and processes each waiting task. Calculate an immediate utility value :

[0144] ;

[0145] in, Score the data priority of the task; The time a task waits in the queue; the longer it waits, the greater the starvation penalty, preventing low-priority tasks from starving indefinitely. The degree of matching between the resources required for the mission and the current state of resources on the satellite; This refers to the dynamic weighting coefficients for task priority. The dynamic weighting coefficient for the hunger penalty;

[0146] In each scheduling cycle, select The task with the highest value is put into operation.

[0147] (2) Computing power adaptive parallel control strategy: When the onboard computing power is sufficient, multi-modal and multi-batch images are processed in parallel, and the parallelism is dynamically adjusted by real-time monitoring of resource fragmentation; specifically including:

[0148] The on-board computing power is divided into multiple resource blocks;

[0149] Check if the current highest priority task can be split; if it can be split, proceed to the next step.

[0150] Parallel distribution: if immediate utility value If the resource density is high and the system has excess resource blocks, the scheduler will distribute the blocks to idle CPU cores or GPUs for streaming parallel processing.

[0151] Synchronously monitor memory bandwidth usage. If multiple tasks running in parallel cause memory bandwidth saturation, automatically reduce the number of parallel tasks and switch to serial pipeline mode.

[0152] (3) Preemptive scheduling and resource reclamation strategy: When a sudden high-priority task arrives and the system resources are already full, a preemptive mechanism is triggered.

[0153] Downgrade and suspension process:

[0154] 1) Scanning for victims: The algorithm searches for currently running tasks. The lowest level of task (usually a P2 level background task).

[0155] 2) State Snapshot: Instead of terminating the task, the system packages its current register state, program counter (PC), and the address of the data cache that is halfway processed into a checkpoint and stores it in non-volatile memory (NVMe / Flash).

[0156] 3) Resource preemption: Forcefully reclaim the computing power and memory resources occupied by the task.

[0157] 4) High-priority injection: The P0 task is immediately executed by injecting it into the newly freed resource block.

[0158] Once execution is complete, jump to the checkpoint to continue execution;

[0159] (4) Resume interruption mechanism: In order to avoid the loss of processing progress due to single event flip (SEU) caused by cosmic rays or abnormal power reset, the system is designed with atomized resume interruption mechanism.

[0160] This invention employs a three-stage commit storage:

[0161] Prepare: Writes intermediate processing data to a temporary buffer;

[0162] Snapshot: Records metadata about the current task progress;

[0163] Commit: Updates the status bit, marking the task as "stage completed";

[0164] When the system restarts or recovers from an anomaly, the bootloader first checks the anomaly database. If it finds a task in an interrupted state, it reads its last checkpoint, restores the state, and re-requests resources from the scheduler to continue execution from where it left off, instead of starting from the beginning. This greatly saves valuable computing time on the satellite.

[0165] Furthermore, step S3 enables automatic preprocessing of multimodal images. For different modal images, the system employs a lightweight algorithm that is more practical in engineering than theoretically optimal, ensuring that the entire process requires no manual intervention. Specifically, this includes:

[0166] 1. Modal adaptive adaptation

[0167] This invention abandons the high-overhead deep learning method and instead adopts a lightweight decision logic based on statistical features. By quickly calculating the entropy, kurtosis, and number of bands of an image, the system can distinguish between optical, SAR, infrared, and hyperspectral data in milliseconds. For example, it can identify SAR images using the long-tail characteristic of histograms or determine infrared images based on dynamic range distribution. The entire process does not require loading large models and relies only on basic mathematical operations to complete the modal adaptation of optical, SAR, infrared, and hyperspectral data.

[0168] 2. Apply the corresponding modality preprocessing algorithm to the modality-adapted data. The preprocessing algorithm includes:

[0169] (1) Automatically perform brightness normalization, Gaussian noise reduction and geometric correction on optical images;

[0170] (2) Automatically perform Lee filtering to remove speckles and normalize the backscattering coefficient for SAR images;

[0171] (3) Automatically perform radiometric correction and band alignment for infrared and hyperspectral images;

[0172] It is worth noting that:

[0173] Optical imaging: Geometric correction abandons the complex physical model orthorectification and simplifies it to resampling based on lookup table (LUT), completely avoiding trigonometric function calculations; the denoising stage adopts fixed-point 3x3 mean filtering and is accelerated by using SIMD instruction set; brightness normalization is achieved by linear stretching through a single traversal.

[0174] SAR imagery: Classic Lee filtering is used for speckle removal, and local variance is quickly estimated through a sliding window, avoiding the iterative overhead of complex diffusion models; backscattering coefficient normalization is achieved by logarithmic transformation combined with lookup table (LUT), eliminating the need to solve complex exponential functions.

[0175] Infrared and hyperspectral imagery: Radiometric correction uses a two-point correction method (gain and bias) instead of nonlinear fitting, ensuring that the computational complexity is reduced to O(N); Band alignment only performs offset correction at the integer pixel level, without time-consuming resampling.

[0176] All preprocessing algorithms have undergone lightweight optimization and computational cost reduction. To facilitate dynamic scheduling, all preprocessing uses a "tile streaming" architecture. Images are divided into fixed-size blocks (e.g., 256x256 pixels) and fed into the processing pipeline one by one. This approach keeps peak memory usage at a constant level, preventing memory overflow caused by loading large images. Simultaneously, the algorithm integrates a computational graph pruning mechanism: when RSAM detects a decrease in CPI, the system automatically skips unnecessary enhancement steps, such as performing only linear stretching and pausing denoising when the battery is low.

[0177] 3. After preprocessing, key image indicators are automatically extracted for initial inspection, including:

[0178] Sharpness detection: The Laplacian variance operator is used, which is the optimal balance between computational cost and sensitivity. The system applies a 3x3 integer convolution kernel to the image and calculates the variance of the response value. If the variance is lower than the dynamic threshold (e.g., optical images <120), it is judged as blurry.

[0179] Contrast evaluation: Instead of complex histogram segmentation, the dynamic range or RMS contrast of pixel values ​​is directly calculated. If overexposed or hazy data is detected (due to excessively narrow dynamic range), automatic retry logic is immediately triggered.

[0180] Registration accuracy: The downsampled image is used to perform fast cross-correlation calculation. Only the root mean square error (RMSE) of the feature point residuals is counted. Once it exceeds the set pixel-level threshold (e.g., 1.5 pixels), it is judged as a registration failure.

[0181] If the initial inspection fails, the system automatically reprocesses the data (e.g., adjusting denoising parameters or redoing registration). If a single data set still fails after reaching the maximum number of retries, the system marks it as "Quality_Exception," stores it in the exception database, and logs specific failure indicators. Subsequently, this data is automatically reported to the ground center via satellite link, awaiting manual intervention or re-shooting instructions, thus forming a complete closed loop of "processing-detection-feedback."

[0182] Images that fail the initial inspection will be automatically reprocessed after adjusting the denoising parameters or redoing the registration. If the images still fail after multiple preprocessing attempts, they will be marked as abnormal data, stored in the abnormal database, and automatically reported to the ground later.

[0183] Furthermore, step S4, based on the on-board lightweight AI processing platform, automatically initiates the corresponding intelligent processing flow according to preset task requirements to achieve image feature extraction, ground feature classification, target detection, and other processing, specifically including:

[0184] 1. Tasks are triggered on demand: Based on the satellite mission plan (preset in the onboard system), the system automatically determines the intelligent processing tasks (such as ground feature classification and target detection) that need to be performed on the current image without the need for ground command intervention;

[0185] 2. Lightweight intelligent processing: Employs quantized and compressed AI models (such as lightweight CNN and Transformer models) to complete image feature extraction, ground feature classification, target detection, and other processing on the on-board computing platform. During the processing, the computing power consumption is automatically monitored to avoid computing power overload.

[0186] 3. Processing result caching: After intelligent processing is completed, the processing results (feature map, classification result, target coordinates, etc.) are automatically associated with and stored with the original image and preprocessed image to form a complete processing archive.

[0187] Furthermore, step S5 establishes a multi-level quality inspection mechanism to achieve automated quality inspection throughout the entire process of preprocessing and intelligent processing. After passing the inspection, data distribution or storage is automatically completed. Specifically, this includes:

[0188] 1. Multi-level quality inspection: Level 1 inspection (after preprocessing) checks the quality of image preprocessing; Level 2 inspection (after intelligent processing) checks the accuracy of the processing results (such as classification accuracy and target detection recall); Level 3 inspection (overall inspection) checks the integrity and correlation of data to ensure that the processing flow is free of abnormalities;

[0189] 2. Anomaly Handling: Detects unqualified data, automatically marks the anomaly type (such as preprocessing failure, processing result deviation), stores it in the anomaly database, and simultaneously activates the anomaly early warning mechanism. When the space-to-ground communication link is clear, it automatically reports the anomaly information to the ground without manual intervention.

[0190] 3. Automatic data output: Qualified data is automatically distributed or stored according to preset rules: High-priority data (such as emergency image processing results) is automatically downlinked to the ground receiver via the satellite-to-ground communication link; Medium and low-priority data are automatically stored in the satellite's large-capacity storage module and downlinked in batches when the satellite-to-ground communication is idle; At the same time, lightweight compression of processing results is supported to reduce transmission and storage overhead.

[0191] Furthermore, step S6 introduces a hierarchical fault diagnosis and self-healing mechanism to achieve autonomous fault identification, diagnosis, and recovery throughout the entire onboard processing flow, ensuring the continuity of the process in unattended operation. (See [link to relevant documentation]). Figure 3 As shown, it specifically includes:

[0192] 1. Full-process fault monitoring: Real-time monitoring of the operational status of each stage (acquisition, scheduling, preprocessing, intelligent processing, quality inspection, and output), and capture of fault signals (such as acquisition interruption, processing timeout, storage overflow, module failure, etc.).

[0193] 2. Layered fault diagnosis: Adopting a layered diagnosis approach at the module level, single machine level, and system level, faults are accurately located and differentiated into minor faults (such as temporary computing power overload), general faults (such as single module abnormality), and serious faults (such as system crash).

[0194] 3. Autonomous self-healing: For minor faults, the system automatically initiates resource adjustments (such as releasing redundant computing power and clearing the cache) to resume process operation; for general faults, the system automatically initiates redundant module replacement (such as switching to the backup module if the preprocessing module fails) and resumes processing progress from breakpoints; for serious faults, the system automatically restarts, restores to the state before the fault, continues to execute processing tasks, and records and reports serious fault information to the ground.

[0195] 4. Redundancy backup guarantee: The core processing modules (such as the adaptive task scheduling module and the fault self-healing module) adopt a dual-board hot standby solution, and the core programs and data adopt triple redundancy backup to ensure that they can switch quickly in the event of a failure and avoid interruption of the processing flow.

[0196] Furthermore, it also includes step S7: after the ground update algorithm and mission planning instructions are received, they are automatically uploaded to the satellite system through the satellite-ground communication link. The satellite system automatically verifies the legality of the updated content. After the verification is passed, the update is automatically updated. After the update is completed, there is no need to restart. It is directly applied to the subsequent processing flow to achieve continuous optimization of the system.

[0197] The following is a specific application example of the present invention:

[0198] Step 1: Automatic acquisition and caching of multimodal image data (unmanned operation)

[0199] 1. Autonomous wake-up: The on-board system automatically wakes up the optical payload and SAR payload according to the preset satellite orbit parameters (such as overpass time and observation area). The wake-up time error is ≤10s and no ground command is required to trigger it. If the satellite receives an emergency observation command from the ground, it will prioritize waking up the corresponding payload and start the emergency acquisition mode.

[0200] 2. Data Acquisition: The optical payload acquires RGB+near-infrared 4-band images with a resolution of 1m at a frame rate of 10 frames / second; the SAR payload acquires dual-polarized (HH, HV channels) images with a resolution of 2m at a frame rate of 5 frames / second; during the acquisition process, metadata such as the geographic coordinates (WGS84 coordinate system), imaging time, and payload operating parameters (such as exposure time and polarization mode) of the images are recorded synchronously to ensure data integrity.

[0201] 3. Format parsing and caching: The system automatically parses optical images (TIFF format) and SAR images (HDF5 format), with a parsing time of ≤50ms / frame. The parsed data is stored in a 128GB DDR5 cache module, using a cyclic overwrite strategy. High-priority data (emergency area images) is locked in storage, while low-priority data (routine observation images) is automatically overwritten. When the remaining space in the cache module is ≤10%, the oldest unprocessed low-priority data is automatically cleaned up to avoid storage overflow. All cached data is encrypted with AES-256 to ensure data security.

[0202] Step 2: Adaptive task scheduling (unmanned, dynamically adjusted)

[0203] 1. Real-time Resource Awareness: The resource monitoring tool collects on-board resource status every 100ms, including CPU utilization, GPU / FPGA computing power utilization, cache remaining capacity, and energy consumption status. A resource status assessment model is established. When the CPI value is in the range of [0.8, 1], resources are abundant; when it is in the range of [0, 0.2], an alarm is triggered. When the CPU utilization is ≥80% and the cache remaining capacity is ≤20%, resources are considered scarce; when the CPU utilization is ≤50% and the cache remaining capacity is ≥50%, resources are considered abundant.

[0204] 2. Data Priority Classification: The system automatically classifies image data into three priorities based on the image imaging area (preset key monitoring areas and emergency disaster areas) and image quality (resolution ≥ 0.8 is acceptable): high priority (emergency disaster areas, images with resolution ≥ 0.9), medium priority (key monitoring areas, images with resolution 0.8-0.9), and low priority (routine observation areas, images with resolution below 0.8). The priority classification time is ≤ 10ms / frame.

[0205] 3. Dynamic Scheduling Strategy: A greedy algorithm is used to implement dynamic scheduling, with the following specific rules:

[0206] (1) When resources are sufficient: Parallel scheduling of optical and SAR image preprocessing and intelligent processing tasks, processing 10 frames of images per batch to ensure processing efficiency;

[0207] (2) When resources are scarce: pause low-priority data processing, prioritize scheduling high-priority data, allocate computing power and cache resources to high-priority tasks, and automatically resume low-priority data processing after resources are released (CPU utilization ≤ 60%, cache remaining capacity ≥ 30%).

[0208] (3) Task interruption resume: If a temporary failure (such as computing power overload) causes the processing to be interrupted, the system will automatically read the processing log after the failure is recovered and continue processing from the breakpoint without having to start over, ensuring that the processing progress is not lost;

[0209] (4) Scheduling cycle: The scheduling strategy is adjusted every 500ms to ensure that resource allocation matches task requirements and scheduling delay is ≤100ms.

[0210] Step 3: Automated preprocessing of multimodal images (unmanned, modality adaptive)

[0211] 1. Modal Adaptive Matching: The system automatically identifies image modalities (optical / SAR) with an accuracy rate of ≥99.5%. It calls the corresponding preprocessing algorithm based on the modal type, eliminating the need for manual configuration.

[0212] 2. Optical image preprocessing: Automatically performs the following operations:

[0213] (1) Brightness normalization: Normalize the image pixel values ​​to the [0,1] interval. The calculation formula is x_norm=(x-x_min) / (x_max-x_min). The processing time is ≤20ms / frame.

[0214] (2) Gaussian noise removal: A 3×3 convolution kernel with a standard deviation of 0.5 is used to remove Gaussian noise from the image. The processing time is ≤15ms / frame.

[0215] (3) Geometric correction: Based on the preset satellite orbit parameters and ground control points, the geometric correction is automatically completed with a correction error of ≤1 pixel and a processing time of ≤30ms / frame.

[0216] 3. SAR image preprocessing: Automatically perform the following operations:

[0217] (1) Lee filtering for speckle removal: A 5×5 window is used to suppress speckle noise, and the processing time is ≤25ms / frame;

[0218] (2) Backscattering coefficient normalization: The backscattering coefficient is normalized to the interval [-1,1], and the calculation formula is y_norm=2×(y-y_min) / (y_max-y_min)-1, with a processing time ≤20ms / frame;

[0219] (3) Polarization channel alignment: Align the HH and HV channels with an alignment error of ≤0.5 pixels and a processing time of ≤15ms / frame.

[0220] 4. Preprocessing quality initial inspection: Extract three key indicators: image sharpness, contrast, and registration accuracy. Sharpness ≥ 0.8, contrast ≥ 0.3, and registration accuracy ≤ 1 pixel are considered qualified. Images that fail the initial inspection will be automatically reprocessed, up to 3 times. If they still fail, they will be marked as abnormal data, stored in the abnormal database, and the abnormality type (such as preprocessing failure) will be recorded.

[0221] Step 4: Intelligent Image Processing (Automation on Demand, Lightweight)

[0222] 1. Tasks are triggered on demand: The system automatically determines the intelligent processing tasks that need to be performed on the current image based on the preset satellite mission plan (such as key area land cover classification and emergency target detection). High-priority images are given priority to perform target detection tasks, while medium and low-priority images are given priority to perform land cover classification tasks, without the need for ground command intervention.

[0223] 2. Intelligent processing operation:

[0224] (1) Target detection: The lightweight YOLOv8 model (8-bit quantization) after quantization compression is used to detect emergency targets (such as houses, roads, and water bodies) in the image. The recall rate is ≥90%, the precision is ≥88%, and the processing time is ≤50ms / frame.

[0225] (2) Land cover classification: The lightweight MobileNetV4 model is used to classify the images into four categories: buildings, vegetation, water bodies and roads. The classification accuracy is ≥92% and the processing time is ≤40ms / frame.

[0226] (3) Processing monitoring: Real-time monitoring of computing power usage. If the GPU usage rate is ≥90%, the processing batch will be automatically reduced (from 10 frames / batch to 5 frames / batch) to avoid computing power overload.

[0227] 3. Processing result caching: The processing results (target coordinates, classification mask, feature map) are associated with the original image and preprocessed image and stored in a naming rule of "modality_imaging time_geographic coordinates_processing type". The data is stored on a 1TB NVMe solid disk and a processing log (processing time, processing parameters, result accuracy) is recorded at the same time.

[0228] Step 5: Fully automated quality inspection and data output (unmanned operation)

[0229] 1. Multi-level quality inspection:

[0230] (1) Level 1 inspection (after preprocessing): The image sharpness, contrast and registration accuracy after preprocessing are inspected. The pass criteria are the same as the initial inspection after preprocessing. The inspection time is ≤10ms / frame.

[0231] (2) Secondary detection (after intelligent processing): Detection target recall rate and ground feature classification accuracy. A recall rate ≥90% and a classification accuracy rate ≥92% are considered qualified. The detection time is ≤15ms / frame.

[0232] (3) Level 3 inspection (overall inspection): Inspection data integrity (original image, pre-processed image, processing result, metadata are complete), correlation (data of each link are consistent), inspection pass rate ≥99%, inspection time ≤20ms / frame.

[0233] 2. Anomaly Handling: Detects unqualified data, automatically marks the anomaly type (preprocessing failure, processing result deviation, data incompleteness), and stores it in the anomaly database. The anomaly database uses an independent storage partition and can store ≥1000 anomaly data entries. At the same time, an anomaly early warning mechanism is activated to record the time, stage, and specific cause of the anomaly. Once the space-to-ground communication link is clear, the anomaly information is automatically reported to the ground without manual intervention.

[0234] 3. Automatic data output:

[0235] (1) High-priority data: The processing results (target coordinates, classification results) are compressed in a lightweight manner (compression ratio 10:1) and sent down to the ground receiver via the satellite-to-ground communication link. The downlink rate is ≥1Gbps and the downlink delay is ≤30s.

[0236] (2) Medium and low priority data: The processing results are stored together with the original image and the pre-processed image to the NVMe solid disk. When the satellite-to-ground communication is idle (the satellite-to-ground link occupancy rate is ≤30%), the data is downlinked in batches of 50 frames / batch.

[0237] (3) Storage management: When the remaining space of the solid-state drive is ≤10%, the low-priority processing results that have not been downloaded for the longest time will be automatically deleted, while the original images and abnormal data will be retained to ensure sufficient storage space.

[0238] Step 6: Autonomous Fault Diagnosis and Self-Healing (Unmanned Operation, Ensuring Continuity)

[0239] 1. Full-process fault monitoring: The system monitors the operating status of each link in real time and collects fault signals every 50ms. The types of faults that can be identified include: data acquisition interruption, preprocessing timeout, processing module failure, storage overflow, computing power overload, communication link interruption, power supply abnormality, etc.

[0240] 2. Layered Fault Diagnosis: A layered diagnosis approach is adopted, consisting of module-level, unit-level, and system-level fault diagnosis.

[0241] (1) Module-level diagnosis: For a single processing module (such as a preprocessing module or an intelligent processing module), diagnose whether the fault is due to an abnormality of the module itself (such as an algorithm error or a module not responding).

[0242] (2) Standalone-machine level diagnosis: For standalone systems on the satellite, diagnose whether the fault is due to resource abnormalities (such as computing power overload, storage overflow).

[0243] (3) System-level diagnosis: For the entire on-board processing system, diagnose whether the fault is a serious problem such as system crash or power abnormality, with a diagnosis accuracy of ≥99% and a diagnosis time of ≤100ms.

[0244] 3. Self-healing mechanism:

[0245] (1) Minor faults (computing power overload, insufficient cache): Automatically release redundant computing power (shut down non-core auxiliary modules), clear cache (delete low-priority temporary data), and restore process operation. Self-healing time ≤ 1s;

[0246] (2) General faults (single module failure, data acquisition interruption): Automatically start redundant module replacement (e.g., if the preprocessing module fails, switch to the backup module), read the processing log, resume the processing progress from the breakpoint, and the self-healing time is ≤5s; if the data acquisition is interrupted, automatically wake up the load and restart the data acquisition process, up to 3 times.

[0247] (3) Serious fault (system crash, power abnormality): The system will be automatically restarted. The restart time is ≤30s. After restarting, the system will be restored to the processing state before the fault and continue to perform processing tasks. At the same time, the serious fault information will be recorded and reported to the ground. If the restart fails, the emergency mode will be activated, and only the core fault monitoring and communication functions will be retained, waiting for ground intervention.

[0248] 4. Redundancy backup guarantee: The core processing module adopts dual-board hot standby. When the main module fails, the backup module automatically switches over with a switching time of ≤100ms. The core program and data adopt triple redundancy backup and are stored in different storage partitions to avoid data loss.

[0249] Step 7: Automatic Updates via Satellite-Ground Collaboration (Assisted Optimization, No Human Intervention)

[0250] 1. Update Trigger: When the ground needs to update the processing algorithm, task planning, or fault diagnosis rules, the update packet (encrypted transmission) is sent up to the satellite system via the satellite-ground communication link. The satellite system automatically detects the link status and receives the update packet when the link is clear.

[0251] 2. Automatic verification: The system automatically verifies the legality of the update package (verification code verification). After successful verification, it automatically backs up the current algorithm, task planning, and fault diagnosis rules to avoid system abnormalities caused by update failure.

[0252] 3. Automatic Update: After backup is complete, the corresponding content will be automatically updated without interrupting the current processing flow. After the update is completed, there is no need to restart the system, and it can be directly applied to subsequent processing tasks. The update time is ≤30 seconds. If the update fails, it will automatically restore to the backup version and report the update failure information to the ground.

[0253] Although the present invention has been described herein with reference to illustrative embodiments, the above embodiments are merely preferred embodiments of the present invention, and the implementation of the present invention is not limited to the above embodiments. It should be understood that those skilled in the art can devise many other modifications and implementations, which will fall within the scope and spirit of the principles disclosed in this application.

Claims

1. An unattended on-board automated processing method for satellite remote sensing image data, characterized in that, include: Step S1: Automatically acquire, preprocess, and cache multimodal image data; Step S2: Based on the on-board resource status, data priority, and task deadline, perform adaptive dynamic task scheduling; Step S3: For remote sensing image data of different modalities, a lightweight algorithm that is more practical than the theoretical optimum is adopted to automatically perform modal adaptive adaptation, core preprocessing operations and preprocessing quality inspection, ensuring that no manual intervention is required throughout the entire process; Step S4: Based on the on-board lightweight AI processing platform, the corresponding intelligent processing flow is automatically started according to the preset task requirements to realize image feature extraction, land cover classification and target detection. The processing results are automatically associated and stored with the original image and the pre-processed image to form a complete processing archive. Step S5: Construct a multi-level quality inspection mechanism to achieve automatic quality inspection throughout the entire process of preprocessing and intelligent processing. After passing the inspection, the data will be automatically distributed or stored. Step S6: Introduce a hierarchical fault diagnosis and self-healing mechanism to achieve autonomous fault identification, diagnosis and recovery throughout the entire onboard processing process, ensuring the continuity of the process in an unattended state.

2. The unattended on-board automated processing method for satellite remote sensing image data according to claim 1, characterized in that, Step S1 specifically includes: Based on the satellite's preset orbital parameters and mission planning instructions, the remote sensing payload of the corresponding mode is automatically activated; The remote sensing payload acquires remote sensing image data in optical, SAR, infrared and hyperspectral modes in real time, and records metadata of the remote sensing image data simultaneously. The system automatically parses the acquired raw image data, adapts the format of data of different modalities, stores the parsed data on the satellite, and adopts a cyclic overwrite strategy to prioritize the retention of high-priority data to avoid storage overflow, while also performing data encryption caching.

3. The unattended on-board automated processing method for satellite remote sensing image data according to claim 1, characterized in that, Step S2 specifically includes: The on-board computing power is collected in real time, a resource status assessment model is established, and the on-board processing capacity is dynamically judged. The on-board computing power includes processor utilization, storage capacity, and energy consumption status. The resource status assessment model is used to perform weighted normalization, bottleneck factor identification, and dynamic weight adjustment processing on the on-board computing power data, and output a comprehensive status classification. Based on the imaging area, imaging quality, and task type of the remote sensing image data, the remote sensing image data is divided into priority levels. By combining comprehensive status classification and priority levels, processing resources are dynamically allocated and the processing order is adjusted.

4. The unattended on-board automated processing method for satellite remote sensing image data according to claim 3, characterized in that, The strategy of dynamically allocating processing resources and adjusting processing order by combining comprehensive status classification and priority levels includes: prioritizing the allocation of currently available resources to the highest priority and fastest-progressing tasks, specifically: Employing a dynamic scheduler kernel based on a greedy algorithm, the scheduler scans the ready queue at each heartbeat and processes each waiting task. Calculate an immediate utility value : ； in, Score the data priority of the task; The time a task waits in the queue; the longer it waits, the greater the starvation penalty, preventing low-priority tasks from starving indefinitely. The degree of matching between the resources required for the mission and the current state of resources on the satellite; This refers to the dynamic weighting coefficient for task priority; The dynamic weighting coefficient for the hunger penalty; In each scheduling cycle, select The task with the highest value is put into operation.

5. The unattended on-board automated processing method for satellite remote sensing image data according to claim 4, characterized in that, The strategy of dynamically allocating processing resources and adjusting processing order by combining comprehensive status classification and priority levels also includes one or more of the following: (1) Computing power adaptive parallel control strategy: When the on-board computing power is sufficient, multi-modal and multi-batch images are processed in parallel, and the parallelism is dynamically adjusted by real-time monitoring of resource fragments; (2) Preemptive scheduling and resource reclamation strategy: When a sudden high-priority task arrives and the system resources are already full, a preemptive mechanism is triggered, and the system searches for the best performing task currently in operation. The lowest priority task is not terminated. Instead, its current register state, program counter, and half-processed data cache address are packaged into a checkpoint and stored in non-volatile memory. The computing power and memory resources occupied by the task are forcibly reclaimed, and the sudden high-priority task is immediately put into the newly freed resource block for execution. After execution, the process jumps to the checkpoint to continue execution. (3) Resumable download mechanism: Three-phase commit storage is adopted: Prepare: Writes intermediate processing data to a temporary buffer; Snapshot: Records metadata about the current task progress; Commit: Updates the status bit, marking the task as "stage completed"; When the system restarts or recovers from an exception, the bootloader first checks the exception database. If it finds a task that is interrupted, it reads its last checkpoint, restores the state, and requests resources from the scheduler again to continue execution from where it was interrupted, instead of starting from the beginning.

6. The unattended on-board automated processing method for satellite remote sensing image data according to claim 5, characterized in that, When onboard computing power is sufficient, multi-modal and multi-batch image processing is performed in parallel, and the parallelism is dynamically adjusted by real-time monitoring of resource fragmentation. Specifically, this includes: Divide the on-board computing power into multiple resource blocks; Check if the current highest priority task can be split; if it can be split, proceed to the next step. Parallel dispatch: if immediate utility value If the resource density is high and the system has excess resource blocks, the scheduler will distribute the blocks to idle CPU cores or GPUs for streaming parallel processing. Synchronously monitor memory bandwidth usage. If multiple tasks running in parallel cause memory bandwidth saturation, automatically reduce the number of parallel tasks and switch to serial pipeline mode.

7. The unattended on-board automated processing method for satellite remote sensing image data according to claim 1, characterized in that, Step S3 specifically includes: A lightweight decision logic based on statistical features is used to quickly calculate the entropy, kurtosis and number of bands of an image, and to complete the modal adaptation of optical, SAR, infrared and hyperspectral data. The modality-adapted data is then processed using the corresponding modality preprocessing algorithm, which includes: (1) Automatically perform brightness normalization, Gaussian noise reduction and geometric correction on optical images; (2) Automatically perform Lee filtering to remove speckles and normalize the backscattering coefficient for SAR images; (3) Automatically perform radiometric correction and band alignment for infrared and hyperspectral images; All preprocessing algorithms are optimized for lightweight operation and computational power reduction. They all adopt a tile-based streaming processing architecture, where the image to be processed is divided into small blocks of fixed size and flows into the processing pipeline one by one. At the same time, the preprocessing algorithms integrate a computational graph pruning mechanism. After preprocessing, key image indicators are automatically extracted for initial inspection, including: Sharpness detection: The Laplacian variance operator is used to achieve the best balance between computational cost and sensitivity; a 3×3 integer convolution kernel is applied to the image to calculate the variance of the response value, and if it is lower than the dynamic threshold, it is judged as blurry; Contrast evaluation: Directly calculate the dynamic range or RMS contrast of pixel values. If overexposed or hazy data is found, trigger automatic retry logic immediately. Registration accuracy: The downsampled image is used to perform fast cross-correlation calculation. Only the root mean square error of the feature point residuals is counted. Once it exceeds the set pixel-level threshold, it is judged as a registration failure. Images that fail the initial inspection will be automatically reprocessed after adjusting the denoising parameters or redoing the registration. If the images still fail after multiple preprocessing attempts, they will be marked as abnormal data, stored in the abnormal database, and automatically reported to the ground later.

8. The unattended on-board automated processing method for satellite remote sensing image data according to claim 1, characterized in that, Step S5 specifically includes: The quality of image preprocessing, the accuracy of intelligent processing results, and the completeness and correlation of overall detection data are checked to ensure that the processing flow is free of abnormalities. Data that fails the test is automatically marked as abnormal and stored in the abnormal database. At the same time, the abnormal early warning mechanism is activated. When the space-to-ground communication link is clear, the abnormal information is automatically reported to the ground without manual intervention. Data that passes the test is automatically distributed or stored according to preset rules: high-priority data is automatically downlinked to the ground receiver via the satellite-to-ground communication link; medium and low-priority data are automatically stored in the satellite's large-capacity storage module and then downlinked to the ground receiver in batches when the satellite-to-ground communication is idle.

9. The unattended on-board automated processing method for satellite remote sensing image data according to claim 1, characterized in that, Step S6 specifically includes: Real-time monitoring of the operational status of each component and capture of fault signals; By adopting a hierarchical diagnostic approach at the module, unit, and system levels, faults can be accurately located, and minor, general, and serious faults can be distinguished. The system enables autonomous fault healing, including: Minor faults will automatically trigger resource adjustments and restore process operation. For general faults, the system automatically starts to replace redundant modules and resumes interrupted data transmission. In the event of a serious malfunction, the system will automatically restart and return to its pre-malfunction state to continue processing tasks. At the same time, the serious malfunction information will be recorded and reported to the ground. A redundant backup mechanism is employed to ensure rapid switching in the event of a failure, thus avoiding interruption of the processing flow.

10. The unattended on-board automated processing method for satellite remote sensing image data according to claim 1, characterized in that, It also includes step S7: After the ground update algorithm and mission planning instructions are received, they are automatically uploaded to the satellite system through the satellite-ground communication link. The satellite system automatically verifies the legality of the updated content. After the verification is passed, the update is automatically updated. After the update is completed, there is no need to restart. It can be directly applied to the subsequent processing flow to achieve continuous optimization of the system.

Images