Generalized dynamic guidance-based star-ground collaborative intelligent remote sensing computing method and system

Abstract

The application discloses a kind of star-ground collaborative intelligent remote sensing computing method and system based on generalized dynamic guidance, it is related to spatial information technology, artificial intelligence and edge computing cross technical field.The method is through ground guidance center to contain intelligent space mask, task self-adapting prompt vector and collaborative reasoning instruction generalized dynamic guidance information;Satellite-borne processing unit is combined with real-time resource state according to guidance information, and using local focusing calculation and conditional early retirement mechanism realizes self-adapting focusing processing;While constructing the closed loop optimization system of heaven and earth, continuously optimizes guiding strategy by result feedback.The application realizes the mode change from full image processing to focusing calculation, significantly improves the on-board processing efficiency, reduces the amount of calculation by 70-90% while ensuring accuracy, reduces star-ground communication data volume by more than 95%, effectively solves the problem of limited on-board resources and data transmission bandwidth bottleneck, and has important value in remote sensing application scenarios such as disaster emergency and environmental monitoring.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of space information technology, artificial intelligence and edge computing, and specifically to a satellite-ground collaborative intelligent remote sensing computing method and system based on generalized dynamic guidance. Background Technology

[0002] With the rapid development of remote sensing satellite constellations, remote sensing data acquisition capabilities have shown an exponential growth trend. However, the traditional "on-board acquisition, ground processing" model is facing severe challenges. The most prominent contradiction lies in the difficulty of downlinking massive amounts of raw data due to data transmission bandwidth bottlenecks, and the resulting problems such as poor processing timeliness, making it difficult to meet the real-time application needs such as emergency response.

[0003] Direct data processing on satellites is an effective means of achieving efficient data application. While existing technologies have explored on-board processing to some extent, most remain at the level of simple static division of labor, where on-board and ground processing tasks are pre-defined, lacking dynamic adaptability. It is worth noting that the guidance information used in existing satellite-ground collaborative computing models is often limited to simple prior information such as binary space masks and image compression, lacking deep semantic understanding capabilities. Furthermore, the lack of effective intelligent interaction mechanisms and closed-loop optimization capabilities between satellite and ground systems leads to low resource utilization efficiency and an inability to adaptively schedule tasks based on real-time mission requirements and resource status.

[0004] These technical deficiencies severely restrict the further improvement of on-orbit remote sensing information extraction efficiency. Therefore, there is an urgent need in this field for a new remote sensing processing architecture that can achieve intelligent satellite-ground collaboration, dynamic guidance, and resource self-adaptation. Summary of the Invention

[0005] This invention aims to overcome the shortcomings of existing technologies and provide a space-ground collaborative intelligent remote sensing computing method and system based on generalized dynamic guidance. The core innovation of this invention lies in proposing the concept of "generalized dynamic guidance," achieving efficient processing of remote sensing data through a space-ground collaborative intelligent computing architecture. The essential difference between this invention and existing technologies is that it does not simply migrate ground processing tasks to satellite, but rather achieves "precise control" of the onboard computing process through "generalized dynamic guidance information" generated by the ground guidance center. This guidance is dynamic, semantic, and optimizable. It innovates the space-ground collaborative model, transforming the traditional static division of labor into integrated collaboration, where the ground is responsible for planning and decision-making, and the satellite is responsible for precise execution.

[0006] The technical solution adopted in this invention is as follows:

[0007] A satellite-ground collaborative intelligent remote sensing computing method based on generalized dynamic guidance includes the following steps:

[0008] Step 1: The ground guidance center processes and analyzes the stored historical multi-source information based on the remote sensing interpretation mission on the satellite to generate generalized dynamic guidance information;

[0009] Step 2: The ground guidance center transmits the generalized dynamic guidance information to the satellite's onboard processing unit via the space-to-ground uplink;

[0010] Step 3: The onboard processing unit receives and parses the generalized dynamic guidance information. Based on the type of the generalized dynamic guidance information and the satellite's resource status, it dynamically generates a focusing calculation mask and a model early retreat threshold.

[0011] Step 4: The onboard processing unit performs remote sensing intelligent calculations based on generalized dynamic guidance information, focused calculation mask map, and model early termination threshold to generate structured results;

[0012] Step 5: The onboard processing unit transmits the processed structured result data back to the ground guidance center via the space-to-ground downlink and uses it to optimize the subsequent generation of generalized dynamic guidance information.

[0013] Furthermore, the remote sensing interpretation tasks on the satellite in step 1 include three tasks: target detection, land cover classification, and change detection; the historical multi-source information in step 1 includes one or more of the following: historical remote sensing images, remote sensing thematic maps, vector geographic information data, and text data.

[0014] Furthermore, the generalized dynamic guidance information mentioned in step 1 transcends binary space masks and is a semantically rich information carrier, specifically including intelligent spatial masks. Task adaptive cue vector and collaborative reasoning instructions The generation method is as follows:

[0015] The ground guidance center first analyzes historical remote sensing images using a land cover classification model, a target detection model, and a change detection model, generating analysis results including target detection results, land cover classification results, and change detection results. After saving the results, statistical analysis is performed by region, including region latitude and longitude, image time, target type and quantity, land cover type, and the number of changes. Then, an intelligent spatial mask is generated according to the following steps. Task adaptive cue vector and collaborative reasoning instructions ;

[0016] a) Intelligent spatial mask For different remote sensing interpretation tasks, the spatial sub-regions of the satellite image are calculated in the following ways. weight :

[0017] Target detection task: The target area refers to the area where the target may appear.

[0018] Change detection task: Fixed areas refer to regions that are unlikely to change based on historical remote sensing image patterns; potentially changeable areas refer to regions where changes occur based on the analysis results of historical remote sensing images. ,in It is the total number of historical remote sensing images that have changed. All historical remote sensing imagery; the mandatory area refers to the area that must be analyzed in this mission;

[0019] Land feature classification task: The classification-invariant region refers to the region in historical remote sensing imagery that has not undergone classification change, while the classification-change region refers to the region in historical remote sensing imagery where the classification result has changed at different times. The historical remote sensing image closest in time to the current task is selected, and the probability of classification change is calculated using its time interval t. , , This refers to a time interval sequence of changes within historical remote sensing images. It is a mean function;

[0020] b) Task-adaptive cue vector Its generation is a dynamic process based on a meta-learning model, specifically including the following steps:

[0021] b1) Task semantic encoding: Using a multilayer perceptron model, the current remote sensing interpretation task, data latitude and longitude range, and data acquisition time elements are encoded and mapped into task embedding vectors. ;

[0022] b2) Historical Data Analysis: Using the same multilayer perceptron model as b1, the results of historical remote sensing image analysis are encoded to generate a historical task experience set. ,in This is the number for the analysis of historical remote sensing images. A single analysis of a single historical remote sensing image is recorded as 1.

[0023] b3) Historical data fusion: Calculating task embedding vectors Experience with historical missions The similarity is calculated, and weights are generated using an attention mechanism. , , where s is the similarity function; then, the context vector is obtained by weighted summation. ;

[0024] b4) Meta-hint synthesis: [The following text appears to be a separate, unrelated sentence:] and Perform nonlinear fusion to output task-adaptive cue vectors. ;in Implemented using a multilayer perceptron;

[0025] c) Collaborative reasoning instructions For the multimodal interpretation model deployed on the satellite, a structured text is generated as a prompt message. It is a set of structured machine-readable instructions in the format of {operation type, target area, parameters}.

[0026] Furthermore, the process of generating the focused calculation mask image and the model early termination threshold in step 3 is as follows:

[0027] a) The process of generating the focused computational mask image is as follows:

[0028] Smart Space Mask Based on the latitude and longitude range and spatial resolution of the satellite image, a focus calculation mask map of the area to be processed is generated. The specific method is from A region with the same latitude and longitude range as the satellite image is cropped out, and nearest neighbor interpolation is performed on the cropped region according to the resolution of the satellite image to obtain the focus calculation mask map. ;

[0029] Adaptive cue vectors for tasks First, a spatial importance heatmap is generated using a lightweight convolutional network. :

[0030] Where: ⊕ represents concatenation and broadcasting of channel dimensions, Broadcast to the same spatial size as image I on the satellite. This is the sigmoid function, with a range of [0,1]. elements in The weight representing the importance of processing at position (i,j);

[0031] Then, based on satellite resource constraints, the area to be processed is dynamically determined, and a focused computational mask map of the area to be processed is generated. :

[0032] ,in, The determination is based on a comprehensive consideration of onboard computing resources, satellite power status, and onboard image size.

[0033] b) The process for generating the model's early termination threshold is as follows:

[0034] b1) Early Departure Decision Point Setting: Setting an early departure decision point in the key layer of the intelligent interpretation model deployed on the onboard processing unit. Among them, the key layer is set according to the requirements, and k is the number of key layers;

[0035] b2) For each early departure decision-maker Dynamically generate early termination threshold for the model ,in For task importance modulation function, These are empirical parameters, set according to the task requirements. This is a resource scarcity modulation function. This represents the current percentage of computing resources being used. For energy resources that have already been consumed, The maximum energy allowed for a satellite to consume. and This is a proportionality coefficient, set based on experience.

[0036] Furthermore, the specific process of remote sensing intelligent computing performed by the spaceborne processing unit in step 4 is as follows:

[0037] a) Input: Smart Spatial Mask and focused calculation mask map ;

[0038] The intelligent interpretation model deployed on the onboard processing unit performs focused computation of the mask map during model forward propagation. The selected region is calculated with full precision; for the unselected region, approximate calculations are performed using channel pruning or low-bit quantization.

[0039] b) Input: Task adaptive cue vector Focused computational mask image and model early termination threshold ;

[0040] The feature map F of the intelligent interpretation model deployed on the onboard processing unit is modulated as follows: The scaling parameter and bias parameters Task-adaptive cue vector and network adaptation generate, Its size is 1×1×Nc, where Nc is the number of channels. This indicates channel-by-channel multiplication; where, adapter network It consists of a two-layer fully connected network, with the input being a task-adaptive cue vector. The hidden layer uses the ReLU activation function to introduce non-linearity, with a dimension of 2*Nc. The first Nc values ​​are scaled using the sigmoid function. The last Nc values ​​are used as bias parameters. ;

[0041] The intelligent interpretation model deployed on the onboard processing unit performs focused computation of the mask map during model forward propagation. The selected region is calculated with full precision; for the unselected region, approximate calculations are performed using channel pruning or low-bit quantization.

[0042] When the onboard processing unit performs model inference, When true, an early termination is triggered, where, To predict confidence levels, Indicates the first The feature maps of each key layer, where ⊕ represents the splicing and broadcasting of channel dimensions;

[0043] c) Input is: Collaborative Reasoning Instructions ;

[0044] Directly use collaborative reasoning instructions The multimodal intelligent interpretation model is deployed as a prompt command input to the onboard processing unit.

[0045] A space-ground collaborative intelligent remote sensing computing system for implementing the above method includes a ground guidance center, a spaceborne processing unit, and a space-ground two-way communication link;

[0046] The ground guidance center includes a multi-source information analysis module and a guidance information generation module; the onboard processing unit includes a command parsing and adaptation module, a focused computing module, a model early termination decision module, and an intelligent computing module; the two-way communication link between the ground guidance center and the onboard processing unit is used to support data communication between the ground guidance center and the onboard processing unit.

[0047] The multi-source information analysis module is used to perform target detection, land cover classification, and change detection analysis on historical remote sensing images, remote sensing thematic maps, vector geographic information data, and text data, and to perform statistical analysis on the results;

[0048] The guidance information generation module is used to generate generalized dynamic guidance information, including intelligent space masks, based on the results of the multi-source information analysis module and the remote sensing interpretation tasks on the satellite. Task adaptive cue vector and collaborative reasoning instructions And transmit the generalized dynamic guidance information to the onboard processing unit through a two-way communication link between space and ground;

[0049] The instruction parsing and adaptation module is used to receive generalized dynamic guidance information sent by the ground guidance center, parse the vector content, and determine the vector type;

[0050] The focused computation module is used to generate a focused computation mask map for guiding computation based on the vector content of the generalized dynamic guidance information;

[0051] The model early termination decision module is used to generate the model early termination threshold based on the vector content of the generalized dynamic guidance information;

[0052] The intelligent computing module deploys an intelligent interpretation model to perform intelligent computation based on generalized dynamic guidance information, focused computation mask map, and model early termination threshold, generating structured results.

[0053] The beneficial effects of this invention are as follows:

[0054] (1) From full-image processing to focused calculation, computational efficiency is greatly improved. Traditional on-board processing requires indiscriminate traversal calculation of the entire massive remote sensing image, while this invention realizes intelligent focused calculation on demand and according to priority through generalized dynamic guidance information. It can reduce the downlink data volume by more than 90%, solving the core bottleneck of satellite-to-ground transmission. The uplink also only needs to transmit a very small amount of guidance information, and the resource consumption is negligible.

[0055] (2) From static configuration to dynamic scheduling, improve system adaptability. The system can dynamically generate generalized prompt vectors according to the task scenario and perceive the status of on-board resources (computing power, memory, power, etc.) in real time. Based on this, it can dynamically adjust the data processing strategy to achieve dynamic and intelligent scheduling of resources.

[0056] (3) The system has self-optimization capabilities. This invention also establishes a closed-loop optimization mechanism, which has the ability to learn from experience and continuously improve. The processing results of satellite uplink and downlink are compared with the ground reference results to calculate the performance loss, and the parameters of the ground guidance information generation model are updated through the backpropagation algorithm. This allows the performance of the entire system to continuously improve itself over time and with the accumulation of tasks. Attached Figure Description

[0057] Figure 1 This is an overall flowchart of the satellite-ground collaborative intelligent remote sensing computing method based on generalized dynamic guidance of the present invention.

[0058] Figure 2 This is a diagram showing the composition of the satellite-ground collaborative intelligent remote sensing computing system of the present invention. Detailed Implementation

[0059] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The embodiments simulate a ground guidance center and a spaceborne processing unit; the system flow and composition are as follows. Figure 1 , Figure 2 As shown, the invention will be described in detail through three typical scenarios.

[0060] A satellite-ground collaborative intelligent remote sensing computing method based on generalized dynamic guidance, such as Figure 1 As shown, it includes the following steps:

[0061] Step 1: The ground guidance center processes and analyzes the stored historical multi-source information based on the remote sensing interpretation mission on the satellite to generate generalized dynamic guidance information.

[0062] The remote sensing interpretation tasks on the satellite include three tasks: target detection, land cover classification, and change detection; the historical multi-source information includes one or more of the following: historical remote sensing images, remote sensing thematic maps, vector geographic information data, and text data.

[0063] Generalized dynamic guidance information transcends binary space masks and is a semantically rich information carrier, specifically including intelligent spatial masks. Task adaptive cue vector and collaborative reasoning instructions The generation method is as follows:

[0064] The ground guidance center first analyzes historical remote sensing images using a land cover classification model, a target detection model, and a change detection model, generating analysis results including target detection results, land cover classification results, and change detection results. After saving the results, statistical analysis is performed by region, including region latitude and longitude, image time, target type and quantity, land cover type, and the number of changes. Then, an intelligent spatial mask is generated according to the following steps. Task adaptive cue vector and collaborative reasoning instructions ;

[0065] a) Intelligent spatial mask For different remote sensing interpretation tasks, the spatial sub-regions of the satellite image are calculated in the following ways. weight :

[0066] Target detection task: The target area refers to the area where the target may appear.

[0067] Change detection task: Fixed areas refer to regions that are unlikely to change based on historical remote sensing image patterns; potentially changeable areas refer to regions where changes occur based on the analysis results of historical remote sensing images. ,in It is the total number of historical remote sensing images that have changed. All historical remote sensing imagery; the mandatory area refers to the area that must be analyzed in this mission;

[0068] Land feature classification task: The classification-invariant region refers to the region in historical remote sensing imagery that has not undergone classification change, while the classification-change region refers to the region in historical remote sensing imagery where the classification result has changed at different times. The historical remote sensing image closest in time to the current task is selected, and the probability of classification change is calculated using its time interval t. , , This refers to a time interval sequence of changes within historical remote sensing images. It is a mean function.

[0069] b) Task-adaptive cue vector Its generation is a dynamic process based on a meta-learning model, specifically including the following steps:

[0070] b1) Task semantic encoding: Using a multilayer perceptron model, the current remote sensing interpretation task, data latitude and longitude range, and data acquisition time elements are encoded and mapped into task embedding vectors. ;

[0071] b2) Historical Data Analysis: Using the same multilayer perceptron model as b1, the results of historical remote sensing image analysis are encoded to generate a historical task experience set. ,in This is the number for the analysis of historical remote sensing images. A single analysis of a single historical remote sensing image is recorded as 1.

[0072] b3) Historical data fusion: Calculating task embedding vectors Experience with historical missions The similarity is calculated, and weights are generated using an attention mechanism. , , where s is the similarity function; then, the context vector is obtained by weighted summation. ;

[0073] b4) Meta-hint synthesis: [The following text appears to be a separate, unrelated sentence:] and Perform nonlinear fusion to output task-adaptive cue vectors. ;in It is implemented using a multilayer perceptron.

[0074] c) Collaborative reasoning instructions For the multimodal interpretation model deployed on the satellite, a structured text is generated as a prompt message. It is a set of structured machine-readable instructions in the format of {operation type, target area, parameters}. For example, {“detect”, “water_area”, “ship”} means to perform ship detection on the water area.

[0075] Step 2: The ground guidance center transmits the generalized dynamic guidance information to the satellite's onboard processing unit via the space-to-ground uplink.

[0076] Step 3: The onboard processing unit receives and parses the generalized dynamic guidance information. Based on the type of the generalized dynamic guidance information and the satellite's resource status, it dynamically generates a focusing calculation mask and a model early retreat threshold.

[0077] The process of focusing on the generation of the computational mask and the model early termination threshold is as follows:

[0078] a) The process of generating the focused computational mask image is as follows:

[0079] Smart Space Mask Based on the latitude and longitude range and spatial resolution of the satellite image, a focus calculation mask map of the area to be processed is generated. The specific method is from A region with the same latitude and longitude range as the satellite image is cropped out, and nearest neighbor interpolation is performed on the cropped region according to the resolution of the satellite image to obtain the focus calculation mask map. ;

[0080] Adaptive cue vectors for tasks First, a spatial importance heatmap is generated using a lightweight convolutional network. :

[0081] Where: ⊕ represents concatenation and broadcasting of channel dimensions, Broadcast to the same spatial size as image I on the satellite. This is the sigmoid function, with a range of [0,1]. elements in The weight representing the importance of processing at position (i,j);

[0082] Then, based on satellite resource constraints, the area to be processed is dynamically determined, and a focused computational mask map of the area to be processed is generated. :

[0083] ,in, The determination is based on a comprehensive assessment of onboard computing resources, satellite energy status, and onboard image size.

[0084] b) The process for generating the model's early termination threshold is as follows:

[0085] b1) Early Departure Decision Point Setting: Setting an early departure decision point in the key layer of the intelligent interpretation model deployed on the onboard processing unit. Among them, the key layer is set according to the requirements, and k is the number of key layers;

[0086] b2) For each early departure decision-maker Dynamically generate early termination threshold for the model ,in For task importance modulation function, These are empirical parameters, set according to the task requirements. This is a resource scarcity modulation function. This represents the current percentage of computing resources being used. For energy resources that have already been consumed, The maximum energy allowed for a satellite to consume. and This is a proportionality coefficient, set based on experience.

[0087] Step 4: The onboard processing unit performs remote sensing intelligent calculations based on generalized dynamic guidance information, focused calculation mask map and model early retreat threshold, and generates structured results.

[0088] The specific process of remote sensing intelligent computing carried out by the onboard processing unit is as follows:

[0089] a) Input: Smart Spatial Mask and focused calculation mask map ;

[0090] The intelligent interpretation model deployed on the onboard processing unit performs focused computation of the mask map during model forward propagation. The selected region is calculated with full precision; for the unselected region, approximate calculations are performed using channel pruning or low-bit quantization.

[0091] b) Input: Task adaptive cue vector Focused computational mask image and model early termination threshold ;

[0092] The feature map F of the intelligent interpretation model deployed on the onboard processing unit is modulated as follows: The scaling parameter and bias parameters Task-adaptive cue vector and network adaptation generate, Its size is 1×1×Nc, where Nc is the number of channels. This indicates channel-by-channel multiplication; where, adapter network It consists of a two-layer fully connected network, with the input being a task-adaptive cue vector. The hidden layer uses the ReLU activation function to introduce non-linearity, with a dimension of 2*Nc. The first Nc values ​​are scaled using the sigmoid function. The last Nc values ​​are used as bias parameters. ;

[0093] The intelligent interpretation model deployed on the onboard processing unit performs focused computation of the mask map during model forward propagation. The selected region is calculated with full precision; for the unselected region, approximate calculations are performed using channel pruning or low-bit quantization.

[0094] When the onboard processing unit performs model inference, When true, an early termination is triggered, where, To predict confidence levels, Indicates the first The feature maps of each key layer, where ⊕ represents the splicing and broadcasting of channel dimensions;

[0095] c) Input is: Collaborative Reasoning Instructions ;

[0096] Directly use collaborative reasoning instructions The multimodal intelligent interpretation model is deployed as a prompt command input to the onboard processing unit.

[0097] Step 5: The onboard processing unit transmits the processed structured result data back to the ground guidance center via the space-to-ground downlink and uses it to optimize the subsequent generation of generalized dynamic guidance information.

[0098] This invention also provides a space-ground collaborative intelligent remote sensing computing system, such as... Figure 2 As shown, the system used to implement the above method includes a ground guidance center, a spaceborne processing unit, and a two-way space-to-ground communication link.

[0099] The ground guidance center includes a multi-source information analysis module and a guidance information generation module; the onboard processing unit includes a command parsing and adaptation module, a focused computing module, a model early termination decision module, and an intelligent computing module; the two-way communication link between the ground guidance center and the onboard processing unit is used to support data communication between the ground guidance center and the onboard processing unit.

[0100] The multi-source information analysis module is used to perform target detection, land cover classification, and change detection analysis on historical remote sensing images, remote sensing thematic maps, vector geographic information data, and text data, and to perform statistical analysis on the results;

[0101] The guidance information generation module is used to generate generalized dynamic guidance information, including intelligent space masks, based on the results of the multi-source information analysis module and the remote sensing interpretation tasks on the satellite. Task adaptive cue vector and collaborative reasoning instructions And transmit the generalized dynamic guidance information to the onboard processing unit through a two-way communication link between space and ground;

[0102] The instruction parsing and adaptation module is used to receive generalized dynamic guidance information sent by the ground guidance center, parse the vector content, and determine the vector type;

[0103] The focused computation module is used to generate a focused computation mask map for guiding computation based on the vector content of the generalized dynamic guidance information;

[0104] The model early termination decision module is used to generate the model early termination threshold based on the vector content of the generalized dynamic guidance information;

[0105] The intelligent computing module deploys an intelligent interpretation model to perform intelligent computation based on generalized dynamic guidance information, focused computation mask map, and model early termination threshold, generating structured results.

[0106] Example 1: Detection of Maritime Ship Targets

[0107] In this embodiment, the ground guidance center first receives the ship detection mission command. The system generates an intelligent spatial mask by analyzing historical remote sensing data and marine geographic information, focusing on areas where ships frequently appear, such as ports and waterways, based on the generated generalized dynamic guidance information in the form of an intelligent spatial mask.

[0108] After receiving the generalized dynamic guidance information, the onboard processing unit activates the ship detection model. The model first determines the processing area based on an intelligent spatial mask, applying the ship detection algorithm to these areas while ignoring data from other areas. Actual testing shows that, depending on the size of the water area, this method can reduce processing time by 65%-90% while maintaining detection accuracy. Furthermore, by eliminating land-based interference, the false alarm rate for target detection can be reduced by 10%-30%.

[0109] Example 2: Post-disaster building damage assessment (change detection)

[0110] This embodiment demonstrates the system's application in disaster emergency response. Upon receiving a disaster assessment task, the ground guidance center immediately activates the multi-source information analysis module. This module integrates satellite imagery, geographic information system data, and disaster reports to generate targeted, generalized dynamic guidance information. Based on historical remote sensing imagery and change detection results, the ground guidance center uses a meta-learning model to generate a task-adaptive prompt vector for this task. Information focusing on residential areas, roads, factories, etc., is encoded into the task-adaptive prompt vector, and a pre-configured early termination decision mechanism is implemented to terminate the calculation prematurely when a predetermined threshold is reached.

[0111] Practical results show that, based on the ResNet-18 model, in a simulated on-board computing environment (Jetson AGXOrin), when processing a 20000×20000 image, the on-board model can ignore large areas of non-human activity such as vegetation and water bodies, focusing on key areas of concern such as residential areas, effectively reducing the data computation scope. The model's early termination mechanism can exit when the accuracy reaches 85%. Without using the strategy of this invention, the on-board model's depth calculation time was 127 minutes, and the data extraction accuracy was 87.1% after manual verification. After using the strategy of this invention, the on-board model performs focused calculations in 6.5 minutes, with a data extraction accuracy of 85.5%, and complete coverage of key areas of concern. The method of this invention can improve efficiency by approximately 20 times while sacrificing a slight amount of accuracy.

[0112] Example 3: Classification of Natural Resource Features

[0113] This task also validated the system's adaptive optimization capabilities. The ground guidance center generated time-adaptive, generalized dynamic guidance information based on vegetation growth cycles. The onboard processing unit continuously optimized its ability to represent crop characteristics through a continuous learning mechanism. The system employed a collaborative reasoning instruction mechanism to decompose the complex land cover classification task into sub-tasks such as vegetation extraction, water body extraction, and building extraction. Each sub-task had corresponding processing priorities and accuracy requirements. The onboard processing module executed the instructions sequentially, adjusting the processing strategy at each step based on real-time resource conditions. Results showed that, under different task modes, using the collaborative reasoning instruction mechanism could improve data processing efficiency by 3-15 times. The specific efficiency improvement varied depending on task complexity and the proportion of land cover; for analyzing water bodies, which accounted for a small proportion, the improvement reached up to 15 times; for classifying vegetation, which accounted for a large proportion, the improvement was approximately 3 times. Through the system's optimization mechanism, the accuracy of guidance information generation can be effectively improved.

[0114] Experimental conclusion:

[0115] Experiments demonstrate that this invention exhibits significant advantages over existing technologies. In terms of computational efficiency, the focused processing strategy improves average computational efficiency by approximately one order of magnitude. Regarding processing accuracy, precise guidance from the target information reduces the false alarm rate by 10-30%, and in early termination mode, accuracy loss in critical areas is within 3%. Furthermore, the significantly reduced data transmission volume significantly alleviates the bandwidth pressure on satellite-to-ground communication, reducing data transmission volume by an average of 95%.

[0116] Of particular note is the system's self-learning capability, which enables it to continuously adapt to new task requirements. Long-term operational data shows that, after multiple optimization iterations, the system demonstrates excellent generalization ability when handling novel remote sensing tasks.

Claims

1. A satellite-ground collaborative intelligent remote sensing computing method based on generalized dynamic guidance, characterized in that, Includes the following steps: Step 1: The ground guidance center processes and analyzes the stored historical multi-source information based on the remote sensing interpretation mission on the satellite to generate generalized dynamic guidance information; Step 2: The ground guidance center transmits the generalized dynamic guidance information to the satellite's onboard processing unit via the space-to-ground uplink; Step 3: The onboard processing unit receives and parses the generalized dynamic guidance information. Based on the type of the generalized dynamic guidance information and the satellite's resource status, it dynamically generates a focusing calculation mask and a model early retreat threshold. Step 4: The onboard processing unit performs remote sensing intelligent calculations based on generalized dynamic guidance information, focused calculation mask map, and model early termination threshold to generate structured results; Step 5: The onboard processing unit transmits the processed structured result data back to the ground guidance center via the space-to-ground downlink and uses it to optimize the subsequent generation of generalized dynamic guidance information; Among them, the generalized dynamic guidance information mentioned in step 1 is a semantically rich information carrier that transcends binary space masks, and its specific forms include intelligent spatial masks. Task adaptive cue vector and collaborative reasoning instructions The generation method is as follows: The ground guidance center first analyzes historical remote sensing images using a land cover classification model, a target detection model, and a change detection model, generating analysis results including target detection results, land cover classification results, and change detection results. After saving the results, statistical analysis is performed by region, including region latitude and longitude, image time, target type and quantity, land cover type, and the number of changes. Then, an intelligent spatial mask is generated according to the following steps. Task adaptive cue vector and collaborative reasoning instructions ; a) Intelligent spatial mask For different remote sensing interpretation tasks, the spatial sub-regions of the satellite image are calculated in the following ways. weight : Target detection task: The target area refers to the area where the target may appear. Change detection task: Fixed areas refer to regions that are unlikely to change based on historical remote sensing image patterns; potentially changeable areas refer to regions where changes occur based on the analysis results of historical remote sensing images. ,in It is the total number of historical remote sensing images that have changed. All historical remote sensing imagery; the mandatory area refers to the area that must be analyzed in this mission; Land feature classification task: The classification-invariant region refers to the region in historical remote sensing imagery that has not undergone classification change, while the classification-change region refers to the region in historical remote sensing imagery where the classification result has changed at different times. The historical remote sensing image closest in time to the current task is selected, and the probability of classification change is calculated using its time interval t. , , This refers to a time interval sequence of changes within historical remote sensing images. It is a mean function; b) Task-adaptive cue vector Its generation is a dynamic process based on a meta-learning model, specifically including the following steps: b1) Task semantic encoding: Using a multilayer perceptron model, the current remote sensing interpretation task, data latitude and longitude range, and data acquisition time elements are encoded and mapped into task embedding vectors. ; b2) Historical Data Analysis: Using the same multilayer perceptron model as b1, the results of historical remote sensing image analysis are encoded to generate a historical task experience set. ,in This is the number for the analysis of historical remote sensing images. A single analysis of a single historical remote sensing image is recorded as 1. b3 Historical Data Fusion: Computing Task Embedded Vectors Experience with historical missions The similarity is calculated, and weights are generated using an attention mechanism. , , where s is the similarity function; then, the context vector is obtained by weighted summation. ; b4) Meta-hint synthesis: [The following text appears to be a separate, unrelated sentence:] and Perform nonlinear fusion to output task-adaptive cue vectors. ;in Implemented using a multilayer perceptron; c) Collaborative reasoning instructions For the multimodal interpretation model deployed on the satellite, a structured text is generated as a prompt message. It is a set of structured machine-readable instructions in the format of {operation type, target area, parameters}.

2. The satellite-ground collaborative intelligent remote sensing calculation method based on generalized dynamic guidance according to claim 1, characterized in that, The remote sensing interpretation tasks on the satellite in step 1 include three tasks: target detection, land cover classification, and change detection; the historical multi-source information in step 1 includes one or more of the following: historical remote sensing images, remote sensing thematic maps, vector geographic information data, and text data.

3. The satellite-ground collaborative intelligent remote sensing calculation method based on generalized dynamic guidance according to claim 1, characterized in that, The process of generating the focused calculation mask map and the model early termination threshold in step 3 is as follows: a) The process of generating the focused computational mask image is as follows: Intelligent spatial mask Based on the latitude and longitude range and spatial resolution of the satellite image, a focus calculation mask map of the area to be processed is generated. The specific method is from A region with the same latitude and longitude range as the satellite image is cropped out, and nearest neighbor interpolation is performed on the cropped region according to the resolution of the satellite image to obtain the focus calculation mask map. ; Adaptive cue vectors for tasks First, a spatial importance heatmap is generated using a lightweight convolutional network. : Where: ⊕ represents concatenation and broadcasting of channel dimensions, Broadcast to the same spatial size as image I on the satellite. This is the sigmoid function, with a range of [0,1]. elements in The processing importance weights for positions (p, q) are represented. Then, based on satellite resource constraints, the area to be processed is dynamically determined, and a focused computational mask map of the area to be processed is generated. : ,in, The determination is based on a comprehensive consideration of onboard computing resources, satellite power status, and onboard image size. b) The process for generating the model's early termination threshold is as follows: b1) Early Departure Decision Point Setting: Setting an early departure decision point in the key layer of the intelligent interpretation model deployed on the onboard processing unit. Among them, the key layer is set according to the requirements, and k is the number of key layers; b2) For each early departure decision-maker Dynamically generate early termination threshold for the model ,in For task importance modulation function, These are empirical parameters, set according to the task requirements. This is a resource scarcity modulation function. This represents the current percentage of computing resources being used. For energy resources that have already been consumed, The maximum energy allowed for a satellite to consume. and This is a proportionality coefficient, set based on experience.

4. The satellite-ground collaborative intelligent remote sensing calculation method based on generalized dynamic guidance according to claim 3, characterized in that, The specific process of remote sensing intelligent computing performed by the spaceborne processing unit in step 4 is as follows: a) Input: Smart Spatial Mask and focused calculation mask map ; The intelligent interpretation model deployed on the onboard processing unit performs focused computation of the mask map during model forward propagation. The selected region is calculated with full precision; for the unselected region, approximate calculations are performed using channel pruning or low-bit quantization. b) Input: Task adaptive cue vector Focused computational mask image and model early termination threshold ; The feature map F of the intelligent interpretation model deployed on the onboard processing unit is modulated as follows: The scaling parameter and bias parameters Task-adaptive cue vector and network adaptation generate, Its size is 1×1×Nc, where Nc is the number of channels. This indicates channel-by-channel multiplication; where, adapter network It consists of a two-layer fully connected network, with the input being a task-adaptive cue vector. The hidden layer uses the ReLU activation function to introduce non-linearity, with a dimension of 2*Nc. The first Nc values ​​are scaled using the sigmoid function. The last Nc values ​​are used as bias parameters. ; The intelligent interpretation model deployed on the onboard processing unit performs focused computation of the mask map during model forward propagation. The selected region is calculated with full precision; for the unselected region, approximate calculations are performed using channel pruning or low-bit quantization. When the onboard processing unit performs model inference, When true, an early termination is triggered, where, To predict confidence levels, represents the feature map of the l-th key layer, and ⊕ represents the concatenation and broadcasting of channel dimension l; c) Input is: Collaborative Reasoning Instructions ; Directly use collaborative reasoning instructions The multimodal intelligent interpretation model is deployed as a prompt command input to the onboard processing unit.

5. A satellite-ground collaborative intelligent remote sensing computing system for implementing the method of any one of claims 1-4, characterized in that, This includes a ground guidance center, onboard processing unit, and a two-way communication link between the satellite and the ground. The ground guidance center includes a multi-source information analysis module and a guidance information generation module; the onboard processing unit includes a command parsing and adaptation module, a focused computing module, a model early termination decision module, and an intelligent computing module; the two-way communication link between the ground guidance center and the onboard processing unit is used to support data communication between the ground guidance center and the onboard processing unit. The multi-source information analysis module is used to perform target detection, land cover classification, and change detection analysis on historical remote sensing images, remote sensing thematic maps, vector geographic information data, and text data, and to perform statistical analysis on the results; The guidance information generation module is used to generate generalized dynamic guidance information, including intelligent space masks, based on the results of the multi-source information analysis module and the remote sensing interpretation tasks on the satellite. Task adaptive cue vector and collaborative reasoning instructions And transmit the generalized dynamic guidance information to the onboard processing unit through a two-way communication link between space and ground; The instruction parsing and adaptation module is used to receive generalized dynamic guidance information sent by the ground guidance center, parse the vector content, and determine the vector type; The focused computation module is used to generate a focused computation mask map for guiding computation based on the vector content of the generalized dynamic guidance information; The model early termination decision module is used to generate the model early termination threshold based on the vector content of the generalized dynamic guidance information; The intelligent computing module deploys an intelligent interpretation model to perform intelligent computing based on generalized dynamic guidance information, focused computation mask map, and model early termination threshold, generating structured results. Among them, intelligent spatial mask Task adaptive cue vector and collaborative reasoning instructions The generation method is as follows: The ground guidance center first analyzes historical remote sensing images using a land cover classification model, a target detection model, and a change detection model, generating analysis results including target detection results, land cover classification results, and change detection results. After saving the results, statistical analysis is performed by region, including region latitude and longitude, image time, target type and quantity, land cover type, and the number of changes. Then, an intelligent spatial mask is generated according to the following steps. Task adaptive cue vector and collaborative reasoning instructions ; a) Intelligent spatial mask For different remote sensing interpretation tasks, the spatial sub-regions of the satellite image are calculated in the following ways. weight : Target detection task: The target area refers to the area where the target may appear. Change detection task: Fixed areas refer to regions that are unlikely to change based on historical remote sensing image patterns; potentially changeable areas refer to regions where changes occur based on the analysis results of historical remote sensing images. ,in It is the total number of historical remote sensing images that have changed. All historical remote sensing imagery; the mandatory area refers to the area that must be analyzed in this mission; Land feature classification task: The classification-invariant region refers to the region in historical remote sensing imagery that has not undergone classification change, while the classification-change region refers to the region in historical remote sensing imagery where the classification result has changed at different times. The historical remote sensing image closest in time to the current task is selected, and the probability of classification change is calculated using its time interval t. , , This refers to a time interval sequence of changes within historical remote sensing images. It is a mean function; b) Task-adaptive cue vector Its generation is a dynamic process based on a meta-learning model, specifically including the following steps: b1) Task semantic encoding: Using a multilayer perceptron model, the current remote sensing interpretation task, data latitude and longitude range, and data acquisition time elements are encoded and mapped into task embedding vectors. ; b2) Historical Data Analysis: Using the same multilayer perceptron model as b1, the results of historical remote sensing image analysis are encoded to generate a historical task experience set. ,in This is the number for the analysis of historical remote sensing images. A single analysis of a single historical remote sensing image is recorded as 1. b3 Historical Data Fusion: Computing Task Embedded Vectors Experience with historical missions The similarity is calculated, and weights are generated using an attention mechanism. , , where s is the similarity function; then, the context vector is obtained by weighted summation. ; b4) Meta-hint synthesis: [The following text appears to be a separate, unrelated sentence:] and Perform nonlinear fusion to output task-adaptive cue vectors. ;in Implemented using a multilayer perceptron; c) Collaborative reasoning instructions For the multimodal interpretation model deployed on the satellite, a structured text is generated as a prompt message. It is a set of structured machine-readable instructions in the format of {operation type, target area, parameters}.

Images