Satellite-ground cooperative intelligent remote sensing calculation method and system based on generalized dynamic guidance

Abstract

The invention discloses a satellite-ground cooperative intelligent remote sensing calculation method and system based on generalized dynamic guidance, and relates to the technical field of space information technology, artificial intelligence and edge calculation crossing. According to the method, generalized dynamic guidance information including an intelligent space mask, a task adaptive prompt vector and a collaborative reasoning instruction is formed through a ground guidance center; the satellite-borne processing unit realizes self-adaptive focusing processing by adopting local focusing calculation and a conditional premature exit mechanism according to the guide information and in combination with a real-time resource state; and meanwhile, a space-ground closed-loop optimization system is constructed, and a guide strategy is continuously optimized through result feedback. According to the method, the mode conversion from full-image processing to focusing calculation is realized, the on-satellite processing efficiency is remarkably improved, the calculation amount is reduced by 70-90% while the precision is ensured, the satellite-ground communication data volume is reduced by more than 95%, the problems of on-satellite resource limitation and data transmission bandwidth bottleneck are effectively solved, and the method is suitable for large-scale popularization and application. The method has important value in remote sensing application scenes such as disaster emergency and environment monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of spatial information technology, artificial intelligence and edge computing, and particularly relates to a star-ground collaborative intelligent remote sensing computing method and system based on generalized dynamic guidance. BACKGROUND

[0002] With the rapid development of remote sensing satellite constellation, the remote sensing data acquisition capability has shown an exponential growth trend, however, the traditional "on-board collection, ground processing" mode is facing severe challenges. The most prominent contradiction is reflected in the difficulty of massive raw data downlink caused by the bottleneck of data transmission bandwidth, and the poor timeliness caused thereby, which is difficult to meet the real-time application requirements such as emergency response.

[0003] Direct data processing on satellite is an effective means to realize efficient application of data. Although the existing technical solutions have made certain explorations in on-board processing, most of them still remain in the simple static division mode, that is, the on-board and ground processing tasks are fixed in advance, and lack of dynamic adaptability. It is worth noting that the guidance information used in the existing star-ground collaborative computing model is often limited to simple prior information such as binary space mask and image compression, and lacks deep semantic understanding ability. In addition, there is a lack of effective intelligent interaction mechanism and closed-loop optimization capability between star-ground systems, resulting in low resource utilization efficiency, and the inability to adaptively schedule according to real-time task requirements and resource status.

[0004] These technical defects seriously restrict the further improvement of on-orbit remote sensing information extraction efficiency, and therefore there is an urgent need in the field for a new type of remote sensing processing architecture that can realize star-ground intelligent collaboration, dynamic guidance and resource adaptation. SUMMARY

[0005] The present application aims to overcome the deficiencies of the prior art and provides a star-ground collaborative intelligent remote sensing computing method and system based on generalized dynamic guidance. The core innovation of the present application is the concept of "generalized dynamic guidance", which realizes efficient processing of remote sensing data through an intelligent computing architecture of star-ground collaboration. The essential difference between the present application and the prior art is that, instead of simply migrating ground processing tasks to the satellite, the "generalized dynamic guidance information" generated by the ground guidance center is used to realize "precise control" of the on-board computing process. This guidance is dynamic, semantic and optimizable. The star-ground collaborative mode is innovated, which changes from traditional static division to integrated collaboration, in which the ground is responsible for planning and decision-making, and the satellite is responsible for precise execution.

[0006] The technical scheme adopted by the present application is as follows:

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

[0008] Step 1: The ground guidance center processes and analyzes the stored historical multi-source information according to the remote sensing interpretation tasks 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-borne processing unit through the space-earth uplink;

[0010] Step 3: The satellite-borne processing unit receives and analyzes the generalized dynamic guidance information, and dynamically generates a focused calculation mask and a model early retirement threshold according to the type of the generalized dynamic guidance information and the resource state of the satellite;

[0011] Step 4: The satellite-borne processing unit carries out remote sensing intelligent calculation according to the generalized dynamic guidance information, the focused calculation mask and the model early retirement threshold, to generate structured results;

[0012] Step 5: The satellite-borne processing unit transmits the structured results data obtained by processing back to the ground guidance center through the space-earth downlink, and uses it to optimize the generation of subsequent generalized dynamic guidance information.

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

[0014] Further, the generalized dynamic guidance information in step 1 is an information carrier beyond the binary space mask and rich in semantics, and the specific forms include intelligent space mask , task adaptive prompt vector and collaborative reasoning instruction , and the generation method is as follows:

[0015] The ground guidance center first analyzes the historical remote sensing images through the feature classification model, target detection model and change detection model, to generate analysis results including target detection results, feature classification results and change detection results, saves the results, and then performs statistical analysis according to regions. The contents of statistical analysis include regional latitude and longitude, image time, target type and quantity, feature type and change occurrence number; and the intelligent space mask , task adaptive prompt vector and collaborative reasoning instruction are generated according to the following steps;

[0016] a) Intelligent space mask , for different remote sensing interpretation tasks, the weight of the spatial sub-region of the satellite image is calculated in the following manner:

[0017] Target detection task: wherein the target region refers to the region where the target may appear;

[0018] Change detection task: wherein the fixed region refers to the region that cannot change according to the historical remote sensing image rule; the possible change region refers to the region that may change according to the historical remote sensing image analysis result, wherein is the total number of images that have changed in the historical remote sensing image, is all historical remote sensing images; the must-check region refers to the region that must be analyzed in this task;

[0019] Feature classification task: wherein the classification unchanged region refers to the region that has not changed in classification in the historical remote sensing image, and the classification change region refers to the region that has changed in classification in the historical remote sensing image at different times; the historical remote sensing image closest in time to the current task is selected to calculate the classification change probability with a time interval t , , is the time interval sequence in which changes occur in the historical remote sensing image, is the mean function;

[0020] b) Task adaptive prompt vector The generation thereof is a dynamic process based on a meta-learning model, and specifically includes the following steps:

[0021] b1) Task semantic encoding: a multi-layer perception model is used to encode the current remote sensing interpretation task, the data latitude and longitude range, and the data acquisition time element, and map them into a task embedding vector ;

[0022] b2) Historical data analysis: the same multi-layer perception model as b1 is used to encode the results of historical remote sensing image analysis, and generate a historical task experience set wherein is the number of historical remote sensing image analysis, and a single analysis of a single historical remote sensing image is recorded as 1 time;

[0023] b3) Historical data fusion: the similarity between the task embedding vector and the historical task experience set is calculated, and the attention mechanism is used to generate a weight , wherein s is a similarity function; and then the context vector is obtained by weighted summation;

[0024] b4) Meta prompt synthesis: the and Nonlinear fusion is performed to output a task adaptive prompt vector ; wherein is implemented by using a multi-layer perception;

[0025] c) Collaborative reasoning instructions : For the multi-modal interpretation model deployed on the satellite, generate a structured text as a prompt information, which is a set of structured machine-readable instructions in the format {operation type, target area, parameter}.

[0026] Further, the focus calculation mask map and the generation process of the model early return threshold in step 3 are as follows:

[0027] a) The generation process of the focus calculation mask map is as follows:

[0028] Mask for intelligent space According to the latitude and longitude range and spatial resolution of the on-board image, a focus calculation mask map of the to-be-processed area is generated ; The specific method is to cut out an area with the same latitude and longitude range as the on-board image from , and perform nearest neighbor interpolation on the cut-out area according to the resolution of the on-board image, that is, the focus calculation mask map is obtained.

[0029] Task adaptive prompt vector , first generate a spatial importance heat map :

[0030] , wherein: represents the concatenation and broadcast of the channel dimension, and is broadcast to the same spatial size as the on-board image I, is a sigmoid function, and the value range is [0, 1], , the element in represents the processing importance weight of position (i, j);

[0031] Then, according to the satellite resource constraint, the area to be processed is dynamically determined, and a focus calculation mask map of the area to be processed is generated :

[0032] , wherein is determined according to the on-board computing resources, the satellite energy state, and the size of the on-board image;

[0033] b) The generation process of the model early return threshold is as follows:

[0034] b1) Early return decision point setting: Set an early return decision maker at the key layer of the intelligent interpretation model deployed on the satellite-borne processing unit ; wherein the key layer is set according to requirements, and k is the number of layers of the key layer;

[0035] b2) for each early exit decision maker , dynamically generate a model early exit threshold , wherein is a task importance modulation function, is an empirical parameter, which is set according to the task situation, is a resource tension modulation function, is the occupation ratio of the current computing resource, is the energy resource that has been consumed, is the maximum energy that the satellite allows to consume, and are proportional coefficients, which are set according to experience.

[0036] Further, the specific process of the intelligent remote sensing calculation carried out by the on-board processing unit in step 4 is as follows:

[0037] a) input is: intelligent space mask and focus calculation mask map ;

[0038] When the intelligent interpretation model deployed on the on-board processing unit is forward propagated, full-precision calculation is performed on the selected area of the focus calculation mask map ; for the unselected area, approximate calculation is performed in the way of channel pruning or low-bit quantization;

[0039] b) input is: task adaptive prompt vector , focus calculation mask map and model early exit threshold ;

[0040] For the feature map F of the intelligent interpretation model deployed on the on-board processing unit, the following modulation is performed: , wherein the scaling parameter and the bias parameter are generated by the task adaptive prompt vector and the adapter network , whose size is 1x1xNc, Nc is the number of channels, represents channel-by-channel multiplication; wherein the adapter network is composed of a two-layer fully connected network, the input is the task adaptive prompt vector , the hidden layer uses the ReLU activation function to introduce nonlinearity, the dimension is 2*Nc, the first Nc values are taken as the scaling parameter through the sigmoid function, and the last Nc values are taken as the bias parameter ;

[0041] The intelligent interpretation model deployed on the spaceborne processing unit triggers early exit when the confidence of the prediction is true, where, The selected region is calculated with full precision; for the unselected region, approximate calculation is performed in the way of channel pruning or low-bit quantization;

[0042] When the spaceborne processing unit performs model inference, is true, early exit is triggered, where, is the prediction confidence, represents the feature map of the i-th key layer, and represents the concatenation and broadcast of the channel dimension;

[0043] c) The input is: collaborative inference instruction ;

[0044] The collaborative inference instruction is directly input to the multi-modal intelligent interpretation model deployed on the spaceborne processing unit as a prompt instruction.

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

[0046] The ground guidance center comprises a multi-source information analysis module and a guidance information generation module, the spaceborne processing unit comprises an instruction analysis and adaptation module, a focused computing module, a model early exit decision module and an intelligent computing module; the space-ground bidirectional communication link is used to support data communication between the ground guidance center and the spaceborne processing unit;

[0047] The multi-source information analysis module is used to perform target detection, feature 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 according to the results of the multi-source information analysis module and the remote sensing interpretation task on the satellite, including intelligent spatial mask , task adaptive prompt vector and collaborative inference instruction , and transmits the generalized dynamic guidance information to the spaceborne processing unit through the space-ground bidirectional communication link;

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

[0050] The focused computing module is used to generate a focused computing mask graph for guiding the calculation according to the vector content of the generalized dynamic guidance information;

[0051] ​The model early exit decision module is configured to generate a model early exit threshold according to the vector content of the generalized dynamic guidance information;

[0052] The intelligent computing module deploys an intelligent interpretation model configured to perform intelligent computing according to the generalized dynamic guidance information, the focus computing mask map and the model early exit threshold to generate a structured result.

[0053] The present application has the following advantages:

[0054] (1) From full-image processing to focus computing, the computing efficiency is greatly improved. Traditional on-board processing requires non-discriminatory traversal computing on the entire mass remote sensing image, while the present application realizes intelligent focus computing on demand and according to priority through generalized dynamic guidance information. The downlink data volume can be reduced by more than 90%, solving the core bottleneck of satellite-ground transmission. The uplink only needs to transmit very small guidance information, occupying negligible resources.

[0055] (2) From static configuration to dynamic scheduling, the system adaptability is improved. The system can dynamically generate a generalized prompt vector according to the task scene, and real-time perceive the on-board resource state (computing power, memory, power, etc.), and dynamically adjust the data processing strategy accordingly, realizing dynamic and intelligent scheduling of resources.

[0056] (3) The system has self-optimization capability. The present application also establishes a closed-loop optimization mechanism, which has the ability to learn from experience and continuously improve. The processing result transmitted from the satellite is compared with the ground reference result, the computing efficiency loss is calculated, and the parameters of the ground guidance information generation model are updated through the back propagation algorithm. The performance of the entire system can be continuously improved over time and with task accumulation. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 is a whole flowchart of the present application based on generalized dynamic guidance intelligent remote sensing computing method.

[0058] Figure 2 is a composition diagram of the present application intelligent remote sensing computing system. DETAILED DESCRIPTION

[0059] The present application will be further described in detail below in combination with the drawings and specific implementation examples. The system flow and composition of the embodiment simulating the ground guidance center and the on-board processing unit are shown in Figure 1 , Figure 2 . The present application is described in detail through three typical scenarios.

[0060] A generalized dynamic guidance based star-ground collaborative intelligent remote sensing computing method, as shown in Figure 1 , includes the following steps:

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

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

[0063] The generalized dynamic guidance information is a carrier of information beyond binary space mask and rich in semantics, including intelligent space mask , task adaptive prompt vector , and collaborative reasoning instructions , and the generation method is as follows:

[0064] The ground guidance center first analyzes the historical remote sensing images through feature classification models, target detection models, and change detection models to generate analysis results, including target detection results, feature classification results, and change detection results. After saving the results, statistical analysis is performed according to regions, and the contents of statistical analysis include regional latitude and longitude, image time, target type and quantity, feature type, and change occurrence number; and the intelligent space mask , task adaptive prompt vector , and collaborative reasoning instructions are generated according to the following steps.

[0065] a) Intelligent space mask , for different remote sensing interpretation tasks, the weight of the spatial sub-region of the image on the satellite is calculated as follows:

[0066] Target detection task: where the target region refers to the region where the target may appear;

[0067] Change detection task: where the fixed region refers to the region that cannot change according to the historical remote sensing image rules; the possible change region refers to the region that may change according to the historical remote sensing image analysis results, where is the total number of images that have changed in the historical remote sensing images, is all historical remote sensing images; the must-check region refers to the region that must be analyzed in this task;

[0068] Feature classification task: ​wherein the classification invariant region refers to the region that has no classification change in historical remote sensing images, and the classification change region refers to the region that has classification change in different time classification results in historical remote sensing images; a historical remote sensing image closest to the current task in time is selected to calculate the classification change probability with a time interval t , , is a sequence of time intervals in which the historical remote sensing image changes, is a mean function.

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

[0070] b1) Task semantic encoding: using a multi-layer perceptron model to encode the current remote sensing interpretation task, data latitude and longitude range, and data acquisition time elements, and mapping them into a task embedding vector ;

[0071] b2) Historical data analysis: using the same multi-layer perceptron model as b1, encoding the results of historical remote sensing image analysis to generate a historical task experience set , wherein is the number of historical remote sensing image analysis, and a single analysis of a single historical remote sensing image is recorded as 1 time;

[0072] b3) Historical data fusion: calculating the similarity of the task embedding vector and the historical task experience set , and using the attention mechanism to generate the weight , , wherein s is a similarity function; and then weighted sum to obtain the context vector ;

[0073] b4) Meta prompt synthesis: nonlinearly fusing and to output the task adaptive prompt vector ; wherein is implemented by a multi-layer perceptron.

[0074] c) Collaborative reasoning instruction : for the multi-modal interpretation model deployed on the satellite, generate a structured text as a prompt information, which is a set of structured machine-readable instructions in the format of {operation type, target region, parameter}; for example, {“detect”, “water_area”, “ship”} represents performing ship detection on the water area.

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

[0076] Step 3: The on-board processing unit receives and parses the generalized dynamic guidance information, dynamically generates a focus calculation mask map and a model early return threshold according to the type of generalized dynamic guidance information and the resource state of the satellite.

[0077] Wherein, the generation process of focus calculation mask map and model early return threshold is as follows:

[0078] a) The generation process of focus calculation mask map is as follows:

[0079] For intelligent space mask , according to the latitude and longitude range and spatial resolution of the on-board image, the focus calculation mask map of the region to be processed is generated ; The specific method is to cut out the region with the same latitude and longitude range as the on-board image from , and perform nearest neighbor interpolation on the cut-out region according to the resolution of the on-board image, that is, the focus calculation mask map is obtained;

[0080] For task adaptive prompt vector , first, a light convolutional network is used to generate a spatial importance heat map :

[0081] , wherein: represents the concatenation and broadcast of channel dimension, broadcasts to the same spatial size as the on-board image I, is a sigmoid function, with a value range of [0, 1], , the element in represents the processing importance weight of position (i, j);

[0082] Then, according to the satellite resource constraint, the region to be processed is dynamically determined, and the focus calculation mask map of the region to be processed is generated :

[0083] , wherein, is determined according to the on-board computing resources, satellite energy state and on-board image size.

[0084] b) The generation process of model early return threshold is as follows:

[0085] b1) Early return decision point setting: set early return decision maker in the key layer of the intelligent interpretation model deployed on the on-board processing unit ; Wherein, the key layer is set according to the demand, and k is the number of key layers;

[0086] b2) for each early exit decision maker , dynamically generating a model early exit threshold , wherein is a task importance modulation function, is an empirical parameter, set according to the task situation, is a resource tension modulation function, is the occupation ratio of the current computing resource, is the energy resource that has been consumed, is the maximum energy that the satellite allows to consume, and are proportional coefficients, set according to experience.

[0087] Step 4: The on-board processing unit carries out remote sensing intelligent computing according to the generalized dynamic guidance information, the focused computing mask map and the model early exit threshold, and generates a structured result.

[0088] Wherein, the specific process of the on-board processing unit carrying out remote sensing intelligent computing is as follows:

[0089] a) Input is: intelligent space mask and focused computing mask map ;

[0090] When the intelligent interpretation model deployed on the on-board processing unit is forward propagated, full-precision calculation is performed on the selected area of the focused computing mask map ; for the unselected area, approximate calculation is performed in the way of channel pruning or low-bit quantization;

[0091] b) Input is: task adaptive prompt vector , focused computing mask map and model early exit threshold ;

[0092] For the feature map F of the intelligent interpretation model deployed on the on-board processing unit, the following modulation is performed: , wherein the scaling parameter and the bias parameter are generated by the task adaptive prompt vector and the adapter network , , the size of which is 1x1xNc, Nc is the number of channels, indicates channel-by-channel multiplication; wherein, the adapter network is composed of a two-layer fully connected network, the input of which is the task adaptive prompt vector , the hidden layer uses the ReLU activation function to introduce nonlinearity, the dimension of which is 2*Nc, the first Nc values are taken as the scaling parameter through the sigmoid function, and the last Nc values are taken as the bias parameter ;

[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 analysis and adaptation module is configured to receive the generalized dynamic guidance information sent by the ground guidance center, and analyze the vector content to determine the vector type.

[0103] The focus calculation module is configured to generate a focus calculation mask map for guidance calculation according to the vector content of the generalized dynamic guidance information.

[0104] The model early exit decision module is configured to generate a model early exit threshold according to the vector content of the generalized dynamic guidance information.

[0105] The intelligent calculation module deploys an intelligent interpretation model, which is configured to perform intelligent calculation according to the generalized dynamic guidance information, the focus calculation mask map, and the model early exit threshold, and generate a structured result.

[0106] Embodiment one: offshore ship target detection

[0107] In this embodiment, the ground guidance center first receives a ship detection task instruction. The system generates an intelligent space mask by analyzing historical remote sensing data and marine geographic information, focusing on areas such as ports and shipping lanes where ships frequently appear, according to the form one: intelligent space mask of the generated generalized dynamic guidance information.

[0108] After the satellite processing unit receives the generalized dynamic guidance information, it starts the ship detection model. The model first determines the processing area according to the intelligent space mask, and uses the ship detection algorithm in these areas while ignoring the data in other areas. Actual tests show that this method can reduce processing time by 65%-90% while ensuring detection accuracy, and the target detection false alarm rate can be reduced by 10%-30% due to the removal of land interference.

[0109] Embodiment two: post-disaster building damage assessment (change detection)

[0110] This embodiment demonstrates the application of the system in disaster emergency response. When the ground guidance center receives a disaster assessment task, it immediately starts the multi-source information analysis module. This module integrates satellite images, geographic information system data, and disaster reports to generate targeted generalized dynamic guidance information. The ground guidance center uses the meta-learning model to generate a task-adaptive prompt vector for this task based on historical remote sensing images and change detection results, encodes information such as residential areas, roads, and factories into the task-adaptive prompt vector, and pre-fabricates a model early exit decision maker that terminates calculation early when a predetermined threshold is reached.

[0111] The actual results show that, based on the ResNet-18 model, in the simulated on-board computing environment (Jetson AGX Orin), for processing an image of 20000x20000, the on-board model can ignore large areas of non-human activity such as vegetation and water bodies, and focus on key areas of interest such as residential areas, effectively reducing the data calculation range. The early exit mechanism can exit when the accuracy reaches 85%. Without using the strategy of the present application, the on-board model depth calculation time is 127 minutes, and the data extraction accuracy is 87.1% after manual verification. After using the strategy of the present application, the on-board model performs focused calculation, and the calculation time is 6.5 minutes, and the data extraction accuracy is 85.5%, and the key area is completely covered. The method of the present application can improve the efficiency by about 20 times with a slight loss of accuracy.

[0112] Example three: natural resource feature classification

[0113] In this task, the verification of the system's adaptive optimization ability is also carried out. The ground guidance center generates generalized dynamic guidance information with time adaptability according to the growth cycle of vegetation. The on-board processing unit gradually optimizes the representation ability of crop features through continuous learning mechanism. The system uses collaborative reasoning instruction mechanism to decompose complex feature classification tasks into vegetation extraction, water body extraction, building extraction and other subtasks. Each subtask has corresponding processing priority and accuracy requirement. The on-board processing module executes according to the instruction sequence, and adjusts the processing strategy according to the real-time resource situation at each step. The results show that, under different task modes, using the collaborative reasoning instruction mechanism can improve the data processing efficiency by 3-15 times. The specific efficiency improvement varies with the complexity of the task and the proportion of the feature. When analyzing the water body with small proportion, the improvement can reach 15 times; when classifying the vegetation with large proportion, the improvement is about 3 times. Through the optimization mechanism of the system, the accuracy of the guidance information generation can be effectively improved.

[0114] Experimental conclusion:

[0115] Experiments show that the present application has significant advantages over the prior art. In terms of computing efficiency, the average computing efficiency can be improved by about one order of magnitude due to the use of focused processing strategy. In terms of processing accuracy, through the accurate guidance of the guidance information, the target false alarm rate can be reduced by 10-30%, and in the early exit mode, the accuracy loss of the key area is within 3%. In addition, due to the significant reduction of data transmission volume, the bandwidth pressure of satellite-ground communication is significantly relieved, and the average data transmission volume can be reduced by 95%.

[0116] It is particularly worth noting that the self-learning ability of the system enables it to continuously adapt to new task requirements. Long-term running data shows that after multiple optimization iterations, the system shows good generalization ability when processing new remote sensing tasks.

Claims

1. A method for satellite-ground collaborative intelligent remote sensing computing based on generalized dynamic guidance, characterized in that, The method comprises the following steps: Step 1: The ground guidance center processes and analyzes the stored historical multi-source information according to the remote sensing interpretation task on the satellite to generate generalized dynamic guidance information; Step 2: The ground guidance center transmits the generalized dynamic guidance information to the satellite-borne processing unit through the space-earth uplink; Step 3: The satellite-borne processing unit receives and analyzes the generalized dynamic guidance information, dynamically generates a focused computing mask graph and a model early retirement threshold according to the type of the generalized dynamic guidance information and the resource state of the satellite; Step 4: The satellite-borne processing unit carries out remote sensing intelligent computing according to the generalized dynamic guidance information, the focused computing mask graph and the model early retirement threshold to generate structured results; Step 5: The satellite-borne processing unit transmits the structured result data obtained by processing back to the ground guidance center through the space-earth downlink and uses it to optimize the generation of subsequent generalized dynamic guidance information. 2.The star-ground collaborative intelligent remote sensing computing method based on generalized dynamic guidance according to claim 1, wherein, The remote sensing interpretation task on the satellite in step 1 includes target detection, feature classification and change detection; the historical multi-source information in step 1 includes one or more of historical remote sensing images, remote sensing thematic maps, vector geographic information data and text data. 3.The star-ground collaborative intelligent remote sensing computing method based on generalized dynamic guidance according to claim 2, wherein, The generalized dynamic guidance information in step 1 is beyond the binary space mask and is a rich semantic information carrier, and specific forms include an intelligent space mask , a task adaptive prompt vector , and a collaborative reasoning instruction , and the generation method is as follows: The ground guidance center first analyzes the historical remote sensing images through the feature classification model, target detection model and change detection model to generate analysis results including target detection results, feature classification results and change detection results, saves the results and then statistically analyzes them according to regions. The statistical analysis content includes regional latitude and longitude, image time, target type and quantity, feature type and change occurrence number. And generate an intelligent space mask according to the following steps , task adaptive prompt vector and collaborative reasoning instructions ; a) Intelligent spatial mask For different remote sensing interpretation tasks, the weight of the spatial sub-region of the on-board image is calculated in the following manner : : Target detection task: wherein the target region refers to a region where the target can appear; Change detection task: Wherein the fixed area refers to the area that cannot change according to the historical remote sensing image rule; the possible change area refers to the case that there is a change for the historical remote sensing image analysis result, , wherein is the total number of images that have changed in the historical remote sensing image, is all historical remote sensing images; the must-check area refers to the area that must be analyzed this time. The ground object classification task is: Wherein the classification unchanged region refers to a region without classification change in historical remote sensing images, and the classification changed region refers to a region with classification change at different times in historical remote sensing images; a historical remote sensing image closest to the current task in time is selected to calculate the classification change probability with a time interval t , , is a time interval sequence in which the classification change occurs in historical remote sensing images, is a mean function; b) Task-adaptive prompt vector The generation is a dynamic process based on a meta-learning model, specifically including the following steps: b1) task semantic encoding: encode the current remote sensing interpretation task, data latitude and longitude range, and data acquisition time elements using a multi-layer perception model to map into a task embedding vector ; b2) historical data analysis: using the same multi-layer perception model as b1, the results of historical remote sensing image analysis are encoded to generate a historical task experience set wherein is the number of historical remote sensing image analysis, and a single analysis of a single historical remote sensing image is recorded as 1 time; b3) historical data fusion: compute task embedding vector with the set of historical task experiences and generate weights with attention mechanism , where s is the similarity function; and then weighted sum to get the context vector ; b4) meta prompt synthesis: the and are nonlinearly fused to output a task adaptive prompt vector ; wherein is implemented by a multi-layer perception. c) Cooperative inference instructions : For the multimodal interpretation model deployed on the satellite, generate a piece of structured text as prompt information, which is a set of structured machine-readable instructions in the format {operation type, target area, parameter}.

4. The generalized dynamic guidance-based star-ground collaborative intelligent remote sensing computing method according to claim 1, characterized in that, The generation process of the focused computing mask graph and the model early retirement threshold in step 3 is as follows: a) The generation process of the focused computing mask graph is as follows: Masking for intelligent space , according to the latitude and longitude range and spatial resolution of the on-board image, a focus calculation mask map of the to-be-processed region is generated ; the specific method is to cut out a region with the same latitude and longitude range as the on-board image from , and perform nearest neighbor interpolation on the cut-out region according to the resolution of the on-board image, that is, a focus calculation mask map is obtained ; Task-adaptive prompting vectors , first generate a spatial importance heat map by a light-weight convolutional network : where: denotes concatenation and broadcast of channel dimension, and is broadcasted to the same spatial dimension as the image I on the star, is a sigmoid function with value range [0, 1], is the element in the i-th row and j-th column of the matrix denotes the processing importance weight at position (i, j); Then, according to the satellite resource constraint, dynamically determine the region needing to be processed, and generate a focused calculation mask map of the region needing to be processed : wherein, determined in accordance with on-board computing resources, satellite energy status, and on-board image size synthesis; b) The generation process of the model early retirement threshold is as follows: b1) Early exit decision point setting: key layer setting of the intelligent interpretation model deployed on the on-board processing unit sets an early exit decision maker ; wherein the key layer is set according to requirements, and k is the number of layers of the key layer b2) for each early retirement decision maker dynamically generating a model early retirement threshold wherein is a task importance modulation function, is an empirical parameter, set according to the task situation, is a resource tightness modulation function, is the current occupation ratio of computing resources, is the energy resources already consumed, is the maximum energy allowed to be consumed by the satellite, and are proportional coefficients, set empirically.

5. The generalized dynamic guidance-based star-ground collaborative intelligent remote sensing computing method according to claim 4, characterized in that, The specific process of the satellite-borne processing unit carrying out remote sensing intelligent computing in step 4 is as follows: a) input is: smart space mask and focus computation mask map ; The intelligent interpretation model deployed on the spaceborne processing unit performs full-precision calculation on the selected regions and approximate calculation on the unselected regions in the manner of channel pruning or low-bit quantization during model forward propagation The selected regions are calculated with full precision; for the unselected regions, approximate calculation is performed in the manner of channel pruning or low-bit quantization. b) input is: task-adaptive prompt vector , focus computation mask map and model early exit threshold ; For the feature map F of the intelligent interpretation model deployed on the satellite processing unit, the following modulation is performed: wherein the scaling parameter and the bias parameter are generated by the task adaptive prompt vector and the adapter network , with a size of 1x1xNc, Nc being the number of channels, denotes channel-by-channel multiplication; wherein the adapter network is composed of a two-layer fully connected network, the input is the task adaptive prompt vector , the hidden layer uses a ReLU activation function to introduce nonlinearity, the dimension is 2*Nc, the first Nc values are taken as the scaling parameter through a sigmoid function, and the last Nc values are taken as the bias parameter ; The intelligent interpretation model deployed on the spaceborne processing unit performs full-precision calculation on the selected regions and approximate calculation on the unselected regions in the manner of channel pruning or low-bit quantization during model forward propagation The intelligent interpretation model deployed on the spaceborne processing unit performs full-precision calculation on the selected regions and approximate calculation on the unselected regions in the manner of channel pruning or low-bit quantization during model forward propagation When the on-board processing unit performs model inference, when is true, an early exit is triggered, wherein, is a prediction confidence, denotes the feature map of the th key layer, and denotes the concatenation and broadcast of channel dimensions. c) input is: a co-inference instruction ; Directing collaborative reasoning instructions As a prompt instruction to a multi-modal intelligent interpretation model deployed in a spaceborne processing unit. 6.A satellite-ground collaborative intelligent remote sensing computing system for implementing the method of any one of claims 1-5, characterized in that, The system comprises a ground guidance center, a satellite-borne processing unit and a space-earth bidirectional communication link; The ground guidance center comprises a multi-source information analysis module and a guidance information generation module, the satellite-borne processing unit comprises an instruction analysis and adaptation module, a focused computing module, a model early retirement decision module and an intelligent computing module, and the space-earth bidirectional communication link is used to support data communication between the ground guidance center and the satellite-borne processing unit; The multi-source information analysis module is used to analyze the target detection, feature classification and change detection of the historical remote sensing images, remote sensing thematic maps, vector geographic information data and text data and statistically analyze the results; The guiding information generation module is configured to generate generalized dynamic guiding information according to the results of the multi-source information analysis module and remote sensing interpretation tasks on the satellite, including intelligent spatial masks , task adaptive prompt vectors , and cooperative reasoning instructions , and transmit the generalized dynamic guiding information to the on-board processing unit through a space-ground bidirectional communication link. The instruction analysis and adaptation module is used to receive the generalized dynamic guidance information sent by the ground guidance center, analyze the vector content and determine the vector type; The focused computing module is used to generate a focused computing mask graph for guiding the computation according to the vector content of the generalized dynamic guidance information; The model early retirement decision module is used to generate a model early retirement threshold according to the vector content of the generalized dynamic guidance information; The intelligent computing module deploys an intelligent interpretation model and is used to carry out intelligent computing according to the generalized dynamic guidance information, the focused computing mask graph and the model early retirement threshold to generate structured results.

Images