Over-the-ground remote sensing task parameter generation method for on-satellite autonomous task planning
By constructing a satellite-to-ground remote sensing task knowledge graph and automatically generating task feature values using the map embedded representation algorithm, the problem of incomplete tasks in satellite autonomous task planning is solved, and the automation and accuracy of autonomous task planning is achieved.
Patent Information
- Application Number
- CN202510440626.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-09
- Publication Date
- 2025-08-15
AI Technical Summary
In the prior art, the autonomous planning of satellite ground remote sensing tasks is due to the increasing differences in satellite functions and diversified task requirements, which leads to high learning costs and long-term time-consuming for business personnel, and the autonomous task planning on satellites lacks effective algorithm support.
By constructing a knowledge graph of satellite ground remote sensing tasks, using the graph embedded representation algorithm to learn historical tasks, generate embedded representations of ground remote sensing tasks, and automatically generate task feature values through a classification prediction model, including load type, resolution and usage mode.
It realizes the automatic completion and accuracy of task elements in the autonomous task planning on the star, generates a reasonable and complete task structure, and reduces the time-consuming and learning costs of manual operations.
Smart Images
Figure CN120493000A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of remote sensing technology, relates to satellite earth remote sensing autonomous mission planning, and specifically relates to a method for generating earth remote sensing mission parameters for on-board autonomous mission planning. Background Art
[0002] Currently, satellite-based Earth remote sensing missions, such as visible light imaging, are generated by personnel individually selecting and configuring the relevant elements based on their imaging needs and accumulated experience. On the one hand, the increasing number of satellites, the increasing diversity of satellite functional characteristics, and the diversification of mission requirements have led to a significant increase in the learning and memorization costs for personnel, and the time required for manual operation to complete satellite imaging missions has increased. On the other hand, with the increase in onboard processing power, autonomous mission planning within large-scale constellations urgently requires corresponding algorithms to replace manual operation and generate the corresponding tasks on board.
[0003] The continuous accumulation of historical satellite missions has generated a wealth of comprehensive and accurate historical Earth remote sensing data. This data embodies the expert knowledge of operators, manifesting in the interrelationships between mission elements and closely tied to the scheduling of Earth remote sensing missions. This historical mission data provides the relevant knowledge for automatically generating reasonable and complete Earth remote sensing missions. Summary of the Invention
[0004] To overcome the incompleteness of mission elements automatically generated by satellites during onboard autonomous mission planning, the present invention aims to provide a method for generating parameters for Earth remote sensing missions for onboard autonomous mission planning. This method stores historical Earth remote sensing missions in the form of a knowledge graph. The knowledge graph stores knowledge in a network-like structure, enabling the construction of complex knowledge structures and possessing powerful knowledge storage, expression, and reasoning capabilities. The remote sensing missions and their elements are then characterized through graph-embedded representation. Finally, unknown element values are automatically generated through classification prediction, meeting the requirements for complete mission generation in onboard autonomous mission planning.
[0005] The technical solution of the present invention is:
[0006] A method for generating parameters of an Earth remote sensing mission for autonomous onboard mission planning, comprising the following steps:
[0007] 1) Construct a knowledge graph of satellite Earth remote sensing missions based on the selected batch of historical satellite Earth remote sensing missions; use entities in the historical satellite Earth remote sensing missions as nodes in the knowledge graph of satellite Earth remote sensing missions, where the entity types include Earth remote sensing mission, target, payload type, resolution, and usage mode; for entities of the Earth remote sensing mission type, use their mission number, mission name, mission start time, and mission end time as their attributes; and for entities of the target type, use their number, name, and background as their attributes;
[0008] 2) digitizing each node and its attributes in the satellite earth remote sensing mission knowledge graph to obtain an initial numerical representation of the corresponding node;
[0009] 3) Generate an embedded representation of the corresponding node based on the initial numerical representation of the node and the associated information of the node;
[0010] 4) Set up a corresponding classification prediction task for each missing information in the Earth remote sensing mission; for each classification prediction task, use the embedded representation of each node to train a corresponding prediction classifier to predict the missing information in the Earth remote sensing mission;
[0011] 5) Select a corresponding prediction classifier based on the missing information in the earth remote sensing task to be processed, and generate an embedded representation x of the earth remote sensing task to be processed through a graph convolutional model; then input the embedded representation x into the selected classifier to predict the missing information in the earth remote sensing task to be processed.
[0012] Furthermore, the one-hot encoding method is used to encode the node to obtain the numerical representation of the corresponding node, and the sine-cosine position encoding method is used to encode the attribute of the node to obtain the numerical representation corresponding to the attribute; the initial numerical representation of the corresponding node is generated according to the numerical representation of the node and the numerical representation of the node attribute.
[0013] Furthermore, the method of generating the embedded representation of each node is: the numerical representation z of node u is u 、The numerical representation z of the neighborhood node v of node u v , the weight α of the neighborhood node v, the numerical representation z of the attribute value b of node u b , the weight β of the attribute value b of node u is input into the graph convolution model to generate the embedded representation of node u
[0014] Furthermore, the graph convolutional model generates an embedded representation of node u The method is:
[0015] 1) According to the numerical representation z v Compute the comprehensive embedded representation n of the neighborhood nodes of node uu , according to the numerical representation of the attribute z b Compute attribute information comprehensive embedded representation p u ;
[0016] 2) According to the numerical representation z u , comprehensive embedded representation of neighborhood nodes u and attribute information integrated embedded representation p u Embedded representation of computing node u
[0017] 3) Unitize the embedded representation of node u
[0018] Furthermore, the trained graph convolutional model is used to generate the embedded representation of node u When training the graph convolutional model, the embedded representation z of the relevant nodes is reduced pi distance while expanding the embedded representation z of non-related nodes ni The distance is used as the objective function to train and optimize the graph convolution model.
[0019] Furthermore, the missing information in the earth remote sensing mission includes payload type, resolution, and usage mode; the prediction classifier includes a payload type prediction classifier, a resolution prediction classifier, and a usage mode prediction classifier.
[0020] Furthermore, the method for constructing a knowledge graph of satellite earth remote sensing missions is as follows: first, modeling the entities and relationships in the historical satellite earth remote sensing missions to obtain corresponding entity models and relationship models, and then constructing the satellite earth remote sensing mission knowledge graph based on the entity models and relationship models.
[0021] Furthermore, the triples in the satellite earth remote sensing mission knowledge graph include (target, subordinate, earth remote sensing mission), (resolution, subordinate, earth remote sensing mission), (usage mode, subordinate, earth remote sensing mission), and (payload type, subordinate, earth remote sensing mission).
[0022] A server, characterized in that it includes a memory and a processor, the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program includes instructions for executing the above method.
[0023] A computer-readable storage medium stores a computer program thereon, wherein the computer program implements the above method when executed by a processor.
[0024] The present invention first formalizes a graph representation of a large number of historical Earth remote sensing missions. It then uses a graph-embedded representation algorithm to learn the knowledge contained in these historical missions, obtaining representations of satellite Earth remote sensing missions and their element entities within the graph. Finally, the embedded representation results are combined with a classification model to automatically generate values for the element values. By automatically generating values for different elements of satellite Earth remote sensing missions based on graph-embedded representation, the present invention achieves a more rational task structure for autonomous onboard mission planning, resolving issues such as a lack of task attributes, incomplete content, and illogical task descriptions that are often generated automatically by autonomous onboard mission planning.
[0025] The advantages of the present invention are as follows:
[0026] (1) For autonomous missions on board satellites, the present invention can automatically complete unknown planned mission elements to ensure the accuracy and implementation of autonomous missions.
[0027] (2) The present invention relates to a knowledge graph structure based on mission planning data in combination with satellite mission planning element parameters, including entities, entity attributes and relationships.
[0028] (3) The present invention designs an entity embedded representation containing attribute features for the satellite planning mission element features, and uses the sine and cosine methods to encode the time type attributes.
[0029] (4) For satellite mission planning entities with known embedded representations, this patent uses different classifiers to predict different unknown factors, and realizes the prediction of multi-factor parameters while sharing the underlying embedded representation model. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 Automatically generate flow charts for Earth observation missions.
[0031] Figure 2 Flowchart of the method for generating unknown feature values based on knowledge graph embedded representation.
[0032] Figure 3 Schematic diagram of temporal attribute encoding. DETAILED DESCRIPTION
[0033] The present invention will be described in further detail below with reference to the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.
[0034] exist Figure 1In the automatic generation of Earth observation tasks shown in [1], the first step is to build a knowledge graph of satellite Earth remote sensing tasks based on batches of historical satellite Earth remote sensing tasks and complete the corresponding data preparation. Secondly, the embedded representation of the satellite Earth remote sensing tasks and the embedded representation of the task element entities in the knowledge graph are calculated, and the element values of the Earth remote sensing tasks are automatically generated based on the representations. The specific steps include the following:
[0035] (1) Constructing a knowledge graph of satellite earth remote sensing missions
[0036] To build a knowledge graph for satellite Earth remote sensing missions, the first step is to model the entities and relationships within the graph. Then, batches of satellite remote sensing missions are organized according to the entity and relationship model structures and stored in a graph database. This patent defines the entity and relationship models based on the common elements of Earth remote sensing missions, as shown in Tables 1 and 2.
[0037] The entity categories are earth remote sensing mission, observation target, resolution, payload type, usage mode, etc. The relationship category is subordination, which includes (target, subordination, earth remote sensing mission), (resolution, subordination, earth remote sensing mission), (usage mode, subordination, earth remote sensing mission), and (payload type, subordination, earth remote sensing mission) in the form of triples.
[0038] Table 1 Entity model information table
[0039]
[0040] Table 2 Relationship model information table
[0041] main body relation object Target subordinate Earth remote sensing mission Load Type subordinate Earth remote sensing mission Resolution subordinate Earth remote sensing mission Usage Mode subordinate Earth remote sensing mission
[0042] (2) Automatic generation of earth remote sensing mission element values based on knowledge graph embedded representation
[0043] In the satellite earth remote sensing mission map, the satellite shooting mission entity (i.e., earth remote sensing mission) is not only related to its associated entities, such as resolution, target, etc., but also has a relationship with the attributes of the mission, such as the start time and end time. In order to simultaneously integrate the information provided by the attributes into the entity representation, this patent introduces an embedded representation of the attribute information in the embedded calculation process of the model. At the same time, in order to explore the relationship between the autonomously generated factor values and the known factors, multiple classification prediction tasks of unknown factors are added. Therefore, the model can simultaneously learn the embedded representation of satellite knowledge graph entities and the prediction of unknown factor values of earth remote sensing missions, which not only makes the information contained in the entity representation more complete, but also further explores the relationship between the attributes and the associated entities, providing a basis for automatic task generation for autonomous mission planning on board.
[0044] The construction of the knowledge graph of the earth remote sensing mission provides a data basis for the embedded representation and prediction of unknown element values. Then, the initial numerical representation of the graph knowledge is obtained through the data digitization in the following step 1. Then, the embedded representation of the entity node is obtained through the graph convolution model in step 2. Finally, the prediction result of the unknown element value is obtained by combining the element classification prediction of the earth remote sensing mission representation in step 3. The specific process is as follows: Figure 2 shown.
[0045] Step 1: Digitalization of Earth Remote Sensing Mission Maps
[0046] 1) Entity one-hot encoding
[0047] The entities in the Earth remote sensing mission knowledge graph include enumerable nodes such as payload type and satellite usage mode, as well as attribute fields that do not have enumeration meaning, such as the start time and end time of the Earth remote sensing mission. The initial numerical representation of all entities is encoded using one-hot encoding, and the dimension of the vector is the same as the number of entities. The dimension of the one-hot encoding vector corresponding to each entity is N, and for entity e l Perform one-hot encoding, l is the entity index, and the generated vector is e l =[x1,x2,…,x j ,…,x N ], where when l=j, x j =1, the rest x j Both are 0.
[0048] 2) Attribute encoding
[0049] The start and end times of Earth remote sensing missions describe the time range of satellite imagery execution. These are non-enumerable time values, and the probability of different missions having the same start and end times is very small. Therefore, the one-hot encoding method used for quantizing entities is not suitable for quantizing these attribute values. This method uses sine and cosine positional encoding for the time attribute.
[0050] Sin-cosine positional encoding is for an input sequence of length L and requires the position information of the kth object in the sequence. The positional encoding is given by sine and cosine functions of different frequencies:
[0051]
[0052] Among them, k is the position of an object in the input sequence, and 0 ≤ k < L, d is the dimension of the vector space. P(k, j) is a position function used to map the position k in the input sequence to (k, j) in the position matrix. n is a hyperparameter scalar with a value of 10,000. i is the column index mapped to the column, 0 ≤ i < d / 2, and a single value of i is mapped to both the sine and cosine functions simultaneously. It can be seen from the expression that even positions correspond to the sine function and odd positions correspond to the cosine function.
[0053] The start time and end time of the ground remote sensing mission include year, month, day, hour, minute, and second, in the format of "yyyy-MM-dd HH:dd:ss". Since when encoding time information using the sine-cosine position encoding method, it is necessary to encode the time information of different dimensions separately and then splice them. Since the "year" information does not have variability in the batch information and often has the same year requirement, this method only encodes the month, day, hour, minute, and second, and the encoding dimension is 32 for all.
[0054] As Figure 3 shown, the month, day, hour, minute, and second are extracted from the time attribute. After passing through the sine-cosine encoder, vector encodings v1, v2, v3, v4, v5 with a dimension of 32 are formed respectively. The above 5 vector encodings are spliced to form the time attribute encoding V = [v1, v2, v3, v4, v5]. A satellite shooting requirement includes a start time attribute encoding V start and an end time attribute encoding V end . The time attribute initialization representation of this satellite shooting entity is the vector splicing value of the start time initialization representation (i.e., the start time attribute encoding V start ) and the end time attribute initialization representation (the end time attribute encoding V end ) as [V start , V end .
[0055] Step 2: Knowledge graph embedded representation
[0056] The nodes and node attributes in the knowledge graph have been converted into numerical representations in Step 1, and then the final node representations are obtained through two-layer graph convolution calculations in this step. In the graph convolution calculation, the current node representation calculation includes the representations of its neighboring nodes and attribute information. The specific calculation method is the graph convolution algorithm:
[0057] Input: The numerical representation z of node u u ; The numerical representation of the neighboring node v of node u The weight α of the neighboring node v; The numerical representation of the attribute value b of node u The weight β of the attribute value b of node u; where and They represent the set of neighborhood nodes of node u and the set of attribute values of node u respectively; γ(·) represents the pooling function, and this method adopts uniform pooling.
[0058] Output: New embedded representation of node u
[0059] Step 1: Calculate the comprehensive embedded representation n of the neighborhood nodes of node u respectively u and attribute information integrated embedded representation p u :
[0060] Step 2: Calculate the embedded representation of node u
[0061] Step 3: Unitize the embedded representation of node u
[0062] Among them, ReLu() is the activation function, W is the weight matrix, and CONCAT() is the concatenation operation for feature fusion.
[0063] The learning objective function of the embedded representation is to maximize the margin loss, that is, to reduce the embedded representation z of the relevant nodes pi distance while expanding the embedded representation z of non-related nodes ni Distance is the purpose, specifically expressed as:
[0064]
[0065] Step 3: Prediction of unknown factor values
[0066] The prediction model used in this patent is a classifier that takes as input an embedded representation vector of an Earth remote sensing mission. The mission vector to be predicted is represented by x, with dimension n. The classifier learning parameter weights are ω, with dimension m×n, and a bias b. Here, m is the number of classifier labels, and f is a nonlinear transformation function. The classifier output is f(ωx+b).
[0067] This patent predicts the unknown element values in the prediction task, such as load type, resolution, usage mode, etc. Because different elements involve different classification labels and different hidden layer information, three prediction classifiers are designed respectively.
[0068] (1) Load type prediction
[0069] The payload type is one of the requirements for satellite photography, which is specifically divided into optical payload, infrared payload and SAR payload. Therefore, the dimension m of the weight ω1 of the payload type prediction classifier f(ω1x+b1) is 3.
[0070] (2) Resolution prediction
[0071] Imaging satellites have varying resolutions, and the range of resolutions is discrete. Therefore, this patent segmented the range of resolutions into distinct resolution intervals, assigning fixed labels to each interval. The relationship between resolutions and labels is shown in Table 3. Therefore, the dimension m of the weight ω2 of the resolution prediction classifier f(ω2x+b2) is 7.
[0072] Table 3 Resolution range and label mapping table
[0073] Resolution range Label (0,1] 1 (1,3] 2 (3,5] 3 (5,10] 4 (10,30] 5 (30,) 6 none 0
[0074] (3) Usage pattern prediction
[0075] The value range of the usage pattern element is an enumerable string, such as stripe pattern, array pattern, etc. This patent classifies and predicts common usage patterns. Therefore, the dimension m of the weight ω2 of the usage pattern prediction classifier f(ω2x+b2) is 10.
[0076] The loss function of the classifier is cross entropy loss as shown in the formula:
[0077]
[0078] Among them, P is the actual category distribution, that is, the label, and Q is the probability distribution predicted by the model.
[0079] While specific embodiments of the present invention have been disclosed for illustrative purposes, intended to facilitate understanding and implementation of the present invention, those skilled in the art will appreciate that various substitutions, variations, and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the disclosure of the preferred embodiments, and the scope of protection claimed in the present invention shall be determined by the scope of the claims.
Claims
1. A method for generating parameters for Earth remote sensing missions for autonomous onboard mission planning, comprising the following steps: 1) Construct a knowledge graph of satellite earth remote sensing missions based on the selected batch of historical satellite earth remote sensing missions; The entities in the historical satellite earth remote sensing missions are used as nodes in the satellite earth remote sensing mission knowledge graph. The entity types include earth remote sensing mission, target, payload type, resolution and usage mode. For entities of earth remote sensing mission type, their mission number, mission name, mission start time and mission end time are used as their attributes. For entities of target type, their number, name and background are used as their attributes. 2) digitizing each node and its attributes in the satellite earth remote sensing mission knowledge graph to obtain an initial numerical representation of the corresponding node; 3) Generate an embedded representation of the corresponding node based on the initial numerical representation of the node and the associated information of the node; 4) Set up a corresponding classification prediction task for each missing information in the Earth remote sensing mission; for each classification prediction task, use the embedded representation of each node to train a corresponding prediction classifier to predict the missing information in the Earth remote sensing mission; 5) Select a corresponding prediction classifier based on the missing information in the earth remote sensing task to be processed, and generate an embedded representation x of the earth remote sensing task to be processed through a graph convolutional model; then input the embedded representation x into the selected classifier to predict the missing information in the earth remote sensing task to be processed.
2. The method according to claim 1, characterized in that The one-hot encoding method is used to encode the node to obtain the numerical representation of the corresponding node, and the sine-cosine position encoding method is used to encode the node attribute to obtain the numerical representation corresponding to the attribute; the initial numerical representation of the corresponding node is generated according to the numerical representation of the node and the numerical representation of the node attribute.
3. The method according to claim 2, characterized in that The method of generating the embedded representation of each node is: the numerical representation z of node u is u 、The numerical representation z of the neighborhood node v of node u v , the weight α of the neighborhood node v, the numerical representation z of the attribute value b of node u b , the weight β of the attribute value b of node u is input into the graph convolution model to generate the embedded representation of node u 4. The method according to claim 3, characterized in that The graph convolutional model generates an embedded representation of node u The method is: 1) According to the numerical representation z v Compute the comprehensive embedded representation n of the neighborhood nodes of node u u , according to the numerical representation of the attribute z b Compute attribute information comprehensive embedded representation p u ; 2) According to the numerical representation z u , comprehensive embedded representation of neighborhood nodes u and attribute information integrated embedded representation p u Embedded representation of computing node u 3) Unitize the embedded representation of node u 5. The method according to claim 3 or 4, characterized in that Generate the embedded representation z of node u using the trained graph convolutional model u NEW ; When training the graph convolutional model, the embedded representation z of the relevant nodes is reduced pi distance while expanding the embedded representation z of non-related nodes ni The distance is used as the objective function to train and optimize the graph convolution model.
6. The method according to claim 1, 2 or 3, characterized in that: The missing information in the earth remote sensing mission includes payload type, resolution, and usage mode; the prediction classifier includes a payload type prediction classifier, a resolution prediction classifier, and a usage mode prediction classifier.
7. The method according to claim 1, characterized in that The method for constructing a knowledge graph of satellite earth remote sensing missions is as follows: first, model the entities and relationships in the historical satellite earth remote sensing missions to obtain corresponding entity models and relationship models, and then construct the satellite earth remote sensing mission knowledge graph based on the entity models and relationship models.
8. The method according to claim 1, characterized in that The triples in the satellite earth remote sensing mission knowledge graph include (target, subordinate, earth remote sensing mission), (resolution, subordinate, earth remote sensing mission), (usage mode, subordinate, earth remote sensing mission), and (payload type, subordinate, earth remote sensing mission).
9. A server, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program comprises instructions for executing the method according to any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.