Optical remote sensing agile satellite imaging method and system facing dynamic cloud shielding

Abstract

The invention discloses an optical remote sensing agile satellite imaging method and system for dynamic cloud shielding, and relates to the technical field of satellite remote sensing, and the method comprises the steps: carrying out the rough prediction of dynamic cloud information, and carrying out the fine prediction; calculating an effective imaging time window of the agile satellite under the condition of cloud shielding; establishing a dynamic cloud shielding agile satellite imaging model; improving and optimizing a multi-objective evolutionary algorithm; and generating an optimal imaging scheme of the agile satellite under the dynamic cloud shielding condition. According to the method, coarse-to-fine grading prediction is carried out on the cloud information, the effective imaging time window of the agile satellite is calculated according to the cloud information, and the model prediction speed is increased; the constructed dynamic cloud shielding-oriented agile satellite imaging model gives consideration to the maximum imaging benefit and the minimum maneuvering consumption of the agile satellite; a multi-objective optimization algorithm optimization solution method is adopted, and optimal configuration of evolution parameters is carried out through a deep reinforcement learning algorithm, so that each generation of evolution has the maximum benefit, and the algorithm optimization efficiency is greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of satellite remote sensing technology, and in particular to an optical remote sensing agile satellite imaging method and system facing dynamic cloud occlusion. Background Art

[0002] Optical remote sensing satellite images are widely used in urban planning, land monitoring, agriculture, ecological environment assessment and other fields due to their advantages such as high resolution, wide coverage and easy acquisition. However, fog has a serious impact on satellite imaging, with the global average cloud cover being about 65%, resulting in a decrease in image quality. Currently, remote sensing image post-processing technology (such as deep learning and pixel substitution methods) is mainly used to remove cloud effects, but it is still difficult to fully restore the real image.

[0003] Although the problem of cloud obstruction on imaging has existed for many years, there are few studies on avoiding the influence of clouds before imaging. The reason is that satellite mobility is limited and the imaging mission planning space is narrow. Effectively predicting the dynamic information of clouds (such as shape, position and height) is the key to obtaining cloud-free images. However, cloud prediction is difficult and the prediction accuracy decays over time.

[0004] At present, satellite imaging mission planning mainly focuses on three ideas: one is to regard clouds as fixed polygons for occlusion calculation, but real-time cloud information is difficult to obtain; the second is to regard clouds as random events for probability estimation, lacking a specific description of the clouds; the third is based on cloud detection methods, using cloud detection sensors to identify clouds in advance, but this method relies on satellites with cloud detection capabilities and lacks universality. Summary of the invention

[0005] The object of the present invention is to provide an optical remote sensing agile satellite imaging method and system for dynamic cloud occlusion to solve the problems raised in the prior art.

[0006] To achieve the above object, the present invention provides the following technical solution: an optical remote sensing agile satellite imaging method for dynamic cloud occlusion, the method comprising the steps of:

[0007] S1, from rough prediction to fine prediction of dynamic cloud information;

[0008] S2, calculate the effective imaging time window of agile satellites under cloud cover;

[0009] S3, establish a dynamic cloud occlusion agile satellite imaging model;

[0010] S4. Improve and optimize the multi-objective evolutionary algorithm;

[0011] S5. Generate the best imaging solution for agile satellites under dynamic cloud cover.

[0012] According to the above scheme, the rough prediction utilizes the high temporal resolution characteristics of geostationary meteorological satellites to obtain the cloud information closest to each imaging time window, thereby achieving a rough prediction of the cloud information at the imaging time.

[0013] The detailed prediction uses geostationary meteorological satellite image data combined with graph neural network technology to make more accurate predictions of cloud information;

[0014] The rough prediction to fine prediction of dynamic cloud information utilizes the high temporal resolution characteristics of geostationary meteorological satellites to obtain the cloud information closest to each imaging time window, thereby realizing a rough prediction of cloud information at the imaging moment. Considering the changes in clouds from the rough prediction moment to the satellite imaging moment, a cloud prediction network model is constructed in combination with graph neural network technology to perform short-term forecasts on highly dynamic cloud information, thereby realizing a fine prediction of cloud shape, position and altitude at the satellite imaging moment.

[0015] According to the above scheme, the cloud prediction network model is an encoder-processor-decoder structure;

[0016] Input satellite image data into the cloud prediction network model;

[0017] By means of the encoder, the satellite image data is mapped into a graph space and converted into a topological graph of a specified size; each node of the topological graph represents an observation point;

[0018] The processor comprises a stack of M spatiotemporal blocks; each of the spatiotemporal blocks performs feature extraction and information transmission on a topological graph in a graph space;

[0019] The spatiotemporal block includes a dynamic correlation capture model and a spatiotemporal correlation capture model;

[0020] The dynamic correlation capture model includes a spatial attention model and a temporal attention model; the spatial attention model captures the dynamic correlation dependency characteristics between topological map nodes through a spatial attention mechanism, and the spatial attention model adjusts the relationship between topological map nodes according to the correlation of spatial information to capture the dynamic characteristics of cloud space changes; the temporal attention model constructs a temporal attention matrix, and based on the attention mechanism, captures the dynamic correlation of cloud information in the time dimension, helping the cloud prediction network model understand the changing rules of the cloud layer at different time points;

[0021] The spatiotemporal correlation capture model includes graph wavelet convolution and one-dimensional dilated causal convolution; the graph wavelet convolution is used to extract the spatial correlation features between graph nodes and capture the spatial dependency between different nodes in the graph structure; the one-dimensional dilated causal convolution extracts the temporal correlation features of each observation value in the time dimension, and the causal convolution ensures that information is only propagated through past time steps to maintain the causal relationship of the time series;

[0022] The decoder reconstructs cloud prediction features based on the features extracted by the processor and converts the spatiotemporal features into a refined prediction of highly dynamic cloud information; the refined prediction includes cloud shape, position and height.

[0023] According to the above scheme, step S2 includes:

[0024] S2-1. Calculate the ground coverage of the agile satellite based on the imaging time, position, speed, maneuverability and sensor parameters of the agile satellite;

[0025] S2-2. Based on the refined prediction results of cloud information, determine whether the imaging area of ​​the agile satellite is blocked by clouds at the current moment; if cloud blocking is formed, adjust the maneuvering state of the agile satellite based on the refined prediction results of cloud information, and recalculate the ground coverage of the agile satellite; if cloud blocking is not formed, convert the ground coverage of the agile satellite into a plane latitude and longitude coordinate system;

[0026] S2-3. Compare the ground target position with the ground coverage of the agile satellite to determine whether there is overlap; if there is overlap, record the agile satellite imaging time and determine whether the agile satellite imaging time exceeds the maximum imaging time limit; if there is no overlap, directly determine whether the agile satellite imaging time exceeds the maximum imaging time limit;

[0027] S2-4. If the imaging time does not exceed the maximum imaging time limit, return to S2-1; if the imaging time exceeds the maximum imaging time limit, output the agile satellite imaging sequence, that is, the effective imaging time window of the agile satellite under cloud cover.

[0028] This process is repeated until the imaging time limit is met, and finally the effective imaging time window of the agile satellite under cloud cover is obtained.

[0029] According to the above scheme, a dynamic cloud-blocking agile satellite imaging model is established, with maximum observation gain and minimum agile satellite maneuverability as the objective function, three-axis maneuvers of agile satellites and whether imaging occurs in the imaging window at each imaging time as decision variables, and agile satellite attitude conversion constraints and agile satellite attitude maneuver constraints as constraints.

[0030] According to the above solution, the decision variables include a first decision variable, a second decision variable, a third decision variable and a fourth decision variable;

[0031] The first decision variable is expressed as: K = {k1, k2, ..., k i ,...,k M};

[0032] Among them, K represents the first decision variable, that is, the imaging situation sequence of the agile satellite in the visible time window; i represents the visible time window index, i∈[1,M]; M represents the total number of visible time windows; k i It is represented by the imaging of the agile satellite in the i-th visible time window; if the agile satellite is imaged in the i-th visible time window, then k i = 1, if the agile satellite does not image in the ith visible time window, then k i =0;

[0033] The second decision variable is expressed as: X = {x1, x2, ..., x i ,...,x M};

[0034] Where X represents the second decision variable, that is, the pitch angle sequence of the agile satellite imaging in each time window; x i It is represented as the pitch angle of the agile satellite in the i-th visible time window; if the agile satellite pitches forward during imaging, then x i Take a positive value. If the agile satellite pitches backward during imaging, then x i Take negative values;

[0035] The third decision variable is expressed as: Y = {y1, y2, ..., y i ,...,y M};

[0036] Where Y represents the third decision variable, that is, the side swing angle sequence imaged by the agile satellite in each time window; y i It is represented as the side swing angle of the agile satellite in the i-th visible time window; if the agile satellite swings to the left of the flight direction during imaging, then y i Take a positive value. If the agile satellite swings to the right of the flight direction during imaging, then y i Take negative values;

[0037] The fourth decision variable is expressed as: Z = {z1, z2, ..., z i ,...,z M};

[0038] Wherein, Z represents the fourth decision variable, i.e., the yaw angle sequence of the agile satellite imaging in each time window; i It is represented as the yaw angle of the agile satellite in the i-th visible time window; if the yaw direction of the agile satellite during imaging is clockwise, then z i Take a positive value. If the yaw direction of the agile satellite during imaging is counterclockwise, then z i Take negative values;

[0039] x i,y i and z i It is represented as the three-axis maneuvering angle of the agile satellite in the i-th visible time window;

[0040] The objective function includes a first objective function and a second objective function;

[0041] The first objective function is expressed as follows:

[0042]

[0043] Among them, max f(K,X,Y,Z) is expressed as the first objective function, that is, maximizing the imaging area coverage to ensure the maximum imaging area coverage; S cov (K, X, Y, Z) represents the effective area coverage of all agile satellites involved in imaging; S obj It is expressed as the area of ​​the target region, i.e., the ground area that the agile satellite needs to observe;

[0044] The second objective function is expressed as follows:

[0045]

[0046] Where min g(X,Y,Z) is expressed as the second objective function, that is, minimizing the agile satellite maneuver to ensure that the agile satellite maneuver is minimized; |x i -x i-1 | represents the difference between two consecutive pitch angles of the agile satellite; |y i -y i-1 | represents the difference between two adjacent side swing angles of the agile satellite; |z i -z i-1 | represents the difference between two consecutive yaw angles of the agile satellite; |x i -x i-1 |、|y i -y i-1 | and |z i -z i-1 | represents the difference of the three-axis maneuvers of the agile satellite in two consecutive maneuvers. The smaller the difference, the smaller the maneuver of the agile satellite.

[0047] The constraint conditions include a first constraint condition and a second constraint condition;

[0048] The first constraint condition is the agile satellite attitude conversion constraint, and the expression is as follows:

[0049]

[0050] in, and They represent the time required for attitude conversion of the pitch axis, roll axis and yaw axis of the agile satellite respectively; T h It is expressed as the maximum attitude transition time requirement of agile satellite;

[0051] The first constraint condition is the agile satellite attitude maneuver constraint, and the expression is as follows:

[0052]

[0053] Among them, x max ,y max and z max They are respectively expressed as the maximum attitude maneuvering capabilities of the agile satellite's pitch axis, roll axis and yaw axis.

[0054] According to the above scheme, step S4 includes:

[0055] S4-1, initializing with the decision variables in the dynamic cloud occlusion agile satellite imaging model to generate an initial population;

[0056] S4-2, encoding the population state, wherein the population state includes various decision variables, the percentage of evolutionary generations, and the hypervolume value of each generation of the population;

[0057] S4-3, using the encoded population state as the input of the neural network, and outputting the evolution parameters of the multi-objective evolutionary algorithm through the neural network; the evolution parameters include parent selection parameters, real crossover parameters, binary crossover parameters, real mutation parameters and binary mutation parameters;

[0058] S4-4, executing evolutionary operations according to the evolutionary parameters output by the neural network to generate offspring populations;

[0059] S4-5. Calculate the evolution reward based on the super volume value;

[0060] S4-6, determine whether the maximum evolutionary generation has been reached; if the maximum evolutionary generation has been reached, calculate the cumulative reward and advantage estimate; if the maximum evolutionary generation has not been reached, return to step S4-2 and continue iterative evolution;

[0061] S4-7, use PPO algorithm to update evolutionary parameters;

[0062] S4-8, repeating steps S4-1 to S4-7 until an evolution parameter configuration with a maximum evolution benefit is obtained;

[0063] S4-9. Execute a multi-objective evolutionary algorithm according to the evolutionary parameter configuration with the maximum evolutionary benefit.

[0064] According to the above scheme, the neural network adopts an Actor-Critic framework, including a shared structure, an Actor output layer and a Critic output layer;

[0065] The group state encoding is input into the neural network. The group state encoding is P×G×C, where P represents the population size, G represents the number of genes of each individual, and C represents the number of features encoded in the state; C' represents the number of output channels after three pooling replication convolution submodules; C in and C out They represent the number of input and output data channels of the pooled replicated convolution submodule respectively; Represented as a concatenation operation;

[0066] The group state encoding performs maximum pooling in the population dimension and maximum pooling in the gene dimension respectively through the shared structure; the maximum pooling in the population dimension extracts the individual with the largest characteristic value of each channel in the population; the maximum pooling in the gene dimension extracts the element with the largest characteristic value in the gene dimension of each individual;

[0067] After completing the maximum pooling of the population dimension and the maximum pooling of the gene dimension, a copy operation is performed to obtain global features;

[0068] The global features and local features are spliced ​​together through a splicing operation, the number of data channels after splicing is 3 times the original number, and a 1×1 convolution kernel is used to perform a two-dimensional convolution operation to obtain shared hidden features; the number of neurons in each layer of the neural network is 64;

[0069] In the shared structure, after each two-dimensional convolution operation, the LeakyReLU function is used as the activation function of the activation layer;

[0070] The Actor output layer performs maximum pooling, copying and two-dimensional convolution operations on the shared hidden features again to obtain the vector value of the Actor;

[0071] The critic output layer performs gene-dimensional maximum pooling on the shared hidden features, and performs maximum pooling, copying operations, and two-dimensional convolution operations again, and obtains a scalar estimate of the critic by calculation and summation;

[0072] The vector value of the Actor and the scalar estimated value of the Critic are the evolution parameters of the multi-objective evolutionary algorithm output by the neural network.

[0073] According to the above scheme, the accumulated rewards are used to measure the total benefits generated by the current evolutionary parameters in future generations, helping to optimize the evolutionary results of multiple generations;

[0074] The advantage estimate is used to measure the rationality of selecting a set of evolutionary parameters under a certain population state, and affects the evolutionary selection at each decision-making stage.

[0075] According to the hypervolume value HV, the evolution reward is calculated as follows:

[0076] R t =HV t -HV t-1 ;

[0077] Among them, R t It is represented as the evolution reward of the tth generation, that is, the reward based on the hypervolume value in the evolution process of the tth generation; t represents the evolutionary generation; HV t It is expressed as the hypervolume value of the tth generation;

[0078] The cumulative reward formula is as follows:

[0079] V(s t )=R t+1 +γR t+2 +γ 2 R t+3 +…+γ T-t R T-t+1 ;

[0080] Among them, V(s t ) is represented as in state s t The cumulative reward under R t is the evolution reward of the tth generation calculated using the hypervolume; T is the total number of evolution generations; γ is the discount factor;

[0081] The advantage estimate is as follows:

[0082]

[0083] in, Expressed as an advantage estimate; δ t+i It is represented as the immediate advantage, i.e. the immediate advantage of the t+jth generation; λ is represented as the equilibrium parameter;

[0084] The instant advantage is as follows:

[0085]

[0086] in, Represented as a scalar estimate of the t-th generation Critic output using the neural network.

[0087] An optical remote sensing agile satellite imaging system for dynamic cloud occlusion, the system includes: a cloud information prediction module, an effective imaging time window calculation module, a model building module, a multi-objective evolutionary algorithm improvement module and a solution generation module;

[0088] The cloud information prediction module is used to achieve cloud information prediction from coarse to fine; coarse prediction uses data from geostationary orbit meteorological satellites to obtain approximate cloud information; fine prediction combines graph neural network technology to accurately predict the cloud layer based on high-resolution image data to obtain information on cloud shape, position and height at the time of imaging;

[0089] The effective imaging time window calculation module takes the maximum observation benefit and the minimum agile satellite maneuver as the objective function, the agile satellite three-axis maneuver and whether the imaging window of each imaging time is imaged as the decision variables, and the agile satellite attitude conversion constraint and the agile satellite attitude maneuver constraint as the constraint conditions to establish a dynamic cloud-blocking agile satellite imaging model;

[0090] The model building module takes the maximum observation benefit and the minimum satellite maneuver as the objective function, the satellite three-axis maneuver and whether the imaging window is imaged at each imaging time as the decision variables, and the satellite attitude conversion constraint and attitude maneuver constraint as the constraint conditions to establish a dynamic cloud occlusion agile satellite imaging model;

[0091] The improved module of the multi-objective evolutionary algorithm initializes decision variables and generates an initial population; encodes the population state and uses it as the input of the neural network; outputs the evolution parameters of the multi-objective evolutionary algorithm through the neural network; performs evolution operations, generates offspring populations, and calculates evolution rewards; and uses the PPO algorithm to update the evolution parameters until the evolution parameter configuration with the maximum evolution benefit is obtained;

[0092] The scheme generation module generates the best imaging scheme for agile satellites under dynamic cloud occlusion conditions based on the optimized multi-objective evolutionary algorithm.

[0093] Compared with the prior art, the present invention has the following beneficial effects:

[0094] 1. The present invention implements a hierarchical prediction of cloud information from coarse to fine, further improving the prediction accuracy of dynamic cloud information;

[0095] 2. The present invention calculates the effective imaging time window prediction of agile satellites based on cloud information, reduces the difficulty of model solution and improves the calculation speed;

[0096] 3. The multi-objective optimization algorithm optimization solution method adopted in the present invention performs optimal configuration of evolution parameters through a deep reinforcement learning algorithm, so that each generation of evolution has the maximum benefit, which greatly improves the algorithm optimization efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 A flowchart of the steps of the optical remote sensing agile satellite imaging method for dynamic cloud occlusion of the present invention;

[0098] Figure 2 It is a structural schematic diagram of a dynamic cloud occlusion agile satellite imaging model of an optical remote sensing agile satellite imaging method for dynamic cloud occlusion according to the present invention;

[0099] Figure 3 The flowchart of the steps of improving and optimizing the multi-objective evolutionary algorithm for the optical remote sensing agile satellite imaging method for dynamic cloud occlusion of the present invention;

[0100] Figure 4 The figure is a schematic diagram of the neural network structure of the optical remote sensing agile satellite imaging method for dynamic cloud occlusion of the present invention. DETAILED DESCRIPTION

[0101] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0102] Example: Figure 1-Figure 4 As shown, the present invention provides a technical solution, an optical remote sensing agile satellite imaging method for dynamic cloud occlusion, using one agile satellite and one piece of cloud information for one regional target;

[0103] Among them, the agile satellite parameters are shown in the following table:

[0104] Agile satellite parameters value Semi-major axis of track 6878.14 (km) Orbital eccentricity 0 Orbital inclination 97.4065(°) Ascending node right ascension 236.54(°) Argument of perigee 0(°) Mean anomaly 0(°) Maximum attitude maneuvering angle 60(°) Maximum attitude maneuvering angular velocity 6(° / s) Half field of view 1.6(°)

[0105] Regional information of regional targets: boundary point coordinates: (35.777,114.42); (35.943,114.42); (35.943,115.032); (35.777,115.032);

[0106] The cloud information is set to simulated image, the cloud shape is set to ellipse, the height range is 1-13km, and the size is 0.5-20km 2 Random variation within

[0107] First, the dynamic cloud information is predicted from coarse to fine;

[0108] The rough prediction utilizes the high temporal resolution characteristics of geostationary meteorological satellites to obtain the cloud information closest to each imaging time window, thereby achieving a rough prediction of the cloud information at the imaging time.

[0109] The detailed prediction uses geostationary meteorological satellite image data combined with graph neural network technology to make more accurate predictions of cloud information;

[0110] The rough prediction to fine prediction of dynamic cloud information utilizes the high temporal resolution characteristics of geostationary meteorological satellites to obtain the cloud information closest to each imaging time window, thereby realizing a rough prediction of cloud information at the imaging moment. Considering the changes in clouds from the rough prediction moment to the satellite imaging moment, a cloud prediction network model is constructed in combination with graph neural network technology to perform short-term forecasts on highly dynamic cloud information, thereby realizing a fine prediction of cloud shape, position and altitude at the satellite imaging moment.

[0111] The cloud prediction network model is an encoder-processor-decoder structure;

[0112] Input satellite image data into the cloud prediction network model;

[0113] By means of the encoder, the satellite image data is mapped into a graph space and converted into a topological graph of a specified size; each node of the topological graph represents an observation point;

[0114] The processor comprises a stack of M spatiotemporal blocks; each of the spatiotemporal blocks performs feature extraction and information transmission on a topological graph in a graph space;

[0115] The spatiotemporal block includes a dynamic correlation capture model and a spatiotemporal correlation capture model;

[0116] The dynamic correlation capture model includes a spatial attention model and a temporal attention model; the spatial attention model captures the dynamic correlation dependency characteristics between topological map nodes through a spatial attention mechanism, and the spatial attention model adjusts the relationship between topological map nodes according to the correlation of spatial information to capture the dynamic characteristics of cloud space changes; the temporal attention model constructs a temporal attention matrix, and based on the attention mechanism, captures the dynamic correlation of cloud information in the time dimension, helping the cloud prediction network model understand the changing rules of the cloud layer at different time points;

[0117] The spatiotemporal correlation capture model includes graph wavelet convolution and one-dimensional dilated causal convolution; the graph wavelet convolution is used to extract the spatial correlation features between graph nodes and capture the spatial dependency between different nodes in the graph structure; the one-dimensional dilated causal convolution extracts the temporal correlation features of each observation value in the time dimension, and the causal convolution ensures that information is only propagated through past time steps to maintain the causal relationship of the time series;

[0118] The decoder reconstructs cloud prediction features based on the features extracted by the processor and converts the spatiotemporal features into a refined prediction of highly dynamic cloud information; the refined prediction includes cloud shape, position and height.

[0119] Calculate the ground coverage of the agile satellite based on the imaging time, position, speed, maneuverability and sensor parameters of the agile satellite;

[0120] Based on the refined prediction results of cloud information, determine whether the imaging area of ​​the agile satellite is blocked by clouds at the current moment; if cloud blocking is formed, adjust the maneuvering state of the agile satellite based on the refined prediction results of cloud information, and recalculate the ground coverage of the agile satellite; if cloud blocking is not formed, convert the ground coverage of the agile satellite into a plane latitude and longitude coordinate system;

[0121] Compare the ground target position with the ground coverage of the agile satellite to determine whether there is overlap; if there is overlap, record the agile satellite imaging time and determine whether the agile satellite imaging time exceeds the maximum imaging time limit; if there is no overlap, directly determine whether the agile satellite imaging time exceeds the maximum imaging time limit;

[0122] If the imaging time does not exceed the maximum imaging time limit, the ground coverage of the agile satellite is recalculated; if the imaging time exceeds the maximum imaging time limit, the agile satellite imaging sequence is output, which is the effective imaging time window of the agile satellite under cloud cover;

[0123] This process is repeated until the imaging time limit is met, and finally the effective imaging time window of the agile satellite under cloud cover is obtained.

[0124] An agile satellite imaging model with dynamic cloud occlusion is established, with the maximum observation gain and the minimum agile satellite maneuver as the objective function, the three-axis maneuver of the agile satellite and whether the imaging window at each imaging time is imaged as the decision variables, and the agile satellite attitude conversion constraint and the agile satellite attitude maneuver constraint as the constraint conditions.

[0125] The decision variables include a first decision variable, a second decision variable, a third decision variable and a fourth decision variable;

[0126] The first decision variable is expressed as: K = {k1, k2, ..., k i ,...,k M};

[0127] Among them, K represents the first decision variable, that is, the imaging situation sequence of the agile satellite in the visible time window; i represents the visible time window index, i∈[1,M]; M represents the total number of visible time windows; k i It is represented by the imaging of the agile satellite in the i-th visible time window; if the agile satellite is imaged in the i-th visible time window, then k i = 1, if the agile satellite does not image in the ith visible time window, then k i =0;

[0128] The second decision variable is expressed as: X = {x1, x2, ..., xi ,...,x M};

[0129] Where X represents the second decision variable, that is, the pitch angle sequence of the agile satellite imaging in each time window; x i It is represented as the pitch angle of the agile satellite in the i-th visible time window; if the agile satellite pitches forward during imaging, then x i Take a positive value. If the agile satellite pitches backward during imaging, then x i Take negative values;

[0130] The third decision variable is expressed as: Y = {y1, y2, ..., y i ,...,y M};

[0131] Where Y represents the third decision variable, that is, the side swing angle sequence imaged by the agile satellite in each time window; y i It is represented as the side swing angle of the agile satellite in the i-th visible time window; if the agile satellite swings to the left of the flight direction during imaging, then y i Take a positive value. If the agile satellite swings to the right of the flight direction during imaging, then y i Take negative values;

[0132] The fourth decision variable is expressed as: Z = {z1, z2, ..., z i ,...,z M};

[0133] Wherein, Z represents the fourth decision variable, i.e., the yaw angle sequence of the agile satellite imaging in each time window; i It is represented as the yaw angle of the agile satellite in the i-th visible time window; if the yaw direction of the agile satellite during imaging is clockwise, then z i Take a positive value. If the yaw direction of the agile satellite during imaging is counterclockwise, then z i Take negative values;

[0134] x i ,y i and z i It is represented as the three-axis maneuvering angle of the agile satellite in the i-th visible time window;

[0135] The objective function includes a first objective function and a second objective function;

[0136] The first objective function is expressed as follows:

[0137]

[0138] Among them, max f(K,X,Y,Z) is expressed as the first objective function, that is, maximizing the imaging area coverage to ensure the maximum imaging area coverage; S cov (K, X, Y, Z) represents the effective area coverage of all agile satellites involved in imaging; S obj It is expressed as the area of ​​the target region, i.e., the ground area that the agile satellite needs to observe;

[0139] The second objective function is expressed as follows:

[0140]

[0141] Where min g(X,Y,Z) is expressed as the second objective function, that is, minimizing the agile satellite maneuver to ensure that the agile satellite maneuver is minimized; |x i -x i-1 | represents the difference between two consecutive pitch angles of the agile satellite; |y i -y i-1 | represents the difference between two adjacent side swing angles of the agile satellite; |z i -z i-1 | represents the difference between two consecutive yaw angles of the agile satellite; |x i -x i-1 |、|y i -y i-1 | and |z i -z i-1 | represents the difference of the three-axis maneuvers of the agile satellite in two consecutive maneuvers. The smaller the difference, the smaller the maneuver of the agile satellite.

[0142] The constraint conditions include a first constraint condition and a second constraint condition;

[0143] The first constraint condition is the agile satellite attitude conversion constraint, and the expression is as follows:

[0144]

[0145] in, and They represent the time required for attitude conversion of the pitch axis, roll axis and yaw axis of the agile satellite respectively; T h It is expressed as the maximum attitude transition time requirement of agile satellite;

[0146] The first constraint condition is the agile satellite attitude maneuver constraint, and the expression is as follows:

[0147]

[0148] Among them, x max ,y max and z maxThey are respectively expressed as the maximum attitude maneuvering capabilities of the agile satellite's pitch axis, roll axis and yaw axis.

[0149] Initializing with the decision variables in the dynamic cloud occlusion agile satellite imaging model to generate an initial population;

[0150] Encoding a population state, wherein the population state includes various decision variables, a percentage of evolutionary generations, and a hypervolume value of each generation of the population;

[0151] The encoded population state is used as the input of the neural network, and the evolution parameters of the multi-objective evolutionary algorithm are output through the neural network; the evolution parameters include parent selection parameters, real crossover parameters, binary crossover parameters, real mutation parameters and binary mutation parameters;

[0152] According to the evolutionary parameters output by the neural network, the evolutionary operation is performed to generate the offspring population;

[0153] Calculate the evolution reward based on the hypervolume value;

[0154] Determine whether the maximum evolutionary generation has been reached; if the maximum evolutionary generation has been reached, calculate the cumulative reward and advantage estimate; if the maximum evolutionary generation has not been reached, return to step S4-2 and continue iterative evolution;

[0155] Use PPO algorithm for evolutionary parameter update;

[0156] Repeat the above evolution process until the evolution parameter configuration with the maximum evolution benefit is obtained;

[0157] A multi-objective evolutionary algorithm is executed according to the evolutionary parameter configuration with the maximum evolutionary benefit.

[0158] Generate the best imaging solution for agile satellites under dynamic cloud occlusion conditions; the best imaging solution for agile satellites under dynamic cloud occlusion conditions is shown in the following table:

[0159]

[0160] The present invention provides another technical solution, an optical remote sensing agile satellite imaging system for dynamic cloud occlusion, the system comprising: a cloud information prediction module, an effective imaging time window calculation module, a model building module, a multi-objective evolutionary algorithm improvement module and a solution generation module;

[0161] The cloud information prediction module is used to achieve cloud information prediction from coarse to fine; coarse prediction uses data from geostationary orbit meteorological satellites to obtain approximate cloud information; fine prediction combines graph neural network technology to accurately predict the cloud layer based on high-resolution image data to obtain information on cloud shape, position and height at the time of imaging;

[0162] The effective imaging time window calculation module takes the maximum observation benefit and the minimum agile satellite maneuver as the objective function, the agile satellite three-axis maneuver and whether the imaging window of each imaging time is imaged as the decision variables, and the agile satellite attitude conversion constraint and the agile satellite attitude maneuver constraint as the constraint conditions to establish a dynamic cloud-blocking agile satellite imaging model;

[0163] The model building module takes the maximum observation benefit and the minimum satellite maneuver as the objective function, the satellite three-axis maneuver and whether the imaging window is imaged at each imaging time as the decision variables, and the satellite attitude conversion constraint and attitude maneuver constraint as the constraint conditions to establish a dynamic cloud occlusion agile satellite imaging model;

[0164] The improved module of the multi-objective evolutionary algorithm initializes decision variables and generates an initial population; encodes the population state and uses it as the input of the neural network; outputs the evolution parameters of the multi-objective evolutionary algorithm through the neural network; performs evolution operations, generates offspring populations, and calculates evolution rewards; and uses the PPO algorithm to update the evolution parameters until the evolution parameter configuration with the maximum evolution benefit is obtained;

[0165] The scheme generation module generates the best imaging scheme for agile satellites under dynamic cloud occlusion conditions based on the optimized multi-objective evolutionary algorithm.

[0166] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

Claims

1. An optical remote sensing agile satellite imaging method for dynamic cloud occlusion, characterized by: The method comprises the steps of: S1, from rough prediction to fine prediction of dynamic cloud information; S2, calculate the effective imaging time window of agile satellites under cloud cover; S3, establish a dynamic cloud occlusion agile satellite imaging model; S4. Improve and optimize the multi-objective evolutionary algorithm; S5. Generate the best imaging solution for agile satellites under dynamic cloud cover.

2. The optical remote sensing agile satellite imaging method for dynamic cloud occlusion according to claim 1 is characterized by: The rough prediction utilizes the high temporal resolution characteristics of geostationary meteorological satellites to obtain the cloud information closest to each imaging time window, thereby achieving a rough prediction of the cloud information at the imaging time. The detailed prediction uses geostationary meteorological satellite image data combined with graph neural network technology to make more accurate predictions of cloud information; The rough prediction to fine prediction of dynamic cloud information utilizes the high temporal resolution characteristics of geostationary meteorological satellites to obtain the cloud information closest to each imaging time window, thereby realizing a rough prediction of cloud information at the imaging moment. Considering the changes in clouds from the rough prediction moment to the satellite imaging moment, a cloud prediction network model is constructed in combination with graph neural network technology to perform short-term forecasts on highly dynamic cloud information, thereby realizing a fine prediction of cloud shape, position and altitude at the satellite imaging moment.

3. The optical remote sensing agile satellite imaging method for dynamic cloud occlusion according to claim 1 is characterized by: The cloud prediction network model is an encoder-processor-decoder structure; Input satellite image data into the cloud prediction network model; By means of the encoder, the satellite image data is mapped into a graph space and converted into a topological graph of a specified size; each node of the topological graph represents an observation point; The processor comprises a stack of M spatiotemporal blocks; each of the spatiotemporal blocks performs feature extraction and information transmission on a topological graph in a graph space; The spatiotemporal block includes a dynamic correlation capture model and a spatiotemporal correlation capture model; The dynamic correlation capture model includes a spatial attention model and a temporal attention model; the spatial attention model captures the dynamic correlation dependency characteristics between topological map nodes through a spatial attention mechanism, and the spatial attention model adjusts the relationship between topological map nodes according to the correlation of spatial information to capture the dynamic characteristics of cloud space changes; The temporal attention model constructs a temporal attention matrix, and based on the attention mechanism, captures the dynamic correlation of cloud information in the time dimension, helping the cloud prediction network model to understand the changing rules of the cloud layer at different time points; The spatiotemporal correlation capture model includes graph wavelet convolution and one-dimensional dilated causal convolution; the graph wavelet convolution is used to extract the spatial correlation features between graph nodes and capture the spatial dependency between different nodes in the graph structure; The one-dimensional dilated causal convolution extracts the time correlation characteristics of each observation value in the time dimension. The causal convolution ensures that information is only propagated through the past time steps, maintaining the causal relationship of the time series; The decoder reconstructs cloud prediction features based on the features extracted by the processor and converts the spatiotemporal features into a refined prediction of highly dynamic cloud information; the refined prediction includes cloud shape, position and height.

4. The optical remote sensing agile satellite imaging method for dynamic cloud occlusion according to claim 1, characterized in that: Step S2 includes: S2-1. Calculate the ground coverage of the agile satellite based on the imaging time, position, speed, maneuverability and sensor parameters of the agile satellite; S2-2. Based on the refined prediction results of cloud information, determine whether the imaging area of ​​the agile satellite is blocked by clouds at the current moment; if cloud blocking is formed, adjust the maneuvering state of the agile satellite based on the refined prediction results of cloud information, and recalculate the ground coverage of the agile satellite; if cloud blocking is not formed, convert the ground coverage of the agile satellite into a plane latitude and longitude coordinate system; S2-3. Compare the ground target position with the ground coverage of the agile satellite to determine whether there is overlap; if there is overlap, record the agile satellite imaging time and determine whether the agile satellite imaging time exceeds the maximum imaging time limit; if there is no overlap, directly determine whether the agile satellite imaging time exceeds the maximum imaging time limit; S2-4. If the imaging time does not exceed the maximum imaging time limit, return to S2-1; if the imaging time exceeds the maximum imaging time limit, output the agile satellite imaging sequence, that is, the effective imaging time window of the agile satellite under cloud cover.

5. The optical remote sensing agile satellite imaging method for dynamic cloud occlusion according to claim 1, characterized in that: The dynamic cloud-blocking agile satellite imaging model is established, with maximum observation benefit and minimum agile satellite maneuver as objective functions, three-axis maneuvers of agile satellites and whether imaging occurs in an imaging window at each imaging time as decision variables, and agile satellite attitude conversion constraints and agile satellite attitude maneuver constraints as constraint conditions, to establish a dynamic cloud-blocking agile satellite imaging model.

6. The optical remote sensing agile satellite imaging method for dynamic cloud occlusion according to claim 5 is characterized by: The decision variables include a first decision variable, a second decision variable, a third decision variable and a fourth decision variable; The first decision variable is expressed as: K = {k1, k2, ..., k i ,...,k M }; Among them, K represents the first decision variable, that is, the imaging situation sequence of the agile satellite in the visible time window; i represents the visible time window index, i∈[1,M]; M represents the total number of visible time windows; k i It is represented by the imaging of the agile satellite in the i-th visible time window; if the agile satellite is imaged in the i-th visible time window, then k i = 1, if the agile satellite does not image in the ith visible time window, then k i =0; The second decision variable is expressed as: X = {x1, x2, ..., x i ,...,x M }; Where X represents the second decision variable, that is, the pitch angle sequence of the agile satellite imaging in each time window; x i It is represented as the pitch angle of the agile satellite in the i-th visible time window; if the agile satellite pitches forward during imaging, then x i Take a positive value. If the agile satellite pitches backward during imaging, then x i Take negative values; The third decision variable is expressed as: Y = {y1, y2, ..., y i ,...,y M }; Where Y represents the third decision variable, that is, the side swing angle sequence imaged by the agile satellite in each time window; y i It is represented as the side swing angle of the agile satellite in the i-th visible time window; if the agile satellite swings to the left of the flight direction during imaging, then y i Take a positive value. If the agile satellite swings to the right of the flight direction during imaging, then y i Take negative values; The fourth decision variable is expressed as: Z = {z1, z2, ..., z i ,...,z M }; Wherein, Z represents the fourth decision variable, i.e., the yaw angle sequence of the agile satellite imaging in each time window; i It is represented as the yaw angle of the agile satellite in the i-th visible time window; if the yaw direction of the agile satellite during imaging is clockwise, then z i Take a positive value. If the yaw direction of the agile satellite during imaging is counterclockwise, then z i Take negative values; x i ,y i and z i It is represented as the three-axis maneuvering angle of the agile satellite in the i-th visible time window; The objective function includes a first objective function and a second objective function; The first objective function is expressed as follows: Among them, max f(K,X,Y,Z) is expressed as the first objective function, that is, maximizing the imaging area coverage to ensure the maximum imaging area coverage; S cov (K, X, Y, Z) represents the effective area coverage of all agile satellites involved in imaging; S obj It is expressed as the area of ​​the target region, i.e., the ground area that the agile satellite needs to observe; The second objective function is expressed as follows: Where min g(X,Y,Z) is expressed as the second objective function, that is, minimizing the agile satellite maneuver to ensure that the agile satellite maneuver is minimized; |x i -x i-1 | represents the difference between two consecutive pitch angles of the agile satellite; |y i -y i-1 | represents the difference between two adjacent side swing angles of the agile satellite; |z i -z i-1 | represents the difference between two consecutive yaw angles of the agile satellite; |x i -x i-1 |、|y i -y i-1 | and |z i -z i-1 | represents the difference of the three-axis maneuvers of the agile satellite in two consecutive maneuvers. The smaller the difference, the smaller the maneuver of the agile satellite. The constraint conditions include a first constraint condition and a second constraint condition; The first constraint condition is the agile satellite attitude conversion constraint, and the expression is as follows: in, and They represent the time required for attitude conversion of the pitch axis, roll axis and yaw axis of the agile satellite respectively; T h It is expressed as the maximum attitude transition time requirement of agile satellite; The first constraint condition is the agile satellite attitude maneuver constraint, and the expression is as follows: Among them, x max ,y max and z max They are respectively expressed as the maximum attitude maneuvering capabilities of the agile satellite's pitch axis, roll axis and yaw axis.

7. The optical remote sensing agile satellite imaging method for dynamic cloud occlusion according to claim 1, characterized in that: Step S4 includes: S4-1, initializing with the decision variables in the dynamic cloud occlusion agile satellite imaging model to generate an initial population; S4-2, encoding the population state, wherein the population state includes various decision variables, the percentage of evolutionary generations, and the hypervolume value of each generation of the population; S4-3, using the encoded population state as the input of the neural network, and outputting the evolution parameters of the multi-objective evolutionary algorithm through the neural network; the evolution parameters include parent selection parameters, real crossover parameters, binary crossover parameters, real mutation parameters and binary mutation parameters; S4-4, executing evolutionary operations according to the evolutionary parameters output by the neural network to generate offspring populations; S4-5. Calculate the evolution reward based on the super volume value; S4-6, determine whether the maximum evolutionary generation has been reached; if the maximum evolutionary generation has been reached, calculate the cumulative reward and advantage estimate; if the maximum evolutionary generation has not been reached, return to step S4-2 and continue iterative evolution; S4-7, use PPO algorithm to update evolutionary parameters; S4-8, repeating steps S4-1 to S4-7 until an evolution parameter configuration with a maximum evolution benefit is obtained; S4-9. Execute a multi-objective evolutionary algorithm according to the evolutionary parameter configuration with the maximum evolutionary benefit.

8. The optical remote sensing agile satellite imaging method for dynamic cloud occlusion according to claim 7, characterized in that: The neural network adopts an Actor-Critic framework, including a shared structure, an Actor output layer and a Critic output layer; Inputting the group state encoding into the neural network, the group state encoding is P×G×C, where P represents the population size, G represents the number of genes of each individual, and C represents the number of features encoded in the state; The group state encoding performs maximum pooling in the population dimension and maximum pooling in the gene dimension respectively through the shared structure; the maximum pooling in the population dimension extracts the individual with the largest characteristic value of each channel in the population; the maximum pooling in the gene dimension extracts the element with the largest characteristic value in the gene dimension of each individual; After completing the maximum pooling of the population dimension and the maximum pooling of the gene dimension, a copy operation is performed to obtain global features; The global features and the local features are spliced ​​by a splicing operation, and a two-dimensional convolution operation is performed using a convolution kernel to obtain a shared hidden feature; In the shared structure, after each two-dimensional convolution operation, the LeakyReLU function is used as the activation function of the activation layer; The Actor output layer performs maximum pooling, copying and two-dimensional convolution operations on the shared hidden features again to obtain the vector value of the Actor; The critic output layer performs gene-dimensional maximum pooling on the shared hidden features, and performs maximum pooling, copying operations, and two-dimensional convolution operations again, and obtains a scalar estimate of the critic by calculation and summation; The vector value of the Actor and the scalar estimated value of the Critic are the evolution parameters of the multi-objective evolutionary algorithm output by the neural network.

9. The optical remote sensing agile satellite imaging method and system for dynamic cloud occlusion according to claim 7, characterized in that: The accumulated reward is used to measure the total benefits generated by the current evolutionary parameters in future generations, helping to optimize the evolutionary results of multiple generations; The advantage estimate is used to measure the rationality of selecting a set of evolutionary parameters under a certain population state, and affects the evolutionary selection at each decision-making stage.

10. An optical remote sensing agile satellite imaging system for dynamic cloud occlusion, applied to an optical remote sensing agile satellite imaging method for dynamic cloud occlusion as claimed in any one of claims 1 to 9, characterized in that: The system includes: cloud information prediction module, effective imaging time window calculation module, model building module, multi-objective evolutionary algorithm improvement module and solution generation module; The cloud information prediction module is used to achieve cloud information prediction from coarse to fine; coarse prediction uses data from geostationary orbit meteorological satellites to obtain approximate cloud information; fine prediction combines graph neural network technology to accurately predict the cloud layer based on high-resolution image data to obtain information on cloud shape, position and height at the time of imaging; The effective imaging time window calculation module takes the maximum observation benefit and the minimum agile satellite maneuver as the objective function, the agile satellite three-axis maneuver and whether the imaging window of each imaging time is imaged as the decision variables, and the agile satellite attitude conversion constraint and the agile satellite attitude maneuver constraint as the constraint conditions to establish a dynamic cloud-blocking agile satellite imaging model; The model building module takes the maximum observation benefit and the minimum satellite maneuver as the objective function, the satellite three-axis maneuver and whether the imaging window is imaged at each imaging time as the decision variables, and the satellite attitude conversion constraint and attitude maneuver constraint as the constraint conditions to establish a dynamic cloud occlusion agile satellite imaging model; The improved module of the multi-objective evolutionary algorithm initializes decision variables and generates an initial population; encodes the population state and uses it as the input of the neural network; outputs the evolution parameters of the multi-objective evolutionary algorithm through the neural network; performs evolution operations, generates offspring populations, and calculates evolution rewards; and uses the PPO algorithm to update the evolution parameters until the evolution parameter configuration with the maximum evolution benefit is obtained; The scheme generation module generates the best imaging scheme for agile satellites under dynamic cloud occlusion conditions based on the optimized multi-objective evolutionary algorithm.

Images