Deep learning point target denoising method for remote sensing task planning

Abstract

The application discloses a deep learning point target denoising method for remote sensing task planning and belongs to the technical field of mega remote sensing constellation task planning.The application is aimed at the problem that the existing method is slow in calculating the satellite imaging time window for a point target in mega remote sensing constellation task planning and is time-consuming.The application comprises the following steps: orbit recursion is carried out according to the initial coordinate position of each satellite to determine the eccentric anomaly and the true anomaly at a given time;the orbit six numbers and the longitude, latitude and height of the task target are expressed as coordinate vectors in the ECEF coordinate system to calculate the imaging side swing angle;the golden section search method is adopted to determine the imaging time window;the imaging time and the imaging side swing angle are compared to determine the mark L and obtain a training data set;the point task denoising neural network is trained by using the network training data set, and the cross-entropy loss function is used for training and fitting of the neural network to obtain the trained point task denoising neural network.The application is used for point target denoising in remote sensing task planning.

Description

Technical Field

[0001] This invention relates to a deep learning-based point target denoising method for remote sensing mission planning, belonging to the field of giant remote sensing constellation mission planning technology. Background Technology

[0002] Low Earth Orbit (LEO) mega-constellations refer to constellations of low-Earth orbit (LEO) satellites with hundreds or even thousands of satellite nodes. They possess significant advantages such as wide-area coverage and interoperability, making them a key project in space infrastructure development in recent years. Since Elon Musk proposed the Starlink constellation plan in 2015, the deployment of LEO mega-constellations has gradually been put on the agenda, and related launch missions have already begun. Classified by type, LEO mega-constellations can be divided into remote sensing, communication, and hybrid constellations. Among them, remote sensing constellations are the primary means of acquiring ground data and intelligence information, possessing significant strategic value.

[0003] Unlike traditional satellite constellations, mega-constellation systems are characterized by their sheer number, dynamic and ever-changing missions, and highly time-varying inter-satellite topology. Among them, the global coverage capability of remote sensing mega-constellations offers significant advantages in responding to international military shifts, sudden natural disasters, routine Earth observation, and early warning, tracking, and designation of highly sensitive military targets. Furthermore, their inter-satellite interconnectivity allows for direct connection with users to quickly obtain their needs and rapidly process and transmit critical operational information via network, which is of great strategic significance. Therefore, how to efficiently utilize mega-constellation resources to meet the requirements of remote sensing missions is a topic worthy of further research.

[0004] Current remote sensing mission planning methods are limited by the scale of satellites and the method of mission collection, thus generally addressing the planning problem of a small number of satellites and a small number of missions. Related research mainly focuses on mission planning for constellations and small clusters, and the proposed methods are more suitable for long-term, centralized planning. Related algorithms are maturing and have been applied in engineering practice. However, from the perspective of building and managing mega-constellations, the sheer number of satellites within a giant constellation presents challenges for satellite management and operation. Traditional methods work well for limited-scale satellite groups, but their limitations are obvious when dealing with mega-constellations, large-scale constellations, and sudden missions. They lack flexibility, have long planning times, and are difficult to handle short-term missions. If, after constellation construction, a management and planning approach similar to that of single satellites or satellite clusters is still adopted, it will greatly waste constellation resources, place enormous pressure on the ground planning system, and even cause mission response delays due to low planning efficiency. Therefore, it is necessary to research new methods tailored to the characteristics of mega-constellations and the needs of new applications to improve mission planning efficiency and maximize the constellation's observation potential.

[0005] In the mission planning problem of giant remote sensing constellations, it is necessary to calculate the visible time window for satellite imaging of the mission, and then use relevant algorithms to perform mission planning within the calculated time window. In traditional methods, calculating the time window often requires recursion through orbital dynamics to calculate the visible time and imaging angle of each point target mission relative to the satellite. Existing methods for planning giant remote sensing constellations have the following drawbacks:

[0006] 1) The calculation process for the imaging window will take up a significant amount of time in the entire task planning process;

[0007] 2) It consumes excessive computing and time resources;

[0008] 3) It reduced the efficiency of task planning. Summary of the Invention

[0009] To address the problem of slow speed and long processing time in calculating the satellite imaging time window for point targets using existing methods in the planning of large remote sensing constellations, this invention provides a deep learning-based point target denoising method for remote sensing mission planning.

[0010] This invention provides a deep learning-based point target denoising method for remote sensing task planning, comprising:

[0011] The orbit is recursively calculated based on the initial coordinates of each satellite to determine the eccentricity of each satellite at a given time; then the true anomaly of each satellite at a given time is calculated based on the eccentricity of each satellite at a given time.

[0012] The orbital six elements of each satellite at a given time are represented as satellite coordinate vector r in the ECEF fixed coordinate system. ECEF The longitude, latitude, and altitude of each mission objective are represented as mission coordinate vector T in the ECEF fixed coordinate system. ECEF The mission imaging side-swing angle of satellite i to mission target j at a given time is calculated.

[0013] The search start and end times for the mission target are set. The golden section search method is used to iteratively determine the imaging time node corresponding to the minimum mission imaging side swing angle. The imaging time window is determined by the imaging time node. Then, the actual imaging side swing angle of the mission target is calculated based on the imaging time node.

[0014] The imaging time window is compared with the given imaging time of the current mission target, and the actual imaging side tilt angle of the mission target is compared with the given imaging side tilt angle of the current mission target. If the imaging time window is within the given imaging time of the current mission target and the actual imaging side tilt angle of the mission target is within the given imaging side tilt angle of the current mission target, then the current satellite is visible to the current mission target, and L is marked as 1; otherwise, it is invisible, and L is marked as 0.

[0015] The satellite coordinate vector r ECEF Task coordinate vector T ECEF The given imaging time of the task target is used as the input data of the point task denoising neural network, and the label L is used as the output data to obtain the original dataset. Data augmentation is performed on the original dataset to obtain the network training dataset.

[0016] A point task denoising neural network was trained using a network training dataset, and the cross-entropy loss function was used to train and fit the neural network, resulting in a trained point task denoising neural network used for point target denoising in remote sensing task planning.

[0017] According to the deep learning-based point target denoising method for remote sensing mission planning of the present invention, the method for calculating the eccentric perigee angle of a satellite at a given time is as follows:

[0018] Let the semi-major axis of the orbit be a0, the eccentricity be e0, and the initial true anomaly angle be f. 00 Calculate the initial eccentricity near point angle as E. 00 :

[0019]

[0020] Then calculate the initial mean anterior angle M. 00 :

[0021] M 00 =E 00 -e0·sin(E 00 (2)

[0022] Then calculate the mean anterior angle M at a given time t. 0t for:

[0023]

[0024] In the formula, μ represents the gravitational constant multiplied by the mass of the Earth;

[0025] Based on the eccentricity e0 and the average near angle M at a given time t 0t By establishing the Kepler equations and solving them, the eccentricity angle E at a given time t can be obtained. 0t :

[0026] M 0t=E 0t -e0·sin(E 0t (4)

[0027] According to the deep learning point target denoising method for remote sensing mission planning of the present invention, by solving formula (4), the eccentric near point angle E at a given time t is obtained. 0t The method is as follows:

[0028] The Newton-Raphson method was used to solve formula (4);

[0029] Initialize and set intermediate variables x0, function value v1, and derivative value v2, with x0 as the eccentricity anomaly angle E. 0t Initial solution:

[0030] Set x0 = M 0t (5)

[0031]

[0032] Update x0 using the following formula, and the updated x0 will be represented as x1:

[0033]

[0034] Then, taking x1 as x0, we perform the next iteration using formulas (6) and (7) until the following convergence condition is met:

[0035]

[0036] In the formula, ∈2 is the error tolerance of the difference, and ∈3 is the error tolerance of the function value v1;

[0037] The final obtained x1 is taken as the eccentric approximation angle E at a given time t. 0t .

[0038] The deep learning-based point target denoising method for remote sensing mission planning according to the present invention is based on the eccentric near angle E at a given time t. 0t The method for obtaining the true anterior angle at a given time is as follows:

[0039]

[0040] In the formula f 0t Let be the true anterior angle at a given time t.

[0041] According to the deep learning-based point target denoising method for remote sensing mission planning of the present invention, the six orbital elements of each satellite at a given time include the semi-major axis a0, eccentricity e0, orbital inclination i0, perigee argument ω0, right ascension of the ascending node Ω0, and true anomaly f at a given time t. 0t ,

[0042] The six orbital elements of the satellite at a given time are expressed as satellite coordinate vector r in the ECEF fixed coordinate system. ECEF The method is as follows:

[0043] Set intermediate variable r m and θ:

[0044]

[0045] θ=ω0+f 0t (11)

[0046] Represent the six orbital elements of the satellite at a given time in the J2000 inertial coordinate system as r J2000 :

[0047]

[0048] The transformation relationship between the inertial J2000 coordinate system and the ECEF fixed coordinate system is T:

[0049]

[0050] In the formula, jd is the Julian day at a given time t, and J2000 represents the given Julian day of the inertial J2000 coordinate system;

[0051] Greenwich Mean Time θ is calculated from T. g :

[0052]

[0053] Thus, the transformation matrix R is obtained:

[0054]

[0055] Then the satellite coordinate vector r ECEF for:

[0056] r ECEF =R·r J2000 (16)

[0057] The deep learning-based point target denoising method for remote sensing mission planning according to the present invention, wherein the mission coordinate vector T ECEF for:

[0058]

[0059] In the formula, p is an intermediate variable:

[0060]

[0061] e is the first eccentricity of the Earth's ellipsoid. 2=0.00669; h is the altitude of the mission target, lon is the latitude of the mission target, and lat is the longitude of the mission target.

[0062] According to the deep learning-based point target denoising method for remote sensing mission planning of the present invention, the mission imaging side-swing angle of satellite numbered i for mission target numbered j at a given time is θ. i,j :

[0063]

[0064] According to the deep learning-based point target denoising method for remote sensing task planning of the present invention, the search start time of the task target is set as a0, the search end time is set as b0, and the method for determining the imaging time window is as follows:

[0065] The golden section search method is used to calculate the start time a1 and end time b1 after the segmentation:

[0066]

[0067] Formula (18) is used to calculate the mission imaging side angles θ1 and θ2 of the current satellite to the current mission target after segmentation, starting time a1 and ending time b1. If θ1 ≥ θ2, the search start time a0 remains unchanged, and the search end time b0 is updated. The recalculated starting time a1 and ending time b1 after segmentation are as follows:

[0068]

[0069] If θ1 < θ2, then the search end time b0 remains unchanged. The search start time a0 is updated, and the recalculated start time a1 and end time b1 after segmentation are as follows:

[0070]

[0071] Calculate the difference between a1 and b1. If the difference between a1 and b1 exceeds the preset time threshold, return to formula (19) to continue calculating the start time a1 and end time b1 after segmentation based on the updated search start time a0 and search end time b0, until the difference between a1 and b1 is within the preset time threshold. Then, a1 and b1 are taken as the imaging time node corresponding to the minimum side swing angle of the task imaging, and the time period formed by a1 and b1 is taken as the imaging time window.

[0072] The method for denoising point targets in deep learning for remote sensing task planning according to the present invention includes the following steps for obtaining the network training dataset:

[0073] The original input data corresponding to the output data labeled L as 1 in the original dataset is oversampled multiple times to obtain oversampled input data; the oversampled input data is then added to the original dataset to obtain the network training dataset.

[0074] According to the deep learning point target denoising method for remote sensing task planning of the present invention, the point task denoising neural network is a multi-layer neural network; the training and fitting process is performed using the tool function library TensorFlow or PyTorch.

[0075] The beneficial effects of the present invention are as follows: The method of the present invention first determines whether a random point target is visible to a satellite within a certain period of time through orbital dynamics recursion, thereby generating training set data. After certain processing, the training set data can be used for neural network training, and the trained neural network can be directly used for point target denoising.

[0076] The method of this invention optimizes the calculation speed of time windows in the planning of giant remote sensing constellations by denoising point tasks; and achieves fast denoising of point tasks through deep learning using neural networks.

[0077] The method of this invention, when applied to mission planning for giant remote sensing constellations, can pre-filter out obviously invisible point tasks, avoiding the need for recursive determination of visibility using orbital dynamics, thus saving significant time and improving computational efficiency. Employing the deep learning-based point task identification and denoising method of this invention can further accelerate denoising efficiency and improve the overall performance of mission planning. Attached Figure Description

[0078] Figure 1 This is a schematic diagram of the satellite nadir point trajectory and imaging time window in the deep learning point target denoising method for remote sensing mission planning described in this invention. Detailed Implementation

[0079] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0080] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0081] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0082] Specific Implementation Method 1: Combination Figure 1As shown, this invention provides a deep learning-based point target denoising method for remote sensing task planning, which mainly consists of the following parts:

[0083] 1) Generation of the original dataset;

[0084] 2) Process the original dataset to obtain the network training dataset;

[0085] 3) Training the point-to-point task denoising neural network;

[0086] Specifically, including,

[0087] The orbit is recursively calculated based on the initial coordinates of each satellite to determine the eccentricity of each satellite at a given time; then the true anomaly of each satellite at a given time is calculated based on the eccentricity of each satellite at a given time.

[0088] The orbital six elements of each satellite at a given time are represented as satellite coordinate vector r in the ECEF fixed coordinate system. ECEF The longitude, latitude, and altitude of each mission objective are represented as mission coordinate vector T in the ECEF fixed coordinate system. ECEF The mission imaging side-swing angle of satellite i to mission target j at a given time is calculated.

[0089] The search start and end times for the mission target are set. The golden section search method is used to iteratively determine the imaging time node corresponding to the minimum mission imaging side swing angle. The imaging time window is determined by the imaging time node. Then, the actual imaging side swing angle of the mission target is calculated based on the imaging time node.

[0090] The imaging time window is compared with the given imaging time of the current mission target, and the actual imaging side tilt angle of the mission target is compared with the given imaging side tilt angle of the current mission target. If the imaging time window is within the given imaging time of the current mission target and the actual imaging side tilt angle of the mission target is within the given imaging side tilt angle of the current mission target, then the current satellite is visible to the current mission target, and L is marked as 1; otherwise, it is invisible, and L is marked as 0.

[0091] The satellite coordinate vector r ECEF Task coordinate vector T ECEF The given imaging time of the task target is used as the input data of the point task denoising neural network, and the label L is used as the output data to obtain the original dataset. Data augmentation is performed on the original dataset to obtain the network training dataset.

[0092] A point task denoising neural network was trained using a network training dataset, and the cross-entropy loss function was used to train and fit the neural network, resulting in a trained point task denoising neural network used for point target denoising in remote sensing task planning.

[0093] In this embodiment, the original dataset is generated by orbital dynamics recursion. The orbital recursion determines whether random point targets are visible to the satellite within a certain period of time, thereby forming the original training set.

[0094] To determine the position of each satellite at a given moment, orbital recursion is required based on the satellite's initial coordinates. Here, the main focus is on determining visibility, so a more precise orbital recursion method is unnecessary; the faster Kepler orbital recursion method can be used instead.

[0095] Furthermore, the method for calculating the eccentricity of a satellite at a given time is as follows:

[0096] Let the semi-major axis of the orbit be a0, the eccentricity be e0, and the initial true anomaly angle be f. 00 Calculate the initial eccentricity near point angle as E. 00 :

[0097]

[0098] Then calculate the initial mean anterior angle M. 00 :

[0099] M 00 =E 00 -e0·sin(E 00 (2)

[0100] Then calculate the mean anterior angle M at a given time t. 0t for:

[0101]

[0102] In the formula, μ represents the gravitational constant multiplied by the mass of the Earth;

[0103] Based on the eccentricity e0 and the average near angle M at a given time t 0t By establishing the Kepler equations and solving them, the eccentricity angle E at a given time t can be obtained. 0t :

[0104] M 0t =E 0t -e0·sin(E 0t (4)

[0105] Solving equation (4) yields the eccentricity angle E at a given time t. 0t The method is as follows:

[0106] The Newton-Raphson method was used to solve formula (4);

[0107] First, initialize the intermediate variable x0, function value v1, and derivative value v2, and set x0 as the eccentricity anomaly angle E. 0t Initial solution:

[0108] Set x0 = M 0t (5)

[0109]

[0110] Update x0 using the following formula, and the updated x0 will be represented as x1:

[0111]

[0112] Then, taking x1 as x0, we perform the next iteration using formulas (6) and (7) until the following convergence condition is met:

[0113]

[0114] In the formula, ∈2 is the error tolerance of the difference, and ∈3 is the error tolerance of the function value v1;

[0115] The final obtained x1 is taken as the eccentric approximation angle E at a given time t. 0t .

[0116] Alternatively, the iteration process can end when the preset maximum number of iterations is reached.

[0117] Given the eccentricity angle E at time t 0t The method for obtaining the true anterior angle at a given time is as follows:

[0118]

[0119] In the formula f 0t Let be the true anterior angle at a given time t, and mod is the modulo operator.

[0120] Furthermore, the six orbital parameters for each satellite at a given time include the semi-major axis a0, eccentricity e0, orbital inclination i0, perigee argument ω0, right ascension of the ascending node Ω0, and true anomaly f at a given time t. 0t ,

[0121] The six orbital elements of the satellite at a given time are expressed as satellite coordinate vector r in the ECEF fixed coordinate system. ECEF The method is as follows:

[0122] Set intermediate variable r m and θ:

[0123]

[0124] θ=ω0+f 0t (11)

[0125] Represent the six orbital elements of the satellite at a given time in the J2000 inertial coordinate system as r J2000 :

[0126]

[0127] The transformation relationship between the inertial J2000 coordinate system and the ECEF fixed coordinate system is T:

[0128]

[0129] In the formula, jd is the Julian day at a given time t, and J2000 represents the given Julian day of the inertial J2000 coordinate system;

[0130] Greenwich Mean Time θ is calculated from T. g :

[0131]

[0132] Thus, the transformation matrix R is obtained:

[0133]

[0134] Then the satellite coordinate vector r ECEF for:

[0135] r ECEF =R·r J2000 (16)

[0136] In this embodiment, the task coordinate vector T ECEF for:

[0137]

[0138] In the formula, p is an intermediate variable:

[0139]

[0140] e is the first eccentricity of the Earth's ellipsoid. 2 =0.00669; h is the altitude of the mission target, lon is the latitude of the mission target, and lat is the longitude of the mission target.

[0141] The mission imaging side tilt angle of satellite i for mission target j at a given time is θ. i,j :

[0142]

[0143] The satellite considered in this embodiment is not an agile satellite; its maneuvers only consider the lateral tilt angle. When the satellite passes near the mission target, it will tilt laterally to scan and image. Therefore, the imaging time window can be considered when the imaging lateral tilt angle is at its minimum, i.e., when the normal to the satellite's nadir velocity crosses the mission target. Figure 1 As shown.

[0144] Furthermore, the golden section search method can be used to find the time point that minimizes the imaging side swing angle:

[0145] Based on the task scenario, the search start time for the task objective is set as a0, and the search end time is set as b0. The method for determining the imaging time window is as follows:

[0146] The golden section search method is used to calculate the start time a1 and end time b1 after the segmentation:

[0147]

[0148] Formula (18) is used to calculate the mission imaging side angles θ1 and θ2 of the current satellite to the current mission target after segmentation, starting time a1 and ending time b1. If θ1 ≥ θ2, the search start time a0 remains unchanged, and the search end time b0 is updated. The recalculated starting time a1 and ending time b1 after segmentation are as follows:

[0149]

[0150] If θ1 < θ2, then the search end time b0 remains unchanged. The search start time a0 is updated, and the recalculated start time a1 and end time b1 after segmentation are as follows:

[0151]

[0152] Calculate the difference between a1 and b1. If the difference between a1 and b1 exceeds the preset time threshold, return to formula (19) to continue iteratively calculating the start time a1 and end time b1 after segmentation based on the updated search start time a0 and search end time b0 until the difference between a1 and b1 is within the preset time threshold. Then, a1 and b1 are taken as the imaging time node corresponding to the minimum side swing angle of the task imaging, and the time period formed by a1 and b1 is taken as the imaging time window.

[0153] In this embodiment, the original dataset consists of multiple inputs and a single output. The inputs are the six orbital elements of the satellite, the latitude and longitude of the mission target, and a specified time period. The output is the visibility marker L of the mission target relative to the satellite within the specified time period.

[0154] For specific use cases, the input can be augmented with additional dimensions, such as "satellite type", "imaging constraint angle", and "recursion start time".

[0155] In practical applications, most point mission targets are invisible to remote sensing satellites, so the majority of the output labels L in the generated raw training data are 0.

[0156] Imbalanced training data cannot be directly used for neural network training and requires oversampling. In the vast majority of the training data, point mission targets are invisible to the satellites, with only a small number of point mission targets visible to the satellites, resulting in an output L of 1.

[0157] As an example, methods for obtaining a network training dataset include:

[0158] The original input data corresponding to the output data labeled L as 1 in the original dataset is oversampled multiple times to obtain oversampled input data. This oversampled input data is then added back to the original dataset to obtain the network training dataset. This reduces the imbalance of the training set samples, thereby improving the training effect of the neural network.

[0159] As an example, depending on the specific task scenario, the point task denoising neural network is constructed as a multi-layer neural network; the training and fitting process is performed using the TensorFlow or PyTorch utility libraries, and the specific neural network parameters and optimizer settings should be designed based on the training results. After multiple training adjustments, the loss curve and accuracy curve are finally converged, completing the training of the neural network.

[0160] Various types of neural networks can be used depending on the specific effects and convergence speed, rather than being limited to a single network type.

[0161] The method of this invention uses a deep learning neural network for denoising, which is highly efficient and greatly improves the accuracy of point target removal.

[0162] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A deep learning-based point target denoising method for remote sensing task planning, characterized in that... include, The orbit is recursively calculated based on the initial coordinates of each satellite to determine the eccentric perigee angle of each satellite at a given time. Then, the true anomaly angle of each satellite at a given time is calculated based on the eccentric anomaly angle of each satellite at a given time; The orbital six roots of each satellite at a given time are represented as satellite coordinate vectors in the ECEF fixed coordinate system. The longitude, latitude, and altitude of each mission objective are represented as mission coordinate vectors in the ECEF fixed coordinate system. The mission imaging side-swing angle of satellite i to mission target j at a given time is calculated. The search start and end times for the mission target are set. The golden section search method is used to iteratively determine the imaging time node corresponding to the minimum mission imaging side swing angle. The imaging time window is determined by the imaging time node. Then, the actual imaging side swing angle of the mission target is calculated based on the imaging time node. The imaging time window is compared with the given imaging time of the current mission target, and the actual imaging side tilt angle of the mission target is compared with the given imaging side tilt angle of the current mission target. If the imaging time window is within the given imaging time of the current mission target and the actual imaging side tilt angle of the mission target is within the given imaging side tilt angle of the current mission target, then the current satellite is visible to the current mission target, and L is marked as 1. Otherwise, it is invisible, and the marker L is set to 0; satellite coordinate vector Task coordinate vector The given imaging time of the task target is used as the input data of the point task denoising neural network, and the label L is used as the output data to obtain the original dataset. Data augmentation is performed on the original dataset to obtain the network training dataset. A point task denoising neural network was trained using a network training dataset, and the cross-entropy loss function was used to train and fit the neural network, resulting in a trained point task denoising neural network used for point target denoising in remote sensing task planning. Set the start time of the search for the task objective to be The search ended at [time]. The method for determining the imaging time window is as follows: The golden section search method is used to calculate the start time after the segmentation. and the end time after splitting : （19）， Calculate the start time after segmentation and the end time after splitting Current satellite's mission imaging side tilt angle for the current mission target and ,like Then the search begins at time . Unchanged, update the search end time. And recalculate the start time after the segmentation and the end time after splitting as follows: （20）， like The search ends at that time. Unchanged, update search start time And recalculate the start time after the segmentation and the end time after splitting as follows: （21）， Calculate the current and The difference, if and If the difference exceeds the preset time threshold, then return to the formula (19) based on the updated search start time. and search end time Continue calculating the start time after the segmentation. and the end time after splitting Until now and If the difference is within a preset time threshold, then... and This serves as the imaging time node corresponding to the minimum side swing angle of the mission imaging. and The time period in which the image is formed is used as the imaging time window.

2. The deep learning-based point target denoising method for remote sensing task planning according to claim 1, characterized in that, The method for calculating the eccentricity angle of a satellite at a given time is as follows: Set the semi-major axis of the track as The eccentricity is The initial true anterior angle is Calculate the initial eccentric near point angle as follows: : （1）， Calculate the initial mean anterior angle again. : （2）， Then calculate the mean anterior angle at a given time t. for: （3）， In the formula This represents the gravitational constant multiplied by the Earth's mass; According to eccentricity and the average anterior angle at a given time t By establishing the Kepler equations and solving them, the eccentricity angle at a given time t can be obtained. : （4）。 3. The deep learning-based point target denoising method for remote sensing task planning according to claim 2, characterized in that, Solving formula (4) yields the eccentric approximation angle at a given time t. The method is as follows: The Newton-Raphson method is used to solve formula (4); Initialize and set intermediate variables function value and derivative value ,Will As the eccentric near point angle Initial solution: set up (5), （6）， Update according to the following formula The updated Represented as : （7）， Then As Then, use formulas (6) and (7) for the next iteration calculation until the following convergence condition is met: （8）， In the formula This is the tolerance for the difference error. function value Error tolerance; The final result The eccentric approximation angle at a given time t .

4. The deep learning-based point target denoising method for remote sensing task planning according to claim 3, characterized in that, The eccentricity angle at a given time t The method for obtaining the true anterior angle at a given time is as follows: （9）， In the formula Let be the true anterior angle at a given time t.

5. The deep learning-based point target denoising method for remote sensing task planning according to claim 4, characterized in that, The orbital six-element number for each satellite at a given time includes the orbital semi-major axis. eccentricity rate Track inclination Perigeal argument Right ascension of ascending node The true anterior angle at a given time t , Represent the six orbital elements of the satellite at a given time as satellite coordinate vectors in the ECEF fixed coordinate system. The method is as follows: Setting intermediate variables and : （10）， （11）， Express the six orbital elements of the satellite at a given time in the J2000 inertial coordinate system as follows: : （12）， The transformation relationship between the inertial J2000 coordinate system and the ECEF fixed coordinate system is T: （13）， In the formula For a Julian day at a given time t, The given Julian day represents the inertial J2000 coordinate system; Greenwich Mean Time (GMT) calculated from T : （14）， Thus, the transformation matrix is ​​obtained. : （15）， Then satellite coordinate vector for: （16）。 6. The deep learning-based point target denoising method for remote sensing task planning according to claim 5, characterized in that, Task coordinate vector for: （17）， In the formula As an intermediate variable: ， The first eccentricity of the Earth's ellipsoid. ; For the height of the mission objective, The dimension of the mission objective. The longitude of the mission objective.

7. The deep learning-based point target denoising method for remote sensing task planning according to claim 6, characterized in that, The mission imaging lateral tilt angle of satellite i for mission target j at a given time is... : （18）。 8. The deep learning-based point target denoising method for remote sensing task planning according to claim 1, characterized in that, Methods for obtaining network training datasets include: The original input data corresponding to the output data labeled L as 1 in the original dataset is oversampled multiple times to obtain oversampled input data; the oversampled input data is then added to the original dataset to obtain the network training dataset.

9. The deep learning-based point target denoising method for remote sensing task planning according to claim 1, characterized in that, The point task denoising neural network is a multi-layer neural network; The training and fitting process is performed using the utility libraries TensorFlow or PyTorch.

Images