Deep neural network collaborative reasoning method and system for low earth orbit satellite cluster

Abstract

The invention discloses a deep neural network collaborative reasoning method and system for a low earth orbit satellite cluster. The method comprises the following steps: acquiring a remote sensing image; resource state parameter information of all satellites is collected, the satellites with the resource states meeting the preset threshold condition serve as slave satellites, and parameter information of the deep neural network model configured on the slave satellites is collected; constructing a collaborative reasoning load distribution optimization model; solving the collaborative reasoning load distribution optimization model to obtain load size information corresponding to the collaborative satellite; and segmenting the remote sensing image according to the load size information corresponding to the collaborative satellites to obtain remote sensing image block data corresponding to each collaborative satellite, enabling the collaborative satellites to call the deep neural network model to perform reasoning on the remote sensing image block data, and aggregating reasoning results of all the remote sensing image block data to obtain a reasoning result of the remote sensing image. According to the invention, satellite resources can be integrated, inter-satellite cooperative reasoning is realized, and satellite processing of remote sensing image data is realized.

Description

Technical Field

[0001] The present invention relates to the field of remote sensing data processing technology, and in particular to a deep neural network collaborative reasoning method and system for low-orbit satellite clusters. Background Art

[0002] As an important method for obtaining surface data, satellite remote sensing has the characteristics of no regional restrictions, wide coverage, long observation period, and strong data objectivity. It is widely used in agriculture, meteorological observation, military and national defense. The rapid reasoning and recognition of key information in remote sensing images is the basis of the above applications. With the development of hardware equipment, satellites generally have certain on-orbit processing capabilities, but due to size and cost constraints, satellite computing, storage, network and other resources are limited, and it is still difficult to meet the processing needs of on-orbit tasks.

[0003] At present, the following two methods are mainly used to infer and identify remote sensing image data: The first is the reasoning method based on the ground cloud computing center, which is also the commonly used satellite on-orbit data processing method. Specifically, since the ground cloud computing center has powerful computing power and storage resources, the satellite's remote sensing image data is transmitted to the ground cloud computing center through the satellite-to-ground link, and the remote sensing image data is directly processed by the ground cloud computing center. The second is a reasoning method based on satellite-ground-cloud-edge collaboration, which is a collaborative processing method using on-board computing resources, ground cloud computing center resources, and communication links between satellites and ground. Specifically, part of the data is processed using on-board resources, and the remaining data is transmitted to the ground cloud computing center through the satellite-to-ground link for processing.

[0004] However, since the amount of remote sensing image data is usually large, if all or most of the data is transmitted to the ground cloud computing center, it will occupy too much bandwidth resources and waste satellite-to-ground communication resources; and since the satellite can only establish a satellite-to-ground link with the ground cloud computing center when it is in transit, the transmission delay between the satellite and the ground is large, and the satellite-to-ground link is unstable, which is not conducive to the execution of time-sensitive and real-time tasks. Summary of the invention

[0005] In order to solve some or all of the technical problems existing in the above-mentioned prior art, the present invention provides a deep neural network collaborative reasoning method and system for low-orbit satellite clusters.

[0006] The technical solution of the present invention is as follows:

[0007] In a first aspect, a deep neural network collaborative reasoning method for a low-orbit satellite cluster is provided, the method comprising:

[0008] Acquisition of remote sensing images;

[0009] The master satellite collects resource status parameter information of each satellite in the low-orbit satellite cluster, takes other satellites whose resource status meets the preset threshold condition as slave satellites, and collects parameter information of the deep neural network model configured on each slave satellite, wherein the master satellite represents a low-orbit satellite that obtains remote sensing images;

[0010] The master satellite constructs a collaborative reasoning load distribution optimization model according to remote sensing image size information, resource status parameters of the collaborative satellite, and parameter information of a deep neural network model configured on the collaborative satellite, wherein the collaborative satellite includes the master satellite and the slave satellite;

[0011] The main satellite solves the collaborative reasoning load distribution optimization model to obtain the load size information corresponding to each collaborative satellite;

[0012] The main satellite segments the remote sensing image according to the payload size information corresponding to the collaborative satellite, obtains the remote sensing image block data corresponding to each collaborative satellite, and sends the remote sensing image block data to the corresponding collaborative satellites respectively so that the collaborative satellites call the deep neural network model thereon to infer the remote sensing image block data and send the inference results to the main satellite, and the main satellite aggregates the inference results of all the remote sensing image block data to obtain the inference results of the remote sensing image.

[0013] In some optional implementations, the resource status parameter information of the satellite includes: CPU cycle number and frequency, memory space, data computing power, data transmission power, data transmission rate, and resource status parameter acquisition time.

[0014] In some optional implementations, the deep neural network model is a convolutional neural network model, and the parameter information of the deep neural network model includes: the number of layers and configuration parameters of each layer;

[0015] The configuration parameters of each layer include: convolution kernel size, number of input channels, number of output channels, convolution step size, and convolution supplementary data size.

[0016] In some optional implementations, the constructed collaborative reasoning load distribution optimization model is expressed as:

[0017] minT

[0018] E≤E max

[0019] a i ≥p i+1 ,i∈N

[0020] subject to i ≥0,a i ∈Z,i∈N

[0021] ∑i∈N a i =H

[0022] r li ≤m i ,i∈N,l∈M

[0023] Where T represents the total latency of collaborative reasoning, E represents the total energy consumption of collaborative reasoning, and E max Indicates the preset energy consumption threshold, a i represents the payload size corresponding to the i-th cooperative satellite, p i+1 represents the convolution supplementary data size of the i+1th cooperative satellite, N = {1, 2, ..., n}, n represents the total number of cooperative satellites, Z represents an integer set, H represents the height size of the remote sensing image, r li represents the computational load data volume of the lth layer of the deep neural network model of the ith collaborative satellite, m i represents the available memory space of the ith collaborative satellite, M = {1, 2, …, L}, and L represents the total number of layers of the deep neural network model.

[0024] In some optional implementations, the collaborative reasoning total delay is calculated using the following formula:

[0025]

[0026] in, represents the computational delay of the lth layer of the deep neural network model of the ith collaborative satellite, represents the communication delay of the lth layer of the deep neural network model of the ith collaborative satellite, ρ i represents the number of CPU cycles used by the i-th collaborative satellite to process data, f i represents the CPU frequency used by the i-th collaborative satellite to process data, b i,i represents the data transmission rate of the ith cooperative satellite, p li represents the convolutional supplementary data size of the lth layer of the deep neural network model of the i-th collaborative satellite, b i,i+1 represents the data transmission rate between the i-th cooperative satellite and the i+1-th cooperative satellite.

[0027] In some optional implementations, the total energy consumption of collaborative reasoning is calculated by the following formula:

[0028]

[0029] in, represents the computational energy consumption of the lth layer of the deep neural network model of the ith collaborative satellite, represents the communication energy consumption of the lth layer of the deep neural network model of the ith collaborative satellite, represents the data calculation power of the i-th collaborative satellite, represents the data transmission power of the i-th cooperative satellite.

[0030] In some optional implementations, the remote sensing image is segmented according to the payload size information corresponding to the coordinated satellite to obtain the remote sensing image block data corresponding to each coordinated satellite, including:

[0031] According to the load size corresponding to each cooperative satellite, the load distribution ratio corresponding to each cooperative satellite is determined, and the remote sensing image is segmented along the height edge of the remote sensing image according to the load distribution ratio to obtain the remote sensing image block data corresponding to each cooperative satellite.

[0032] Secondly, a deep neural network collaborative reasoning system for low-orbit satellite clusters is also provided, which can be configured on low-orbit satellites, including:

[0033] An image acquisition module, used for acquiring remote sensing images;

[0034] An information collection module is used to collect resource status parameter information of each satellite in the low-orbit satellite cluster, and use satellites whose resource status meets the preset threshold conditions as slave satellites to collect parameter information of the deep neural network model configured on each slave satellite;

[0035] A modeling module is used to build a collaborative reasoning load distribution optimization model according to remote sensing image size information, resource status parameters of collaborative satellites, and parameter information of a deep neural network model configured on the collaborative satellites, wherein the collaborative satellites include a master satellite and a slave satellite, and the master satellite is a low-orbit satellite configured with the system;

[0036] The load distribution solution module is used to solve the collaborative reasoning load distribution optimization model and obtain the load size information corresponding to each collaborative satellite;

[0037] The collaborative reasoning execution module is used to segment the remote sensing image according to the payload size information corresponding to the collaborative satellite, obtain the remote sensing image block data corresponding to each collaborative satellite, aggregate the reasoning results of all the remote sensing image block data, and obtain the reasoning results of the remote sensing image.

[0038] The main advantages of the technical solution of the present invention are as follows:

[0039] The deep neural network collaborative reasoning method and system for low-orbit satellite clusters of the present invention can effectively integrate on-board resources and realize inter-satellite collaborative reasoning, and can make full use of on-board resources to realize on-board processing of remote sensing image data, avoid problems such as unstable satellite-to-ground links and long transmission time, and meet the execution requirements of time-sensitive tasks and real-time tasks; it can perform adaptive load distribution optimization of remote sensing image data according to the on-board resource status, and can adapt to the dynamic changes of on-board computing power and network; and it can minimize the collaborative reasoning running time within a given energy consumption requirement range. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The drawings described herein are used to provide a further understanding of the embodiments of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0041] Figure 1 This is a flow chart of a deep neural network collaborative reasoning method for a low-orbit satellite cluster according to an embodiment of the present invention;

[0042] Figure 2 This is a schematic diagram of the structure of a deep neural network collaborative reasoning system for a low-orbit satellite cluster according to an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. 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 making creative work are within the scope of protection of the present invention.

[0044] The technical solution provided by the embodiments of the present invention is described in detail below with reference to the accompanying drawings.

[0045] See also Figure 1 In a first aspect, an embodiment of the present invention provides a deep neural network collaborative reasoning method for a low-orbit satellite cluster, the method is used to realize collaborative processing of remote sensing images on a low-orbit satellite cluster, and includes the following steps S1-S5:

[0046] Step S1, acquiring remote sensing images.

[0047] Specifically, a low-orbit satellite, ie, a main satellite, used to acquire the remote sensing image to be processed is determined according to actual needs, and the remote sensing image to be processed acquired by the main satellite and its size information are determined.

[0048] In step S2, the master satellite collects resource status parameter information of each satellite in the low-orbit satellite cluster, takes other satellites whose resource status meets the preset threshold conditions as slave satellites, and collects parameter information of the deep neural network model configured on each slave satellite.

[0049] Specifically, in actual application, each low-orbit satellite broadcasts its own resource status parameter information through the inter-satellite communication payload configured on the satellite, and the main satellite receives the broadcast signal of each satellite in the low-orbit satellite cluster through the inter-satellite communication payload configured on the satellite, and then obtains the resource status parameter information of each satellite.

[0050] In one embodiment of the present invention, the resource status parameter information of the satellite includes: the number of CPU cycles and frequency, memory space, data calculation power, data transmission power, data transmission rate, and resource status parameter acquisition time. The resource status parameter acquisition time specifically indicates the time when the main satellite acquires the resource status parameter information.

[0051] Low-orbit satellites can determine their own corresponding resource status parameter information except for the acquisition time through the data processing payload configured on the satellite to provide computing resources, storage resources and network resources.

[0052] Furthermore, based on the resource status parameter information collected from each satellite, other satellites whose resource status meets the preset threshold conditions are used as slave satellites to obtain a collaborative satellite set including master satellites and slave satellites for subsequent collaborative reasoning of remote sensing images.

[0053] In one embodiment of the present invention, the preset threshold conditions are specifically set according to actual needs, for example, the CPU frequency is not less than 1 MHz, the data transmission rate is not less than 2 Mb / s, the memory space is not less than 1 Gb, etc.

[0054] Furthermore, based on the determined collaborative satellites, parameter information of the deep neural network model configured on each collaborative satellite is collected.

[0055] In one embodiment of the present invention, a deep neural network model is pre-configured on each low-orbit satellite in the low-orbit satellite cluster.

[0056] The deep neural network model is generally a convolutional neural network model. The parameter information of the deep neural network model includes: the number of layers and the configuration parameters of each layer; wherein, the configuration parameters of each layer include: convolution kernel size, number of input channels, number of output channels, convolution step size, and convolution supplementary data size.

[0057] Specifically, in actual application, each low-orbit satellite broadcasts the parameter information of its own deep neural network model through the inter-satellite communication payload configured on the satellite, and the master satellite receives the broadcast signals of each satellite in the low-orbit satellite cluster through the inter-satellite communication payload configured on the satellite, and then obtains the parameter information of the deep neural network model configured on each slave satellite.

[0058] Step S3: The main satellite constructs a collaborative reasoning load distribution optimization model based on the remote sensing image size information, the resource status parameters of the collaborative satellite, and the parameter information of the deep neural network model configured on the collaborative satellite.

[0059] In one embodiment of the present invention, the cooperative satellite includes a master satellite and a slave satellite.

[0060] Specifically, in one embodiment of the present invention, the main satellite performs communication model modeling, energy consumption model modeling and constraint model modeling according to the size information of the remote sensing image to be processed, the resource status parameters of the collaborative satellite and the parameter information of the deep neural network model configured on the collaborative satellite, and constructs a collaborative reasoning load distribution optimization model based on the communication model, energy consumption model and constraint model.

[0061] In one embodiment of the present invention, the communication model is constructed based on the resource status parameters of the collaborative satellite and the parameter information of the deep neural network model configured on the collaborative satellite, which is specifically expressed as follows:

[0062]

[0063] Where T represents the total latency of collaborative reasoning, represents the computational delay of the lth layer of the deep neural network model of the ith collaborative satellite, represents the communication delay of the lth layer of the deep neural network model of the ith collaborative satellite, N = {1, 2, ..., n}, n represents the total number of collaborative satellites, M = {1, 2, ..., L}, L represents the total number of layers of the deep neural network model, ρ i represents the number of CPU cycles used by the i-th collaborative satellite to process data, r li represents the computational load data volume of the lth layer of the deep neural network model of the ith collaborative satellite, f i represents the CPU frequency used by the i-th collaborative satellite to process data, a i represents the payload size corresponding to the i-th cooperative satellite, b i,i represents the data transmission rate of the ith cooperative satellite, p li represents the convolutional supplementary data size of the lth layer of the deep neural network model of the i-th collaborative satellite, b i,i+1 represents the data transmission rate between the i-th cooperative satellite and the i+1-th cooperative satellite.

[0064] Furthermore, the energy consumption model is constructed based on the resource state parameters of the collaborative satellite and the above communication model, which is specifically expressed as follows:

[0065]

[0066] Where E represents the total energy consumption of collaborative reasoning, represents the computational energy consumption of the lth layer of the deep neural network model of the ith collaborative satellite, represents the communication energy consumption of the lth layer of the deep neural network model of the ith collaborative satellite, represents the data calculation power of the i-th collaborative satellite, represents the data transmission power of the i-th cooperative satellite.

[0067] Furthermore, the constraint model is used to represent the constraint conditions that the load allocated by the collaborative satellite needs to meet, and is constructed based on the resource state parameters of the collaborative satellite and the parameter information of the deep neural network model configured on the collaborative satellite, and is specifically expressed as:

[0068] a i ≥p i+1 ,i∈N

[0069] a i ≥0,a i ∈Z,i∈N

[0070]

[0071] r li ≤m i ,i∈N,l∈M

[0072] Among them, p i+1 represents the convolution supplementary data size of the i+1th collaborative satellite, Z represents an integer set, H represents the height size of the remote sensing image, m i Represents the available memory space of the i-th collaborative satellite.

[0073] In the above constraint model, the first constraint condition indicates that the load size allocated by the collaborative satellite should be larger than the convolution supplementary data size on the neighboring collaborative satellite; the second constraint condition indicates that the load size allocated by each collaborative satellite should be a non-negative integer; the third constraint condition indicates that the sum of the load sizes allocated by each collaborative satellite should be equal to the height of the remote sensing image to be processed; the fourth constraint condition indicates that the amount of computational load data of each layer of the deep neural network model must meet the capacity limit of the available memory space.

[0074] Furthermore, based on the communication model, energy consumption model and constraint model constructed above, the following collaborative reasoning load distribution optimization model is established:

[0075] minT

[0076]

[0077] Among them, E max Indicates the preset energy consumption threshold. The specific value of the energy consumption threshold is set according to actual needs.

[0078] The collaborative reasoning load distribution optimization model constructed above means solving the load size corresponding to each collaborative satellite while meeting a given energy consumption threshold so as to minimize the collaborative reasoning running time. The collaborative reasoning running time is the total collaborative reasoning delay.

[0079] Step S4: The main satellite solves the collaborative reasoning load distribution optimization model to obtain the load size information corresponding to each collaborative satellite.

[0080] In one embodiment of the present invention, the above-mentioned collaborative reasoning load distribution optimization model is an integer linear programming problem, which is a type of problem that has been widely studied and currently has many mature and efficient solution algorithms. By iteratively modifying the constraint range of the collaborative reasoning load distribution optimization model and solving it, the optimal solution of the collaborative reasoning load distribution optimization model can be gradually approached.

[0081] In one embodiment of the present invention, based on the size information of the remote sensing image to be processed, the resource status parameters of the collaborative satellite and the parameter information of the deep neural network model configured on the collaborative satellite, the collaborative reasoning load distribution optimization model is directly solved by using the existing linear programming solution tool to obtain the load size information S = [a1, a2, ..., a n ] Among them, the linear programming solving tool is, for example, the CVXPY toolkit and the IBM CPLEX toolkit.

[0082] In step S5, the main satellite divides the remote sensing image according to the payload size information corresponding to the collaborative satellite, obtains the remote sensing image block data corresponding to each collaborative satellite, and sends the remote sensing image block data to the corresponding collaborative satellite so that the collaborative satellite calls the deep neural network model thereon to infer the remote sensing image block data and sends the inference result to the main satellite, aggregates the inference results of all the remote sensing image block data, and obtains the inference result of the remote sensing image.

[0083] In one embodiment of the present invention, the main satellite segments the remote sensing image according to the payload size information corresponding to the cooperative satellite to obtain the remote sensing image block data corresponding to each cooperative satellite, specifically including:

[0084] According to the load size corresponding to each cooperative satellite, the load distribution ratio corresponding to each cooperative satellite is determined, and the remote sensing image is segmented along the height edge of the remote sensing image according to the load distribution ratio to obtain the remote sensing image block data corresponding to each cooperative satellite.

[0085] Specifically, the payload size information S corresponding to the collaborative satellite is S = [a1, a2, ..., a n ] as an example, the load distribution ratio corresponding to the first cooperative satellite is a1 / (a1+a2+…+a n ), the load distribution ratio corresponding to the nth cooperative satellite is a n / (a1+a2+…+a n ). In one embodiment of the present invention, the main satellite is used as the first cooperative satellite.

[0086] In one embodiment of the present invention, the master satellite sends the remote sensing image block data corresponding to each cooperative satellite to the corresponding other cooperative satellites, i.e., slave satellites, through the inter-satellite communication payload configured on the satellite, so that the other cooperative satellites that receive the remote sensing image block data call the deep neural network model on them to infer the remote sensing image block data, and the master satellite also infers according to its corresponding remote sensing image block data; after completing the inference of the remote sensing image block data, the other cooperative satellites send the inference results to the master satellite through the inter-satellite communication payload configured on the satellite, and the master satellite aggregates the inference results of all remote sensing image block data to obtain the inference results of the remote sensing image. In the inference process, the master satellite and the slave satellite can communicate through the inter-satellite communication payload configured on the satellite to exchange the intermediate calculation data required in the inference process.

[0087] refer to Figure 2 In a second aspect, an embodiment of the present invention further provides a deep neural network collaborative reasoning system for a low-orbit satellite cluster, which can be configured on a low-orbit satellite. The system includes:

[0088] An image acquisition module 100 is used to acquire remote sensing images;

[0089] The information collection module 200 is used to collect resource status parameter information of each satellite in the low-orbit satellite cluster, take the satellites whose resource status meets the preset threshold conditions as slave satellites, and collect parameter information of the deep neural network model configured on each slave satellite;

[0090] The modeling module 300 is used to build a collaborative reasoning load distribution optimization model according to the remote sensing image size information, the resource status parameters of the collaborative satellite and the parameter information of the deep neural network model configured on the collaborative satellite, wherein the collaborative satellite includes a master satellite and a slave satellite;

[0091] It should be noted that the low-orbit satellite configured with this system is the main satellite;

[0092] The load distribution solving module 400 is used to solve the collaborative reasoning load distribution optimization model to obtain the load size information corresponding to each collaborative satellite;

[0093] The collaborative reasoning execution module 500 is used to segment the remote sensing image according to the payload size information corresponding to the collaborative satellite, obtain the remote sensing image block data corresponding to each collaborative satellite, aggregate the reasoning results of all the remote sensing image block data, and obtain the reasoning results of the remote sensing image.

[0094] In one embodiment of the present invention, the above modules are devices corresponding to the steps of the above method. The specific working principle of each module and its beneficial effects can be found in the above method and will not be repeated here.

[0095] The deep neural network collaborative reasoning method and system for low-orbit satellite clusters provided in one embodiment of the present invention can effectively integrate on-board resources to realize inter-satellite collaborative reasoning, make full use of on-board resources, realize on-board processing of remote sensing image data, avoid problems such as unstable satellite-to-ground links and long transmission time, and meet the execution requirements of time-sensitive tasks and real-time tasks; can adaptively optimize the load distribution of remote sensing image data according to the on-board resource status, and can adapt to the dynamic changes in on-board computing power and network; and can minimize the collaborative reasoning running time within a given energy consumption requirement.

[0096] It should be noted that, in this article, relational terms such as "first" and "second" are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep neural network collaborative reasoning method for low-orbit satellite clusters, characterized in that: include: Acquisition of remote sensing images; The master satellite collects resource status parameter information of each satellite in the low-orbit satellite cluster, takes other satellites whose resource status meets the preset threshold condition as slave satellites, and collects parameter information of the deep neural network model configured on each slave satellite, wherein the master satellite represents a low-orbit satellite that obtains remote sensing images; The master satellite constructs a collaborative reasoning load distribution optimization model according to remote sensing image size information, resource status parameters of the collaborative satellite, and parameter information of a deep neural network model configured on the collaborative satellite, wherein the collaborative satellite includes the master satellite and the slave satellite; The main satellite solves the collaborative reasoning load distribution optimization model to obtain the load size information corresponding to each collaborative satellite; The main satellite segments the remote sensing image according to the payload size information corresponding to the collaborative satellite, obtains the remote sensing image block data corresponding to each collaborative satellite, and sends the remote sensing image block data to the corresponding collaborative satellites respectively so that the collaborative satellites call the deep neural network model thereon to infer the remote sensing image block data and send the inference results to the main satellite, and the main satellite aggregates the inference results of all the remote sensing image block data to obtain the inference results of the remote sensing image.

2. The deep neural network collaborative reasoning method for low-orbit satellite clusters according to claim 1 is characterized in that: The resource status parameter information of the satellite includes: the number of CPU cycles and frequency, memory space, data calculation power, data transmission power, data transmission rate, and resource status parameter acquisition time.

3. The deep neural network collaborative reasoning method for low-orbit satellite clusters according to claim 2 is characterized in that: The deep neural network model is a convolutional neural network model, and the parameter information of the deep neural network model includes: the number of layers and the configuration parameters of each layer; The configuration parameters of each layer include: convolution kernel size, number of input channels, number of output channels, convolution step size, and convolution supplementary data size.

4. The deep neural network collaborative reasoning method for low-orbit satellite clusters according to claim 3 is characterized in that: The constructed collaborative reasoning load distribution optimization model is expressed as: Where T represents the total latency of collaborative reasoning, E represents the total energy consumption of collaborative reasoning, and E max Indicates the preset energy consumption threshold, a i represents the payload size corresponding to the i-th cooperative satellite, p i+1 represents the convolution supplementary data size of the i+1th cooperative satellite, N = {1, 2, ..., n}, n represents the total number of cooperative satellites, Z represents an integer set, H represents the height size of the remote sensing image, r li represents the computational load data volume of the lth layer of the deep neural network model of the ith collaborative satellite, m i represents the available memory space of the ith collaborative satellite, M = {1, 2, …, L}, and L represents the total number of layers of the deep neural network model.

5. The deep neural network collaborative reasoning method for low-orbit satellite clusters according to claim 4 is characterized in that: The total delay of collaborative reasoning is calculated by the following formula: in, represents the computational delay of the lth layer of the deep neural network model of the ith collaborative satellite, represents the communication delay of the lth layer of the deep neural network model of the ith collaborative satellite, ρ i represents the number of CPU cycles used by the i-th collaborative satellite to process data, f i represents the CPU frequency used by the i-th collaborative satellite to process data, b i,i represents the data transmission rate of the ith cooperative satellite, p li represents the convolutional supplementary data size of the lth layer of the deep neural network model of the i-th collaborative satellite, b i,i+1 represents the data transmission rate between the i-th cooperative satellite and the i+1-th cooperative satellite.

6. The deep neural network collaborative reasoning method for low-orbit satellite clusters according to claim 5 is characterized in that: The total energy consumption of collaborative reasoning is calculated by the following formula: in, represents the computational energy consumption of the lth layer of the deep neural network model of the ith collaborative satellite, represents the communication energy consumption of the lth layer of the deep neural network model of the ith collaborative satellite, represents the data calculation power of the i-th collaborative satellite, represents the data transmission power of the i-th cooperative satellite.

7. The deep neural network collaborative reasoning method for low-orbit satellite clusters according to any one of claims 1 to 6, characterized in that: According to the payload size information corresponding to the cooperative satellite, the remote sensing image is segmented to obtain the remote sensing image block data corresponding to each cooperative satellite, including: According to the load size corresponding to each cooperative satellite, the load distribution ratio corresponding to each cooperative satellite is determined, and the remote sensing image is segmented along the height edge of the remote sensing image according to the load distribution ratio to obtain the remote sensing image block data corresponding to each cooperative satellite.

8. A deep neural network collaborative reasoning system for low-orbit satellite clusters, characterized in that: Capable of being deployed on low-orbit satellites, including: An image acquisition module, used for acquiring remote sensing images; An information collection module is used to collect resource status parameter information of each satellite in the low-orbit satellite cluster, and use satellites whose resource status meets the preset threshold conditions as slave satellites to collect parameter information of the deep neural network model configured on each slave satellite; A modeling module is used to build a collaborative reasoning load distribution optimization model according to remote sensing image size information, resource status parameters of collaborative satellites, and parameter information of a deep neural network model configured on the collaborative satellites, wherein the collaborative satellites include a master satellite and a slave satellite, and the master satellite is a low-orbit satellite configured with the system; The load distribution solution module is used to solve the collaborative reasoning load distribution optimization model and obtain the load size information corresponding to each collaborative satellite; The collaborative reasoning execution module is used to segment the remote sensing image according to the payload size information corresponding to the collaborative satellite, obtain the remote sensing image block data corresponding to each collaborative satellite, aggregate the reasoning results of all the remote sensing image block data, and obtain the reasoning results of the remote sensing image.

Images