Introduction
Low earth orbit (LEO) satellite constellation refers to the satellite system composed of numerous satellites operating in earth's orbit, at altitudes between 500 and 2000 km. Systems like Landsat [1], [2], Skysat [3], and Starlink [4], [5] are extensively used in remote sensing and cellular telecommunication applications due to their global coverage, flexible networking capabilities, and short satellite-to-ground transmission distance [6]. In recent years, the construction and management of mega constellations have drawn considerable interest from researchers and governments alike, given their significant potential to drive economic growth and bolster national security.
Traditional remote sensing LEO satellite usually forward the collected images to ground cloud center for centralized analysis. Due to the fast-passing territory time and long revisit period, the transmission latency of large data sets from LEO satellites to the ground cloud center is high [7]. It is difficult to effectively support time-sensitive applications such as earthquake relief. Therefore, the concept of satellite on-orbit computing is proposed. Based on the captured remote sensing images, the analysis task is directly executed on the satellite, and only the results are sent back to the ground can significantly reduce the processing delay [8].
In this project, we combine the idea of distributed cloud computing [9] with satellite on-orbit computing, and propose the architecture scheme of space cloud. Cloud computing is a service model where cloud data centers offer storage and computing resources to users through network. Infrastructure, guest operating systems, and applications are abstracted from the underlying hardware to create virtual instances such as virtual machines and containers [10], [11]. Within the space cloud architecture, each LEO satellite equipped with on-orbit computing capabilities serves as a distributed computing node. The remote sensing images captured by these satellites are stored as a local dataset. Users send their tasks to the satellite control center. These tasks are then transmitted to the master node, a geostationary earth orbit (GEO) satellite, through the ground-satellite link. For each task, the master node GEO calls the currently available LEO satellites with suitable data to perform collaborative on-orbit computing through intersatellite links. Finally, the results are returned to the ground users through the ground-satellite network.
As shown in Fig. 1, by integrating the storage/computing resources of multiple LEO satellites, as well as the intersatellite communication link resources in the spatial information network, the entire LEO satellite constellation transforms into a single cloud resource pool. The direct on-orbit processing and high-bandwidth intersatellite link of space cloud significantly enhance the efficiency of deployment, operation, and maintenance for time-sensitive remote sensing and computing services on a global scale. As a new spatial information infrastructure, space cloud aids in processing the growing volume of big data sensed from earth surfaces, and supports global geographic information applications such as geomorphology detection, smart cities, earthquake relief, etc.
Space cloud for remote sensing is built with distributed storage/computing resources on multiple on-orbit LEO satellites. The satellite-captured datasets are integrated as a cloud resource pool. Users send their requests to satellite control center through ground network. The control center uploads the missions to satellite cloud through ground-satellite radio links. Collaborative orbit computing is realized through the intersatellite links. The control and dataflow patterns are shown both satellite constellation and terrestrial network facilities. Some real-life data fusion, image understanding, object detection, and AI applications are shown in the dashed box conceptually.
Our contributions are highlighted in the following four technical aspects.
We proposed a novel space cloud architecture deployed in outer space. This cloud platform operates on a LEO satellite constellation, aiming at providing time-sensitive remote sensing and computing services on a global scale.
We developed a Kubernetes-based platform to schedule and orchestrate the virtualized resources of on-orbit satellites for cloud control and management.
We conducted simulation experiments using real satellite operational data to demonstrate that our space cloud framework achieves superior computing efficiency and faster response times compared with traditional modes.
We highlighted the innovative features of our self-developed Luojia-3-01 satellite, which is designed for intelligent remote sensing and on-orbit computing services. This satellite is set to become the initial test platform for the space cloud construction.
Related Works
A. Luojia-3-01: Prototype Machine for Space Cloud Construction
LEO satellite orbits around earth in the height of 500 to 2000 km with global coverage and low time delay. Due to the development of information technologies in recent years, LEO satellites are used as either a space router for cellular telecommunication, or an Internet of Things (IoT) sensor for remote sensing image capturing. Fig. 2 shows the architecture of a custom-designed LEO satellite Luojia-3-01 by Wuhan University and Aerospace Dongfanghong Satellite Company, Ltd. [12].
3-D schematic diagram of Luojia-3-01 satellite built and launched by China's Wuhan University and Aerospace Dongfanghong Satellite Company, Ltd., for intelligent remote sensing applications on January 15th, 2023 [12].
This satellite is designed for research for intelligent remote sensing applications. Based on high-performance GPU computing hardware and large-capacity memory hardware, an open software platform is designed on-board. This platform includes Linux operating system environment, user API function interface, basic database of image processing, deep learning software framework, etc.
Luojia-3-01 supports on-orbit autonomous mission planning. Users upload the longitude/latitude of the target point and related information. Then satellite independently completes imaging task planning, image acquisition, on-orbit processing, and image downloading tasks. Luojia-3-01 also supports on-orbit APP installation, operation, uninstallation, update, etc. It provides end-to-end near-real-time image information services to individual user mobile phone terminals. This satellite has already been successfully launched on January 15th, 2023, and will perform as a prototype machine for the space cloud system construction to execute initial testing.
B. 5G Cloud and IoT Based on Virtualization
Cloud computing is mainly realized through virtualization technologies such as virtual machine and container [10], [13], [14]. It is one of the key technologies adopted in the fifth-generation standard for broadband cellular networks (5G). In 5G-cloud radio access network, network elements of base band units are virtualized as centralized resource pool to offer high computation and communication loads [15]. Network function virtualization is used to virtualize the entire physical infrastructure of network node functions [16]. Network slicings of enhanced mobile broadband, ultrareliable low latency communications, massive machine type of communication are built with different virtual network resources to meet the requirement of multiple services [17]. Software-defined network further separate the control plane and data plane of network to provide external data control [18].
With the enhancement of hardware device performance, cloud sites are gradually deployed from the traditional centralized cloud to the distributed edge sides closer to users [9], [19]. The quick accessible cloud resources bring the advantages of time-sensitivity, parallelism, and redundancy. Multiple distributed nodes may execute a single task collaboratively. It integrates the disperse computing and storage resources in different locations to handle large-scale computing tasks such as smart city and smart healthcare [20], [21]. For instance, federated learning is deployed in the distributed cloud environment to efficiently execute joint machine learning tasks while preserving the local data security and personal data privacy [22], [23].
These technologies also promote the proliferation of IoT in terms of large-scale sensing, communication, and computing in 5G and future 6G [11]. With the support of edge intelligence [24], space-air-ground-underwater communications [25], terahertz communications [26], etc., IoT becomes a more intelligent network to connect millions of heterogeneous devices. It supports the data collection and local processing for various smart life scenarios, such as smart healthcare [27], unmanned vehicle [28], intelligent industry [29], etc.
C. LEO Satellites for Remote Sensing and On-Orbit Computing
For telecommunication, the LEO satellite constellation is deployed as a space-based Internet system to provide global communication services. For example, the Starlink project proposed by SpaceX intends to provide satellite internet connectivity to underserved areas with an expected service rate of 50–150 Mbps [4], [5]. In total, nearly 12 000 satellites are planned to be deployed by 2027. China also proposed a Hongyan Constellation System with 60 nodes and global data processing center. It intends to provide users with global real-time data communications and integrated information services under complex weather and terrain conditions [30].
For remote sensing, the LEO satellite observes and gathers earth resource data on low-earth orbit. Such data presents the features and changes of ground objects, and is often used for geographical applications [31], [32], [33]. For example, farmland monitoring promotes agricultural harvesting production. Tornado path tracking contributes to disaster relief, etc., which is of great significance to national economy development.
Deep learning image understanding algorithms play a crucial role in the analysis of remote sensing images. These algorithms enable the automatic extraction, analysis, and understanding of useful information from a single image or a sequence of images. They can identify patterns, detect changes, and classify objects, which greatly improve the efficiency and accuracy of tasks such as land use classification, environmental monitoring, and disaster management [34], [35], [36], [37].
The rest of this article is organized as follows. Section III introduces the innovative features of our self-developed Luojia-3-01 satellite for intelligent on-orbit remote sensing. Section IV illustrates the dynamic topology of satellite constellation. Section V proposes a Kubernetes cloud orchestration platform to schedule the multisatellite resources in the dynamic constellation. Section VI presents the experimental settings and analysis performance results of space cloud for data filtering and analysis. Then, the advantages in terms of data sensing, AI computing, data storage, and telecommunication efficiencies in global-scale and real-time sensing applications using space cloud are discussed. Finally, Section VII concludes this article.
Luojia-3-01 Satellite for Intelligent Remote Sensing
Traditional remote sensing satellite operation mode usually takes the following main steps: satellite data acquisition, data downloading, ground decompression, ground preprocessing, ground analysis, and information extraction. The information acquisition process is complex and the transmission link is long. The effective information acquisition time reaches the hour level, which is difficult to meet the growing demand for minute or even second-level information acquisition in various fields.
To solve this problem, it is necessary to carry out task-driven data acquisition, real-time processing, and information-intelligent extraction tasks directly using the satellite on-orbit storage and computing capabilities. In addition, the remote sensing tasks can be integrated with the communication, transmission, and navigation resources of spatial information networks to achieve real-time intelligent services.
In this article, we report the design of a new type of intelligent satellite for integrated telecommunication, navigation, and remote sensing applications. The design parameters of Loujia-3-01 satellites and key applications are listed as follows: We will use this LEO satellite to illustrate all space cloud new ideas and key remote sensing applications throughout this article.
Basic deployment data: 500 km height with 240 ± 10 kg. Semimajor axis 6870.071 km. Inclination 97.41°.
Communication: Ka-band intersatellite links and X-band satellite-to-ground link. Optical intersatellite link will be used in the future.
Remote Sensing Key Services:
Remote sensing: Imaging mode of video, pushframe, and scanning with resolution 0.7 m@500 km and spectral range 0.42–0.7 μm. Video frame rate: from 0.1 to 15 fps.
Intelligent processing: 500Gflops processing capacity with 4 GB cache per board.
Software interface: Open software API interface supporting remote control, telemetry, and third-party apps loading and running
On-orbit autonomous mission planning: Autonomous planning, instruction generation, and observation implementation according to the target latitude and longitude.
The Luojia-3-01 satellite was recently deployed in space on January 15th, 2023. At 21:48 on January 15th, 2023, the ground satellite receiving station at Wuhan University began to receive the digital data from Luojia 3-01. The satellite team has completed intelligent remote sensing scientific experiments such as multimode imaging task arrangement, intelligent task planning, on-orbit processing APP, data receiving and decoding, and target information extraction. Some transmitted and processed images are illustrated in Fig. 3. This satellite provides various functions: high-resolution remote sensing imaging, on-orbit real-time intelligent processing, navigation reception and enhancement, and satellite-to-satellite communication as briefly introduced in the following.
Some transmitted and processed remote sensing images by Luojia-3-01 satellite on January 15th, 2023. (a) Kano City, Nigeria, 17:52, January 19th, 2023. (b) Berdan's Port, United Arab Emirates, 17:53, January 18th, 2023.
High-resolution remote sensing imaging function: We configure the high-resolution color area array imaging cameras on the Luojia-3-01 satellite to obtain submeter resolution optical remote sensing data. It can obtain Bayer format images with spatial resolution better than 0.75 m. Due to the flexible attitude maneuverability, the satellite can realize imaging modes: video gaze, area array push frame, and area array push scan. With different on-board processing processes, different imaging modes can be designed to obtain dynamic and static remote sensing data. The remote sensing capability meets various application requirements such as static target detection, dynamic target tracking, and public life service.
On-orbit real-time intelligent processing function: The intelligent processing units of the Luojia-3-01 satellite adopt heterogeneous channels: GPU and field programmable gate array for main computing and backup compression, respectively. At the same time, the main and standby channels are equipped with large-capacity storage modules to support the on-orbit storage. Based on the high-performance hardware platform, we configure a Linux open software platform to realize the on-demand and real-time processing of remote sensing data. This system involves basic operating system environment, user API function interface, image processing basic library, and deep learning software framework for on-orbit intelligent processing.
The Luojia-3-01 satellite integrates a large number of processing algorithms, including autonomous task planning, intelligent screening of regions of interest, high-quality real-time imaging, high-precision real-time geometric positioning, intelligent information processing, intelligent compression, etc. Users only need to send information such as latitude and longitude of target points to the satellite. The satellite can independently complete imaging mission planning, image acquisition, on-orbit processing, and image downloading. It realizes real-time, intelligent, and high-precision processing of task-driven imaging data and efficient data compression.
Navigation reception and enhancement function: The navigation signal can be received, processed, and launched by configuring navigation enhancement system on the satellite. On the one hand, integrated intelligent satellite receives the satellite navigation signals emitted by navigation satellites (Global Positioning System or Beidou system). Using the on-orbit real-time intelligent processing unit, the precise time synchronization in orbit can be calculated. It improves the precise navigation and positioning accuracy of navigation system abroad, thus avoiding the establishment of ground stations.
On the other hand, the ranging signal can be independently generated. This signal can also be integrated with the existing satellite navigation signal to shorten the convergence time of precision positioning. It helps to meet the accuracy of meter-level real-time navigation positioning and improve the service performance of the existing satellite navigation signal system.
Satellite-to-satellite communication and transmission function: X-band direct satellite-to-ground transmission and Ka-band intersatellite transmission are configured on Luojia-3-01 to realize satellite-to-satellite transmission. It establishes a fast channel among multiple satellites and ground stations to ensure the rapid circulation of collaborative processing procedures and parameters, and remote sensing data.
Through dynamic intersatellite transmission networking, data can be downloaded in time, which can effectively solve the delay problem of satellite overhead transmission. At the same time, the integrated intelligent satellite platform with intersatellite transmission function can also be used as an effective supplement to satellite communication network.
Space Cloud Constellation Coverage With Dynamic Connectivity
A. Space Cloud Satellite Constellation Coverage
Multiple LEO satellites deployed on different orbital planes form a satellite constellation. Compared with a single satellite, satellite constellation further enhances the coverage capability to provide low-latency services in a global scale.
In this project, we illustrate the original concept of space cloud for global remote sensing and on-orbit cloud computing services. We design a satellite constellation for space cloud based on the orbital parameters of Luojia-3-01. The constellation architecture is illustrated in Fig. 4. The full system will be deployed with 6 orbital planes, with 12 satellites on each orbit. The global remote sensing is enabled by having the dynamic satellite connectivity capability.
Satellite constellation of space cloud for global remote sensing and cloud computing services. This constellation is extended from the orbital parameter of Luojia-3-01. It is deployed with 6 orbits (orange lines), with 12 satellites on each orbit. The green lines illustrate the dynamic intersatellite links.
Within this constellation, each LEO satellite works as not only an IoT sensor for data collection, but also a cloud node for computing. Ground users deploy their tasks in this satellite cloud, and the results are returned to ground users through ground-to-satellite links to support time-sensitive geographical applications. Directly processing computing tasks on satellites and exchanging the data through spatial network significantly enhances the processing efficiency for remote sensing applications.
We show the coverage percentage of our satellite cloud in Fig. 5. For a single satellite, it takes 23h34m to cover 74.16% ground area. For constellation of 6 orbital planes with 12 satellites on each, only 2h59m is needed to cover 98.97% ground area. We further consider 12 orbital planes with 18 satellites on each, it takes only 55m41s to cover 98.07% ground area, which means that for any single target on earth, this system may take at most 1 h to capture the local image, which provides low-latency services for the global range.
Earth ground coverage and accumulated coverage of a single satellite and space cloud constellation in a global range. For a single satellite, it takes 23h34m to cover 74.16% ground area. For constellation of 6 orbital planes with 12 satellites on each, only 2h59m is needed to cover 98.97% ground area. (a) Ground coverage percentage of a single satellite. (b) Ground coverage percentage of space cloud, which has 6 orbital planes with 12 satellites on each.
B. Dynamic Connectivity of the Satellite Constellation
Space cloud operation mode puts forward high requirements for the dynamic allocation, flexible expansion, and rapid deployment of LEO computing and storage resources. It requires us to design a new satellite cloud orchestration platform to control and schedule the distributed LEO resources within the dynamic constellation topology.
Satellites connect with each other through intersatellite link, also known as crosslink, which realizes information transmission and exchange among multiple satellites. Intersatellite links are divided into radio frequency links and laser links. Radio frequency link technology has high maturity and low failure risk, but the low radio bandwidth limits its data transmission efficiency. Laser link can achieve larger communication capacity and higher data transmission rate about 10 Gbps [38], [39]. Because of the unsynchronized cycle of LEO satellites and earth movements, the constellation topology is constantly changing. It leads to the high dynamic and time-varying characteristics of spatial network.
We use the space cloud satellite constellation in Fig. 4 as an example. As shown in Fig. 6, at first, satellite 4-3 connects to 5-8 and 5-9 within 00:00:00 to 00:03:46. Since the satellite is moving on the orbit, 4-3 will then connect to 1-6 and 1-7 within 00:05:20 to 00:11:53. Due to this dynamic constellation topology, each LEO satellite may connect to different satellites at different time. Because of the periodicity and predictability of satellite on-orbit operation, this time-varying connection relationship can be calculated as shown in Table I, which is stored on-orbit as a routing table.
All LEO satellites may connect with each other through intersatellite links. For each LEO satellite, it may connect to different satellites at different time due to the dynamic constellation connectivity. (a) Satellite 4-3 connects to 5-8 and 5-9 at 00:02:34. (b) Satellite 4-3 connects to (c) Satellite 4-3 connects to 1-6 and 1-7 at 00:06:38 1-5, 1-6, 2-4, 2-5 at 00:13:12.
Space Cloud Resources Orchestration Using Kubernetes Containers
Fig. 7 shows the functional blocks of our cloud with Kubernetes. GEO satellite is the master node. It orchestrates multiple LEO worker node resources through intersatellite links. We use the Kubernetes container cluster management system to schedule and orchestrate the resources of multiple satellites. The Kubernetes is an open-source container cluster management system developed by Google [13]. It is responsible for the management, control, and resource allocation of containers. It provides the functions of resource scheduling, load balancing, and cross-physical host management for container application services, which is the so-called container orchestration [40].
Architecture of space cloud with Kubernetes. GEO satellite is the master node. It orchestrates multiple LEO worker node resources through intersatellite links.
Kubelet and Kube-proxy run on each worker node. Kubelet receives the task information from the API Server of GEO Master through the intersatellite link. Then it establishes a Pod by loading related Docker containers. These Docker containers can be loaded from the local Docker Registry. Users can also deploy their own Dockers on each worker node. These Pods will be used for handling various remote sensing applications. And multiple pods are available to communicate with each other through the intersatellite links with Kube-proxy to execute collaborative computing.
Each LEO stores captured remote sensing images as local dataset. These images are called by different applications through the Container Storage Interface (CSI) of Kubernetes. Multiple pods on different LEO satellites cooperate with each other through intersatellite links to complete the computing task. Then the passing territory satellite returns the results to the users on the ground. All business logic states of the platform will be stored in a distributed and consistent key-value storage. It will also be mapped to the ground control center. Through the key-value storage etcd, ground users conveniently query the details of jobs and services in different scheduling states.
This master-worker management mode allows the ground satellite control center to monitor the operation of each satellite in the LEO satellite constellation in real time. It also allows ground users to access, compile, and operate satellite resources interactively. In the dynamic constellation network topology, even if the network connection is disconnected after the deployment of management nodes, LEO satellite agents can still operate containers and arrange resources. Once the network is restored, through the message broker (such as MQTT protocol), the management node and the work node will complete the synchronization of information data. During this period, the working node still performs joint computing work with other working nodes that can establish connections through intersatellite links.
Experimental Results on On-Orbit Cloud Processing of Remote Sensing Data
In this section, we show how to execute on-orbit distributed computing in our space cloud, and how the high bandwidth intersatellite links help to enhance the overall efficiency. We also show that space cloud provides data returning services with high transmission rate same as ground edge clouds.
A. Remote Sensing Data Filtering and Processing
As shown in Fig. 8, the imaging region of remote sensing satellite presents a continuous strip shape while capturing the ground area [41]. The width of this strip is limited, only 10–50 km under narrow-swath mode for high-resolution imaging. It usually takes 3–5 days to cover the whole earth. When an emergency event happens somewhere on the earth, such as an earthquake or a volcanic eruption, the satellite usually cannot capture photos in time, or it cannot send back the newly captured photos to the ground receiving station immediately due to out of receiving range.
Imaging mode of remote sensing satellites. The imaging region of the satellite presents a continuous “strip” shape. The continuous imaging strip takes 3–5 days to cover the whole earth, which is difficult to handle time-sensitive services.
For traditional remote sensing, the received raw data must undergo a series of data filtering and enhancement before putting into use. Cloud and snow cover on each image should be discriminated [42]. Geometric correction, radiometric calibration, and atmospheric correction [43] are executed to enhance the image appearance. For large area which cannot be covered by a single image, seamless mosaic is used to splice multiple images [44]. Transmitting entire raw data package without filtering is time-consuming. Analyzing these images in a centralized cloud prolongs the total processing time beyond tolerance. As shown in Fig. 9, after the parallel processing of atmospheric correction, an enhanced image covers the whole Shenzhen area is generated through seamless mosaic.
Radiometric calibration, atmospheric correction, and seamless mosaic on two remote sensing images of Shenzhen area (red box). This operation can be executed directly on-orbit in parallel to enhance the processing efficiency for time-sensitive remote sensing applications.
We use the real satellite transmission records of QM-1 satellite to show the data transmission process. QM-1 transmits 1.05 GB data from satellite to earth receiving station on June 27th, 2022, from 11:13:48 to 11:21:22. The total transmission time is 7m34s. Suppose the raw data contains four high resolution images with 1.05 GB each. The intersatellite transmission rate is set as 10 Gbps [38], while the ground transmission time is measured through remote cloud transmission. The simulated experiment results are summarized in Fig. 10.
Computing and transmission time for data enhancement tasks with centralized cloud and space cloud. Space cloud with satellite computing realized much less computing time and transmission time to provide low-latency services for time-sensitive applications.
For traditional centralized computing, it needs at least 30m16s to transmit all four images from satellite to receiving station. Data package of this size needs to be transmitted two to three rounds from space to ground, since the satellite passing territory time is limited. Furthermore, receiving station takes 4m28s to transmit the data to centralized cloud. Centralized cloud then executes further analysis with 3m26s. The total processing time is at least 38m10s.
However, satellite cloud applies parallel processing on four LEO satellites with 1m48s computing time on each node. Each satellite transmits the processed images to a final node through high-bandwidth intersatellite link with 42 s, respectively. After seamless mosaic of 7 s, this final node takes 1m48s to transmit only one image of analyzed result back to earth. The total processing time is 10m11s. Compared with traditional mode, space cloud saves at least 73.3% of total time in terms of both computing and transmission delays.
B. On-Orbit Federated Learning for Scene Classification
Scene classification detects ground objects with image understanding methods. Here we use federated learning for scene classification to show the on-orbit collaborative computing efficiency compared with ground computing. The dataset is RSD46-WHU [45], [46], a large-scale dataset for remote sensing scene classification. This dataset contains 117 000 images with 46 classes as shown in Fig. 11.
RSD46-WHU dataset for scene classification in remote sensing with 46 classes. Different ground objects can be recognized with image understanding deep learning algorithms. Using this dataset, we execute federated learning of scene classification in a simulated dynamic constellation to show the efficiency of on-orbit collaborative computing.
We allocate 6719 data samples on each satellite node, which contains all 46 classes (some data may be reused). For the first global epoch, we only execute local training with satellite 4-3 to infer the time used for one global epoch. At the beginning of global epochs 2, satellite 4-3 checks the current connecting satellites and choose the ones with continuing connecting time longer than the inferred time as clients.
As shown in Table I, both satellite 5-8 and 5-9 are connected to 4-3 at 00:02:26, the beginning of global epoch 2. But 5-9 will disconnect with 4-3 at 00:03:46. So 4-3 will only choose 5-8 as client and send the global model to 5-8 to execute training. The time of global epochs 1 and 2 will be averaged to guide the next global epoch. We repeat this operation until the federated learning model converges. The period and participants for each global epoch under Resnet18 model [34] are given in Table II.
We execute experiments with different Resnet models. with 500–3000 images for each class. Federated learning is a distributed machine learning paradigm with multiple clients [22], [23]. Each node locally executes gradient descent with their local dataset as follows:
\begin{equation*}
\forall k,\ w_{t + 1}^k \leftarrow {\bar{w}}_t - \eta {\nabla }_k\left({{{\bar{w}}}_t,\ {S}_k} \right). \tag{1}
\end{equation*}
In (1),
\begin{equation*}
\ {\bar{w}}_{t + 1} = \mathop \sum \limits_{k = 1}^K \frac{{{s}_k}}{{{s}_1 + \ldots + {s}_K}}\left[ {\left[ {w_{t + 1}^k} \right]} \right]. \tag{2}
\end{equation*}
We compare the results of ground distributed cloud with same dynamic arrangement, and centralized cloud with 18 364 data samples. The results are shown in Figs. 12 and 13. All three platforms realize the classification accuracy of 90%. Satellite computing and ground distributed computing achieves more stable convergence than centralized computing. These two methods also achieve higher processing speed than ground centralized cloud due to multinode parallel processing.
Total processing time of satellite computing, ground centralized, and distributed computing under different deep learning model settings. Compared with ground computing, space cloud with satellite computing realized the lowest total processing time in terms of different deep learning model.
Accuracy and loss curve of Resnet101 with satellite computing, ground centralized, and distributed computing. Satellite computing realized the same level of training accuracy and loss as ground computing, with the lowest total processing time.
For ground distributed computing, transmission media and queening time at routers delay the network speed. However, intersatellite link of space cloud has high transmission rate with less affecting factors in space. This phenomenon is more obvious for complex model with larger transmitted model file size. For example, for Resnet101 model with 163 MB data size, satellite-cloud constellation takes only 66 min to complete computing, resulting in 30.64% reduction in total processing time compared with ground distributed computing.
Satellite cloud computing can be further improved through allocating more data on each node, optimize the intersatellite link range, or building a larger constellation system. With more satellites on orbital planes, each satellite may connect to more satellites at different time, which will involve more satellite to execute collaborative computing.
C. On-Orbit Object Detection Results
Edge clouds are often placed close to users to enable time-sensitive services [10]. AI inference requires edge clouds to execute a pretrained model on new data samples to generate analysis results, and return the results back to users to support decision making. This operation is often deployed at the scenarios with constant new data generation in smart transportation, healthcare, and remote sensing object detection.
In what follows, we show that space cloud provides data returning services with high data rate as ground edge clouds. We use the remote sensing object detection task as an example [48]. Object detection extracts the ground object information of appearance, location range, and changing from the remote sensing images. Fig. 14 shows the result of executing object detection task on a Luojia-3-01 captured image. This task can be completed using neural network algorithms [35].
Object detection on Luojia-3-01 captured high resolution remote sensing images to track storage tanks.
In the space cloud scenario, the pretrained object detection model is deployed on LEO satellite. For a new capture remote sensing image, after data filtering and enhancement, the model is used on this image to infer the object detection results.
Then, the result will be sent back to ground users through the intersatellite links. Satellite data is received by ground receiving station or user terminals through satellite-to-ground links. The transmission rate of satellite-to-ground link is determined by free space propagation loss and near-earth atmosphere. While ground network transmission is affected by transmission media, switching routing, and queuing.
The real data transmission records of three remote sensing data within 2022/6/27 to 2022/7/03 are shown in Table III. This data is provided by the satellite receiving station of Wuhan University. Here, QM-1 is a multimode remote sensing satellite with visible and noctilucent multispectral. AQUA is a US-made satellite measuring the impact of ocean currents, clouds, and water on the environment. ***-1 is a commercial satellite. Due to the confidential issues, some sensitive information is blocked. These satellites transmit data with different bandwidth on different frequency channels. We use (3) to calculate the real data transmission rate (RDTR).
\begin{equation*}
RDTR\ \left({Mbps} \right) = \frac{{8 \times 1024 \times Data\ Size\ \left({GB} \right)}}{{Transmission\ Time\left({seconds} \right)}}. \tag{3}
\end{equation*}
We calculate the mean value of real data transmission rate for each satellite. We compare the satellite transmission rate with central cloud (Beijing) transmission, and edge cloud (Shenzhen) transmission under different bandwidth settlement. The results are shown in Fig. 15.
Real transmission data rate of satellite, edge cloud, and aliyun centralized cloud under different settings. Satellite transmission realizes the same transmission rate as the 1000 m edge cloud, which is also close to the 80 m edge cloud.
Due to distance issues, the 80 m edge cloud achieves the highest transmission rate, while the remote centralized cloud is the lowest under different bandwidths. However, satellite transmission realizes the same level transmission rate as the 1000 m edge cloud, which is also close to the 80 m edge cloud. Even deployed in the 500 km high space, LEO satellite still transmits data with high bandwidth usage and few path losses. It shows that space cloud is available to provide low-latency services same as edge cloud computing.
The satellites with bandwidth of hundreds to thousands of Mbps are under development. For instance, the Luojia-3-01 achieves the Ka-band bandwidth of 300 Mbps. These satellites will contribute more in time-sensitive applications in the future.
Conclusion
We have proposed a new space cloud framework. The entire satellite constellation forms a single cloud platform with dynamic computing nodes. This space cloud provides on-orbit computing capability to perform real-time remote sensing and AI cognition operations in the space. Multisatellite collaborative computing makes up for the lack of computing power of a single node. High-bandwidth intersatellite link further improves the distributed computing efficiency. Ground users deploy their tasks in the cloud, and receive the computing results through the ground-satellite links. In summary, our major contributions are highlighted in four scientific aspects:
We developed a Kubernetes cloud management platform to control, schedule, and orchestrate the virtualized multisatellite resources in order to promote parallelism.
Space cloud achieves higher computing efficiency than ground centralized or distributed cloud platform through multisatellite collaborative computing.
Space cloud returns the on-orbit processed data back to earth users much faster than using the ground cloud for the same purposes.
We developed innovative features of the Luojia-3-01 LEO satellite for intelligent remote sensing and on-orbit computing services. This satellite will become the initial test platform for the space cloud construction.
With the development of information technologies such as 5G/6G, cloud computing, the IoT, and artificial intelligence, a massive amount of data is being generated within the earth's space. Traditional single satellite mode for space data acquisition, processing, and application are no longer able to meet the real-time application needs of the public. Multiple mega-satellite constellation projects have been proposed by different countries to provide global-scale sensing and communication services.
For example, by the end of August 2023, SpaceX in the United States has already launched more than 5000 satellites for the Starlink project, providing high-speed and low-latency internet services to 1.5 million users in more than 50 countries around the world. China Satellite Network Group Company, Ltd., also put forward GW (Guo Wang) satellite constellation plan in 2021, aiming at establishing a low-orbit broadband satellite network consisting of 13 000 satellites. This constellation is expected to complete the first phase of construction in 2025.
In the future, the LEO constellations may also be integrated with other communication and positioning satellites, forming a massive spatial information network system to provide unified sensing, positioning, and communication services. Within this framework, space cloud based on multisatellite collaboration is expected to be an effective framework for processing the generated earth resource spatio-temporal data and execute global resource scheduling.
Using satellite-to-ground and intersatellite link resources, space cloud achieves data storage, forwarding, and sharing in a unified resource pool deployed in outer space, providing real-time intelligent services on a global scale. Our work will benefit multiple remote earth resource applications, such as smart cities, disaster relief, agricultural production, and environmental protection. It will promote ubiquitous connectivity and intelligent services worldwide, and may contribute to the future 6G infrastructure construction.