Applications and Challenges of Artificial Intelligence in Space Missions

Artificial Intelligence (AI) is increasingly finding acceptance in the space community, especially Machine Learning (ML), a subdomain of AI. ML algorithms now find numerous applications in autonomous navigation, spacecraft health monitoring and operational management of satellite constellations. However, a large number of surveys on the applications of AI in space missions can be classified into two categories. The first category suffers from the limitation of being old and not covering some crucial and recent developments in the field; such as the contributions of Deep Learning (DL) and bioinspired AI algorithms. The challenge with the second category lies in its being too detailed with respect to the development and application of specific AI techniques or algorithms. These limitations have necessitated the need to have a concise survey with a wider scope for those interested in the applications and challenges of AI in the space industry, especially those with technical backgrounds in other fields. In this paper, we surveyed the use of traditional AI techniques in various domains of space missions without delving into formal methods. Some bioinspired AI algorithms were also surveyed and their potential application areas highlighted. Unlike similar surveys that focus only on technological challenges, we also addressed some crucial legal drawbacks that emanate from the reliance and use of AI in space. Summarily, while discussing future directions we reviewed some advancements in Internet of Things (IoT) and Blockchain technologies. Our review prioritized three application domains positioned to benefit immensely from the inevitable AI-Blockchain convergence in the space community. These include the Internet of Space Things (IoST), Satellite Communication and Spacecraft Data Security.


I. INTRODUCTION
AI has witnessed a growing interest in the space community over the last two decades. In [1] the authors presented an anatomy of AI spanning over the first 16 years of the 21 st Century (2000)(2001)(2002)(2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015). Coincidentally, both AI and space exploration had their beginnings in the 1950s. In 1955, Newell Shaw and Simon developed Information Processing Language (IPL-11), the first AI language. On October 4, 1957, the Soviet Union launched Sputnik into space-the first human-built spacecraft. In 1958, McCarthy introduced LISP, which soon became the programming language of choice for AI applications after its publication in 1960. In the same year, precisely on January 31, the United States launched and sent into orbit her first satellite-Explorer 1 [2]. The two fields, AI and Space Exploration, continued to develop independently until 1972 when the Jet Propulsion Laboratory (JPL), belonging to the National Aeronautics and Space Administration (NASA), began AI research on her Mars rover. AI is currently being researched in the domain of satellite operations, especially in supporting the operational apparatus of vast satellite constellations and rovers [2]. Considering that it takes about twenty-two minutes for radio waves to travel between Mars and Earth, rovers like NASA's Curiosity rely heavily on AI to make decisions and navigate on their own without commands from mission control.
Computer Vision (CV) techniques applied to satellite imagery still remains one of the most promising applications of AI in the space sector.
In space robotics two key factors have contributed to making AI more applicable. First, owing to the rapid advancement in hardware, the improved computational power of smaller form-factor devices has made it possible for them to run sophisticated algorithms [3]. Furthermore, autonomy and big data are converging to solve new problems in new domains. Robots are making rapid and considerable progress in the field of perception. They are beginning to hear, see, read, and touch in ways not previously possible [4]. As a result, they need to collect far richer data sets than traditional, strictly passive, or even active (viewpointchanging) sensors to enable them to interact with the environment [4]. Although ML is a core technology within robotics, robot learning has suffered considerable challenges due to the theoretical advancement at the boundary between optimization and ML. Currently, one of the biggest challenges facing humanity's space exploration quest is not visiting where no one has visited before, but managing the data generated from space missions. In light of the above, the paramount questions to ask are: How can AI help with the optimization of space missions? Are there technologies that we can leverage to resolve challenges with AI in space?
While multiple surveys on the applications of AI in space missions have been published, many of them are old and do not cover some crucial and recent developments in the field; such as the contributions of Deep Learning (DL) and bioinspired AI algorithms. Few others, although recent, focus on specific domains like remote sensing, Fault Detection Isolation and Recovery (FDIR), and space exploration.
The survey in [5] focused mainly on autonomous planning and scheduling of operations, Anomaly Detection (AD) and FDIR. Izzo et al. [6] surveyed the recent developments in the area of spacecraft guidance dynamics and control, with a focus on evolutionary optimization, tree searches and ML (including DL and RL) as the principal drivers and technologies for current and future research in the field. Similarly, in [7] Kunze et al. surveyed and discussed AI techniques as ‗enablers' for long-term robot autonomy, current advancement in integrating these techniques within long-running robotic systems, and the future challenges and opportunities for AI in Long-Term Autonomy (LTA). Application of Al to remote sensing was reviewed in [8], including onboard data processing and the promise of Al techniques for improving our capability to perform automated analysis of multispectral imagery. Zhu et al. [9] in a more recent review underscore recent advances in remotesensing data analysis with a main focus on analyzing the challenges of employing DL for remote-sensing data analysis. Some major DL concepts pertinent to remotesensing were introduced in [10] and more than 200 publications in this field, most of which were published during the last two years, were reviewed and analyzed.
There exist publications with a different scope in close proximity to our work. For example, the use of AI with distinct focus on Deep Neural Networks (DNNs) onboard spacecraft was discussed by Furano et al. [11] and the possible benefits analyzed in terms of bandwidth downlink. Berquand et al. [12] demonstrated that AI could be employed at the start of the space mission life cycle via an Expert System (ES) deployed as a Design Engineering Assistant (DEA). Girimonte and Izzo [13] provided an overview of AI for space applications and discussed: distributed AI for swarm autonomy, distributed computing for enhanced situation self-awareness and for decision support in spacecraft system design. The state of the art in data mining of satellite telemetry was reviewed in [14] and a framework of necessary processes on data mining to resolving various problems in telemetry data was presented. In [15], DL in space was identified as one of the development directions for mobile and embedded ML and the role of on-device DL in significantly improving spacecraft operation was discussed.
Summarily, most studies surveying AI with respect to space applications have either been reviews concerning closely related aspects of space missions (( [5], [7]) or detailed topical reviews with respect to the development and application of specific AI techniques or algorithms ( [15], [11] and [16]) to a specific field ( [17]). This paper, contrastingly, surveys four unrelated application areas of space missions: Spacecraft Health Monitoring, Remote Sensing, Satellite Communications and Robotic Autonomous Systems (RAS) for space. Challenges and Future Directions were also discussed, especially with respect to the inevitable AI-Blockchain convergence in the space industry. Considering the need to have a simple but concise review, this survey focused more on communication and remote sensing satellites. Furthermore, conscious efforts were made to exclude complex mathematical formulas and abstractions.
The remainder of this paper is organized as follows; Section 2 highlights the need for AI in space missions and provides an overview of AI, ML and Autonomy. In this section we identified nine recently developed bioinspired optimization algorithms and their potential application areas in space missions. Section 3 highlights and discusses four application areas: Spacecraft Health Monitoring, Remote Sensing, Satellite Communications and RAS for space. Challenges and Future Directions are addressed in Section 4 and Section 5 respectively. Finally, Section 6 presents the Conclusion. FIGURE 1. US Challenger Accident [18] Current satellite architectures are built for missions and operations in the hostile environment of space. Usually, satellite operations post-launch are tightly constrained by an FIGURE 2. A Russian Soyuz Rocket Failure [19] Conversely, in unsupervised learning, the algorithm derives a structure from the data after identifying similarities in the inputs. What the results should look like is not known [29]. The principal goal of unsupervised learning is to discover hidden and interesting patterns in unlabeled data. Unlike supervised learning, unsupervised learning uses unlabeled, unclassified, and categorized training data [41]. Popular unsupervised learning algorithms include but are not limited to K-Means Clustering [42] and Dimensionality Reduction Algorithms [39]. In K-Means Clustering, variables in the data are grouped together based on relationships among them. An example of Dimensionality Reduction Algorithms is Principal Component Analysis (PCA) algorithm. The goal of this algorithm is minimizing the dimensionality of large data sets, by transforming a large set of variables into a smaller one while preserving as much ‗variability' (i.e. statistical information) as possible [43].

II. BACKGROUND: THE NEED FOR AI
Summarily, RL differs significantly from both supervised and unsupervised learning. An RL agent has the goal of learning the best way to accomplish a task through repeated interactions with its environment [44]. In RL, no labeled dataset is received by the machine. Instead, information is collected after interacting with the environment through different actions. The agent is rewarded after each action; hence its objective is maximizing this expected average reward where the action would become optimal [29]. The Markov Decision Process (MDP) represents a notable example of an RL model [45]. Other ML algorithms, though with less widely usage, exist [29]. They include: Naïve Bayes [39], Decision Trees [46], Bayesian Regularization [47], Kriging [48], Boosting [49] and more [39].

C. BIOINSPIRED ALGORITHMS AND OPTIMIZATION
Optimization could be defined as the process of finding an optimal set from the set of all possible solutions for a given problem [50]. An algorithm known as the optimization algorithm is conventionally developed to find such a solution. However, many real-life problems can be characterized as multi-objective problems involving multiple conflicting objectives that should be considered simultaneously [51,52]. They usually contain contradictory criteria, where optimization of one objective might also have negative influence on other objectives. In recent years, new Bioinspired Algorithms (BIAs) have been developed to help overcome the limitations of traditional AI algorithms, especially when it concerns the optimization of multiobjective problems [53]. These algorithms naturally tend to have a higher efficiency than the traditional AI methods; with the two most widely accepted categories being evolutionary and swarm algorithms [54,55]. In ML, bioinspired optimization algorithms have been found to address the optimal solutions of complex problems in various science and engineering fields [55]. For example, mobile robot control (a critical aspect of rover design) is one of the main application fields of BIAs. Traditional AI algorithms are usually met with developmental constraints such as the reliance on high-precision sensors and complex computing [54] in this field. BIAs can be also hybridized together to solve the slow convergence speed, low prediction accuracy, and trapping in local optima problems for anomaly detection of artificial satellites [14]. In our survey of BIAs we summarized in Table II eight types of BIAs and domains were they could be potentially relevant in space missions; or used to overcome the limitations of traditional AI techniques.
Ashraf in [55] noted nine recently developed optimization algorithms inspired by the biological behavior of some animals when fighting for food and mates: a. Genetic Bee Colony (GBC) algorithm [56] is a new optimization algorithm that integrates the advantages of the Genetic Algorithms (GA) and Artificial Bee Colony (ABC) for optimizing numerical problems. b. Fish Swarm Algorithm (FSA) is a novel population-based/swarm intelligent algorithm inspired by the natural schooling behavior of fish [57,58]. c. Artificial Algae Algorithm (AAA) is a populationbased optimization algorithm inspired by the behavior of microalgae cells microalgae lifestyles such as the algal tendency, reproduction, and adaptation to the surrounding environments [55,59]. d. Whale Optimization Algorithm (WOA) [ [65], was inspired by the behavioral characteristics of elephant herds and. It divides the agents into two groups: male and female elephants. .

Table II BIAs with Potential Application Areas in Space Missions
Despite the merits of bioinspired optimizers, solving multiobjective problems is a huge challenge as demonstrated in different application domains (e.g. robotics [66], bioinformatics [67], energy and power [68]). According to the No Free Lunch (NFL) theorem [69], the superior performance of an optimizer on a class of applications or problems cannot guarantee a similar performance on other problems. In other words, an optimization algorithm may perform well in a set of problems and fail to solve a different set of problems. For a multi-objective problem no single optimal solution exists, but rather a set of alternative solutions represent the optimal solutions [70]. These solutions are optimal when all objectives are considered and no other solutions in the search space are superior to them. These best trade-off solutions are also known as Pareto Optimal (PO) solutions [71], a concept that was first proposed by Edgeworth and Pareto [72]. It has been shown that evolutionary algorithms such as Non-Dominated Sorting Genetic Algorithm (NSGA-I) [73] and Multi-Objective Evolutionary Algorithm based on Decomposition (MOEA/D) [74], can approximate the true PO solutions of multi-objective problems [70]. Recently Grasshopper Optimization Algorithm (GOA) was proposed and it has been proven to show very fast convergence speed toward the optimal solutions [75].

D. AUTONOMY
Autonomy could be defined as the ability of an agent to accomplish goals through logical decision making based on its knowledge and comprehension of the world, itself, and the situation [114]. An intelligent agent capable of autonomy must exhibit some fundamental properties such as reasoning, learning, and problem-solving [115] . Over the years, the investment in and application of autonomy has yielded significant breakthroughs in areas such as: Power Systems, Mission & Flight Operations, On-Orbit Assembly & Docking [116]. Past missions have demonstrated that using onboard autonomy to enable faster response times improves operational efficiency, optimizes costs and increases system reliability [117]. For example, by using the Autonomous Sciencecraft Experiment to automatically plan and adaptively  [112,113], [111] re-plan observations as required the EO-1 spacecraft was able to cut operational costs by $1 M per year, with a 50% increase in science return [118]. However, as far as decisional capacity is concerned, a fundamental difference exists between automated and autonomous systems. An automated system does not make choices for itself, but relies on potentially sophisticated scripts where choices have both already been made and encoded, or they must be made externally. In contrast, an autonomous system can make informed choices on its own, even when encountering uncertainty or unanticipated events within its environment [119]. In [22], Johnson et al. propose the following necessary conditions for autonomy: Autonomy embodies a range of behaviors associated with agents. Therefore, an autonomous system must possess the traditional qualities of agency, which is the ability to sense, think, and act. It must be able to sense its environment and its state; make decisions about what to do, and act on the decisions through their action. Going by this definition, the chess-playing program Deep Blue is not an autonomous system, even if it exhibits quite an impressive level of intelligence in chess-playing. Given that its range of behaviors is solely limited to the moves of a chess game, it cannot be said to meet all the requirements for agency. 2) GOAL-DRIVEN An autonomous system must be capable of achieving a set of goals by transforming the goals into sequences of actions that accomplish each goal. Instead of just responding to its immediate environment, it should be able to make decisions to achieve a prescribed goal. Based on the above condition, a teleoperated robot would not fall under the classification of an autonomous agent, though it could exhibit a range of goaldirected behaviors. This is because it is command-driven rather than goal-driven.

3) FLEXIBILITY
Finally, an autonomous system should have the flexibility to respond to off-nominal situations by adjusting its actions to enable it to reach a goal successfully. This condition refers to a system's robustness with respect to its environment. Based on this condition, a system that is goal-directed but prone to failure in accomplishing its goals due to unpredicted changes in its immediate environment is not autonomous. Robustness, to some extent, will require the decentralization of control [22,120]. Summarily, the architectures for autonomy consist of a planning layer, a task sequencing layer, and a reactive layer. These three layers are often referred to as deliberative, executive, and functional layers; which can be distinguished in terms of their abstraction from the hardware and their response-time requirements [120]. The European Cooperation for Space Standardization (ECSS) expounds four levels of autonomous capabilities, with level E4 being the most autonomous. The architectures and their descriptions are listed in Table III. Only systems compliant with level E4 can be regarded as genuinely autonomous, whereas levels E1 to E3 deal with manually controlled or automated systems [5]. In general, autonomy is distributed across a system in three ways: no autonomy, partial autonomy and, full autonomy.

I. SPACECRAFT HEALTH MONITORING
Safety and reliability are among some of the most critical concerns to be addressed when planning space missions. In [121], Atkinson et al. describe briefly the spacecraft and ground systems monitoring process at the Jet Propulsion Laboratory, California Institute of Technology, and highlight some of the challenges associated with existing technology used in mission operations.  Due to the limitations of existing technology, Spacecraft Health Automated Reasoning Prototype (SHARP), a new automated system based on AI technology was designed to automate health and status analysis for multi-mission spacecraft and ground data systems operations. By performing real-time analysis of spacecraft and ground data systems engineering telemetry, SHARP effectively detected and analyzed potential spacecraft and ground systems problems.
Recent progress in data mining techniques has also made it possible to use archived spacecraft telemetry data to produce advanced spacecraft health monitoring applications for AD and FD [122]. In his paper, Quan et al. introduced some conventional approaches for AD and FD before proposing feasible approaches using data mining technology. We shall be taking a look at only two of the conventional approaches.

A. ANOMALY DETECTION & FAULT DETECTION
In spacecraft design the technique of identifying faults and isolating them is known as FDIR [123]. Due to the harshness of space environment and the complex structure of a spacecraft it is practically impossible to utterly eliminate the possibility of anomalies or faults that might jeopardize the mission. From the deterioration of electrical circuits and solar cells as a result of high radiation particles to orbital decay due to drag, the spacecraft is also prone to collision with meteoroid debris. Fig. 4 illustrates the impact of radiation particles, drag, and meteoroid debris on a spacecraft. The extreme difficulty or near impossibility of directly repairing or replacing a damaged component necessitates that serious attention is paid to FD and diagnosis. Therefore, it is not an exaggeration to state that designing and implementing FDIR techniques are among the most complex tasks in spacecraft development. All subsystems and their modes of operations must be factored into the design process. It is also essential to factor in safety and reliability at the early stages of mission design. Unfortunately, current FDIR processes are built on the results of Failure Modes, Effects, and Criticality Analysis (FMECA) and Fault Tree Analysis (FTA) [5]. One of the drawbacks of these two approaches is that they can only be applied late in the development process, prohibiting FDIR to become an integral part of the system. [124][125][126] present an extensive survey of the current state of the art of FDIR approaches and emerging techniques, and how to overcome their limitations.
Failures, when not timely and adequately resolved, can result in mission jeopardy. Fig. 5 shows Nigerian Sat1, which was lost because of a failure that prevented one of the solar arrays from deploying after launch. Consequently, the amount of available solar power for recharging the batteries after an eclipse phase was limited. The failed satellite was subsequently replaced with NigComSat-1R (See Fig. 6). Given that most people tend to confuse the term fault with failure, it is pertinent for us to define what we mean by fault and failure as used within the scope of this work. A fault in simple terms is the deviation of at least one system parameter from its desired value. On the contrary, failure is the manifestation of a fault in terms of system functionality that leads to the partial loss of system services [5]. An example of a fault is when a temperature or pressure value is out of limit. In contrast, the inability of a battery charging unit to charge the spacecraft's battery is an example of a failure that could lead to the loss of power to critical subsystems; thereby resulting in the loss of one system service or more. NigComsat Sat-1R [127] Therefore, to guarantee system availability, reliability, and performance faults must be correctly handled so that they do not lead to failure. However, it is not enough for a system to detect a fault without taking further action. Following detection, the fault should be traced to an exact location (e.g., subsystem or memory area) and isolated, after which the system in the recovery step tries to transfer to a safe state of execution in which the fault has been mitigated [5,128].

1) LIMIT CHECKING
Although there have been many conventional approaches to FD and diagnosis, limit checking remains the most fundamental and widely used technique in spacecraft AD. It is used to monitor whether the various values of sensor readings such as current, voltage, and temperature are in the proper ranges, which are predefined by the spacecraft engineers and specified by lower and upper limits [122]. Given the simplicity of this technique, it can only check one sensor value at a time and might be herculean for engineers and operators to build or change the predefined ranges. Though the most significant advantage of this approach is its simplicity, it lacks flexibility and cannot check some small anomalies that occur without violating the limits on the variables [122]. To overcome these limitations Yari et al. proposed a new data-driven health monitoring and AD method for artificial satellites [129].

2) MODEL-BASED DIAGNOSIS
Given the need to overcome the limitations of limit checking techniques some rule-based expert systems and model-based reasoning methods were developed [130]. The model-based approach to automating AD encodes human knowledge into a model, which is then used to automatically detect faults [131,132]. The core idea behind model-based diagnosis is establishing the mathematical model of the spacecraft's subsystems (e.g., thermal control model, propulsion subsystem model) and detecting the possible anomaly or fault through the comparison between the real-time telemetry data and the model output data [122]. However, one major disadvantage of this AD approach is that building models can be labour-intensive. Modeling every subsystem of a highly complex system such as a spacecraft may not be feasible [131]. It also may not be realistically possible to model all possible failure modes mathematically. Furthermore, it may not be able to recognize faults that involve relationships among large numbers of parameters. In summary, all traditional approaches of AD/FD rely heavily on prior knowledge of the system, such as the range of sensor value, mathematical model, and fault rules. With these approaches, it is challenging to acquire accurate and complete models and knowledge of the spacecraft systems beforehand.

C. ANOMALY DETECTION AND FAULT DETECTION BASED ON DATA MINING
The traditional approaches to AD and FD are labourious, hence the need to adopt a data-driven approach that seeks to build a model for detecting anomalies directly from the data, rather than building it based on human expertise. Spacecraft telemetry data involves thousands of sensor values from different subsystems, making it difficult for human experts to pick up faults that involve the relationships among large numbers of variables. However, with the recent advancements in data mining and ML technologies, this new intelligent monitoring approach utilizes the history telemetry data to obtain the knowledge of the system or build the system model dynamically [122,131].
Recent developments in data mining techniques have also made it possible to use archived spacecraft system data to generate advanced system health monitoring applications. In addition to complementing existing approaches, "datadriven" applications are capable of characterizing and monitoring interactions between multiple variables to provide valuable decision support for engineers and mission controllers [131,133].
In contrast to common individual parameter monitoring and model-based schemes, several data-driven software tools have been practically and successfully applied by NASA and JAXA to mission operations for both the Space Shuttle and the International Space Station (ISS). Orca1, a data-mining tool that searches for abnormal data points in multivariate data sets by calculating each data point's distance from adjacent points, has been applied in the shuttle. Similarly, the Inductive Monitoring System (IMS) has been deployed in the ISS. The IMS tool2 uses clustering to analyze archived spacecraft data and characterizes nominal interactions between selected parameters [122,131,133].
Using expert-labeled telemetry anomaly data from the Mars Science Laboratory (MSL) Curiosity rover and the Soil Moisture Active Passive (SMAP) satellite, a team of researchers at NASA demonstrated the effectiveness of Long Short-Term Memory (LSTMs) networks (a type of Recurrent Neural Network (RNN)) in overcoming issues associated with traditional AD methods for spacecraft telemetry [134].
In [122], Quan et al. proposed five feasible approaches using data mining technology for spacecraft AD/FD. We would look briefly at Adaptive Limit checking and Expert System (ES) based on data mining. The reader is hereby referred to Table IV for a summary of the remaining approaches.

1) ADAPTIVE LIMIT CHECKING
Conventional limit checking techniques require the predefinition of the lower and upper limit of sensor value, which can be quite herculean. In contrast, Fig. 7 shows a simplified workflow of the Adaptive Limit Checking method using ML techniques. This method uses ML algorithms (e.g. Gray System Theory, SVM, Rough Set Method, and GA) to dynamically and automatically produce the range limits for sensor value [122]. For training the algorithms the historical spacecraft telemetry data are used as training data. However, different algorithms will use different historical telemetry data and have different efficiency.

2) EXPERT SYSTEM BASED ON DATA-MINING
In an ES based on data mining as illustrated in Fig.8, the knowledge database is not static or predefined. Instead, it is Table IV. PROPOSED ML APPROACHES TO AD/FD produced by the data mining application and dynamically added to the database [122]. At the core of this procedure is a learning algorithm used to extract useful rules for the diagnosis expert system. Though the authors of [122] acknowledge that a diagnostic method based on data mining can maximize the use of the historical data to enhance the conventional diagnosis approach, they fear many problems still need to be solved before they are used in the actual engineering practice. One of such challenges includes the building of the status vector.   [122] In [128], Gao et al. presented a new FD and diagnosis approach hinged on PCA and SVM as illustrated in Fig. 9. The framework of their approach is divided into two phases: the training phase and the real-time detection phase. First, in order to reduce the dimensionality and complexity of the input data PCA is used to extract feature vectors from input data. This is followed by the use of binary SVM to detect fault. After the FD, a multi-class SVM is used to identify fault type.
A probabilistic graphical model employed to represent a set of arbitrary variables whose conditional independencies probabilities can be computed via a directed acyclic graph.

ANN
Neural networks are non-linear statistical data modeling tools; which can be used to model the complex associations between inputs and outputs or to find patterns in data.

Distance-Based Diagnosis
Distance computed according to the real-time telemetry, and the historical data can be used to determine whether there is some fault in the spacecraft in distance-based diagnosis method.

Expert System (ES) Based on Data-Mining
In the ES based on data mining, the knowledge database is dynamic and not predefined or static. Given that the data mining application has its learning algorithm it can produce new rules without operator interference, and knowledge can always be added to the database.

Adaptive-Limit Checking
Adaptive limit checking techniques utilize some ML algorithms like SVM and GA to produce the range limits for sensor value dynamically and automatically. Compared with the solid limit it produces more accuracy and reduces misinformation.
Summarily, in the training phase, a training dataset that can either be obtained from archived telemetry data or simulation data was used to calculate the PCA matrix and train SVMs. In the real-time detection phase, the real-time telemetry data of the spacecraft system is processed. The approach first transforms the input data to a low dimensional feature space using PCA matrix to obtain feature vectors. These feature vectors then pass through the Binary SVM to detect whether there is a fault or not. If the fault is detected, a multi-class SVM is used to identify the fault type. From experimental results, the authors showed that the method was efficient and practical for FD and diagnosis of a spacecraft system.

D. LIMITATIONS OF PROPOSED APPROACHES TO ANOMALY DETECTION AND FAULT DETECTION
Challenges central to AD in multivariate time series data also apply to spacecraft telemetry. This implies that a lack of labeled anomalies necessitates the use of unsupervised or semi-supervised approaches [134]. Having analysed the proposed approaches to AD/FD in [122] it was gathered that the limitations of each approach will emanate from the limitation(s) of the algorithm(s) used for the data-mining. For example, the adaptive limit checking approach used ML algorithms like SVM, Rough Set Method, and GA to dynamically and automatically produce the range limits for sensor value. Each algorithm utilized different historical telemetry data and produced different efficiency. Generally, SVM has some significant weaknesses, among which are: algorithmic complexity that affects the training time of the classifier in large data sets, development of optimal classifiers for multi-class problems and the performance of SVMs in unbalanced data sets [135].
In spite of its weaknesses, SVM has demonstrated highly competitive performance in many real-world applications. However, for large data sets the training kernel matrix grows in quadratic form with the size of the data set, thus making the training of SVM on large data sets a very slow process [135]. This is expected to be the case for spacecraft telemetry data, considering that they are usually highly complex, highly dimensional and multi-dimensional [122].
Gao et al. in [128] used binary SVM to detect anomalies. However, SVM is designed for the classification of two classes, which makes it unsuitable for fault diagnosis. Considering that practical fault diagnosis has to deal with multi-type faults that involve relationships among a large number of parameters, they proposed two approaches for constructing a multi-class SVM: combining several binary SVMs and implementing multi-class classification. Their approach was implemented using telemetry data from an actual in-orbit satellite and simulated by Matlab/Simulink [136] [128]. Using 25000 points of data the accuracy of their approach was verified using two case scenarios. In Case 1, all 25000 points of data were used as training data, after which the same was used as test data to gauge the performance. In Case 2, the 25000 points of data were divided into two sets by cross-extraction. The first set was used for training while the second was used as test data.
The results show a classification accuracy of 99.2% and 97.4% for Case 1 and Case 2 respectively, thus demonstrating that the method is efficient and practically suitable for a spacecraft system's FD and diagnosis. Table V compares Multi-class SVM approach with ANN. From the table, it can be seen that for Case 1 the accuracy of SVM and ANN are quite close. However, for Case 2 SVM performed better that ANN because of SVM's extraordinary generalization capability, along with its optimal solution and discriminative power; especially when the number of input data is small [135]. In addition, SVM has a better performance when the training and test data are not the same [128]. Notwithstanding the dividends, a key challenge of this approach is the use of PCA, a linear dimension-reduction technique not suited for nonlinear systems. Given that some components of the spacecraft system are nonlinear, the authors acknowledged the need of a nonlinear PCA (e.g. KPCA) to be used in extracting features from input data in their future work [128].

II. REMOTE SENSING
Remote sensing refers to the science of identification of earth surface features and estimation of their geo-bio-physical properties using electromagnetic radiation as the medium of interaction [137]. Images acquired remotely over long distances are usually affected by noise and other environmental conditions like cloud coverage; which makes them differ widely in terms of colour variations and textural contrasts [138]. Furthermore, captured data needs to be transmitted to the ground station for aggregation and analysis, which can be expensive. However, a satellite can reduce the amount of data transmitted by employing DL for on-board pre-processing. Parts of the image of no interest (e.g. those occluded by clouds) can be discarded so as to drastically reduce the amount of data transmitted [139], [15]. Fig. 10 shows a diagram illustrating six satellite image processing methods, out of which we shall treat the first two.

1) ENHANCEMENT
Enhancement methods applied to preprocessed data rely on metaheuristic-based algorithms to manipulate the digital pixel values for effective visual interpretation of images. Two algorithms employed in image enhancements are Particle Swarm Optimization (PSO) and GA. However, among the evolutionary algorithms GA and PSO suffer from the limitation of getting trapped in local minima [138]. One way to overcome such limitation is by combining various optimization techniques. Combining Cuckoo Search (CS) algorithm with PSO has shown better results in terms of closeness to optimal solution when compared to PSO or GA individually [71], [138]. In [140] Suresh et al. proposed a new method-Modified Differential Evolution (MDE) algorithm for contrast and brightness enhancement.

2) CLASSIFICATION
In a general sense, image classification can be defined as -the process of categorizing all pixels in an image or raw remotely sensed satellite data to obtain a given set of labels or land cover themes [141]. In supervised classification, available known pixels are used by the analyst to generate representative parameters for each class of interest. Contrastingly, in unsupervised classification pixels are grouped according to the reflectance properties of pixels. These groupings are called clusters. Two popular clustering algorithms are K-means and Expectation Maximization.
One key advantage of the unsupervised classification technique is that it can be used when no sample sites exists. Although image classification is dependent on the type of satellite image used, classification followed by other stages such as de-noising and segmentation can lead to better processing of the image [138]. Commonly used image classification algorithms include fuzzy algorithms, ANN and ES. The main advantage of Radial Basis Function Neural Network (RBFNN) is its immunity to noise signals; perhaps due to its large number of tunable parameters [142].

B. CHALLENGES WITH REMOTE SENSING
Common challenges associated with satellite image processing methods have been reported in literature and include the following: image complexity, large image sizes, presence of unwanted artifacts and background information. In addition, satellite sensor variations (such as radiometric resolution, spatial, and spectra) and a change of viewpoint bring tremendous diversities into satellite image representations [143]. Images of the above said different variations could be integrated by fusion techniques to enrich the chosen study area's available information.

1) VOLUME
The volume of data generated from satellites is already far more than human imagery analysts can effectively analyze. It is growing exponentially, thus necessitating the need for CV techniques applied to satellite imagery. Recent breakthroughs in DL have also improved to the point where DL now outperforms humans in some tasks like classifying objects in images. However, current open-source AI tools are not optimized for satellite imagery [144].
At present, even though a wide range of techniques are available for image processing, it is extremely difficult to settle for a technique which can be generally applied to all types of satellite images. Owing to the different color and textural variations and the limitations of traditional remote sensing using optical sensors, Synthetic Aperture Radar (SAR) imaging, an alternate form of remote sensing, permits the observation of large target areas during day or night under almost any weather condition [145], [146]. Although SAR has shown significant benefits over traditional remote sensing techniques, it comes with additional complexities [147]. Researchers have begun to employ advanced ML techniques (DL) to SAR data to cope with these challenges adequately. Yionel et al. [154] introduced a DL framework for inverse problems in imaging and demonstrated the advantages and applicability of this approach in passive SAR image reconstruction. As a preliminary study, Chen and Wang in [155] used a single layer of Convolutional Neural Network (CNN) to learn features from SAR images automatically. DL is nowadays being frequently utilised in SAR data applications by leading companies like Orbital Insight and Descartes Labs specializing in satellite earth observation analytics [156].

2) COMPLEXITY
Traditional satellite image models are predominantly based on statistical methods like: Linear Discriminant Analysis (LDA); Maximum Likelihood Estimation (MLE); PCA; and other regression-based models. Unfortunately, most traditional algorithms are plagued with a lack of logical reasoning. With ML techniques vibrant progress has been made in incorporating human-level reasoning into these traditional models [157]. Despite the merits of ML methods, supervised ML techniques suffer from excessive computational complexity due to demanding training processes and insufficient ground truth data for labeling. To overcome these challenges, the training phase must be limited. This can be achieved through an unsupervised mode of learning. Furthermore, the accuracy of classification, either in supervised or unsupervised mode, can be improved using hybridization. Parallel processing of data could also be used to improve the issues with the speed of computation through higher-level Graphical Processing Unit (GPU) architectures. However, the development of this method is still in progress [157]. Table VI shows a few statistical models utilized in satellite image analysis. The performance of K-means clustering and neural network with back-propagation for successful image segmentation and classification in satellite images employing dense count values of the pixel intensities were compared in [158]. However, using more advanced intelligence schemes like DNN in satellite image analysis could improve the level of reasoning with a reduction in training complexities [159]. Complexities associated with the training of deep networks with a vast number of hidden layers can be resolved to a greater extent by employing a greedy learning-based algorithm [157]. Greedy algorithm or search is an efficient tool that is usually employed in optimization problems, especially when dealing with large sets of data [160].

3) INEFFICIENCY
A fundamental challenge with statistical models is that they are inefficient in prediction due to the irregularly varying patterns of voluminous data [161]. Thus, ML techniques are better in deriving prediction models as it learns through experience during the training phase. In [162], two ML techniques (SVM and CNN) were developed to extract human settlements from Very High Resolution (VHR) satellite images of 3 provinces in Afghanistan. By comparing the results with analyst-verified reference data information (LandScan Settlement Layer), the authors demonstrated that in terms of overall pixel cells the CNN technique yielded more accurate results overall while based on derived statistics against the reference data the SVM technique performed more accurately in omission. Given the merits of ML techniques in deriving prediction models, the hybridization of different ML methods as shown in Table VII and the integration of statistical and ML methods could acquire more efficient results. In [157], Giorgio Giacinto et al. proposed an approach with a combination of neural and statistical algorithms in a simple design phase to disclose that every algorithm is significant in solving a particular issue, and no single algorithm is proved to be perfect. Castellana et al. [168] performed the combination of both supervised and unsupervised modes of change detection and on a pixel-based method to achieve better classification of remote sensing images. In [169], Pabitra Mitra et al. combined an active learning method with SVM and advanced an active SVM with reduced labeled data for classification [157].

C. PERFOMANCE EVALUATION OF IMAGE PROCESSING ALGORITHMS
Performance evaluation in a universal sense refers to the degree of some required behavior of an algorithm, whether it is attainable accuracy, adaptability, or robustness. It allows the innate characteristics of an algorithm to be emphasized, as well as the evaluation of its advantages and limitations [170].The justification for evaluating an algorithm is to understand its behavior in dealing with different categories of images, and/or help in estimating the best parameters for different applications [171]. Although a wide range of techniques exist for image processing it is immensely cumbersome to arrive at a technique which can be commonly applied to all types of satellite images. This is due to the different color and textural variations of the images.
Assessing the performance of any algorithm in image processing is demanding because performance depends on several factors, as surmised by Heath et al. [172]: (a) the algorithm itself, (b) the algorithm parameters used in the evaluation, (c) the method used for evaluating the algorithm, (d) the nature of images used to measure the performance of the algorithm. In [138] the authors analysed different metrics for evaluating the overall performance of the discussed image processing techniques. However, for the qualitative evaluation of a proposed technique some parameters need to be computed to enable its comparison with available techniques; in addition to evaluating its suitability and reliability over other techniques. Four computable parameters were analysed: Feature Similarity Index (FSIM), Structural Similarity Index (SSIM), Precision and Recall. However, only the first two shall be treated. In a nutshell, FSIM denotes the measure of similarity of features between the input image and the final image. It is computed using equation (1): where the whole image is represented by X, SL(x) represents the similarity in the two images and PC m is the phase congruency map. Structural Similarity Index (SSIM) indicates the structural similarity between the input image and the final image. It is calculated using equation (2): Where and represent the sample means of x and y respectively; and give the sample variances of x and y, respectively; and shows the sample correlation coefficient between x and y, where x and y are local windows in the input images. The calculated SSIM index is always a decimal value with range between -1 and 1. A value of 1 indicates perfect structural similarity while 0 indicates no structural similarity. 1 is only attainable in the case of two identical sets of data. Summarily, performance evaluation ought to rely on the use of performance indicators that convey the qualities of an algorithm. Considering the above suggestion, six typical performance indicators as surmised by Wirth et al in [173] include: (1) sensitivity: an algorithm's response to small changes in features; (2) robustness: an algorithm's capacity for tolerating various conditions; (3) accuracy: how well the algorithm has performed with respect to some reference; (4) adaptability: how the algorithm deals with variability in images; (5) reliability: the extent to which an algorithm, when repeated using the same stable data, yields the same result; (6) efficiency: the practical viability of an algorithm (time and space).

III. SATELLITE COMMUNICATION
In the near future (2020-2025), it is expected that satellite communication systems provide capacities close to 1 Terabit/s [174]. This would require the deployment of the next generation broadband infrastructure, including fixed and mobile internet access (5G) such as the KONNECT VHTS by EUTELSAT, expected to be launched in 2021. Therefore, to meet the requirement of capacity increment there is a trend towards the use of larger, more powerful Geosynchronous Equatorial Orbit (GEO) satellites, and/or flexible payloads.
Very high throughput satellite (VHTS) aim at achieving 1 Terabit/s per satellite in the near future and will play a key role in supporting future 5G and broadcast terrestrial networks [175]. Currently, VHTS systems provide uniform throughput over the entire service area based on multi-spot coverage with frequency and polarization reuse schemes using larger bandwidths in the feeder link in the frequency Q/V bands or by optic links. However, traffic demands are expected to be unevenly distributed over the service area since the user distribution is not uniform within the coverage. This results in a system where some beams lack the required capacity, i.e., not meeting the traffic demands, whereas other beams overcome the required capacity or, simply, wasting resources [176][177][178]. This challenge is addressed using flexible payload architectures to enable the allocation of payload resources in a flexible manner to meet the traffic demand of each beam. New propositions for flexible payloads systems must take into account the analysis of novel efficient cost function to optimize resources allocation and the analysis of the payload architecture and its adaptive allocation methods for new traffic demand scenarios [179].
Notwithstanding the merits of Dynamic Resource Management (DRM) approaches derivable from flexible payload architectures, DRM adds significant complexity to VHTS systems. The authors in [177] analyzed and proposed the use of CNNs to manage the resources available in flexible payload architectures for DRM. A comparison between different payload architectures in terms of DRM response is carried out and the CNN algorithm performance is compared with three other algorithms, previously suggested in the literature to demonstrate the effectiveness of the suggested approach and to examine all the challenges involved.
Optimizing the data return from space missions requires planning, design, standards, and operations coordinated from formulation and development stages throughout the mission. Autonomy, enhanced by cognition and ML, are potential methods for optimizing data return, reducing operational costs, and managing the complexity of automated systems [180]. By 2020-2025 it is envisaged that there will be more than 100 high throughput satellite (HTS) systems in GEO and mega constellations of Low Earth Orbit (LEO) satellites, delivering terabits per second of capacity across the world. These evolved satellite systems will provide Radio Access Networks (RANs), called satellite RANs, which will be integrated into the 5G system together with other wireless technologies, including cellular systems, Wi-Fi, and so on [175]. Considering the ever-increasing traffic distribution on earth the authors in [181] developed an Extreme Learning Machine (ELM) Distributed Routing (ELMDR) strategy. ELM, a fast and efficient ML algorithm, was adopted to forecast the traffic at satellite node. Simulation results demonstrate that in comparison to the conventional ACO algorithm ELMDR not only sufficiently uses underutilized link, but also reduces delay.

IV. ROBOTIC AUTONOMOUS SYSTEMS FOR SPACE
Robots designed for space missions operate in increasingly complex environments. This includes all types of robotic agents designed to explore a planet's surface and those deployed in orbit around extraterrestrial bodies. For those operating in completely known and static environments, the challenge of LTA shrinks down to one of robustness. Similarly, for autonomous agents operating in an unstructured environment, the operational environment could change over the lifetime of the agent [7].
Categorically, space robots can be broadly classified into two groups: Orbital and planetary. Orbital robots find application in the assembly of large space telescopes, satellite repair, and deployment of assets for scientific investigations on platforms such as the ISS. In contrast, planetary robots (e.g., Curiosity and Opportunity) are used for the survey, observation, and close examination of extraterrestrial surfaces like the planets (e.g., Mars).
The rest of this section surveys systems and approaches that address the challenges of LTA using AI techniques. Although AI may have witnessed a growing interest in the space industry over the last two decades staying current on ML advancements and knowing how to best leverage available technologies can be complicated. For instance, by introducing dexterous manipulators to traditional satellite platforms the spacecraft design becomes increasingly complex. Stemming from the high level of interconnectedness between the manipulator and its floating base, the manipulator becomes an intrinsic part of the overall spacecraft design [182]. Fundamentally, the whole satellite turns in into a 'space robot. ' Depending on its application, a space robot is generally required to possess two functional attributes: locomotion and autonomy. Locomotion or mobility impacts it with the ability to manipulate or interact with a sample. Though autonomy may vary from semi-autonomous to fully autonomous, the nature of the mission and distance from the Earth determines to a great extent the level of autonomy designed into the system. With time it is envisaged that future robots would evolve a higher level of autonomy to meet the increasingly challenging goals of space missions. Regardless of the category of autonomy, specific action plans are needed for a robot to complete its tasks. This requires optimization to ensure the conservation of resources such as energy and communication bandwidth [183].
It is worth stressing that advancements in general AI techniques are vital for improvement in many aspects of the autonomous functions of space robots. For instance, ML is often applied to sensing, perception (e.g., machine vision) tasks. It has also been applied to locomotion, such as in the improvement of locomotion strategies or policies and navigation. Planning, system-wide autonomy, scheduling, and resource allocation are also areas of continuing work for ML.
Considering the above scenarios, the consolidation of AI methods in space engineering is undoubtedly an enabling factor for improving the autonomy of space robots [13]. Irrespective of the domain LTA systems need to combine different AI abilities to cope with challenging environments and tasks. Inherently, the combination of abilities such as navigation & mapping, perception, and Knowledge Representation (KR) presents a functional integration challenge [7,184]. The challenge becomes even more significant when other AI abilities like planning, interaction, and learning are thrown into the mix. The following highlight three core subdomains of RAS where AI finds application. FIGURE 11. Key Modules of a Typical Robotic Perception System [185] Robots have contributed immensely in helping humankind explore space, thereby contributing to expanding the frontier of scientific knowledge and our access to the extraterrestrial world. In 1970, the former Soviet Union built the first robot (Lunokhod) that traveled on the surface of an extraterrestrial body. It was remotely operated from Earth and traversed more than 10.5 km on the moon [186]. Within the past four decades, the AI community has witnessed a resurgence of interest in autonomous navigation and planning models for robots, especially the challenge of Simultaneous Localisation and Mapping (SLAM) [187]. When cameras are employed as the only exteroceptive sensor, it is called visual SLAM [188]. The challenge with autonomous navigation of mobile robots is categorized into three principal areas: localization, mapping, and path planning [189]. Localisation entails determining exactly the current pose of a robot in an environment, whereas mapping blends the fragmented observations of the surroundings into a single congruous model (What does the world look like?). Path planning concerns itself with determining the best route in the map to navigate through the environment (How can I reach a given location?) [188].

1) NAVIGATION AND PLANNING
Motion planning is extension of path planning. Although the terms appear to be the same a few differences exist. Path planning seeks to find the path between origin and destination in workspace by strategies like shortest distance or shortest time. Contrastingly, motion planning focuses on generating interactive trajectories in workspace when robots interact with dynamic environment. As a result, motion planning needs to consider kinetics features, velocities and poses of robots and dynamic objects nearby. therefore path is planned from topological level [190]. In 1986, Smith et al. [191] introduced the concept of implementing spatial uncertainty estimation, which gave birth to the predevelopment of SLAM technique. This was followed in 1991 by Leonard et al. in [192] developing a SLAM technique based on [191], which used a probabilistic approach in solving the SLAM problem. This gave birth to the implementation of the Extended Kalman Filter (EKF) method, which laid the foundation for the first SLAM algorithm known as the EKF-SLAM [193]. In 2001, [194] described the substantial implementation of the SLAM algorithm on a vehicle operating in an outdoor environment using Millimeter-Wave (MMW) radar to provide relative map observations. VOLUME XX, 2021 In real environments with many landmarks present Kalman filter-based algorithms require time quadratic in the number of landmarks to incorporate each sensor observation. In light of this drawback Montemerlo et al. [195] introduced FastSLAM, an algorithm that recursively approximates the entire posterior distribution over robot pose and landmark locations, yet scales exponentially with the number of landmarks on the map. FastSLAM uses a hybrid technique that integrates Particle Filter and EKF approach, making it popular for its higher data accuracy [193,195]. FastSLAM evolved to FastSLAM 2.0 a year later with a conceptually simple modification. In FastSLAM, when proposing a new robot pose, the proposal distribution relies only on the motion estimate. In contrast, in FastSLAM 2.0 it relies on both the motion estimate and the most recent sensor measurement. This approach was found to be less wasteful with its samples than the original FastSLAM algorithm, especially in circumstances where the noise in motion is high relative to the measurement noise [196].
Many modifications have been made to SLAM algorithms over the years. These include the Squared Root Smoothing and Mapping (SAM) method proposed in 2006 [197] for improving the mapping process of mobile robots; the UFastSLAM proposed in 2007 [198] to improve the FastSLAM method by using unscented transformation algorithm and the Differential Evolution technique proposed in 2009 [199].
Research in motion planning is currently witnessing a lot of attention as a result of advancements in DL and RL. This is due to their better performance in coping with non-linear and complex problems. Basically, motion planning algorithms can be divided into two broad groups: traditional algorithms and ML-based algorithms according to their principles and the era they were invented [190]. They are further categorized as shown in Table VIII.
Traditional algorithms can be classified into three categories (graph search algorithms, sampling-based algorithms and interpolating curve algorithms) while ML-based planning algorithms are based on ML techniques that include supervised learning (e.g. SVM, optimal value RL and policy gradient RL).

2) PERCEPTION
For deep space missions that suffer from lags and gaps in communication, onboard DL systems would bring a much needed degree of autonomy and robustness to GNC. In terms of LTA and performance in space environments, robotic perception is critical for planetary and orbital robots to make decisions, plan, and operate in real-world environments, utilizing functionalities ranging from occupancy grid mapping to object detection. These days, most robotic perception systems rely on ML techniques, ranging from classical to DL approaches where the learning can be in the form of unsupervised learning or supervised classifiers using handcrafted features, or DL neural networks [185]. As depicted in Fig. 11, the fundamental components of any AI perception system are sensory data processing, data representation (environment modeling), and ML-based algorithms. However, in the majority of applications, the fundamental role of environment mapping is to model data from exteroceptive sensors mounted onboard the robot to facilitate reasoning, and inference regarding the robot's realworld environment [185].  [186] The Mars Exploration Rover (MER) shown in Fig. 12 is equipped with two stereo imaging systems on a camera mount on the bar of the rover mast: Panoramic (Pancam) and Navigation camera (Navcam). The panoramic camera is dedicated to the mapping of medium-to-far objects in panoramic images, while the other is a navigation camera with the best focus at 1 m with a field of view (FoV) of 45 0 [186,213]. Early approaches to mapping of dynamic environments were object-centric methods that identified and removed moving objects from the maps [214] or used them as moving landmarks for self-localisation [215]. Nevertheless, not all dynamic objects really move at the instance of mapping, necessitating that their identification requires long-term observations. Ambrus et al. [216] addressed this challenge by processing several 3D point clouds of the same environment recorded in several weeks to VOLUME XX, 2021 identify and separate movable objects and refine the static environment structure simultaneously.

3) LEARNING AND INTERACTION
Learning is critical for any robotic system deployed in open or dynamic worlds where autonomy is needed to maintain reliability, robustness, and cut operational cost. However, some of the most challenging application domains for longterm running robots, such as the MER, involve interacting autonomously with a diverse range of objects, which implies that being able to learn and adapt from experience is crucial to the success of the mission. Robots or systems that need to interact autonomously with their environment through sensory-motor capabilities must possess the capacity to act deliberately to fulfill their mission [184], especially in environments requiring LTA. Techniques which allow robots to continually learn from experience such as reinforcement learning, or focus on particular experiences (e.g., failures, novelty) such as learning from demonstration should allow online improvement of capabilities [98].
Considering the interest in these research fields in Robotics and AI communities, there are several publications and survey papers that review state of the art for a few focused deliberation functions such as planning, goal reasoning, monitoring, and recognition of actions and plans [184,217].  [218] Robotic perception, as observed by Sünderhauf et al. [219] differs from traditional CV perception. In robotics, the expected outputs of a perception system will result in decisions and actions in the real world. In contrast, most CV applications take images and translate the outputs into information. This shows that perception plays a critical role in the learning and interaction of any goal-driven robotic system.
Like localization and navigation, robot perception functions depend on the robot's operational environment [185]. For artificial agents like robots to interact with their real-world environment, perception and manipulation must complement each other. Once a robot is (self) localized, it can proceed with the execution of its task by taking deliberate action, or a set of interactive actions. In the case of autonomous mobile manipulators in a typical open/dynamic setup, this involves navigating to the region of interest, observing the current scene to build a 3D map for collisionfree grasp planning, and localising target objects in the operating environment and grasping them [185].
The technical feasibility of robotic servicing has been sufficiently demonstrated on the Hubble Space Telescope and the ISS where robots carry out in-orbit activities like inspection, component assembly, and docking [20,220]. Fig.  13 shows Dextre, a two-armed robot or telemanipulator that is part of the Mobile Servicing System on the ISS. Equally known as Special Purpose Dexterous Manipulator (SPDM), Dextre performs repairs that would have otherwise require spacewalks [221]. However, to achieve the vision of on-orbit servicing for satellites, developing a new design and operation paradigms of satellite architectures is necessary [20]. In [17], Nanjangud et al. provide an overview of the RASs technologies that enable O3 on smallsat platforms. Robotic O3 involves a robotic agent (or chaser) operating on a client spacecraft (also called a target), which can be further classified as cooperative or non-cooperative targets.

IV. CHALLENGES AND LIMITATIONS
Despite the risks posed to humans and robotic explorers operating in space's extremely challenging environment, humankind has not given up on the desire to conquer space and other planets. In 2017, seven nations attempted a total of 90 known orbital launch attempts from spaceports in eight different countries. As shown in Fig. 14 we can increase the levels of autonomy and automation, thereby freeing humans to focus on tasks for which they are better suited [223]. However, there are still challenges that VOLUME XX, 2021 stand in the way of achieving full-scale autonomy and automation. To combat these, we would require advances in robotic sensing and perception, mobility and manipulation, rendezvous and docking, onboard and ground-based autonomous capabilities. For the next foreseeable future, it appears we would still need human-robot integration and suite of other data analysis tools to explore space. Although ML, a subdomain of AI, can be used to enable LTA in spacecraft and improve the science return of space missions, the requirements of the missions alone severely limit the use of many current ML approaches [117].

A. AUTONOMY
Space robots are often designed to possess mobility (or locomotion) to manipulate by griping, roving, drilling, and sampling. To a large extent, these functionalities are influenced by the operational environments (either orbital or planetary). Depending on the type of mission and distance from the Earth, these robots are expected to possess a varying autonomy level, ranging from teleoperation to fully autonomous operation [114]. In most cases, the onboard autonomy deployed in spacecraft consists of the use of a planner. For example, the autonomous capabilities of NASA's Opportunity rover came from MAPGEN, a mixedinitiative task planner and an autonomous navigation system. It was used to create daily mission schedules automatically, which were then refined by terrestrial scientists [224].
There are already several known factors that frustrate the realization of fully autonomous operation in space and limit the use of many current ML approaches. First, space missions have an extremely high cost of failure with little or no opportunity for external aid or repair. Frequent failures of GEO satellites result in high economic costs for governments and private companies, thereby resulting in the continuous increase of GEO debris and crowding of the GEO orbit [225]. Stemming from the limitations of the traditional satellite architecture paradigm and the dearth of a maintenance industry for satellites, Saleh et al. [226] propose O3 to provide flexibility to decision-makers in satellite design and operations in the industry. Any autonomy derived from O3 and provided for by AI techniques must be proved to be reliable, robust, and constrained from posing any threat to the spacecraft's core operations, station-keeping and health. In [225], Liang et al. proposed a universal O3 in GEO, consisting of a 7-DOF redundant manipulator (with replaceable end-effectors, a 2-DOF docking mechanism, a set of stereo vision and general subsystems of a traditional spacecraft platform.
In contrast to a classic payload, which is usually detached from the platform, a robotic device attached to a satellite becomes an integral part of the spacecraft itself. By introducing dexterous manipulators to traditional satellite platforms, Jaekel et al. [182] argue that the spacecraft design becomes increasingly sophisticated and complex. The high interdependencies between the manipulator and its floating base automatically turn the whole satellite into a 'space robot.' Consequently, a sophisticated combination of traditional concepts and AI will be needed for system FDIR. When dealing with FDIR equipping a spacecraft with an onboard data analysis system can enable the detection of and reaction to dynamic events. For example, a satellite's timely reaction to an asteroid on a collision path would be possible if the event was detected onboard and the spacecraft equipped to react. In contrast, with ground-based analysis, such real-time reactions are not possible.
Although the integration of AI techniques at the system level is essential for the functional realization of LTA, there is little research to no standard on combining modules from different areas of AI. With the help of robotic software development [227] and robotic middleware such as the Robot Operating System (ROS) [228], researchers and roboticists have been provided with methods to integrate their software components and other people's components in a structured way, thus improving software maintainability and reusability. Examples of such middleware on which frameworks can be built upon include (STRANDS [229]) for long-term navigation planning & task scheduling; (ROSPlan [230]) for planning and execution and (RoboSherlock [231]) for knowledge-enabled perception.
Having a framework to build on makes it easier for integration and the use of different AI methods. However, our extensive survey reveals that there is still a lack of understanding and research in the domain of system-level integration. Although the space robotics market size in 2018 crossed USD 2 billion and the industry is growing rapidly, we believe that system-level integration of AI methods and their evaluation in autonomous systems research is still a significant challenge in academia and industry [232].

B. RADIATION
The high levels of radiation that occur in space can precipitate various problems for sensitive electronic components like solid-state memory, microprocessors, and network interfaces. Adverse effects of radiation on unprotected board computation could range from complete burnout to the occasional bit flips in memory that can corrupt some data. In a study of the effect of radiation-corrupted RAM on different clustering algorithms [233], a method was developed to simulate radiation-induced bit flips and quantitatively assess the sensitivity of clustering and classification algorithms likely to be deployed onboard spacecraft. The findings surmised that the k-means algorithm could withstand radiation in the Earth orbit environment without the need for radiation-hardened memory. It also found out that simpler algorithms (regular k-means clustering and linear SVM) have less sensitivity (more tolerance) than more complex versions (kd-k-means, Gaussian SVM [117,233].
Kd-k-means, a faster version of the clustering algorithm that stores the data set as a kd-tree in memory, was discovered to be much more sensitive to radiation and not recommendable for onboard use on spacecraft. As a result, a VOLUME XX, 2021 kd-tree variant was subsequently developed and restructured to increase its robustness to radiation [234].

C. COMPUTATIONAL CONSTRAINT
A significant challenge arising directly from the highradiation environment is the computational constraints under which remote spacecraft are forced to operate. A few reasons contribute to this problem. First, processors and memory chips used in space need to pass extremely rigorous testing to certify them fit for operation in the harsh environment of space. Unfortunately, the amount of testing and preparation spacecraft systems designers require to get a system ready for spaceflight often takes years. As a result, engineers, designers, and quality assurance experts prefer to rely on legacy processors and components that have been flown and used in older spacecraft. This implies that even more recent spacecraft run on processors and memory that lag a decade behind the desktop state of the art [235].
Since most modern electronic components were not manufactured to operate in high-radiation environments, those meant for use in space must undergo a mandatory radiation hardening process, thereby increasing their cost. For example, the RAD750 processor used by Deep Impact, Mars Reconnaissance Orbiter, the Kepler space telescope, and other current missions and instruments runs at only 133 MHz and costs ∼$200,000 [117]. Alternatively, there are specially shielded packaging of electronic parts to keep radiation at bay. However, this approach is expensive and can lead to an increase in the weight of the spacecraft, which could adversely influence launch costs. Two other ways of dealing with space radiation include using redundant subsystems and selective shielding [236].

D. NATURE OF TELEMETRY DATA
Satellite telemetry data are largely heterogeneous and multimodal, made up of hundreds to thousands of variables and attributes. In addition, telemetry data is usually collected from different sensors in various subsystems with different output formats and operating modes that must be changed from one mode to another over time ( [122] [129], [237]). According to Yairi et al. [14] there has been little work done in artificial satellite telemetry data mining, even though in the last few years the applications of SI in data mining have become increasingly implemented in different fields like bioinformatics and forensic analytics. However, in the research community of ML, some researchers are becoming interested in applying data mining techniques to the health monitoring problems for space missions.
Notwithstanding that some SI algorithms have shown favourable and effective results in many application areas, they still suffer from defects such as weak generalization ability, low classification accuracy and slow convergence among others. Furthermore, evolutionary algorithms require some randomness to proceed [238], all problems to which Chaos theory can be applied to improve the performance of the evolutionary algorithm. In [239], Darwish et al. proposed the hybridization of SI and evolutionary computing algorithms for DL and how these integration solve some problems in satellite health monitoring based on ML data mining. Gehad et al. [238] presented a novel Chaotic WOA (CWOA) to overcome these problems. The performance of CWOA was compared with WOA and ten other optimization algorithms, after which experimental results showed that circle chaotic map was the best chaotic map that significantly boosted the performance of WOA [240].

D. DEEP LEARNING TECHNIQUES
Space systems can be quite expensive to develop and launch into outer space. Due to the gradual degradation of system components and/or devices, it is highly recommended to have an on-board fault-diagnosis system capable of detecting, isolating, identifying or classifying faults in the spacecraft [123]. Historically the majority of the automated control and navigation maneuvering operations were directed by a human, which is burdensome and only attainable for near Earth missions [15]. Considering the need to introduce higher levels of autonomy and automation using AI, DL will continue to play a critical role in developing mission-defined RAS that are capable of surviving the hostile environment of space. However, there is often the temptation among mainstream ML communities to assume that the challenges with DL in domains like CV and Natural Language Processing (NLP) are the same for robotics [219]. Unfortunately, this is not true. Sunderhauf et al. [219] discussed a number of robotics-specific challenges for DL. As shown in Fig. 15, these problems were categorized into three conceptually orthogonal axes: learning, embodiment, and reasoning. For instance, under the learning axis, -Identifying Unknowns‖ is a key challenge because of the common assumption in DL that trained models will be deployed under closed-set conditions [241]. However, robots are often deployed in ever-changing, uncontrolled real-world environments, and are bound to come into contact with environmental conditions that were not covered by the training data, including instances of classes, scenarios and textures [219].
In open-set conditions [242], [243] it is critical to identify the unknowns. The robot's perception system must not assign high-confidence scores to unknown objects, or falsely recognize them as one of the known classes. For instance, if a robot's vision system (object detection) is fooled by data outside of its training data distribution [244], the consequences of acting on false but high-confidence detections can be disastrous.

E. LEGAL CHALLENGES
The current space treaties do not address the use and regulation of AI in space. The legal challenges emanating from the reliance and use of AI in space necessitates ascertaining the existence of linkage between space systems and services using AI to a system of governing rules and guiding legal principles [245]. Harnessing AI and ML technologies in the exploration of outer space will in all likelihood span a broad array of intended and unintended consequences such as privacy and liability issues. These consequences require consideration of a broad range of legal and regulatory concerns that the space industry alone cannot answer. The following are three core topics worthy of interest.

1) LIABILITY OF INTELLIGENT SPACE OBJECTS
The growing delegation of decision making to AI will have repercussions on many areas of law for which mens rea, or intention, is required for a crime to have been committed. As machines increasingly take on tasks and decisions traditionally performed by humans, should we consider giving AI systems 'personhood' and moral or legal agency? [246]. In the legal arena, the term -person‖ generally refers to an entity which is subject to legal rights and duties. Generally, we think of a person as a human being [247]. However, the legal rights and duties imposed on inanimate objects and artificial entities emanate from actions or conduct engaged in by human beings. This is undoubtedly not the case for actions or conduct taken based on AI. Although a machine can learn independently from human input and make decisions based on its learning and available information, that ability does not necessarily equate with natural or legal personhood [245].
Liability under the space law treaty regime is grounded in Outer Space Treaty (OST) Article VIII, which is the genesis of the Liability Convention [245]. The Liability Convention establishes a restricted framework for assessing international liability which only applies to a launching State [248]. As noted, decisions and conduct of legal persons are ultimately decisions made by a human being. Ultimately, since fault liability under the Liability Convention Article III is premised on the fault of a state or the faults of persons, a decision by an intelligent space object will, in all likelihood, not be the -fault of persons‖ [245].
Existing liability models may be inadequate to address the future role of AI in criminal activities [249]. For example, while autonomous agents can carry out the criminal act or omission, the voluntary aspect of actus reus would not be met, since the idea that an autonomous agent can act voluntarily is debatable. This implies that agents, artificial or otherwise, could potentially perform criminal acts or omissions without satisfying the conditions of liability for that particular criminal offence [246]. In the event that criminal liability is fault-based, it also requires mens rea (a guilty mind). The mens rea may comprise an intention to commit the actus reus using an AI-based application, or knowledge that deploying an autonomous agent will or could cause it to perform a criminal action or omission [250].

2) DATA PROTECTION AND ETHICAL CHALLENGES
AI also raises important ethical and privacy concerns that could erode trust in emerging technologies if not addressed thoughtfully. AI requires access to vast amounts of data, but poorly drawn laws and government policies can hinder beneficial access without reducing the risk of AI activities [251]. The fear that satellite imagery can be used to discern car plates, individuals, and -manholes and mailboxes‖ is not fictional. In 2013, police in Oregon, used Google Earth satellite image depicting marijuana growing illegally on a man's property. In January, 2020, the United States imposed an immediate interim export controls regulating the dissemination of AI technology software that possesses the ability to automatically scan aerial images to recognize anomalies or identify objects of interest, such as vehicles, houses, and other structures [252]. Employing AI in satellite imaging presents ethical issues relating to loss of control over one´s personal information and activities, which encompasses the right of individuals to move in their own home (yards and gardens) and/or other non-public places without being identified, tracked or monitored [253]. Nevertheless, a larger influx of data, observation capabilities and high-quality imagery from EO satellites is expected to become more widely available on a timely basis [254].

3) LIMITATIONS OF SPACE LAW
There are no international or space treaties that address or regulate the use of AI in space. This simply implies that domestic legislation must serve as the principal source for the substantive law relating to the use of AI in space [245]. Furthermore, the dearth of international regulation of AI poses potential complex problems relating to the applicable substantive law in disputes involving the use of AI in space. It is not yet clear how ethical and legal concerns, especially around responsibility and analysis of decisions made by AIbased systems can be solved. Adequate policies, regulations, ethical guidance and a legal framework to prevent the misuse of AI in space need to be developed and enforced by regulators [245,255,256].

V. FUTURE DIRECTIONS
AI and blockchain are among the key disruptive drivers behind innovation today [257]. While AI has its fair share of issues with trustworthiness, explainability, and privacy, blockchain on the other hand suffers from shortcomings such as scalability, security and efficiency. By leveraging advances in AI and blockchain platforms we can demonstrate the capabilities of a new framework to collect vast amounts of data. The integration or marriage of these two technologies seems inevitable; especially since they have the potential to complement each other and revolutionize the next digital generation [257,258].
Regardless of the undisputed challenges with AI, several developments in algorithmic improvements have boosted the performance of DL methods and their models' accuracy. However, the large datasets required for trainings tend to be generally proprietary, highly centralized and expensive to recreate. Furthermore, published models soon become obsolete and in need of retraining with new data. In view of the above challenges there seems to be a push towards finding ways to collaboratively improve ML models hosted on public blockchains. Blockchain will bring trustlessness, privacy, and explainability to AI while will in turn help build an ML system on blockchain for better security, scalability, and more effective personalization and governance [257].
Researchers at Microsoft [258] proposed a new framework for collaboratively building a dataset and using smart contracts to host a continuously updated model. A free and open-source implementation of this framework for the Ethereum blockchain is provided at https://github.com/microsoft/0xDeCA10B. Similarly, the space industry has begun experimenting with blockchain technology across its entire supply chain to address some of the challenges faced by agencies like ESA and NASA. Going forward, we shall discuss three domains that stand to benefit the most from the inevitable AI-Blockchain convergence in the space community. They include the Internet of Space Things (IoST), Satellite Communications, and Spacecraft Data Security.

INTERNET OF SPACE THINGS (IoST)
The growing popularity of CubeSats has given rise to the practicability of ubiquitous cyber-physical systems known as the IoST/CubeSats [259]. Given the already mentioned challenges associated with the development and launch of traditional satellites CubeSats have become a viable alternative for building global satellite networks [260,261]. The concept of Internet of Space Things (IoST) utilizes Low Earth Orbit (LEO) satellites as part of a ubiquitous cyberphysical system for implementing true global connectivity. The system leverages Software-Defined Networking (SDN) and Network Function Virtualization (NFV) to integrate onthe-ground data and satellite information. Over the last couple of years research interest in the use of AI and SDR to manage networks and communication systems has gained momentum [262].
In [259] Kak et al. introduced a highly customizable largescale optimal constellation design framework for IoST with the aim of achieving global coverage and robust connectivity. Fortunately, the use of CubeSats offers several advantages. First, they make extensive use of Commercial-Off-The-Shelf (COTS) components which helps bring costs down. Secondly, through the use of sequential redundancy, CubeSats have much shorter development and deployment cycles. Furthermore, CubeSat constellations are more resilient to satellite failures due to the larger number of CubeSats in use. Notwithstanding the many use cases for IoST that focus on on-Earth or near-Earth applications, it is worth stating that IoST can also can find application in deep space exploration through interplanetary data relaying, sensing and monitoring of asteroids, Mars, and Moon [260,263].
For example, NASA aims to establish a human colony on Mars by 2025, which will require connectivity beyond Earth. The IoST, consisting of deep-space CubeSats, is expected to play a crucial role in providing such intra-galactic connectivity [264]. Currently, the utmost goal of IoST is to enable global connectivity beyond planet earth and provide sensing capabilities at a low cost [263]. Remote locations such as Mars that have little satellite coverage and connectivity can benefit from data sharing. Just as Earth's self-driving cars collectively share information about obstacle detection on a road, satellites or rovers could provide relative navigational information to each other and thus improve the accuracy of their space-based positioning and navigation.

SOFTWARE-DEFINED RADIO (SDR) AND SATELLITE COMMUNICATIONS
LEO satellite network plays a critical role in future spaceterrestrial integrated network because of its unique advantages. Notwithstanding, the effective and reliable routing for LEO satellite network is a difficult task due to time-varying topology, imbalanced communication load and frequent link handover [181]. This is where automation enhanced by cognition and ML can help with optimizing data return from space missions, reducing costs of operations and managing the complexity of communication systems. Although autonomous control systems without humans in the loop already exist for ground station network management, adding ML and cognitive algorithms will open up entire new fields of research with the potential to reduce complexity, enhance performance, and undoubtedly minimize the cost of space operations [181] [180]. NASA is currently on the verge of defining and developing future space and ground architecture for optimizing the data return from space missions. In [180] the authors discussed the potential role of ML in the link-to-link aspect of the communication systems. For the first time the advantages and disadvantages of applying ML to space links in actual flight environment was demonstrated in an experiment using NASA's Space Communication and Navigation Testbed onboard the ISS and the ground station located at NASA John H. Glenn Research Center.
Given the successful result, SDR has been identified as a key technology that provides the needed flexibility and configurability for NASA's future cognitive communication systems. Cognitive communication architectures will undoubtedly play a critical role in future space missions, providing seamless internetworking services for communication and navigation from within Earth's orbit out VOLUME XX, 2021 through Mars and other planetary exploration in deep space. However, achieving the above objectives might require integrating future 5G communication frameworks with different radio access technologies and AI-based Dynamic Spectrum Management (DSM) mechanisms. AI-based DSM mechanisms such as spectrum sensing, signal classification and dynamic spectrum access have been proven to achieve superior performance and robustness than conventional schemes. These mechanisms, enabled by Cognitive Radio (CR), Blockchain and AI, also provide more flexible and efficient means of implementing DSM. In the future, the combination of AI techniques and the DSM mechanisms would become a novel and promising research direction [175], [25,180].

SPACECRAFT DATA SECURITY AND THE BLOCKCHAIN FRAMEWORK
Considering that both AI and Blockchain deal with data to create value, the AI-Blockchain convergence is absolutely inevitable in the space industry. While data is central to AI's efficacy blockchain enables collaborative and secure data sharing. Therefore, tokenizing space resources such as satellites, telescopes and orbits in the form of blockchainbased digital tokens will open up new research areas and possibilities. Currently, Space Assets Management, Space Financing and Secure Satellite Communication represent some of the key areas that can hugely benefit from blockchain.
In 2017, NASA awarded a $330,000 grant to Dr. Jin Wei Kocsis of the University of Akron to support the development of an autonomous blockchain-based spacecraft system. The new system, called the Resilient Networking and Computing Paradigm (RNCP), relies on blockchain and represents NASA's first step toward blockchain adoption in space applications [265,266]. The authors in [266] also investigated the adoption of blockchain theory based on the space digital token concept and proposed a new conceptual blockchain space industry framework to address some major challenges facing the space industry.
Ant-inspired algorithms, when applied to the routing problem in wireless communication networks between satellites or planetary sensors achieved great efficiency [267]. However, communication security still remains a major challenge among spacecraft/satellites. Spacecraft and ground-based systems that control them are at risk of both active hacking and denial-of-service attacks. For most existing spacecraft communication security was insufficiently implemented, thus leaving significant attack vectors related to spacecraft control [268].
Blockchain can therefore be employed to help in securing satellites swarms' communications, managing and authenticating space transactions between those swarms and ground stations [269].
In order to understand how blockchain can be employed in this regard it is first necessary to understand possible communication patterns between satellites and blockchain system. In [270] these patterns are classified as four communication models and enumerated as follows: 1. A Satellite/spacecraft works as a blockchain node within the blockchain network; 2) A Satellite/spacecraft works as a validator (i.e. a miner) node within the blockchain network; 3) A satellite/spacecraft read from the blockchain; 4) A satellite requests a specific transactional data to be written to the blockchain.
With respect to the enumerated models above, Multi-Factor Authentication (MFA) can be used to verify a satellite's identity, ground station's identity, or communication pattern validity by requiring multiple security proofs [266]. For example, satellite A and Satellite B in a satellite swarm can use a code to authenticate a satellite's membership in the swarm. Take for instance a scenario where Satellite A requests a specific connection with satellite B. A sends the last block's Nonce code in the blockchain, which is verified by B before establishing the communication link. After terminating the connection a new block with a new Nonce code is provided, and then the new block is verified by all the miners in the blockchain network. Verification by miners (e.g. satellites and ground stations) proves the validity of connection and adds the new block to the blockchain system [266].
Data-sharing enabled by blockchain could also be used to address distrust by creating virtual trusted space zones in which rovers or satellites in swarms identify and update each other in a trustless network environment. For instance, when a specific satellite in its orbit is endangered with a space debris collision, it will update all satellites in the same swarm (i.e. in the same orbit) with the new information; the update being distributed as a digital token.

VI. CONCLUSION
As spacecraft systems become larger and exceedingly complex, AI techniques are needed to help with control, operations, and communications. For instance, the application of ML algorithms to various aspects of remote sensing, spacecraft health monitoring and communication offers the potential to improve throughput and data return to Earth from space missions. In this survey we analyzed the applications of AI in Spacecraft Health Monitoring, Remote Sensing, Satellite Communications, and RAS. Some BIAs were also surveyed and their potential application areas in space missions highlighted.
Regardless of the benefits derivable from the use of AI in space missions, there are still lots of challenges and open issues to be addressed. These hurdles need not to be seen as unassailable obstacles, but opportunities that point the way to where technological advancements are needed. For example, there seems to be a push towards finding ways to collaboratively improve ML models by hosting them on public blockchains. Similarly, SDR has been identified as a key technology that provides the needed flexibility and configurability for future cognitive communication systems. We encourage researchers to look into mathematical methods such as quantum computing and chaotic theory; and hybridizing them with bio-inspired computing to overcome some of the limitations associated with BAIs. Unlike similar surveys focused only on technological challenges, we thought that it was important to also address some of the key legal challenges that emanate from the reliance and use of AI in space.
Summarily, our paper outlined the need for advanced ML methods for space applications. ML has the potential to greatly increase these missions' capabilities, as well as enabling ambitious new autonomy possibilities in almost all fields of spacecraft operations. However, the combination of blockchain technology and AI will revolutionize space missions altogether as never before witnessed. This convergence has the potential to utilize data in ways never before thought possible. Given that data is the key ingredient for the development and enhancement of AI algorithms, blockchain holds the promise of securing this data and allowing us to audit all the transitional steps that these algorithms take to infer conclusions from the data.