JAN. 2021 1

Artificial Intelligence for Satellite Communication: A Review Fares Fourati, Mohamed-Slim Alouini, Fellow, IEEE

Abstract—Satellite communication offers the prospect of ser- 66 LEO satellites and 6 spares, Starlink by SpaceX plans to vice continuity over uncovered and under-covered areas, service have 4425 LEO satellites plus some spares, and O3b has 20 ubiquity, and service scalability. However, several challenges MEO satellites including 3 on-orbit spares [1]. must first be addressed to realize these benefits, as the resource management, network control, network security, spectrum man- Satellite communication use cases can also be split into agement, and energy usage of satellite networks are more chal- three categories: i) service continuity, to provide network lenging than that of terrestrial networks. Meanwhile, artificial access over uncovered and under-covered areas; ii) service intelligence (AI), including machine learning, deep learning, and ubiquity, to ameliorate the network availability in cases of reinforcement learning, has been steadily growing as a research temporary outage or destruction of a ground network due to field and has shown successful results in diverse applications, including wireless communication. In particular, the application disasters; and iii) service scalability, to offload traffic from of AI to a wide variety of satellite communication aspects the ground networks. In addition, satellite communication have demonstrated excellent potential, including beam-hopping, systems could provide coverage to various fields, such as the anti-jamming, network traffic forecasting, channel modeling, transportation, energy, agriculture, business, and public safety telemetry mining, ionospheric scintillation detecting, interference fields [2]. managing, remote sensing, behavior modeling, space-air-ground integrating, and energy managing. This work thus provides a Although satellite communication offers improved global general overview of AI, its diverse sub-fields, and its state-of- coverage and increased communication quality, it has several the-art algorithms. Several challenges facing diverse aspects of challenges. Satellites, especially LEO satellites, have limited satellite communication systems are then discussed, and their on-board resources and move quickly, bringing high dynam- proposed and potential AI-based solutions are presented. Finally, ics to the network access. The high mobility of the space an outlook of field is drawn, and future steps are suggested. segments, and the inherent heterogeneity between the satel- Index Terms—Satellite Communication, Artificial Intelligence, lite layers (GEO, MEO, LEO), the aerial layers (unmanned Machine Learning, Deep Learning, Reinforcement Learning aerial vehicles (UAVs), balloons, airships), and the ground layer make network control, network security, and spectrum I.INTRODUCTION management challenging. In addition, achieving high energy efficiency for satellite communication is more challenging than HE remarkable advancement of wireless communication for terrestrial networks. T systems, quickly increasing demand for new services in Several surveys have discussed different aspects of satellite various fields, and rapid development of intelligent devices communication systems, such as handoff schemes [3], mobile have led to a growing demand for satellite communication satellite systems [4], MIMO over satellite [5], satellites for systems to complement conventional terrestrial networks to the Internet of Remote Things [6], inter-satellite communica- give access over uncovered and under-covered urban, rural, tion systems [7], Quality of Service (QoS) provisioning [8], and mountainous areas, as well as the seas. space optical communication [9], space-air-ground integrated There are three major types of satellites, including the networks [10], small satellite communication [11], physical geostationary Earth orbit, also referred to as a geosynchronous space security [12], CubeSat communications [13], and non- equatorial orbit (GEO), medium Earth orbit (MEO), and low terrestrial networks [2]. Meanwhile, interest in artificial intel- arXiv:2101.10899v1 [eess.SP] 25 Jan 2021 Earth orbit (LEO) satellites. This classification depends on ligence (AI) increased in recent years. AI, including machine three main features, i.e., the altitude, beam footprint size, and learning (ML), deep learning (DL) and reinforcement learning orbit. GEO, MEO, and LEO satellites have an orbit around (RL), has shown successful results in diverse applications in the Earth at an altitude of 35786 km, 7000–25000 km, and science and engineering fields, such as electrical engineering, 300–1500 km, respectively. The beam footprint of a GEO software engineering, bioengineering, financial engineering, satellite ranges from 200 to 3500 km; that of an MEO or and medicine etc. Several researchers have thus turned to AI LEO beam footprint satellite ranges from 100 to 1000 km. techniques to solve various challenges in their respective fields The orbital period of a GEO satellite is equal to that of and have designed diverse successful AI-based applications, to the Earth period, which makes it appear fixed to the ground overcome several challenges in the wireless communication observers, whereas LEO and MEO satellites have a shorter field. period, many LEO and MEO satellites are required to offer Many researchers have discussed AI and its applications continuous global coverage. For example, Iridium NEXT has to wireless communication in general [14]–[17]. Others have focused on the application of AI to one aspect of wireless Fares Fourati and Mohamed Slim Alouini are with King Abdullah Univer- sity of Science and Technology (KAUST), CEMSE Division, Thuwal, 23955- communication, such as wireless communications in the Inter- 6900 KSA, (e-mail: [email protected], [email protected]) net of Things (IoT) [18], network management [19], wireless JAN. 2021 2

AE Autoencoder AI Artificial intelligence AJ Anti-jamming ARIMA Auto regressive integrated moving average ARMA Auto regressive moving average BH Beam hopping CNN Convolutional neural network DL Deep learning DNN Deep neural network DRL Deep reinforcement learning ELM Extreme learning machine EMD Empirical mode decomposition FARIMA Fractional auto regressive integrated moving average FCN Fully convolutional network FDMA Frequency division multiple access FH Frequency hopping GA Genetic algorithms GANs Generative adversarial networks GNSS Global navigation satellite system IoS Internet of satellites kNN k-nearest neighbor LRD Long-range-dependence LSTM Long short-term memory MDP Markov decision process ML Machine learning MO-DRL Multi-objective deep reinforcement learning NNs Neural networks PCA Principal component analysis Fig. 1. Applications of artificial intelligence (AI) for different satellite QoS Quality of service communication aspects RFs Random forests RL Reinforcement learning RNNs Recurrent neural networks RS Remote sensing security [20], emerging robotics communication [21], antenna RSRP Reference signal received power design [22] and UAV networks [23], [24]. Vazquez et al. [25] SAGIN Space-air-ground integrated network briefly discussed some promising use cases of AI for satellite SRD Short range dependence SVM Support vector machine communication, whereas Kato et al. [26] discussed the use of SVR Support vector regression AI for space-air-integrated networks. The use of DL in space SatIot Satellite Internet of Things applications has also been addressed [27]. UE User equipment Overall, several researchers have discussed wireless and VAEs Variational autoencoders TABLE I satellite communication systems, and some of these have ACRONYMSAND ABBREVIATIONS discussed the use of AI for one or a few aspects of satellite communication; however, an extensive survey of AI applica- tions in diverse aspects of satellite communication has yet to be performed. algorithms, challenges, achievements, and outlooks are also This work therefore aims to provide an introduction to AI, addressed. a discussion of various challenges being faced by satellite communication and an extensive survey of potential AI-based A. Artificial Intelligence applications to overcome these challenges. A general overview of AI, its diverse sub-fields and its state-of-the-art algorithms Although AI sounds like a novel approach, it can be are presented in Section II. Several challenges being faced traced to the 1950s and encompasses several approaches and by diverse aspects of satellite communication systems and paradigms. ML, DL, RL and their intersections are all parts potential AI-based solutions are then discussed in Section of AI, as summarized in Fig.2 [28]. Thus, a major part of AI III; these applications are summarized in Fig.1. For ease of follows the learning approach, although approaches without reference, the acronyms and abbreviations used in this paper any learning aspects are also included. Overall, research into are presented in Table I. AI aims to make the machine smarter, either by following some rules or by facilitating guided learning. The former refers to symbolic AI; the latter refers to ML. Here smarter indicates II.ARTIFICIAL INTELLIGENCE (AI) the ability to accomplish complex intellectual tasks normally The demonstration of successful applications of AI in necessitating a human such as classification, regression, clus- healthcare, finance, business, industries, robotics, autonomous tering, detection, recognition, segmentation, planning, schedul- cars and wireless communication including satellites has led it ing, or decision making. In the early days of AI, many believed to become a subject of high interest in the research community, that these tasks could be achieved by transferring human industries, and media. knowledge to computers by providing an extensive set of rules This section therefore aims to provide a brief overview of that encompasses the humans’ expertise. Much focus was thus the world of AI, ML, DL and RL. Sub-fields, commonly used placed on feature engineering and implementing sophisticated JAN. 2021 3

measure the performance of the algorithm [28]. This simple idea of learning a useful representation of data has been useful in multiple applications from image classification to satellite communication. ML algorithms are commonly classified as either deep or non-deep learning. Although DL has gained higher popularity and attention, some classical non-deep ML algorithms are more useful in certain applications, especially when data is lacking. ML algorithms can also be classified as supervised, semi-supervised, unsupervised, and RL classes, as shown in Fig.4. In this subsection, only non-RL, non-deep ML ap- proaches are addressed; RL and DL are addressed in sections II.C and II.D, respectively. Fig. 2. Artificial Intelligence, Machine Learning, Deep Learning and Rein- 1) Supervised, Unsupervised and Semi-supervised Learn- forcement Learning ing: Supervised, unsupervised and semi-supervised learning are all ML approaches that can be employed to solve a broad variety of problems. During supervised learning, all of the training data is labeled, i.e., tagged with the correct answer. The algorithm is thus fully supervised, as it can check its predictions are right or wrong at any point in the training process. During image classification, for example, the algorithm is provided with images of different classes and each image is tagged with the corresponding class. The supervised model learns the patterns from the training data to then be able to predict labels Fig. 3. Machine Learning Approach for non-labeled data during inferencing. Supervised learning has been applied for classification and regression tasks. As labeling can be impossible due to a lack of information handcrafted commands to be explicitly used by the comput- or infeasible due to high costs, unsupervised learning employs ers. Although this symbolic AI has been suitable for many an unlabeled data set during training. Using unlabeled data, applications, it has shown various limitations in terms of both the model can extract hidden patterns or structures in the precision and accuracy for more advanced problems that show data that may be useful to understand a certain phenomenon more complexity, less structure, and more hidden features such or its output could be used as an input for other models. as computer-vision and language-processing tasks. To address Unsupervised learning has been commonly used for clustering, these limitations, researchers turned to a learning approach anomaly detection, association and autoencoders (AEs). known as ML. As a middle ground between supervised and unsupervised learning, semi-supervised learning allows a mixture of non- labelled and labaled portions of training data. Semi-supervised B. Machine Learning (ML) learning is thus an excellent option when only a small part of ML, which encompasses DL and RL, is a subset of AI. the data is labeled and/or the labeling process is either difficult In contrast to symbolic AI, where the machine is provided or expensive. An example of this technique is pseudo-labeling, with all the rules to solve a certain problem, ML requires a which has been used to improve supervised models [33]. learning approach. Thus, rather than giving the rules to solve 2) Probabilistic Modeling: Probabilistic modeling as men- a problem, the machine is provided with the context to learn tioned by its name, involves models using statistical techniques the rules by itself to solve the issue, as shown in Fig.3 and to analyze data and was one of the earliest forms of ML best summarized by the AI pioneer Alan Turing [29]: ”An [30]. A popular example is the Naive Bayes classifier, which important feature of a learning machine is that its teacher uses Bayes’ theorem while assuming that all of the input will often be very largely ignorant of quite what is going on features are independent; as they generally are not, this is a inside, although he may still be able to some extent to predict naive assumption [28]. Another popular example is logistic his pupil’s behavior,” An ML system is trained rather than regression; as the algorithm for this classifier is simple, it is programmed with explicit rules. The learning process requires commonly used in the data science community. data to extract patterns and hidden structures; the focus is 3) Support Vector Machine (SVM): Kernel methods are on finding optimal representations of the data to get closer a popular class of algorithms [28], [31]; where the most to the expected result by searching within a predefined space well-known one of them is the SVM, which aims to find of possibilities using guidance from a feedback signal, where a decision boundary to classify data inputs. The algorithm representations of the data refer to different ways to look at or maps the data into a high dimensional representation where encode the data. To achieve that, three things are mandatory: the decision boundary is expressed as a hyperplane. The input data, samples of the expected output, and a way to hyperplane is then searched by trying to maximize the distance JAN. 2021 4

Fig. 6. Neural Networks Fig. 4. Machine Learning Sub-fields

more robust version of decision trees, random forests (RFs), combines various decision trees to bring optimized results. This involves building many different weak decision trees and then assembling their outputs using bootstrap aggregating (bagging) [37], [38]. Another popular version of decision trees, that is often more effective than RFs is a gradient boosting machine; gradient boosting also combines various decision tree models but differs from RFs by using gradient boosting [39], which is a way to improve ML models by iteratively training new models that focus on the mistakes of the previous models. The XGBoost [40], [41] library is an excellent implementation of the gradient boosting algorithm Fig. 5. Decision Tree that supports C++, Java, Python, R, Julia, Perl, and Scala. RFs and gradient boosting machines are the most popular and robust non-deep algorithms that have been widely used to win between the hyperplane and the nearest data points from each various data science competitions on the Kaggle website [42]. class in a process called maximizing the margin. Although mapping the data into a high dimensional space is theoritically 5) Neural Networks (NNs): NNs contain different layers of straightforward, it requires high computational resources. The interconnected nodes, as shown in Fig.6, where each node is a ’kernel trick’, which is based on kernel functions [32], is thus perceptron that feeds the signal produced by a multiple linear used to compute the distance between points without explicit regression to an activation function that may be nonlinear [43], computation of coordinates, thereby avoiding the computation [44]. A nonlinear activation function is generally chosen to add of the coordinated of a point in a high-dimensional space. more complexity to the model by eliminating linearity. NNs SVMs have been the state-of-the-art for classification for a can be used for regression by predicting continuous values or fairly long time and have shown many successful applications for classification by predicting probabilities for each class. In in several scientific and engineering areas [34]. However a NN, the features of one input (e.g., one image) are assigned SVMs have shown limitations when applied on large datasets. as the input layer. Then, according to a matrix of weights the Furthermore, when the SVM is applied to perceptual problems, next hidden layers are computed using matrix multiplications a feature engineering step is required to enhance the perfor- (linear manipulations) and then non linear activation functions. mance because it is a shallow model; this requires human The training of NNs is all about finding the best weights. expertise. Although it has been surpassed by DL algorithms, To do so, a loss function is designed to compare the output it is still useful because of its simplicity and interpretability. of the model and the ground truth for each output, to find 4) Decision Trees: A decision tree is a supervised learning the weights that minimize that loss function. Backpropagation algorithm that represents features of the data as a tree by algorithms have been designed to train chains of weights defining conditional control statements, as summarized in using optimization techniques such as gradient-descent [45]. Fig.5 [35], [36]. Given its intelligibility and simplicity, it is NNs have been successfully used for both regression and one of the most popular algorithms in ML. Further, decision classification, although they are most efficient when dealing a trees can be used for both regression and classification, as high number of features (input parameters) and hidden layers, decisions could be either continuous values or categories. A which has led to the development of DL. JAN. 2021 5

Fig. 7. Simplified Architecture of a Recurrent Neural Networks

C. Deep Learning (DL)

In contrast to shallow models, this sub-field of ML requires Fig. 8. Autoencoder high computational resources [28], [46]. Recent computational advancements and the automation of feature engineering have paved the way for DL algorithms to surpass classical ML al- gorithms for solving complex tasks, especially perceptual ones such as computer vision and natural language processing. Due to their relative simplicity, shallow ML algorithms, require human expertise and intervention to extract valuable features or to transform the data to make it easier for the model to learn. DL models minimize or eliminate these steps as these transformations are implicitly done within the deep networks. 1) Convolutional Neural Networks (CNN): CNN [47], [48], Fig. 9. Generative Adverserial Networks GANs are a common type of deep NNs (DNNs) that are composed of an input layer, hidden convolution layers, and an output layer RNN models are most commonly used in the fields of and have been commonly used in computer vision applications natural language processing, speech recognition and music such as image classification [50], object detection [51], and composition. object tracking [52]. They have also shown success in other 3) Autoencoders (AEs): AEs are another type of NNs used fields including speech and natural language processing [53]. to learn efficient data representation in an unsupervised way As their name indicates, CNNs are based on convolutions. The [55]. AEs encode the data using the bottleneck technique, hidden layers of a CNN consist of a series of convolutional which comprises dimensionality reduction to ignore the noise layers that convolve. An activation function is chosen and of the input data and an initial data regeneration from the followed by additional convolutions. CNN architectures are encoded data, as summarized in Fig.8. The initial input and defined by by choosing the sizes, numbers, and positions of generated output are then compared to asses the quality of filters (kernels) and the activation functions. Learning then coding. AEs have been widely applied for for dimensionality involves finding the best set of filters that can be applied to reduction [56] and anomaly detection [57]. the input to extract useful information and predict the correct 4) Deep generative models: output. Deep generative models [58] are DL models that involve the automatic discovering and 2) Recurrent Neural Networks (RNNs): RNNs [54] are learning of regularities in the input data in such a way that new another family of neural networks in which nodes form a samples can be generated. These models have shown various directed graph along a temporal sequence where previous out- applications, especially in the field of computer vision. The puts are used as inputs. RNNs are specialized for processing most popular generative models are variational AEs (VAEs) a sequence of values x(0), x(1), x(2), ..., x(T). RNNs use and generative adversarial networks (GANs). their internal memory to process variable-length sequences Of these, VAEs learn complicated data distribution using of inputs. Different architectures are designed based on the unsupervised NNs [59]. Although VAEs are a type of AEs, problem and the data. In general, RNNs are designed as in their encoding distribution is regularized during the training Fig. 7, where for each time stamp t, x(t) represents the input to ensure that their latent space (i.e., representation of com- at that time, a(t) is the activation, and y(t) is the output, W , a pressed data) has good properties for generating new data. W , W , b and b are coefficients that are shared temporarily x y x y GANs are composed of two NNs in competition, where a and g and g are activation functions. 1 2 generator network G learns to capture the data distribution and generate new data and a discriminator model D estimates the a(t) = g1(Wa.a(t − 1) + Wx.x(t) + ba) (1) probability that a given sample came from the generator rather than the initial training data, as summarized in Fig. 9 [60], [61]. The generator thus is used to produce misleading samples y(t) = g2(Wy.a(t) + by) (2) and to that the discriminator can determine whether a given JAN. 2021 6

Fig. 10. Reinforcement Learning sample is fake or real. The generator fools the discriminator by generating almost real samples and the discriminator fools the generator by improving its discriminative capability. Fig. 11. Training and test errors over the training time. Early stopping is common technique to reduce overfitting by stopping the training process at D. Reinforcement Learning (RL) an early stage, i.e. when the test error starts to remarkably increasing This subset of ML involves a different learning method than those using supervised, semi-supervised, or unsupervised can be deep or shallow. As each approach offers something dif- learning [64]. RL is about learning what actions to take in the ferent to the world of AI, interest in each should depend on the hope to maximize a reward signal. The agent must find which given problem; a more-complex approach or algorithm does actions bring the most recompense by trying each action, as not necessarily lead to better results. For example, a common shown in 10. These actions can affect immediate rewards as assumption is that DL is better than shallow learning. Although well as subsequent rewards. Some RL approaches require the this holds in several cases, especially for perceptual problems introduction of DL; such approaches are part of deep RL such as computer vision problems, it is not always applicable, (DRL). as DL algorithms require greater computational resources and One of the challenges encountred during RL is balancing large datasets which are not always available. Supervised the trade-off between exploration and exploitation. To get learning is an effective approach when a fully labeled dataset a maximum immediate reward, an RL agent must perform is available. However, this is not always the case, as data exploitation, i.e., choose actions that it has explored previously can be expensive, difficult or even impossible. Under these and found to be the best. To find such actions, it must explore circumstances, semi-supervised or unsupervised learning or the solution space, i.e., try new actions. RL is more applicable. Whereas unsupervised learning can find All RL agents have explicit goals, are aware of some hidden patterns in non-labeled data, RL learns the best policy aspects of their environment, can take actions that impact their to achieve a certain task. Thus, unsupervised learning is a good environments, and act despite significant uncertainty about tool to extract information from data, Whereas RL is better their environment. Other than the agent and the environment, suited for decision-making tasks. Therefore, the choice of an an RL system has four sub-elements: a policy, a reward signal, approach or an algorithm should not be based on its perceived a value function, and, sometimes, a model of the environment. elegance, but by matching the method to characteristics of Here, learning involves the agent determining the best the problem at hand, including the goal, the quality of the method to map states of the environment to actions to be data, the computational resources, the time constraints, and the taken when in those states. After each action, the environment prospective future updates. Solving a problem may require a sends the RL agent a reward signal, which is the goal of the combination of more than one approach. RL problem. Unlike a reward that brings immediate evaluation After assessing the problem and choosing an approach, an of the action, a value function estimates the total amount of algorithm must be chosen. Although ML has mathematical recompense an agent can anticipate to collect in the longer- foundations, it remains an empirical research field. To choose term. Finally, a model of the environment mimics the behavior the best algorithm, data science and ML researchers and of the environment. These models can be used for planning engineers empirically compare different algorithms for a given by allowing the agent to consider possible future situations problem. Algorithms are compared by splitting the data into before they occur. Methods for solving RL problems that a training set and a test set. The training set is then used to utilize models are called model-based methods, whereas those train the model, whereas the test set is to compare the output without models are referred to as model-free methods. between models. In competitive data science, such as in Kaggle [42] compe- E. Discussion titions, where each incrementation matters, models are often 1) Model Selection: AI is a broad field that encompasses combined to improve their overall results, and various en- various approaches, each of which encompasses several algo- semble techniques such as bagging [38], boosting [39], and rithms. AI could be based on predefined rules or on ML. This adaptive boosting [62] are used. learning can be supervised, semi-supervised, unsupervised, or 2) Model Regularization: After the approach and algorithm reinforcement learning; in each of these categories learning have been selected, hyperparameter tuning is generally done JAN. 2021 7 to improve the output of the algorithm. In most cases, ML algorithms depend on many hyperparameters; choosing the best hyperparameters for a given problem thus allows for higher accuracy. This step can be done manually by intuitively choosing better hyperparameters, or automatically using vari- ous methods such as grid search and stochastic methods [63]. A common trap in ML is overfitting, during which the machine stops learning (generalizing) and instead begins to memorize the data. When this occurs, the model can achieve good results on seen data but fails when confronted with new data, i.e., a decreased training error and an increasing test error, as shown in Fig. Fig.11. Overfitting can be discovered Fig. 12. The demand–capacity mismatch among beams demonstrates the limitation of using fixed and uniformly distributed resources across all beams by splitting the data into training, validation and testing sets, in a multi-beam satellite system where neither the validation nor the testing sets are used to train the model. The training set is used to train the model, the validation set is used to verify the model predictions on unseen data and for hyperparameter tuning, and the testing set is used for the final testing of the model. A variety of methods can be employed to reduce overfitting. It be reduced by augmenting the size of the dataset, which is commonly performed in the field of computer vision. For example, image data could be augmented by applying transfor- mations to the images, such as rotating, flipping, adding noise, or cutting parts of the images. Although useful, this technique is not always applicable. Another method involves using cross- validation rather than splitting the data into a training set and Fig. 13. Simplified architecture of beam hopping (BH) a validation set Early stopping, as shown in Fig.11, consists of stopping the learning process before the algorithm begins to memorize the data. Ensemble learning is also commonly resources; some spot beams have a higher demand than used. the offered capacity, leaving the demand pending (i.e., hot- 3) The hype and the hope: Rapid progress has been made spots), while others present a demand lower than the installed in AI research, including its various subfields, over the last capacity, leaving the offered capacity unused (i.e., cold-spots). ten years as a result of exponentially increasing investments. Thus, to improve multi-beam satellite communication, the on- However, few substantial developments have been made to board flexible allocation of satellite resources over the service address real-world problems; as such, many are doubtful that coverage area is necessary to achieve more efficient satellite AI will have much influence on the state of technology and communication. the world. Chollet [28] compared the progress of AI with Beam hopping (BH) has emerged as a promising technique that of the internet in 1995, the majority of people could not to achieve greater flexibility in managing non-uniform and foresee the true potential, consequences, and pertinence of the variant traffic requests throughout the day, year and lifetime internet, as it had yet to come to pass. As the case with the of the satellite over the coverage area [65], [66]. BH, involves overhyping and subsequent funding crash throughout the early dynamically illuminating each cells with a small number of 2000s before the widespread implementation and application active beams, as summarized in 13, thus using all available of the internet, AI may also become an integral part of global on-board satellite resources to offer service to only a subset of technologies. The authors thus believe that the inevitable beams. The selection of this subset is time-variant and depends progress of AI is likely to have long-term impacts and that AI on the traffic demand, which is based on the time-space will likely be a major part of diverse applications across all dependent BH illumination pattern. The illuminated beams are scientific fields, from mathematics to satellite communication. only active long enough to fill the request for each beam. Thus, the challenging task in BH systems is to decide which beams III.ARTIFICIAL INTELLIGENCEFOR SATELLITE should be activated and for how long, i.e., the BH illumination COMMUNICATION pattern; this responsibility is left to the resource manager who then forwards the selected pattern to the satellite via telemetry, A. Beam hopping tracking and command [67]. 1) Definition & limitations: Satellite resources are expen- Of the various methods that researchers have provided to sive and thus require efficient systems involving optimizing realize BH, most have been based on classical optimization and time-sharing. In conventional satellite systems the re- algorithms. For example, Angeletti et al. [68], demonstrated sources are fixed and uniformly distributed across beams [65]. several advantages to the performance of a system when As a result, conventional large multi-beam satellite systems using BH and proposed the use of genetic algorithm (GA) to have shown a mismatch between the offered and requested design the BH illumination pattern; Anzalchi et al. [69], also JAN. 2021 8 illustrated the merits of BH and compared the performance when applying an optimization algorithm to a large search between BH and non-hopped systems. Alberti et al. [70], space. Thus, the learning-based prediction reduces the search proposed a heuristic iterative algorithm to obtain a solution space, and the optimization can be reduced on a smaller set to the BH illumination design. BH has also been used to of promising BH patterns. decrease the number of transponder amplifiers for Terabit/s Researchers have also employed multi-objective DRL (MO- satellites [71]. An iterative algorithm has also been proposed DRL) for the DVB-S2X satellite. Under real conditions, Zhang to maximize the overall offered capacity under certain beam et al. [81] demonstrated that the low-complexity MO-DRL demand and power constraints in a joint BH design and algorithm could ensure the fairness of each cell, and amelio- spectrum assignment [72]. Alegre et al. [73], designed two rate the throughput better than previous techniques including heuristics to allocate capacity resources basing on the traffic DRL [79] by 0.172%. In contrast, the complexity of GA request per-beam, and then further discussed the long and producing a similar result is about 110 times that of the MO- short-term traffic variations and suggested techniques to deal DRL model. Hu et al. [82] proposed a multi-action selection with both variations [74]. Liu et al. [75], studied techniques technique based on double-loop learning and obtained a multi- for controlling the rate of the arriving traffic in BH systems. dimensional state using a DNN. Their results showed that the The QoS delay fairness equilibrium has also been addressed proposed technique can achieve different objectives simulta- in BH satellites [76]. Joint BH schemes were proposed by neously, and can allocate resources intelligently by adapting Shi et al. [77] and Ginesi et al. [78] to further ameliorate the to user requirements and channel conditions. efficiency of on-board resource allocation. To find the optimal BH illumination design, Cocco et al. [79] used a simulated B. Anti-jamming annealing algorithm. Although employing optimization algorithms has achieved 1) Definition & limitations: Satellite communication sys- satisfactory results in terms of flexibility and delay reduction tems are required to cover a wide area, and provide high-speed, of BH systems, some difficulties remain. As the search space communication and high-capacity transmission. However, in dramatically grow with the number of beams, an inherent tactical communication systems using satellites, reliability and difficulty in designing the BH illumination pattern is finding security are the prime concerns; therefore, an anti-jamming the optimal design rather than one of many local optima [72]. (AJ) capability is essential. Jamming attacks could be launched For satellites with hundreds or thousands of beams, classical toward main locations and crucial devices in a satellite net- optimization algorithms may require long computation times work to reduce or even paralyze the throughput. Several AJ which is impractical in many scenarios. methods have thus been designed to reduce possible attacks Additionally, classical optimization algorithms, including and guarantee secure satellite communication. the GAs or other heuristics, require revision when the scenario The frequency-hopping (FH) spread spectrum method has changes moderately; this leads to a higher computational been preferred in many prior tactical communication systems complexity, which is impractical for on-board resource man- using satellites [83], [84]. Using the dehop–rehop transpon- agement. der method employing FH-frequency division multiple access 2) AI-based solutions: Seeking to overcome these limita- (FH-FDMA) scenarios, Bae et al. [85] developed an efficient tions and enhance the performance of BH, some researchers synchronization method with an AJ capability. have proposed AI-based solutions. Some of these solutions Most prior AJ techniques are not based on learning and have been fully based on the learning approach, i.e., end- thus cannot deal with clever jamming techniques that are to-end learning, in which the BH algorithm is a learning capable of continuously adjusting the jamming methodology algorithm. Others have tried to improve optimization algo- by interaction and learning. Developing AI algorithms offer rithms by adding a learning layer, thus combining learning advanced tools to achieve diverse and intelligent jamming and optimization. attacks based on learning approaches and thus present a To optimize the transmission delay and the system through- serious threat to satellite communication reliability. In two put in multibeam satellite systems, Hu et al [80] formulated such examples, a smart jamming formulation automatically an optimization problem and modeled it as a Markov decision adjusted the jamming channel [86], whereas a smart jammer process (MDP). DRL is then used to solve the BH illumination maximized the jamming effect by adjusting both the jamming design and optimize the long-term accumulated rewards of power and channel [87]. In addition, attacks could be caused the modeled MDP. As a result, the proposed DRL-based BH by multiple jammers simultaneously implementing intelligent algorithm can reduce the transmission delay by up to 52.2% jamming attacks based on learning approaches. Although this and increased the system throughput by up to 11.4% when may be an unlikely scenario, it has not yet been seriously con- compared with previous algorithms. sidered. Further, most researchers have focused on defending To combine the advantages of end-to-end learning ap- against AJ attacks in the frequency-based domain, rather than proaches and optimization approaches, for a more efficient spacebased AJ techniques, such as routing AJ. BH illumination pattern design, Lei et al. [67] suggested a 2) AI-based solutions: By using a long short-term memory learning and optimization algorithm to deal with the beam (LSTM) network, which is a DL RNN, to learn the temporal hopping pattern illumination selection, in which a learning trend of a signal, Lee et al. [88] demonstrated a reduction approach, based on fully connected NNs, was used to predict of overall synchronization time in the previously discussed non-optimal BH patterns and thus address the difficulties faced FH-FDMA scenario [85]. Han et al. [89] proposed the use JAN. 2021 9

Several researchers have performed traffic forecasting for both terrestrial and satellite networks; these techniques have included the Markov [92], autoregressive moving average (ARMA) [93], autoregressive integrated moving average (ARIMA) [94] and fractional ARINA (FARIMA) [95] models. By using empirical mode decomposition (EMD) to decompose the network traffic and then applying the ARMA forecasting model, Gao et al. [96] demonstrated remarkable improvement. The two major difficulties facing satellite traffic forecasting are the LRD of satellite networks and the limited on-board computational resources. Due to the LRD property of satellite networks, short-range-dependence (SRD) models have failed to achieve accurate forecasting. Although previous LRD mod- Fig. 14. Space-based anti-jamming (AJ) routing. The red line represents the found jammed path, and the green one represents the suggested path [89] els have achieved better results than SRD models, they suffer from high complexity. To address these issues, researchers have turned to AI techniques. of a learning approach for AJ to block smart jamming in the 2) AI-based solutions: Katris and Daskalaki [95] combined Internet of Satellites (IoS) using a space-based AJ method, AJ FARIMA with NNs for internet traffic forecasting, whereas routing, summarized in Fig.14. By combining game theory Pan et al. [97] combined a differential evolution with NNs modeling with RL and modeling the interactions between for network traffic prediction. Due to the high complexity of smart jammers and satellite users as a Stackelberg AJ routing classical NNs, a least-square SVM, which is an optimized game, they demonstrated how to use DL to deal with the large version of a SVM, has also been used for forecasting [98]. decision space caused by the high dynamics of the IoS and By applying principal component analysis (PCA), to reduce RL to deal with the interplay between the satellites and the the input dimensions and then a generalized regression NN, smart jamming environment. DRL thus made it possible to Ziluan and Xin [99] achieved higher-accuracy forecasting with solve the routing selection issue for the heterogeneous IoS less training time. Zhenyu et al. [100] used traffic forecasting while preserving an available routing subset to simplify the as a part of their distributed routing strategy for LEO satellite decision space for the Stackelberg AJ routing game. Based on network. An extreme learning machine (ELM) has also been this routing subset, a popular RL algorithm, Q-Learning, was employed for traffic load forecasting of satellite node before then used to respond rapidly to intelligent jamming and adapt routing [101]. Bie et al. [91] used EMD to decompose the AJ strategies. traffic of the satellite with LRD into a series with SRD and at Han et al. [90] later combined game theory modeling one frequency to decrease the predicting complexity and aug- and RL to obtain AJ policies according to the dynamic ment the speed. Their combined EMD, fruit-fly optimization, and unknown jamming environment in the Satellite-Enabled and ELM methodology achieved more accurate forecasting at Army IoT (SatIoT). Here, a distributed dynamic AJ coalition a higher speed than prior approaches. formation game was examined to decrease the energy use in the jamming environment, and a hierarchical AJ Stackelberg D. Channel Modeling game was proposed to express the confrontational interac- tion between jammers and SatIoT devices. Finally, RL-based 1) Definition & limitations: A channel model is a math- algorithms were utilized to get the sub-optimal AJ policies ematical representation of the effect of a communication according to the jamming environment. channel through which wireless signals are propagated; it is modeled as the impulse response of the channel in the frequency or time domain. C. Network Traffic Forecasting A wireless channel presents a variety of challenges for 1) Definition & limitations: Network traffic forecasting reliable high-speed communication, as it is vulnerable to noise, is a proactive approach that aims to guarantee reliable and interference, and other channel impediments, including path high-quality communication, as the predictability of traffic is loss and shadowing. Of these, path loss is caused by the waste important in many satellite applications, such as congestion of the power emitted by the transmitter and the propagation control, dynamic routing, dynamic channel allocation, network channel effects, whereas shadowing is caused by the obstacles planning, and network security. Satellite network traffic is between the receiver and transmitter that absorb power [102]. self-similar and demonstrates long-range-dependence (LRD) Precise channel models are required to asses the perfor- [91]. To achieve accurate forecasting, it is therefore necessary mance of mobile communication system and therefore to to consider its self-similarity. However,forecasting models for enhance coverage for existing deployments. Channel models terrestrial networks based on self-similarity have a high com- may also be useful to forecast propagation in designed de- putational complexity; as the on-board satellite computational ployment outlines, which could allow for assessment before resources are limited, terrestrial models are not suitable for deployment, and for optimizing the coverage and capacity satellites. An efficient traffic forecasting design for satellite of actual systems. For small number of transmitter possible networks is thus required. positions, outdoor extensive environment evaluation could JAN. 2021 10

alized data. Despite the practicality of this method, as it only needs satellite images to forecast the path loss distribution, 2D images will not always be sufficient to characterize the 3D structure. In these cases, more features (e.g., building heights) must be input into the model. Fig. 15. Channel parameters prediction. 2D aerial/satellite images used as input to the deep convolutional neural network (CNN)to to predict channel parameters. The model is trained separately for each parameter. E. Telemetry Mining be done to estimate the parameters of the channel [103], 1) Definition & limitations: Telemetry is the process of [104]. As more advanced technologies have been used in recording and transferring measurements for control and mon- wireless communication, more advanced channel modelling itoring. In satellite systems, on-board telemetry helps mission was required. Therefore the use of stochastic models that are control centers track platform’s status, detect abnormal events, computationally efficient while providing satisfactory results and control various situations. [105]. Satellite failure can be caused by a variety of things; most Ray tracing is used for channel modeling, which requires commonly, failure is due to the harsh environment of space, 3D images that are generally generated using computer vision i.e., heat, vacuum, and radiation. The radiation environment methods including stereo-vision-based depth estimation [106], can affect critical components of a satellite, including the [107], [108], [109]. communication system and power supply. A model is proposed for an urban environment requires Telemetry processing enables tracking of the satellite’s features, including road widths, street orientation angles, and behavior to detect and minimize failure risks. Finding corre- height of buildings [110]. A simplified model was then pro- lations, recognizing patterns, detecting anomalies, classifying, posed, by Fernandes and Soares [111] that required only the forecasting, and clustering are applied to the acquired data for proportion of building occupation between the receiver and fault diagnosis and reliable satellite monitoring. transmitter, which could be computed from segmented images One of the earliest and simplest techniques used in telemetry manually or automatically [112]. analysis is limit checking. The method is based on setting Despite the satisfactory performance of some of the listed a precise range for each feature (e.g., temperature, voltage, techniques, they still have many limitations. For example, the and current), and then monitoring the variance of each feature 3D images required by ray tracing r are not generally available to detect out-of-range events. The main advantage of this and their generation is not computationally efficient. Even algorithm is its simplicity limits, as can be chosen and updated when the images are available, ray tracing is computationally easily to control spacecraft operation. costly and data exhaustive and therefore is not appropriate for Complicated spacecraft with complex and advanced appli- real-time coverage area optimization. Further, the detailed data cations challenges current space telemetry systems. Narrow required for the model presented by Cichon and Kurner [110] wireless bandwidth and fixed-length frame telemetry make is often unavailable. transmitting the rapidly augmenting telemetry volumes dif- 2) AI-based solutions: Some early applications of AI for ficult. In addition, the discontinuous short-term contacts be- path loss forecasting have been based on classical ML al- tween spacecraft and ground stations limit the data transmis- gorithms such as SVM [113], [114], NNs [115]–[120] and sion capability. Analyzing, monitoring and interpreting huge decision trees [121]. Interested readers are referred to a survey telemetry parameters could be impossible due to the high of ML-based path loss prediction approaches for further details complexity of data. [122]. 2) AI-based solutions: In recent years, AI techniques have However, although previous ML efforts have shown great been largely considered in space missions with telemetry. results, many require 3D images. Researchers have recently Satellite health monitoring has been performed using proba- thus shifted their attention to using DL algorithms with 2D bilistic clustering [126], dimensionality reduction, and hidden satellite/aerial images for path loss forecasting. For example, Markov [127], and regression trees [128], whereas others have Ates et al. [123], approximated channel parameters, including developed anomaly detection methods using the K-nearest the standard deviation of shadowing and the path loss expo- neighbor (kNN), SVM, LSTM and testing on the telemetry nent, from satellite images using deep CNN without the use of Centre National d’Etudes Spatiales spacecraft [129]–[131]. of any added input parameters, as shown in Fig.15. Further, the space functioning assistant was further devel- By using a DL model on satellite images and other input pa- oped in diverse space applications using data-driven [132] rameters to predict the reference signal received power (RSRP) and model-based [133] monitoring methods. In their study of for specific receiver locations in a specific scenario/area, the use of AI for fault diagnosis in general and for space Thrane et al. [124] demonstrated a gain improvement of utilization, Sun et al. [134] argued that the most promising ≈ 1 and ≈ 4.7 at 811 MHz and 2630 MHz respectively, direction is the use of DL; suggested its usage for fault over previous techniques, including ray tracing. Similarly diagnosis for space utilization in China. Ahmadien et al. [125], applied DL on satellite images for path By comparing different ML algorithms using telemetry data loss prediction, although they focused only on satellite images from the Egyptsat-1 satellite, Ibrahim et al. [135] demonstrated without any supplemental features and worked on more gener- the high prediction accuracy of LSTM, ARIMA, and RNN JAN. 2021 11

low latitudes, where scintillation is expected to occur [140], [141]. Robust receivers and proper algorithms for scintillation- detecting algorithms are thus both required [142]. To evaluate the magnitude of scintillation impacting a signal, many researchers have employed simple event trig- gers, based on the comparison of the amplitude and phase of two signals over defined interval [143]. Other proposed alternatives, have included using wavelet techniques [144], decomposing the carrier-to-noise density power propostion via adaptive frequency-time techniques [145], and assessing the Fig. 16. Representation of ionospheric scintillation, where distortion occurs histogram statistical properties of collected samples [146]. during signal propagation. The blue, green, and red lines show the line-of-sight Using simple predefined thresholds to evaluate the mag- signal paths from the satellite to the earth antennas, the signal fluctuation, and nitude of scintillation can be deceptive due its complexity. the signal delay, respectively. The loss of the transient phases of events could cause a delay in raising possible caution flags, and weak events with models. They suggested simple linear regression for forecast- high variance could be missed. Further, it can be difficult ing critical satellite features for short-lifetime satellites (i.e., to distinguish between signal distortions caused by other 3–5 years) and NNs for long-lifetime satellites (15-20 years). phenomena, including multi-path. However, other proposed Unlike algorithms designed to operate on the ground in alternatives depend on complex and computationally costly the mission control center, Wan et al. [136] proposed a self- operations or on customized receiver architectures. learning classification algorithm to achieve on-board telemetry 2) AI-based solutions: Recently, studies have proved that data classification with low computational complexity and low AI can be utilized for the detection of scintillation. For time latency. example, Rezende et al. [147], proposed a survey of data mining methods, that rely on observing and integrating GNSS receivers. F. Ionospheric Scintillation Detecting A technique based on the SVM algorithm has been sug- 1) Definition & limitations: Signals transmission by satel- gested for amplitude scintillation detection [148], [149], and lites toward the earth can be notably impacted due to their then later expanded to phase scintillation detection [150], propagation through the atmosphere, especially the iono- [151]. sphere, which is the ionized part of the atmosphere higher By using decision trees and RF to systematically detect layer, and is distinguished by an elevated density of free ionospheric scintillation events impacting the amplitude of the electrons (Fig.16). The potential irregularities and gradients GNSS signals, Linty et al.’s [152] methodology outperformed of ionization can distort the signal phase and amplitude, in a state-of-the art methodologies in terms of accuracy (99.7%) process known as ionospheric scintillation. and F-score (99.4%), thus reaching the levels of a manual In particular, propagation through the ionosphere can cause human-driven annotation. distortion of global navigation satellite system (GNSS) signals, More recently, Imam and Dovis [153] proposed the use of leading to significant errors in the GNSS-based applications. decision trees, to differentiate between ionospheric scintilla- GNSSs are radio-communication satellite systems that allow tion and multi-path in GNSS scintillation data. Their model, a user to compute the local time, velocity, and position in any which annotates the data as scintillated, multi-path affected, place on the Earth by processing signals transferred from the or clean GNSS signal, demonstrated an accuracy of 96% satellites and conducting trilateration [137]. GNSSs can also be used in a wide variety of applications, such as scientific G. Managing Interference observations. 1) Definition & limitations: Interference managing is Because of the low-received power of GNSS waves, any mandatory for satellite communication operators, as interfer- errors significantly threaten the accuracy and credibility of ence negatively affects the communication channel, resulting the positioning systems. GNSS signals propagating through in a reduced QoS, lower operational efficiency and loss of the ionosphere face the possibility of both a temporal delay revenue [154]. Moreover, interference is a common event that and scintillation. Although delay compensation methods are is increasing with the increasing congestion of the satellite applied to all GNSS receivers [137], scintillation is still frequency band as more countries are launching satellites and a considerable issue, as its quasi-random nature makes it more applications are expected. With the growing number of difficult to model [138]. Ionospheric scintillation thus remains users sharing the same frequency band, the possibility of in- a major limitation to high-accuracy applications of GNSSs. terfering augments, as does the risk of intentional interference, The accurate detection of scintillation thus required to improve as discussed in section III.B. the credibility and quality of GNSSs [139]. To observe the Interference managing is a thus essential to preserve high- signals, which are a source of knowledge for interpreting and quality and reliable communication systems; management modeling the atmosphere higher layers, and to raise caution includes detection, classification, and suppression of interfer- and take countermeasures for GNSS-based applications, net- ence, as well as the application of techniques to minimize its works of GNSS receivers, have been installed, both at high and occurrence. JAN. 2021 12

2) AI-based solutions: The revolution in computer vision capabilities caused by DL has led to the increased development of RS by adopting state-of-the-art DL algorithms on satellite images, image classification for RS has become most popular task in computer vision. For example, Kussul et al. [161] used DL to classify land coverage and crop types using RS images from Landsat-8 and Sentinel-1A over a test site in Ukraine. Zhang et al [162] combined DNNs by using a gradient- boosting random CNN for scene classification. More recently, Chirayath et al. [163] proposed the combination of kNN and CNN to map coral reef marine habitats worldwide with RS imaging. RS and AI have also been used in communication theory applications, such as those discussed in section III.D [123], [124] and [125]. Many object detection and recognition applications have Fig. 17. Satellite selection and antenna adjustment been developed using AI on RS images [164]. Recently, Zhou et al. [165] proposed the use of YOLOv3 [166], [167], a CNN- Interference detection is a well-studied subject that has been based object detection algorithm, for vehicle detection in RS addressed in the past few decades [155], [156], especially for images. Others have proposed the use of DL for other object satellite communication [154], [157]. detection tasks, such as, building [168], airplane [169], cloud However, researchers have commonly relied on the decision [170], [171], [172], ship [173], [174], and military target [175] theory of hypothesis testing, in which specific knowledge of detection. AI has also been applied to segment and restore the signal characteristics and the channel model is needed. RS images, e.g., in cloud restorations, during which ground Due, to the contemporary diverse wireless standards, the regions shadowed by clouds are restored. design of specific detectors for each signal category is fruitless Recently, Zheng et al. [176] proposed a two-stage cloud approach. removal method in which U-Net [177] and GANs are used 2) AI-based solutions: To minimize interference, Liu et to perform cloud segmentation and image restoration, respec- al. [158], suggested the use of AI for moving terminals tively. and stations in satellite-terrestrial networks by proposing a AI proposed for on-board scheduling of agile Earth- framework combining different AI approaches including SVM, observing satellites, as autonomy improves their performance unsupervised learning and DRL for satellite selection, antenna and allows them to acquire more images, by relying on on- pointing and tracking, as summarized in Fig.17. board scheduling for quick decision-making. By comparing Another AI-based approach executes automatic real-time the use of RF, NNs, and SVM to prior learning and non- interference detection is based on the forecasting of the follow- learning-based approaches, Lu et al. [178] demonstrated that ing signal spectrum to be received in absence of anomaly, by RF improved both the solution quality and response time. using LSTM trained on historical anomaly-free spectra [159]. I. Behavior Modeling Here the predicted spectra is then compared to the received signal using a designed metric, to detect anomalies. 1) Definition & limitations: Owing to the increasing num- Henarejos et al. [160] proposed the use of two AI-based bers of active and inactive (debris) satellites of diverse orbits, approaches, DNN AEs and LSTM, for detecting and clas- shapes, sizes, orientations and functions, it is becoming in- sifying interference, respectively. In the former, the AE is feasible for analysts to simultaneously monitor all satellites. trained with interference free signals and tested against other Therefore, AI, especially ML, could play a major role by signals without interference to obtain practical thresholds. The helping to automate this process. difference in error in signals with and without interference is 2) AI-based solutions: Mital et al. [179] discussed the then exploited to detect the presence of interference. potential of ML algorithms to model satellite behavior. Super- vised models have been used to determine satellite stability [180], whereas unsupervised models have been used to detect H. Remote sensing (RS) anomalous behavior and a satellites’ location [181], and an 1) Definition & limitations: RS is the process of extracting RNN has been used to predict satellite maneuvers over time information about an area, object or phenomenon by process- [182]. ing its reflected and emitted radiation at a distance, generally Accurate satellite pose estimation, i.e., identifying a satel- from satellite or aircraft. lite’s relative position and attitude, is critical in several space RS has a wide range of applications in multiple fields operations, such as debris removal, inter-spacecraft commu- including land surveying, geography, geology, ecology, me- nication, and docking. The recent proposal for satellite pose teorology, oceanography, military and communication. As RS estimation from a single image via combined ML and geo- offers the possibility of monitoring areas that are dangerous, metric optimization by Chen et al. [183] won the first place difficult or impossible to access, including mountains, forests, in the recent Kelvins pose estimation challenge organized by oceans and glaciers it is a popular and active research area. the European Space Agency [184]. JAN. 2021 13

include the satellites in space, the balloons, airships, and UAVs in the air, and the ground segment, as shown in Fig.18. The multi-layered satellite communication system which consists of GEO, MEO, and LEO satellites, can use multi- cast and broadcast methods to ameliorate the network capacity, crucially easing the augmenting traffic burden [10], [26]. As SAGINs allow packet transmission to destinations via multiple paths of diverse qualities, they can offer different packet transmissions methods to encounter diverse service demands [26]. However, the design and optimization of SAGINs is more challenging than that of conventional ground communica- tion systems owing to their inherent self-organization, time- variability, and heterogeneity [10]. A variety of factors that must be considered when designing optimization techniques have thus been identified [10], [26]. For example, the diverse propagation mediums, the sharing of frequency bands by different communication types, the high mobility of the space and air segments, and the inherent heterogeneity between the three segments, make the network control and spectrum Fig. 18. Space-air-ground integrated networks (SAGINs) [26] management of SAGIN arduous. The high mobility results in frequent handoffs, which makes safe routing more difficult to realize, thus making SAGINs more exposed to jamming. The amount of space debris has augmented immensely over Further, as optimizing the energy efficiency is also more the last few years, which can cause a crucial menace to challenging than in standard terrestrial networks, energy man- space missions due to the high velocity of the debris. It is agement algorithms are also required. thus essential to classify space objects and apply collision 2) AI-based solutions: In their discussion of challenges avoidance techniques to protect active satellites. As such, facing SAGINs, Kato et al. [26] proposed the use of a CNN Jahirabadkar et al. [185] presented a survey of diverse AI for the routing problem to optimize the SAGIN’s overall methodologies, for classification of space objects using the performance using traffic patterns and the remaining buffer curves of light as a differentiating property. size of GEO and MEO satellites. Yadava et al. [186] employed NNs and RL for on-board Optimizing the satellite selection and the UAV location attitude determination and control; their method effectively to optimize the end-to-end data rate of the Source-Satellite- provided the needed torque to stabilize a nanosatellite along UAV-Destination communication is challenging due to the three axes. vast orbiting satellites number and the following time-varying To avoid catastrophic events because of battery failure, network architecture. To address this problem, Lee et al. [188] Ahmed et al. [187] developed an on-board remaining battery jointly optimized the source-satellite-UAV association and the life estimation system using ML and a logical analysis of data location of the UAV via DRL. Their suggested technique approaches. achieved up to a 5.74x higher average data rate than a direct communication baseline in the absence of UAV and satellite. For offloading calculation-intensive applications, a SAGIN J. Space-Air-Ground Integrating edge/cloud computing design has been developed in such 1) Definition & limitations: Recently, notable advances a way that satellites give access to the cloud and UAVs have been made in ground communication systems to pro- allow near-user edge computing. [189]. Here, a joint resource vide users higher-quality internet access. Nevertheless, due to allocation and task scheduling approach is used to allocate the restricted capacity and coverage area of networks, such the computing resources to virtual machines and schedule the services are not possible everywhere at all times, especially offloaded tasks for UAV edge servers, whereas an RL-based for users in rural or disaster areas. computing offloading approach handles the multidimensional Although terrestrial networks have the most resources and SAGIN resources and learns the dynamic network condi- highest throughput, non-terrestrial communication systems tions. Here, a joint resource allocation and task scheduling have a much broader coverage area. However, non-terrestrial approach is used to assign the computing resources to virtual networks have their own limitations; e.g., satellite communica- machines and plan the offloaded functions for UAV edge tion systems have a long propagation latency, and air networks servers, whereas an RL-based computing offloading approach have a narrow capacity and unstable links. handles the multidimensional SAGIN resources and learns the To supply users with better and more-flexible end-to-end dynamic network characteristics. Simulation results confirmed services by taking advantage of the way the networks can the efficiency and convergence of the suggested technique. complement each other, researchers have suggested the use of As the heterogeneous multi-layer network requires advanced space-air-ground integrated networks (SAGINs) [10], which capacity-management techniques, Jiang and Zhu [190] sug- JAN. 2021 14 gested a low-complexity technique for computing the capacity communication endpoints due to the dynamic connectivity among satellites and suggested a long-term optimal capacity patterns of LEO satellites. The management of handoff in LEO assignment RL-based model to maximize the long-term utility satellites varies remarkably from that of terrestrial networks, of the system. since handoffs happen more frequently due to the movement of By formulating the joint resources assignment problem as a satellites [3]. Many researchers have thus focused on handoff joint optimization problem and using a DRL approach, Qiu et management in LEO satellite networks. al. [191] proposed a software-defined satellite-terrestrial net- In general, user equipment (UE) periodically measures the work to jointly manage caching, networking, and computing strength of reference signals of different cells to ensure access resources. to a strong cell, as the handoff decision depends on the signal strength or some other parameters. Moreover, the historical K. Energy Managing RSRP contains information to avoid unnecessary handoff. 1) Definition & limitations: Recent advances in the con- Thus, Zhang [197] converted the handoff decision to a nection between ground, aerial, and satellite networks such as classification problem. Although the historical RSRP is a time SAGIN have increased the demand imposed on satellite com- series, a CNN was employed rather than an RNN because munication networks. This growing attention towards satellites the feature map of historical RSRP has a strong local spatial has led to increased energy consumption requirements. Satel- correlation and the use of an RNN could lead to a series lite energy management thus represents a hot research topic of wrong decisions, as one decision largely impacts future for the further development of satellite communication. decisions. In the proposed AI-based method, the handoff was Compared with a GEO Satellite, an LEO satellite has decreased by more than 25% for more than 70% of the UE, restricted on-board resources and moves quickly. Further, an whereas the commonly used “strongest beam” method only LEO satellite has a limited energy capacity owing to its small reduced the average RSRP by 3%. size [192]; as billions of devices need to be served around 2) Heat Source Layout Design: The effective design of the the world [193], current satellite resource capability can no heat sources used can enhance the thermal performance of longer satisfy demand. To address this shortage of satellite the overall system, and has thus become a crucial aspect of communication resources, an efficient resource scheduling several engineering areas, including integrated circuit design scheme to take full use of the limited resources, must be and satellite layout design. With the increasingly small size designed. As current resource allocation schemes have mostly of components and higher power intensity, designing the heat- been designed for GEO satellites, however, these schemes source layout has become a critical problem [198]. Conven- do not consider many LEO specific concerns, such as the tionally, the optimal design is acquired by exploring the design constrained energy, movement attribute, or connection and space by repeatedly running the thermal simulation to compare transmission dynamics. the performance of each scheme [199]–[201]. To avoid the ex- 2) AI-based solutions: Some researchers have thus turned tremely large computational burden of traditional techniques, to AI-based solutions for power saving. For example, Kothari Sun et al. [202] employed an inverse design method in which et al. [27] suggested the usage of DNN compression before the layout of heat sources is directly generated from a given data transmission to improve latency and save power. In the expected thermal performance based on a DL model called absence of solar light, satellites are battery energy dependent, Show, Attend, and Read [203]. Their developed model was which places a heavy load on the satellite battery and can capable of learning the underlying physics of the design shorten their lifetimes leading to increased costs for satellite problem and thus could efficiently forecast the design of communication networks. To optimize the power allocation in heat sources under a given condition without any performing satellite to ground communication using LEO satellites and simulations. Other DL algorithms have been used in diverse thus extend their battery life, Tsuchida et al. [194] employed design areas, such as mechanics [204], optics [205], fluids RL to share the workload of overworked satellites with near [206], and materials [207]. satellites with lower load. Similarly, implementing DRL for 3) Reflectarray analysis and design: ML algorithms have energy-efficient channel allocation in Satlot allowed for a been employed in the analysis and design of antennas [22], 67.86% reduction in energy consumption when compared including the analysis [208], [209] and design [210], [211] with previous models [195]. Mobile edge computing enhanced of reflectarrays. For example, NNs were used by Shan et SatIoT networks contain diverse satellites and several satellite al. [212] to forecast the phase-shift, whereas kriging was gateways that could be jointly optimized with coupled user as- suggested to forecast the electromagnetic response of reflec- sociation, offloading decisions computing, and communication tarray components [213]. Support vector regression (SVR) resource allocation to minimize the latency and energy cost. has been used to accelerate the examination [214] and to In a recent example, a joint user-association and offloading directly optimize narrowband reflectarrays [215]. To hasten decision with optimal resource allocation methodology based calculations without reducing their precision, Prado et al. on DRL proposed by Cui et al. [196] improved the long-term [216] proposed a wideband SVR-based reflectarray design latency and energy costs. method, and demonstrated its ability to obtain wideband, dual- linear polarized, shaped-beam reflectarrays for direct broadcast L. Other Applications satellite applications. 1) Handoff Optimization: Link-layer handoff occurs when 4) Carrier Signal Detection: As each signal must be sepa- the change of one or more links is needed between the rated before classification, modulation, demodulation, decod- JAN. 2021 15

ing and other signal processing, localization, and detection of [5] P.-D. Arapoglou, K. Liolis, M. Bertinelli, A. Panagopoulos, P. Cottis, carrier signals in the frequency domain is a crucial problem and R. De Gaudenzi, “MIMO over satellite: A review,” IEEE Commun. Surveys Tuts., vol. 13, no. 1, pp. 27-51, 1st Quart. 2011. in wireless communication. [6] M. De Sanctis, E. Cianca, G. Araniti, I. Bisio, and R. Prasad, “Satellite The algorithms used for carrier signal detection have been communications supporting Internet of remote things,” IEEE Internet commonly based on threshold values and required human Things J., vol. 3, no. 1, pp. 113-123, Feb. 2016. intervention [217]–[222], although several improvements have [7] R. Radhakrishnan, W. W. Edmonson, F. Afghah, R. M. Rodriguez-Osorio, F. Pinto, and S. C. Burleigh, “Survey of inter-satellite communication been made including the use of a double threshold [223], for small satellite systems: Physical layer to network layer view,” IEEE [224]. Kim et al. [225] proposed the use of a slope-tracing- Commun. Surveys Tuts., vol. 18, no. 4, pp. 2442-2473, May 2016. based algorithm to separate the interval of signal elements [8] C. Niephaus, M. Kretschmer, and G. Ghinea, “QoS provisioning in converged satellite and terrestrial networks: A survey of the state-of-the- based on signal properties such as amplitude, slope, deflection art,” IEEE Commun. Surveys Tuts., vol. 18, no. 4, pp. 2415-2441, Apr. width, or distance between neighboring deflections. 2016. More recently, DL has been applied to carrier signal detec- [9] H. Kaushal and G. Kaddoum, “Optical communication in space: Chal- lenges and mitigation techniques,” IEEE Commun. Surveys Tuts., vol. 19, tion; for example, Morozov and Ovchinnikov [226] applied no. 1, pp. 57-96, 1st Quart. 2017. a fully connected NN for their detection in FSK signals, [10] J. Liu, Y. Shi, Z. M. Fadlullah, and N. Kato, “Space-Air-Ground whereas Yuan et al. [227] used DL, to morse signals blind Integrated Network: A Survey,” IEEE Communications Surveys & detection in wideband spectrum data. Huang er al. [228] Tutorials, vol. 20, no. 4, pp. 2714-2741, Fourthquarter 2018, doi: 10.1109/COMST.2018.2841996. employed a fully convolutional network (FCN) model to detect [11] S. C. Burleigh, T. De Cola, S. Morosi, S. Jayousi, E. Cianca, and C. carrier signal in the broadband power spectrum. A FCN is a Fuchs, “From connectivity to advanced Internet services: A comprehen- DL method for semantic image segmentation in which the sive review of small satellites communications and networks,” Wireless Commun. Mobile Comput., vol. 2019, pp. 1-17, May 2019. broadband power spectrum is regarded as a 1D image and [12] B. Li, Z. Fei, C. Zhou, and Y. Zhang, “Physical-layer security in space each subcarrier as the target object to transform the carrier information networks: A survey,” IEEE Internet Things J., vol. 7, no. 1, detection problem on the broadband to a semantic 1D image pp. 33-52, Jan. 2020. segmentation problem [229]–[231]. Here, a 1D deep CNN [13] N. Saeed, A. Elzanaty, H. Almorad, H. Dahrouj, T. Y. Al-Naffouri, and M. -S. Alouini, “CubeSat Communications: Recent Advances and Future FCN-based on was designed to categorize each point on a Challenges,” IEEE Communications Surveys & Tutorials, vol. 22, no. 3, broadband power spectrum array into two categories (i.e., pp. 1839-1862, thirdquarter 2020, doi: 10.1109/COMST.2020.2990499. subcarrier or noise), and then position the subcarrier signals’ [14] O. Simeone, “A Very Brief Introduction to Machine Learning With Ap- plications to Communication Systems,” IEEE Transactions on Cognitive location on the broadband power spectrum. After being trained Communications and Networking, vol. 4, no. 4, pp. 648-664, Dec. 2018, and validated using a simulated and real satellite broadband doi: 10.1109/TCCN.2018.2881442. power spectrum dataset, respectively, the proposed deep CNN [15] M. Chen, U. Challita, W. Saad, C. Yin, and M. Debbah, “Artificial Neural Networks-Based Machine Learning for Wireless Networks: A Tutorial,” successfully detected the subcarrier signal in the broadband IEEE Communications Surveys & Tutorials, vol. 21, no. 4, pp. 3039-3071, power spectrum and achieved a higher accuracy than the slope Fourthquarter 2019, doi: 10.1109/COMST.2019.2926625. tracing method. [16] Y. Qian, J. Wu, R. Wang, F. Zhu, and W. Zhang, “Survey on Rein- forcement Learning Applications in Communication Networks,” Journal of Communications and Information Networks, vol. 4, no. 2, pp. 30-39, CONCLUSION June 2020, doi: 10.23919/JCIN.2019.8917870. This review provided an overview of AI and its different [17] EC. Strinati, S. Barbarossa, JL. Gonzalez, D. Ktenas, N. Cassiau, sub-fields, including ML, DL, and RL. Some limitations to L. Maret, and C. Dehos, “6G: The Next Frontier: From Holographic Messaging to Artificial Intelligence Using Subterahertz and Visible Light satellite communication were then presented and their pro- Communication,” IEEE Vehicular Technology Magazine, vol. 14, no. 3, posed and potential AI-based solutions were discussed. The pp. 42-50, Sept. 2019, doi: 10.1109/MVT.2019.2921162. application of AI has shown great results in a wide variety [18] J. Jagannath, N. Polosky, A. Jagannath, F. Restuccia, T. Melo- dia, “Machine learning for wireless communications in the Inter- of satellite communication aspects, including beam-hopping, net of Things: A comprehensive survey,” Ad Hoc Networks, Vol- AJ, network traffic forecasting, channel modeling, telemetry ume 93, Jan. 2019, 101913, ISSN 1570-8705, [online] Available: mining, ionospheric scintillation detecting, interference man- https://doi.org/10.1016/j.adhoc.2019.101913. aging, remote sensing, behavior modeling, space-air-ground [19] G. P. Kumar and P. Venkataram, “Artificial intelligence approaches to network management: recent advances and a integrating, and energy managing. Future work should aim to survey,” Computer Communications, Volume 20, Issue 15, Dec. apply AI, to achieve more efficient, secure, reliable, and high- 1997, Pages 1313-1322, ISSN 0140-3664, [online] Available: quality communication systems. https://doi.org/10.1016/S0140-3664(97)00094-7. [20] Y. Zou, J. Zhu, X. Wang, and L. Hanzo, “A Survey on Wireless Security: Technical Challenges, Recent Advances, and Future Trends,” REFERENCES Proceedings of the IEEE, vol. 104, no. 9, pp. 1727-1765, Sept. 2016, [1] G. Maral, M. Bousquet, and Z. Sun, “Introduction,” in Satellite Commu- doi: 10.1109/JPROC.2016.2558521. nications Systems: Systems, Techniques and Technology, 6th ed. Hoboken, [21] S. H. Alsamhi, O. Ma, and M. S. Ansari. “Survey on artificial intelli- NJ, USA: Wiley, 2020, ch. 1, sec. 3, pp. 3–11. gence based techniques for emerging robotic communication,” Telecom- [2] F. Rinaldi, H. L. Maattanen, J. Torsner, S. Pizzi, S. Andreev, A. Iera, munication Systems: Modelling, Analysis, Design and Management, vol. Y. Koucheryavy, and G. Araniti “Non-Terrestrial Networks in 5G & 72, issue 3, no. 12, pp. 483-503, Mars 2019, doi: 10.1007/s11235-019- Beyond: A Survey,” in IEEE Access, vol. 8, pp. 165178-165200, 2020, 00561-z doi: 10.1109/ACCESS.2020.3022981. [22] H. M. E. Misilmani and T. Naous, “Machine Learning in An- [3] P. Chowdhury, M. Atiquzzaman, and W. Ivancic, “Handover schemes in tenna Design: An Overview on Machine Learning Concept and Al- satellite networks: State-of-the-art and future research directions,” IEEE gorithms,” 2019 International Conference on High Performance Com- Commun. Surveys Tuts., vol. 8, no. 4, pp. 2-14, Aug. 2006. puting & Simulation (HPCS), Dublin, Ireland, 2019, pp. 600-607, doi: [4] P. Chini, G. Giambene, and S. Kota, “A survey on mobile satellite 10.1109/HPCS48598.2019.9188224. systems,” Int. J. Satell. Commun. Netw., vol. 28, no. 1, pp. 29-57, Aug. [23] P. S. Bithas, E. T. Michailidis, N. Nomikos, D. Vouyioukas, and 2009. A. Kanatas “A survey on machine-learning techniques for UAV-based JAN. 2021 16

communications,” Sensors 2019 19.23, Nov. 2019, [online] Available: [51] Z. Zou, Z. Shi, Y. Guo, and J. Ye, “Object detection in 20 years: A https://doi.org/10.3390/s19235170. survey,” arXiv preprint arXiv:1905.05055, 2019. [24] M. A. Lahmeri, M. A. Kishk, and MS. Alouini. “Machine learning for [52] Q. Chu, W. Ouyang, H. Li, X. Wang, B. Liu, and N. Yu “Online UAV-Based networks.” arXiv preprint, 2020, arXiv:2009.11522. multi-object tracking using CNN-based single object tracker with spatial- [25] M. A.´ Vazquez,´ P. Henarejos, A. I. Perez-Neira,´ E. Grechi, A. Voight, JC. temporal attention mechanism,” Proceedings of the IEEE International Gil, I. Pappalardo, FD. Credico, and R. M. Lancellotti, “On the Use of AI Conference on Computer Vision. 2017. for Satellite Communications.” arXiv preprint, 2020, arXiv:2007.10110. [53] K. R. Chowdhary, “Natural language processing,” Fundamentals of [26] N. Kato, ZM. Fadlullah, F. Tang, B. Mao, S. Tani, A. Okamura, and Artificial Intelligence. Springer, New Delhi, 2020. pp. 603–649. J. Liu, “Optimizing Space-Air-Ground Integrated Networks by Artificial [54] I. Goodfellow, Y. Bengio, and A. Courville, “Sequence Mod- Intelligence,” IEEE Wireless Communications, vol. 26, no. 4, pp. 140-147, eling: Recurrent and Recursive Nets,” in Deep learning, Cam- August 2019, doi: 10.1109/MWC.2018.1800365. bridge: MIT press, 2016, ch. 10, pp. 367–415. [online] Available: [27] V. Kothari, E. Liberis, and N. D. Lane. “The Final Frontier: Deep https://www.deeplearningbook.org/ Learning in Space,” Proceedings of the 21st International Workshop on [55] I. Goodfellow, Y. Bengio, and A. Courville, “Autoencoders,” in Deep Mobile Computing Systems and Applications, pp. 45-49., 2020. learning, Cambridge: MIT press, 2016, ch. 14, pp. 499–523. [online] [28] F. Chollet, “What is Deep Learning ?” in Deep Learning with Python, Available: https://www.deeplearningbook.org/ 1st ed. New York, NY, USA: Manning, 2017, ch. 1, pp. 3–24. [56] Y. Wang, Y. Hongxun , and Z. Sicheng, “Auto-encoder based dimen- [29] A. M. Turing, “Computing Machinery and Intelligence,” in Mind, 59th sionality reduction,” Neurocomputing 184, 2016, pp. 232–242. ed., 1950, ch. 1, pp. 433–460. [57] C. Zhou and RC. Paffenroth, “Anomaly detection with robust deep [30] C. M. Bishop, “Linear Models for Classification,” in Pattern Recognition autoencoders,” Proceedings of the 23rd ACM Special Interest Group and Machine Learning, 1st ed. Berlin, Heidelberg, Germany: Springer- on Knowledge Discovery and Data Mining International Conference on Verlag, 2006, ch. 4, pp. 179–224. Knowledge Discovery and Data Mining. 2017. [31] C. M. Bishop, “Kernel Methods,” in Pattern Recognition and Machine [58] I. Goodfellow, Y. Bengio, and A. Courville, “Deep generative models,” Learning, 1st ed. Berlin, Heidelberg, Germany: Springer-Verlag, 2006, in Deep learning, Cambridge: MIT press, 2016, ch. 20, pp. 651–716. ch. 6, pp. 291–325. [online] Available: https://www.deeplearningbook.org/ [32] B. E. Boser, I. M. Guyon, and V. N. Vapnik, “A training algorithm for [59] C. Doersch, “Tutorial on variational autoencoders,” arXiv preprint optimal margin classifiers,” In Proceedings of the fifth annual workshop arXiv:1606.05908. 2016. on Computational learning theory (COLT ’92), New York, NY, USA, [60] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Association for Computing Machinery, pp. 144–152., 1992, [online] Ozair, A. Courville, Y. Bengio, Generative adversarial nets, Advances in Available: https://doi.org/10.1145/130385.130401 neural information processing systems, 2014. [33] F. Fourati., W. Souidene, and R. Attia, “An original framework for Wheat [61] A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, and Head Detection using Deep, Semi-supervised and Ensemble Learning A. A. Bharath, “Generative adversarial networks: An overview,” IEEE within Global Wheat Head Detection (GWHD) Dataset,” arXiv preprint, Signal Processing Magazine 35.1. 2018. pp. 53-65 . 2020, arXiv:2009.11977. [62] DD. Margineantu, and TG. Dietterich,“Pruning adaptive boosting,” [34] J. Cervantesa, F. Garcia-Lamonta, L. Rodr´ıguez-Mazahuab, and A. ICML. Vol. 97. 1997. Lopezc, “A comprehensive survey on support vector machine classifica- [63] J. Snoek, H. Larochelle, and RP. Adams, “Practical bayesian optimiza- tion: Applications, challenges and trends,” Neurocomputing, 2020, 408, tion of machine learning algorithms,” Advances in neural information 189-215. processing systems 25. 2012. 2951–2959. [35] J. R. Quinlan, “Induction of decision trees.” Machine learning 1.1, 1986, [64] RS. Sutton and GB. Andrew, “Reinforcement Learning: An Introduc- 81–106. tion,” A Bradford Book, Cambridge, MA, USA. 2018. [36] C. M. Bishop, “Graphical Models,” in Pattern Recognition and Machine [65] J. Anzalchi, A. Couchman, P. Gabellini, G. Gallinaro, L. D’Agristina, Learning, 1 st ed. Berlin, Heidelberg, Germany: Springer-Verlag, 2006, N. Alagha, and P. Angeletti, “Beam hopping in multi-beam broadband ch. 8, pp. 359–423. satellite systems: System simulation and performance comparison with [37] L. Breiman, “Random forests,” Machine learning 45.1, 2001, pp. 5-32. non-hopped systems,” in Proc. 5th Adv. Satell. Multimedia Syst. Conf. [38] L. Breiman, “Bagging predictors,” Machine learning 24.2, 1996, pp. 11th Signal Process. Space Commun. Workshop, Sep. 2010, pp. 248— 123-140. 255. [39] J. H. Friedman “Greedy function approximation: a gradient boosting [66] A. Freedman, D. Rainish, and Y. Gat, “Beam hopping: How to make it machine,” Annals of statistics, 2001, pp.1189-1232. possible,” in Proc. Broadband Commun. Conf., Oct. 2015, pp. 1—6. [40] [online] Available: https://xgboost.readthedocs.io/en/latest/ [67] L. Lei, E. Lagunas, Y. Yuan, M. G. Kibria, S. Chatzinotas, and B. [41] T. Chen and T. He, “Xgboost: extreme gradient boosting.” Ottersten, “Beam Illumination Pattern Design in Satellite Networks: Package Version: 1.3.2.1, Jan., 2021, [online] Available: Learning and Optimization for Efficient Beam Hopping,” in IEEE Access, https://cran.r-project.org/web/packages/xgboost/vignettes/xgboost.pdf vol. 8, pp. 136655-136667, 2020, doi: 10.1109/ACCESS.2020.3011746. [42] [online] Available: https://www.kaggle.com/ [68] P. Angeletti, D. Fernandez Prim, R. Rinaldo, “Beam hopping in multi- [43] P. Baldi and K. Hornik, “Neural networks and principal component anal- beam broadband satellite systems: system performance and payload ysis: Learning from examples without local minima.” Neural networks architecture analysis,” The 24th AIAA Int. Communications Satellite 2.1, 1989, pp. 53–58. Systems Conf., San Diego, June 2006 [44] C. M. Bishop, “Neural Networks” in Pattern Recognition and Machine [69] J. Anzalchi, A. Couchman, P. Gabellini, G. Gallinaro, L. D’Agristina, Learning, 1st ed. Berlin, Heidelberg, Germany: Springer-Verlag, 2006, N. Alagha, and P. Angeletti, “Beam hopping in multibeam broadband ch. 5, pp. 225–290. satellite systems: system simulation and performance comparison with [45] R. Hecht-Nielsen, “Theory of the backpropagation neural network.” non-hopped systems,” The 2010 5th Advanced Satellite Multimedia Neural networks for perception. Academic Press, 1992, pp. 65–93. Systems Conf. and the 11th Signal Processing for Space Communications [46] I. Goodfellow, Y. Bengio, and A. Courville, “Introduction,” in Deep Workshop, Cagliari, Italy, September 2010, pp. 248—255. learning, Cambridge: MIT press, 2016, ch. 1, pp. 1–26. [online] Avail- [70] X. Alberti, J. M. Cebrian, A. Del Bianco, Z. Katona, J. Lei, M. A able: https://www.deeplearningbook.org/ Vazquez-Castro, A. Zanus, L. Gilbert, and N. Alagha, “System capacity [47] I. Goodfellow, Y. Bengio, and A. Courville, “Convolutional Networks,” optimization in time and frequency for multibeam multi-media satellite in Deep learning, Cambridge: MIT press, 2016, ch. 9, pp. 326–366. systems,” in Proc. 11th Signal Process. Space Commun. Workshop, Sep. [online] Available: https://www.deeplearningbook.org/ 2010, pp. 226—233. [48] S. Albawi, T. A. Mohammed, and S. Al-Zawi, “Understanding of a con- [71] B. Evans and P. Thompson, “Key issues and technologies for a Terabit/s volutional neural network,” International Conference on Engineering and satellite,” The 28th AIAA Int. Communications Satellite Systems Conf. Technology (ICET), Antalya, 2017, pp. 1–6, doi: 10.1109/ICEngTech- (ICSSC 2010), Anaheim, California, USA, June 2010, p. 8713 nol.2017.8308186. [72] J. Lei and M. Vazquez-Castro, “Multibeam satellite frequency/time dual- [49] J. Redmon, S. Divvala, R. Girshick, and A. Farhadi “You only look once: ity study and capacity optimization,” in Proc. IEEE Int. Conf. Commun., Unified, real-time object detection.” Proceedings of the IEEE conference Oct. 2011, vol. 13, no. 5, pp. 471-–480. on computer vision and pattern recognition. 2016. [73] R. Alegre, N. Alagha, MA. Vazquez,´ “Heuristic algorithms for flexible [50] T. He, Z. Zhang, H. Zhang, Z. Zhang, J. Xie, M. Li “Bag of tricks for resource allocation in beam hopping multi-beam satellite systems,” The image classification with convolutional neural networks.” Proceedings of 29th AIAA Int. Communications Satellite Systems Conf. (ICSSC 2011), the IEEE Conference on Computer Vision and Pattern Recognition. 2019. Nara, Japan, July 2011, p. 8001 JAN. 2021 17

[74] R. Alegre, N. Alagha, MA. Vazquez,´ “Offered capacity optimization Approach,” in IEEE Transactions on Services Computing, vol. 9, no. 5, mechanisms for multi-beam satellite systems,” The 2012 IEEE Int. Conf. pp. 796-805, 1 Sept.-Oct. 2016, doi: 10.1109/TSC.2016.2599878. on Communications (ICC), Ottawa, ON, Canada, June 2012, pp. 3180— [95] C. Katris, S. Daskalaki, Comparing forecasting approaches for In- 3184 ternet traffic, Expert Systems with Applications, Volume 42, Is- [75] H. Liu, Z. Yang, Z. Cao, “Max-min rate control on traffic in broadband sue 21, 2015, 8172–8183, ISSN 0957-4174, [online] Available: multibeam satellite communications systems,” IEEE Commun. Lett., https://doi.org/10.1016/j.eswa.2015.06.029. 2013, 17, (7), pp. 1396—1399 [96] B. Gao, Q. Zhang, Y.-S. Liang, N.-N. Liu, C.-B. Huang, and N.-T. [76] H. Han, X. Zheng, Q. Huang, and Y. Lin, “QoS-equilibrium slot allo- Zhang, “Predicting self-similar networking traffic based on EMD and cation for beam hopping in broadband satellite communication systems,” ARMA,” 32. 47-56. 2011. Wirel. Netw., 2015, 21, (8), pp. 2617—2630 [97] X. Pan, W. Zhou, Y. Lu, and N. Sun, “Prediction of Network Traffic of [77] S. Shi, G. Li, Z. Li, H. Zhu, and B. Gao, “Joint power and bandwidth Smart Cities Based on DE-BP Neural Network,” in IEEE Access, vol. 7, allocation for beamhopping user downlinks in smart gateway multibeam pp. 55807–55816, 2019, doi: 10.1109/ACCESS.2019.2913017. satellite systems,” Int. J. Distrib. Sensor Netw., 2017, 13, (5), doi: [98] JX. LIU and ZH. JIA. “Telecommunication Traffic Prediction Based 1550147717709461 on Improved LSSVM,” International Journal of Pattern Recognition and [78] A. Ginesi, E. Re, and P.D. Arapoglou, “Joint beam hopping and Artificial Intelligence. 32. 10.1142/S0218001418500076. 2017. precoding in HTS systems,” Int. Conf. on Wireless and Satellite Systems, [99] L. Ziluan and L. Xin. “Short-term traffic forecasting based on principal OXFORD, GREAT BRITAIN, U.K, 2017, 43-–51 component analysis and a generalized regression neural network for [79] G. Cocco, T. de Cola, M. Angelone, Z. Katona, and S. Erl, “Radio satellite networks,” Journal of China Universities of Posts and Telecom- resource management optimization of flexible satellite payloads for DVB- munications. 25. 15-28+36. 10.19682/j.cnki.1005-8885.2018.0002. 2018. S2 systems,” IEEE Trans. Broadcast., vol. 64, no. 2, Jun. 2018, pp. 266— [100] Z. Na, Z. Pan, X. Liu, Z. Deng, Z. Gao, and Q. Guo, “Dis- 280 tributed Routing Strategy Based on Machine Learning for LEO [80] X. Hu, S. Liu, X. Hu, Y. Wang, L. Xu, Y. Zhang, C. Wang, and W. Satellite Network,” Wireless Communications and Mobile Computing, Wang, “Deep reinforcement learning-based beam Hopping algorithm in vol. 2018, Article ID 3026405, 10 pages, 2018. [online] Available: multibeam satellite systems,” IET Communications. pp. 2485–91, Jan. https://doi.org/10.1155/2018/3026405 2019. [101] GB. Huang, QY. Zhu, C. Siew, (2004). “Extreme learning machine: A [81] Y. Zhang, X. Hu, R. Chen, Z. Zhang, L. Wang, and W. Wang, “Dynamic new learning scheme of feedforward neural networks,” IEEE International Beam Hopping for DVB-S2X Satellite: A Multi-Objective Deep Rein- Conference on Neural Networks - Conference Proceedings. 2. 985–990 forcement Learning Approach,” 2019 IEEE International Conferences on vol.2. 10.1109/IJCNN.2004.1380068. Ubiquitous Computing & Communications (IUCC) and Data Science and [102] A. Goldsmith, “Path Loss and Shadowing,” in Wireless Communica- Computational Intelligence (DSCI) and Smart Computing, Networking tions, Cambridge University Press, 2005, ch. 2, pp. 25–48. and Services (SmartCNS), Shenyang, China, 2019, pp. 164–169, doi: [103] T. S. Rappaport, G. R. MacCartney, M. K. Samimi, and S. Sun, 10.1109/IUCC/DSCI/SmartCNS.2019.00056. “Wideband Millimeter-Wave Propagation Measurements and Channel [82] X. Hu, Y. Zhang, X. Liao, Z. Liu, W. Wang, and F. M. Ghannouchi, Models for Future Wireless Communication System Design,” in IEEE “Dynamic Beam Hopping Method Based on Multi-Objective Deep Rein- Transactions on Communications, vol. 63, no. 9, pp. 3029–3056, Sept. forcement Learning for Next Generation Satellite Broadband Systems,” 2015, doi: 10.1109/TCOMM.2015.2434384. in IEEE Transactions on Broadcasting, vol. 66, no. 3, Sept. 2020, pp. [104] S. Sangodoyin, S. Niranjayan, and A. F. Molisch, “A Measurement- 630–646, doi: 10.1109/TBC.2019.2960940. Based Model for Outdoor Near-Ground Ultrawideband Channels,” in [83] M. K. Simon, J. K. Omura, and R. A. Sholtz “Spread spectrum IEEE Transactions on Antennas and Propagation, vol. 64, no. 2, pp. 740– communications,” vols. 1-3. Computer Science Press, Inc., 1985. 751, Feb. 2016, doi: 10.1109/TAP.2015.2505004. [84] D. Torrieri, “Principles of spread-spectrum communication systems,” [105] C. Wang, J. Bian, J. Sun, W. Zhang, and M. Zhang, “A Survey of 5G Vol. 1. Heidelberg: Springer, 2005. Channel Measurements and Models,” in IEEE Communications Surveys [85] S. Bae, S. Kim, and J. Kim, “Efficient frequency-hopping synchroniza- & Tutorials, vol. 20, no. 4, pp. 3142-3168, Fourthquarter 2018, doi: tion for satellite communications using dehop-rehop transponders,” in 10.1109/COMST.2018.2862141. IEEE Transactions on Aerospace and Electronic Systems, vol. 52, no. [106] B. Ai, K. Guan, R. He, J. Li, G. Li, D. He, Z. Zhong, and K. M. S. Huq, 1, pp. 261–274, Feb. 2016, doi: 10.1109/TAES.2015.150062. “On Indoor Millimeter Wave Massive MIMO Channels: Measurement and [86] F. Yao, L. Jia, Y. Sun, Y. Xu, S. Feng, and Y. Zhu, “A hierarchical Simulation,” in IEEE Journal on Selected Areas in Communications, vol. learning approach to anti-jamming channel selection strategies,” Wirel. 35, no. 7, July 2017, pp. 1678–1690, doi: 10.1109/JSAC.2017.2698780. Netw., vol. 25, no. 1, Jan. 2019, pp. 201–213. [107] G. Liang and H. L. Bertoni, “A new approach to 3-D ray tracing for [87] C. Han and Y. Niu, “Cross-Layer Anti-Jamming Scheme: A Hierarchical propagation prediction in cities,” IEEE Trans. Antennas Propag., vol. 46, Learning Approach,” IEEE Access, vol. 6, pp. 34874-34883, Jun. 2018. no. 6, Jun. 1998, pp. 853—863. [88] S. Lee, S. Kim, M. Seo, and D. Har, “Synchronization of Frequency [108] M. Zhu, A. Singh, and F. Tufvesson, “Measurement based ray launch- Hopping by LSTM Network for Satellite Communication System,” in ing for analysis of outdoor propagation,” 2012 6th European Conference IEEE Communications Letters, vol. 23, no. 11, Nov. 2019, pp. 2054– on Antennas and Propagation (EUCAP), Prague, 2012, pp. 3332–3336, 2058, , doi: 10.1109/LCOMM.2019.2936019. doi: 10.1109/EuCAP.2012.6206329. [89] C. Han, L. Huo, X. Tong, H. Wang, and X. Liu, “Spatial Anti- [109] Z. Yun and M. F. Iskander, “Ray Tracing for Radio Propagation Jamming Scheme for Internet of Satellites Based on the Deep Rein- Modeling: Principles and Applications,” in IEEE Access, vol. 3, 2015, forcement Learning and Stackelberg Game,” in IEEE Transactions on pp. 1089–1100, doi: 10.1109/ACCESS.2015.2453991. Vehicular Technology, vol. 69, no. 5, May 2020, pp. 5331–5342 , doi: [110] D. J. Cichon and T. Kurner,¨ “Propagation prediction models,” Florence, 10.1109/TVT.2020.2982672. Italy, Tech. Rep. COST-231 TD (95) 66, Apr. 1995, pp. 115–207. [90] C. Han, A. Liu, H. Wang, L. Huo, and X. Liang, “Dynamic Anti- [111] L. C. Fernandes and A. J. M. Soares, “Simplified characterization of the Jamming Coalition for Satellite-Enabled Army IoT: A Distributed Game urban propagation environment for path loss calculation,” IEEE Antennas Approach,” in IEEE Internet of Things Journal, vol. 7, no. 11, Nov. 2020, Wireless Propag. Lett., vol. 9, 2010, pp. 24—27. pp.10932–10944, doi: 10.1109/JIOT.2020.2991585. [112] L. C. Fernandes and A. J. M. Soares, “On the use of image segmenta- [91] Y. Bie, L. Wang, Y. Tian, and Z. Hu, “A Combined Forecasting Model tion for propagation path loss prediction,” in IEEE MTT-S Int. Microw. for Satellite Network Self-Similar Traffic,” in IEEE Access, vol. 7, 2019, Symp. Dig., Oct. 2011, pp. 129—133. pp. 152004–152013, doi: 10.1109/ACCESS.2019.2944895. [113] M. Piacentini and F. Rinaldi, “Path loss prediction in urban environ- [92] L. Rossi, J. Chakareski, P. Frossard, and S. Colonnese, “A Poisson ment using learning machines and dimensionality reduction techniques,” Hidden Markov Model for Multiview Video Traffic,” in IEEE/ACM Comput. Manage. Sci., vol. 8, no. 4, Nov. 2011, 371—385. Transactions on Networking, vol. 23, no. 2, April 2015, pp. 547–558, [114] M. Uccellari, F. Facchini, M. Sola, E. Sirignano, G. M. Vitetta, A. doi: 10.1109/TNET.2014.2303162. Barbieri, and S. Tondelli, “On the use of support vector machines for the [93] D. Yan and L. Wang, “TPDR: Traffic prediction based dynamic routing prediction of propagation losses in smart metering systems,” in Proc. IE for LEO&GEO satellite networks,” 2015 IEEE 5th International Confer- [115] S. P. Sotiroudis, S. K. Goudos, K. A. Gotsis, K. Siakavara, and J. ence on Electronics Information and Emergency Communication, Beijing, N. Sahalos, “Application of a composite differential evolution algorithm 2015, pp. 104–107, doi: 10.1109/ICEIEC.2015.7284498. in optimal neural network design for propagation path-loss prediction in [94] F. Xu, Y. Lin, J. Huang, D. Wu, H. Shi, J. Song, and Y. Li, “Big Data mobile communication systems,” IEEE Antennas Wireless Propag. Lett., Driven Mobile Traffic Understanding and Forecasting: A Time Series vol. 12, 2013, pp. 364—367. JAN. 2021 18

[116] S. P. Sotiroudis and K. Siakavara, “Mobile radio propagation path [135] S. K. Ibrahim, A. Ahmed, M. A. E. Zeidan, and I. E. Ziedan, loss prediction using Artificial Neural Networks with optimal input “Machine Learning Methods for Spacecraft Telemetry Mining,” in IEEE information for urban environments,” AEU-Int. J. Electron. Commun., Transactions on Aerospace and Electronic Systems, vol. 55, no. 4, pp. vol. 69, no. 10, Oct. 2015, pp. 1453—1463. 1816–1827, Aug. 2019, doi: 10.1109/TAES.2018.2876586. [117] I. Popescu, I. Nafornita, and P. Constantinou, “Comparison of neural [136] P. Wan, Y. Zhan, and W. Jiang, “Study on the Satellite Telemetry Data network models for path loss prediction,” in Proc. IEEE Int. Conf. Classification Based on Self-Learning,” in IEEE Access, vol. 8, pp. 2656- Wireless Mobile Comput., Netw. Commun., Aug. 2005, pp. 44—49. 2669, 2020, doi: 10.1109/ACCESS.2019.2962235. [118] E. Ostlin, H.-J. Zepernick, and H. Suzuki, “Macrocell path-loss pre- [137] P. W. Ward, J. W. Betz, and C. J. Hegarty, “Satellite signal acquisition, diction using artificial neural networks,” IEEE Trans. Veh. Technol., vol. tracking, and data demodulation in Understanding GPS: Principles and 59, no. 6, Jul. 2010, pp. 2735-–2747. Applications,” Norwood, MA, USA: Artech House, pp. 153–241, 2006. [119] B. J. Cavalcanti, G. A. Cavalcante, L. M. D. Mendonc¸a, G. M. [138] A. V. Dierendonck, J. Klobuchar, and Q. Hua, “Ionospheric scintillation Cantanhede, M. M. de Oliveira, and A. G. D’Assunc¸ao,˜ “A hybrid path monitoring using commercial single frequency C/A code receivers,” in loss prediction model based on artificial neural networks using empirical Proc. 6th Int. Tech. Meet. Satellite Div. Inst. Navig., Salt Lake City, UT, models for LTE and LTE-A at 800 MHz and 2600 MHz,” J. Microw., USA, vol. 93, pp. 1333—1342, Sep. 1993. Optoelectron. Electromagn. Appl., vol. 16, Sep. 2017, pp.708—722. [139] J. Lee, Y. T. J. Morton, J. Lee, H.-S. Moon, and J. Seo, “Monitoring [120] Y. Zhang, J. Wen, G. Yang, Z. He, and X. Luo, “Air-to-air path loss and mitigation of ionospheric anomalies for GNSSbased safety critical prediction based on machine learning methods in urban environments,” systems: A review of up-to-date signal processing techniques,” IEEE Wireless Commun. Mobile Comput., vol. 6, May 2018, Art. no. 8489326. Signal Process. Mag., vol. 34, no. 5, pp. 96-–110, Sep. 2017 [121] C. A. Oroza, Z. Zhang, T. Watteyne, and S. D. Glaser, “A [140] C. Cesaroni, L. Alfonsi, R. Romero, N. Linty, F. Dovis, S. V. Veettil, J. machinelearning-based connectivity model for complex terrain large-scale Park, D. Barroca, M. C. Ortega, and R. O. Perez, “Monitoring Ionosphere lowpower wireless deployments,” IEEE Trans. Cogn. Commun. Netw., Over South America: The MImOSA and MImOSA2 projects,” 2015 vol. 3, no. 4, Dec. 2017 pp. 576—584. International Association of Institutes of Navigation World Congress [122] Y. Zhang, J. Wen, G. Yang, Z. He, and J. Wang, “Path loss prediction (IAIN), Prague, pp. 1–7, 2015, doi: 10.1109/IAIN.2015.7352226. based on machine learning: Principle, method, and data expansion,” Appl. Sci., vol. 9, p. 1908, May 2019. [141] L. Nicola, R. Rodrigo, C. Calogero, D. Fabio, B. Michele, C. J. [123] H. F. Ates, S. M. Hashir, T. Baykas, and B. K. Gunturk, “Path Loss Thomas, F. G. Joaquim, W. Jonathan, L. Gert, R. Padraig, C. Pierre, Exponent and Shadowing Factor Prediction From Satellite Images Using C. Emilia, and A. Lucilla, “Ionospheric scintillation threats to GNSS in Deep Learning,” in IEEE Access, vol. 7, 2019, pp. 101366–101375, doi: polar regions: the DemoGRAPE case study in Antarctica,” in Proc. Eur. 10.1109/ACCESS.2019.2931072. Navig. Conf., pp. 1—7, 2016. [124] J. Thrane, D. Zibar, and H. L. Christiansen, “Model-Aided Deep [142] J. Vila-Valls, P. Closas, C. Fernandez-Prades, and J. T. Curran, “On Learning Method for Path Loss Prediction in Mobile Communication the ionospheric scintillation mitigation in advanced GNSS receivers IEEE Systems at 2.6 GHz,” in IEEE Access, vol. 8, 2020, pp. 7925–7936, doi: Trans.” Aerosp. Electron. Syst., to be published. 10.1109/ACCESS.2020.2964103. [143] S. Taylor, Y. Morton, Y. Jiao, J. Triplett, and W. Pelgrum, “An improved [125] O. Ahmadien, H. F. Ates, T. Baykas, and B. K. Gunturk, “Predict- ionosphere scintillation event detection and automatic trigger for GNSS ing Path Loss Distribution of an Area From Satellite Images Using data collection systems,” in Proc Int. Tech. Meet. Inst. Navig., pp. 1563— Deep Learning,” in IEEE Access, vol. 8, 2020, pp. 64982–64991, doi: 1569, 2012. 10.1109/ACCESS.2020.2985929. [144] W. Fu, S. Han, C. Rizos, M. Knight, and A. Finn, “Real-time iono- [126] T. Yairi, N. Takeishi, T. Oda, Y. Nakajima, N. Nishimura, and N. spheric scintillation monitoring,” in Proc. 12th Int. Tech. Meet. Satellite Takata, “A Data-Driven Health Monitoring Method for Satellite House- Div. Inst. Navig., vol. 99, pp. 14—17, 1999. keeping Data Based on Probabilistic Clustering and Dimensionality Re- [145] S. Miriyala, P. R. Koppireddi, and S. R. “Chanamallu Robust detection duction,” in IEEE Transactions on Aerospace and Electronic Systems, vol. of ionospheric scintillations using MF-DFA technique Earth,” Planets Sp., 53, no. 3, June 2017, pp. 1384–1401, doi: 10.1109/TAES.2017.2671247. vol. 67, no. 98, pp. 1—5, 2015. [127] T. Yairi, T. Tagawa, and N. Takata, “Telemetry monitoring by dimen- [146] R. Romero, N. Linty, F. Dovis, and R. V. Field, “A novel approach to sionality reduction and learning hidden markov model,” in Proceedings ionospheric scintillation detection based on an open loop architecture,” in of International Symposium on Artificial Intelligence, Robotics and Proc. 8th ESA Workshop Satellite Navig. Technol. Eur. Workshop GNSS Automation in Space, 2012. Signals Signal Process., pp. 1—9, Dec. 2016. [128] T. Yairi, M. Nakatsugawa, K. Hori, S. Nakasuka, K. Machida and [147] L. F. C. Rezende, E. R. de Paula, S. Stephany, I. J. Kantor, M. T. A. N. Ishihama, “Adaptive limit checking for spacecraft telemetry data H. Muella, P. M. de Siqueira and K. S. Correa, “Survey and prediction of using regression tree learning,” 2004 IEEE International Conference on the ionospheric scintillation using data mining techniques,” Sp. Weather, Systems, Man and Cybernetics (IEEE Cat. No.04CH37583), The Hague, vol. 8, no. 6, pp. 1—10, 2010. 2004, pp. 5130–5135 vol.6, doi: 10.1109/ICSMC.2004.1401008. [148] Y. Jiao, J. J. Hall, and Y. T. “Morton Performance evaluations of an [129] T. Shahroz, L. Sangyup, S. Youjin, L. Myeongshin, J. Okchul, C. equatorial GPS amplitude scintillation detector using a machine learning Daewon, and S. W. Simon, “Detecting Anomalies in Space using Multi- algorithm,” in Proc 29th Int. Tech. Meet. Satellite Div. Inst. Navig., pp. variate Convolutional LSTM with Mixtures of Probabilistic PCA,” 25th 195—199, Sep. 2016. ACM Special Interest Group on Knowledge Discovery and Data Mining [149] Y. Jiao, J. J. Hall, and Y. T. Morton, “Automatic equatorial GPS International Conference, Alaska, USA, 2019. amplitude scintillation detection using a machine learning algorithm,” [130] K. Hundman, V. Constantinou, C. Laporte, I. Colwell, and T. Soder- IEEE Trans. Aerosp. Electron. Syst., vol. 53, no. 1, pp. 405—418, Feb. strom, “Detecting Spacecraft Anomalies Using LSTMs and Nonparamet- 2017. ric Dynamic Thresholding,” 24th ACM Special Interest Group on Knowl- edge Discovery and Data Mining International Conference. London, UK, [150] Y. Jiao, J. J. Hall, and Y. T. Morton, “Automatic GPS phase scintillation 2018. detector using a machine learning algorithm,” in Proc. Int. Tech. Meet. [131] S. Fuertes, G. Picart, JY. Tourneret, L. Chaari, A. Ferrari, and Inst. Navig., Monterey, CA, USA, pp. 1160—1172, Jan. 2017. C. Richard, “Improving Spacecraft Health Monitoring with Automatic [151] Y. Jiao, J. J. Hall, and Y. T. Morton, “Performance evaluation of an Anomaly Detection Techniques,” 14th International Conference on Space automatic GPS ionospheric phase scintillation detector using a machine- Operations. Daejeon, Korea, 2016. learning algorithm Navigation,” vol. 64, no. 3, pp. 391—402, 2017. [132] D. L. Iverson, R. Martin, M. Schwabacher, L. Spirkovska, W. Taylor, [152] N. Linty, A. Farasin, A. Favenza, and F. Dovis, “Detection of GNSS R. Mackey, J. P. Castle and V. Baskaran, “General Purpose DataDriven Ionospheric Scintillations Based on Machine Learning Decision Tree,” in Monitoring for Space Operations,” Journal of Aerospace Computing IEEE Transactions on Aerospace and Electronic Systems, vol. 55, no. 1, Information & Communication, 9(2):26-44 2012. pp. 303–317, Feb. 2019, doi: 10.1109/TAES.2018.2850385. [133] PI. Robinson, M. H. Shirley, D. Fletcher, R. Alena, D. Duncavage, [153] R. Imam and F. Dovis, “Distinguishing Ionospheric Scintillation from and C. Lee “Applying model-based reasoning to the fdir of the command Multipath in GNSS Signals Using Bagged Decision Trees Algorithm,” and data handling subsystem of the international space station,” in 2020 IEEE International Conference on Wireless for Space and Ex- Proc. of International Symposium on Artificial Intelligence, Robotics and treme Environments (WiSEE), Vicenza, Italy, 2020, pp. 83-88, doi: Automation in Space, 2003. 10.1109/WiSEE44079.2020.9262699. [134] Y. Sun, L. Guo, Y. Wang, Z. Ma, and Y. Niu, “Fault diagnosis for [154] C. Politis, S. MalekiSina, M. Christos, G. TsinosChristos, G. T. Show, space utilisation,” in The Journal of Engineering, vol. 2019, no. 23, pp. “On-board the Satellite Interference Detection with Imperfect Signal 8770-8775, 12 2019, doi: 10.1049/joe.2018.9102. Cancellation,” JAN. 2021 19

[155] A. V. Dandawate and G. B. Giannakis, “Statistical tests for presence [174] L. Zong-ling et al., “Remote Sensing Ship Target Detection and Recog- of cyclostationarity,” in IEEE Transactions on Signal Processing, vol. 42, nition System Based on Machine Learning,” IGARSS 2019 - 2019 IEEE no. 9, pp. 2355–2369, Sept. 1994, doi: 10.1109/78.317857. International Geoscience and Remote Sensing Symposium, Yokohama, [156] O. A. Dobre, A. Abdi, Y. Bar-Ness, and W. Su, “Survey of auto- Japan, pp. 1272–1275, 2019, doi: 10.1109/IGARSS.2019.8898599. matic modulation classification techniques: classical approaches and new [175] H. Bandarupally, H. R. Talusani, and T. Sridevi, “Detection of Mil- trends,” in IET Communications, vol. 1, no. 2, pp. 137–156, April 2007, itary Targets from Satellite Images using Deep Convolutional Neural doi: 10.1049/iet-com:20050176. Networks,” 2020 IEEE 5th International Conference on Computing Com- [157] J. Hu, D. Bian, Z. Xie, Y. Li, and L. Fan, “An approach for narrow munication and Automation (ICCCA), Greater Noida, India, pp. 531–535, band interference detection in satellite communication using morpho- 2020, doi: 10.1109/ICCCA49541.2020.9250864. logical filter,” International Conference on Information Technology and [176] J. Zheng, X. -Y. Liu, and X. Wang, “Single Image Cloud Removal Management Innovation, Shenzhen, China, Sept., Using U-Net and Generative Adversarial Networks,” in IEEE Transactions [158] Q. Liu, J. Yang, C. Zhuang, A. Barnawi, and B. A Alzahrani, “Artificial on Geoscience and Remote Sensing, doi: 10.1109/TGRS.2020.3027819. Intelligence Based Mobile Tracking and Antenna Pointing in Satellite- [177] O. Ronneberger, P. Fischer, and T. Brox. “U-net: Convolutional net- Terrestrial Network,” in IEEE Access, vol. 7, pp. 177497–177503, 2019, works for biomedical image segmentation,” International Conference on doi: 10.1109/ACCESS.2019.2956544. Medical image computing and computer-assisted intervention. Springer, [159] L. Pellaco, N. Singh, and J. Jalden.´ “Spectrum Prediction and Cham, 2015. Interference Detection for Satellite Communications,” arXiv preprint [178] J. Lu, Y. Chen, and R. He, “A Learning-Based Approach for Agile arXiv:1912.04716, 2019. Satellite Onboard Scheduling,” in IEEE Access, vol. 8, pp. 16941-16952, [160] P. Henarejos, M. A.´ Vazquez,´ and A. I. Perez-Neira,´ “Deep Learning 2020, doi: 10.1109/ACCESS.2020.2968051. For Experimental Hybrid Terrestrial and Satellite Interference Manage- [179] R. Mital, K. Cates, J. Coughlin and G. Ganji, “A Machine Learning ment,” 2019 IEEE 20th International Workshop on Signal Processing Approach to Modeling Satellite Behavior,” 2019 IEEE International Advances in Wireless Communications (SPAWC), Cannes, France, 2019, Conference on Space Mission Challenges for Information Technology pp. 1–5, doi: 10.1109/SPAWC.2019.8815532. (SMC-IT), Pasadena, CA, USA, pp.62–69, 2019, doi: 10.1109/SMC- [161] N. Kussul, M. Lavreniuk, S. Skakun, and A. Shelestov, “Deep Learning IT.2019.00013. Classification of Land Cover and Crop Types Using Remote Sensing [180] K. Weasenforth, J. Hollon, T. Payne, K. Kinateder, and A. Kruchten, Data,” in IEEE Geoscience and Remote Sensing Letters, vol. 14, no. 5, “Machine Learning-based Stability Assessment and Change Detection for pp. 778–782, May 2017, doi: 10.1109/LGRS.2017.2681128. Geosynchronous Satellites,” Advanced Maui Optical and Space Surveil- [162] F. Zhang, B. Du, and L. Zhang, “Scene Classification via a Gradient lance Technologies Conference, 2018. Boosting Random Convolutional Network Framework,” in IEEE Transac- [181] B. Jia, K. D. Pham, E. Blasch, Z. Wang, D. Shen, and G. Chen, tions on Geoscience and Remote Sensing, vol. 54, no. 3, pp. 1793–1802, “Space object classification using deep neural networks,” in 2018 IEEE March 2016, doi: 10.1109/TGRS.2015.2488681. Aerospace Conference, Big Sky, MT, pp. 1—8, 2018. [163] A. S. Li, V. Chirayath, M. Segal-Rozenhaimer, J. L. Torres-Perez,´ [182] K. Hundman, V. Constantinou, C. Laporte, I. Colwell, and T. Soder- and J. van den Bergh, “NASA NeMO-Net’s Convolutional Neural Net- strom, “Detecting Spacecraft Anomalies Using LSTMs and Nonparamet- work: Mapping Marine Habitats with Spectrally Heterogeneous Remote ric Dynamic Thresholding,” in Proceedings of the 24th ACM Special Sensing Imagery,” in IEEE Journal of Selected Topics in Applied Earth Interest Group on Knowledge Discovery and Data Mining International Observations and Remote Sensing, vol. 13, pp. 5115–5133, 2020, doi: Conference on Knowledge Discovery & Data Mining - KDD ’18, London, 10.1109/JSTARS.2020.3018719. United Kingdom, pp. 387-–395, 2018. [164] S. A. Fatima, A. Kumar, A. Pratap, and S. S. Raoof, “Object [183] B. Chen, J. Cao, A. Parra, and T. Chin, “Satellite Pose Estimation Recognition and Detection in Remote Sensing Images: A Compara- with Deep Landmark Regression and Nonlinear Pose Refinement,” 2019 tive Study,” 2020 International Conference on Artificial Intelligence IEEE/CVF International Conference on Computer Vision Workshop (IC- and Signal Processing (AISP), Amaravati, India, pp. 1–5, 2020, doi: CVW), Seoul, Korea (South), pp. 2816–2824, 2019, doi: 10.1109/IC- 10.1109/AISP48273.2020.9073614. CVW.2019.00343. [165] L. Zhou, J. Liu, and L. Chen, “Vehicle detection based on remote sens- [184] M. Kisantal, S. Sharma, T. H. Park, D. Izzo, M. Martens,¨ and ing image of Yolov3,” 2020 IEEE 4th Information Technology, Network- S. D’Amico, “Satellite Pose Estimation Challenge: Dataset, Compe- ing, Electronic and Automation Control Conference (ITNEC), Chongqing, tition Design, and Results,” in IEEE Transactions on Aerospace and China, pp. 468–472, 2020, doi: 10.1109/ITNEC48623.2020.9084975. Electronic Systems, vol. 56, no. 5, pp. 4083–4098, Oct. 2020, doi: [166] J. Redmon, et al. “You only look once: Unified, real-time object 10.1109/TAES.2020.2989063. detection,” Proceedings of the IEEE conference on computer vision and [185] S. Jahirabadkar, P. Narsay, S. Pharande, G. Deshpande, and pattern recognition. 2016. A. Kitture, “Space Objects Classification Techniques: A Survey,” [167] J. Redmon and A. Farhadi. “Yolov3: An incremental improvement,” 2020 International Conference on Computational Performance arXiv preprint arXiv:1804.02767, 2018. Evaluation (ComPE), Shillong, India, pp. 786–791, 2020, doi: [168] A. Femin and K. S. Biju, “Accurate Detection of Buildings from 10.1109/ComPE49325.2020.9199996. Satellite Images using CNN,” 2020 International Conference on Electrical, [186] D. Yadava, R. Hosangadi, S. Krishna, P. Paliwal, and A. Jain, “Attitude Communication, and Computer Engineering (ICECCE), Istanbul, Turkey, control of a nanosatellite system using reinforcement learning and neural pp. 1–5, 2020, doi: 10.1109/ICECCE49384.2020.9179232. networks,” 2018 IEEE Aerospace Conference, Big Sky, MT, pp. 1–8, [169] A. Hassan, W. M. Hussein, E. Said and M. E. Hanafy, “A Deep Learn- 2018, doi: 10.1109/AERO.2018.8396409. ing Framework for Automatic Airplane Detection in Remote Sensing [187] A. M. Ahmed, A. Salama, H. A. Ibrahim, M. A. E. Sayed, and S. Satellite Images,” 2019 IEEE Aerospace Conference, Big Sky, MT, USA, Yacout, “Prediction of Battery Remaining Useful Life on Board Satellites pp. 1–10, 2019, doi: 10.1109/AERO.2019.8741938. Using Logical Analysis of Data,” 2019 IEEE Aerospace Conference, Big [170] G. Mateo-Garcia, V. Laparra, D. Lopez-Puigdollers, and L. Gomez- Sky, MT, USA, pp. 1–8, 2019, doi: 10.1109/AERO.2019.8741717. Chova, “Cross-Sensor Adversarial Domain Adaptation of Landsat-8 and [188] JH. Lee, J. Park, M. Bennis, YC. Ko, “Integrating LEO Satellite Proba-V images for Cloud Detection,” in IEEE Journal of Selected Topics and UAV Relaying via Reinforcement Learning for Non-Terrestrial Net- in Applied Earth Observations and Remote Sensing, doi: 10.1109/JS- works,” arXiv preprint arXiv:2005.12521, 2020. TARS.2020.3031741. [189] N. Cheng, F. Lyu, W. Quan, C. Zhou, H. He, W. Shi, and X. [171] Z. Shao, Y. Pan, C. Diao, and J. Cai, “Cloud Detection in Remote Shen, “Space/Aerial-Assisted Computing Offloading for IoT Applica- Sensing Images Based on Multiscale Features-Convolutional Neural Net- tions: A Learning-Based Approach,” in IEEE Journal on Selected Areas work,” in IEEE Transactions on Geoscience and Remote Sensing, vol. in Communications, vol. 37, no. 5, pp. 1117–1129, May 2019, doi: 57, no. 6, pp. 4062–4076, June 2019, doi: 10.1109/TGRS.2018.2889677. 10.1109/JSAC.2019.2906789. [172] M. Tian, H. Chen, and G. Liu, “Cloud Detection and Classifica- [190] C. Jiang and X. Zhu, “Reinforcement Learning Based Capacity Man- tion for S-NPP FSR CRIS Data Using Supervised Machine Learn- agement in Multi-Layer Satellite Networks,” in IEEE Transactions on ing,” IGARSS 2019 - 2019 IEEE International Geoscience and Re- Wireless Communications, vol. 19, no. 7, pp. 4685–4699, July 2020, doi: mote Sensing Symposium, Yokohama, Japan, pp. 9827–9830, 2019, doi: 10.1109/TWC.2020.2986114. 10.1109/IGARSS.2019.8898876. [191] C. Qiu, H. Yao, F. R. Yu, F. Xu, and C. Zhao, “Deep Q-Learning [173] F. Wang, F. Liao, and H. Zhu, “FPA-DNN: A Forward Propagation Aided Networking, Caching, and Computing Resources Allocation in Acceleration based Deep Neural Network for Ship Detection,” 2020 Inter- Software-Defined Satellite-Terrestrial Networks,” in IEEE Transactions national Joint Conference on Neural Networks (IJCNN), Glasgow, United on Vehicular Technology, vol. 68, no. 6, pp. 5871–5883, June 2019, doi: Kingdom, pp. 1–8, 2020, doi: 10.1109/IJCNN48605.2020.9207603. 10.1109/TVT.2019.2907682. JAN. 2021 20

[192] W. Liu, F. Tian, and Z. Jiang, “Beam-hopping based resource allocation statistical learning method”, IEEE Trans. Antennas Propag., vol. 66, no. algorithm in LEO satellite network,” in Proc. Int. Conf. Space Inf. Netw. 8, pp. 3995-4007, Aug. 2018. Singapore: Springer, pp. 113—123, 2018. [214] D. R. Prado, J. A. Lopez-Fern´ andez,´ G. Barquero, M. Arrebola, and [193] Z. Qu, G. Zhang, H. Cao, and J. Xie, “LEO satellite constellation for F. Las-Heras, “Fast and accurate modeling of dual-polarized reflectarray Internet of Things,” IEEE Access, vol. 5, pp. 18391—18401, 2017. unit cells using support vector machines”, IEEE Trans. Antennas Propag., [194] H. Tsuchida, Y. Kawamoto, N. Kato, K. Kaneko, S. Tani, S. Uchida, vol. 66, no. 3, pp. 1258-1270, Mar. 2018. and H. Aruga, “Efficient Power Control for Satellite-Borne Batteries [215] D. R. Prado, J. A. Lopez-Fern´ andez,´ M. Arrebola, and G. Goussetis, Using Q-Learning in Low-Earth-Orbit Satellite Constellations,” in IEEE “Support vector regression to accelerate design and crosspolar optimiza- Wireless Communications Letters, vol. 9, no. 6, pp. 809-812, June 2020, tion of shaped-beam reflectarray antennas for space applications”, IEEE doi: 10.1109/LWC.2020.2970711. Trans. Antennas Propag., vol. 67, no. 3, pp. 1659-1668, Mar. 2019. [195] B. Zhao, J. Liu, Z. Wei, and I. You, “A Deep Reinforcement Learning [216] D. R. Prado, J. A. Lopez-Fern´ andez,´ M. Arrebola, M. R. Pino, Based Approach for Energy-Efficient Channel Allocation in Satellite and G. Goussetis, “Wideband Shaped-Beam Reflectarray Design Using Internet of Things,” in IEEE Access, vol. 8, pp. 62197-62206, 2020, doi: Support Vector Regression Analysis,” in IEEE Antennas and Wireless 10.1109/ACCESS.2020.2983437. Propagation Letters, vol. 18, no. 11, pp. 2287-2291, Nov. 2019, doi: [196] G. Cui, X. Li, L. Xu, and W. Wang, “Latency and Energy Optimization 10.1109/LAWP.2019.2932902. for MEC Enhanced SAT-IoT Networks,” in IEEE Access, vol. 8, pp. [217] P. Henttu and S. Aromaa, “Consecutive mean excision algorithm”, Proc. 55915-55926, 2020, doi: 10.1109/ACCESS.2020.2982356. IEEE 7th Int. Symp. Spread Spectr. Techn. Appl., vol. 2, pp. 450-454, [197] C. Zhang, “An AI-based optimization of handover strategy in non- Sep. 2002. terrestrial networks,” presented at the 12th ITU Academic Conference [218] H. Saarnisaari, “Consecutive mean excision algorithms in narrowband Kaleidoscope Industry-driven digital transformation, Online, Dec. 7-11, or short time interference mitigation”, Proc. PLNS, pp. 447-454, Apr. 2020. 2004. [198] X. Chen, W. Yao, Y. Zhao, X. Chen, and X. Zheng, “A practical [219] H. Saarnisaari and P. Henttu, “Impulse detection and rejection methods satellite layout optimization design approach based on enhanced finite- for radio systems”, Proc. MILCOM, vol. 2, pp. 1126-1131, Oct. 2003. circle method”, Struct. Multidisciplinary Optim., vol. 58, no. 6, pp. 2635- [220] H. G. Keane, “A new approach to frequency line tracking”, Proc. 2653, Dec. 2018. ACSSC, vol. 2, pp. 808-812, Nov. 1991. [199] K. Chen, J. Xing, S. Wang, and M. Song, “Heat source layout opti- [221] R. Eschbach, Z. Fan, K. T. Knox, and G. Marcu, “Threshold modulation mization in two-dimensional heat conduction using simulated annealing and stability in error diffusion”, IEEE Signal Process. Mag., vol. 20, pp. method”, Int. J. Heat Mass Transf., vol. 108, pp. 210-219, May 2017. 39-50, Jul. 2003. [200] Y. Aslan, J. Puskely, and A. Yarovoy, “Heat source layout optimization [222] H. Mustafa, M. Doroslovacki, and H. Deng, “Algorithms for emitter for two-dimensional heat conduction using iterative reweighted L1-norm detection based on the shape of power spectrum”, Proc. CISS, pp. 808- convex minimization,” Int. J. Heat Mass Transf., vol. 122, pp. 432-441, 812, Mar. 2003. Jul. 2018. [223] J. Vartiainen, J. Lehtomaki,¨ S. Aromaa, and H. Saarnisaari, “Local- [201] K. Chen, S. Wang, and M. Song, “Temperature-gradient-aware bionic ization of multiple narrowband signals based on the FCME algorithm,” optimization method for heat source distribution in heat conduction”, Int. Proc. NRS, vol. 1, pp. 5, Aug. 2004. J. Heat Mass Transf., vol. 100, pp. 737-746, Sep. 2016. [224] J. Vartiainen, J. J. Lehtomaki, and H. Saarnisaari, “Double-threshold [202] J. Sun, J. Zhang, X. Zhang, and W. Zhou, “A Deep Learning-Based based narrowband signal extraction,” Proc. VTC, vol. 2, pp. 1288-1292, Method for Heat Source Layout Inverse Design,” in IEEE Access, vol. May 2005. 8, pp. 140038-140053, 2020, doi: 10.1109/ACCESS.2020.3013394. [225] J. Kim, M. Kim, I. Won, S. Yang, K. Lee, and W. Huh, “A biomedical [203] H. Li, P. Wang, C. Shen, and G. Zhang, ”Show, attend and read: A signal segmentation algorithm for event detection based on slope tracing,” simple and strong baseline for irregular text recognition,” Proceedings of Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., pp. 1889-1892, Sep. the AAAI Conference on Artificial Intelligence. Vol. 33. 2019. 2009. [204] Y. Zhang and W. Ye, “Deep learning–based inverse method for layout [226] O. A. Morozov, and P. E. Ovchinnikov, “Neural network detection of design”, Struct. Multidisciplinary Optim., vol. 16, no. 3, pp. 774-788, MSK signals,” Proc. IEEE 13th Digit. Signal Process. Workshop 5th IEEE 2019. Signal Process. Educ. Workshop, pp. 594-596, Jan. 2009. [205] J. Peurifoy, Y. Shen, L. Jing, Y. Yang, F. Cano-Renteria, B. G. DeLacy, [227] Y. Yuan, Z. Sun, Z. Wei, and K. Jia, “DeepMorse: A deep convolutional J. D. Joannopoulos, M. Tegmark, and M. Soljaciˇ c,´ “Nanophotonic particle learning method for blind morse signal detection in wideband wireless simulation and inverse design using artificial neural networks”, Sci. Adv., spectrum,” IEEE Access, vol. 7, pp. 80577-80587, 2019. vol. 4, no. 6, Jun. 2018. [228] H. Huang, J. Li, J. Wang, and H. Wang, “FCN-Based Carrier Signal [206] J. Tompson, K. Schlachter, P. Sprechmann, and K. Perlin, “Accelerating Detection in Broadband Power Spectrum,” in IEEE Access, vol. 8, pp. eulerian fluid simulation with convolutional networks”, Proc. 5th Int. 113042-113051, 2020, doi: 10.1109/ACCESS.2020.3003683. Conf. Learn. Represent. ICLR, pp. 3424-3433, Apr. 2017, [online] [229] E. Shelhamer, J. Long, and T. Darrell, “Fully convolutional networks Available: http://OpenReview.net. for semantic segmentation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. [207] A. Agrawal, P. D. Deshpande, A. Cecen, G. P. Basavarsu, A. N. 39, no. 4, pp. 640-651, Apr. 2017. Choudhary, and S. R. Kalidindi, “Exploration of data science techniques [230] K. He, G. Gkioxari, P. Dollar,´ and R. Girshick, “Mask R-CNN”, Proc. to predict fatigue strength of steel from composition and processing IEEE Int. Conf. Comput. Vis. (ICCV), pp. 2961-2969, Oct. 2017. parameters”, Integrating Mater. Manuf. Innov., vol. 3, no. 1, pp. 90-108, [231] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional net- Dec. 2014. works for biomedical image segmentation,” in Medical Image Computing [208] P. Robustillo, J. Zapata, J. A. Encinar, and J. Rubio, “ANN charac- and Computer-Assisted Intervention, Munich, Germany:Springer, vol. terization of multi-layer reflectarray elements for contoured-beam space 9351, pp. 234-241, 2015. antennas in the Ku-band”, IEEE Trans. Antennas Propag., vol. 60, no. 7, pp. 3205-3214, Jul. 2012. [209] A. Freni, M. Mussetta and P. Pirinoli, “Neural network characterization of reflectarray antennas”, Int. J. Antennas Propag., vol. 2012, pp. 1-10, May 2012. [210] F. Gunes¸,¨ S. Nesil, and S. Demirel, “Design and analysis of Minkowski reflectarray antenna using 3-D CST Microwave Studio-based neural network model with particle swarm optimization”, Int. J. RF Microw. Comput. Eng., vol. 23, no. 2, pp. 272-284, Mar. 2013. [211] P. Robustillo, J. Zapata, J. A. Encinar, and M. Arrebola, “Design of a contoured-beam reflectarray for a Eutelsat European coverage using a stacked-patch element characterized by an artificial neural network”, IEEE Antennas Wireless Propag. Lett., vol. 11, pp. 977-980, 2012 [212] T. Shan, M. Li, S. Xu and F. Yang, “Synthesis of refiectarray based on deep learning technique”, Proc. Cross Strait Quad-Regional Radio Sci. Wireless Technol. Conf., pp. 1-2, Jul. 2018. [213] M. Salucci, L. Tenuti, G. Oliveri, and A. Massa, “Efficient prediction of the EM response of reflectarray antenna elements by an advanced