Industry standard UAV classification [5].
Open access peer-reviewed article
This Article is part of Robotics Section
Version of Record (VOR)
This version of record replaces the original advanced online publication published on 31/03/2025
Article metrics overview
374 Article Downloads
View Full Metrics
Article Type: Review Paper
Date of acceptance: March 2025
Date of publication: March 2025
DoI: 10.5772/acrt.20240052
copyright: ©2025 The Author(s), Licensee IntechOpen, License: CC BY 4.0
Table of contents
The integration of drones in various applications has seen rapid growth in recent years, with machine learning emerging as a critical enabler of this advancement. This study provides a comprehensive survey of the classification and applications of machine learning techniques in drone technology. By reviewing key machine learning paradigms—supervised, unsupervised, and reinforcement learning—and their relevance to drones, this survey highlights real-world applications, including object recognition, route planning, obstacle avoidance, search area optimization, and autonomous navigation. Furthermore, it examines the challenges associated with implementing machine learning in drones, such as data privacy concerns, data quality issues, and computational limitations. By exploring future directions and the transformative potential of machine learning in enhancing drone capabilities across industries, this survey aims to serve as a valuable resource for researchers, practitioners, and students exploring the intersection of machine learning and unmanned aerial vehicle technology.
artificial intelligence
drones
machine learning
supervised learning
unsupervised learning
Author information
Due to their ability to carry out tasks that were previously difficult or impossible for humans, drones, also known as unmanned aerial vehicles (UAVs), have witnessed a surge in popularity in recent years. Drones are being used in a wide range of applications, including delivery, agriculture, and infrastructure inspection, as well as military and surveillance operations. However, the classification and management of drones have also been complicated by the rapid advancement of drone technology. We present the ways in which machine learning (ML) algorithms can be used to improve drone technology in various applications and the classification of drones according to their size, shape, and functionality. In addition to shedding light on the advantages and drawbacks of employing machine learning in drone technology, recent cutting-edge methods and approaches as well as potential new developments are highlighted [1].
Size, shape, and functionality are few of the factors that are used to classify drones. Drones can be categorized according to their size, which can range from nano drones, which are smaller than a human hand, to large drones that are capable of carrying heavy payloads. Shape is another way to categorize drones, which can range from conventional quadcopters to novel designs like the X-wing and VTOL. Lastly, drones are also categorized according to their functions, such as delivery, agriculture, surveillance, and search and rescue [2].
New opportunities for enhancing drone technology have emerged as a result of the rapid development of ML algorithms. One application of ML is object recognition, in which drones with cameras can identify people or objects on the ground. Drones can now perform tasks like traffic monitoring and animal tracking in the wild thanks to ML algorithms that can be trained to recognize and classify specific objects in real time. Autonomous navigation, in which drones can fly without human intervention, is another application of ML. Autonomous navigation is especially useful for tasks like inspecting infrastructure or surveying large areas of land. Data from the drone’s sensors can be used to train ML algorithms that can decide where to fly, how to avoid obstacles, and when to return to the base. Additionally, drone delivery, which is becoming an increasingly popular application for drones, is being improved through the use of ML. Using ML algorithms, delivery drones can optimize their flight paths and delivery schedules, shortening delivery times and costs. During the delivery process, ML can also be used to identify and avoid obstacles, ensuring that packages are delivered safely and effectively [3].
There are several potential future developments that ML could make possible as drone technology continues to advance. The creation of swarming drones, in which multiple drones collaborate to complete a task, is one area of study. Swarming drones may be used for disaster responses, surveillance, and even temporary communication networks in faraway places. The creation of drones that are capable of operating in complex environments, such as indoors or in urban areas, is another area of research. Analyzing the environment and making decisions about navigation and avoid obstacles could be done with ML algorithms [4].
In this review, we examined how size, shape, and functionality can be used to classify drones and how ML algorithms can be used to improve drone technology in various applications as well as the most recent and cutting-edge methods, strategies and potential future developments. We anticipate that this field will continue to expand and innovate in the coming years because ML is an essential tool for improving drone technology. For those who are interested in the relationship between drones and machine learning, this survey is a comprehensive resource.
The integration of ML in drone technology has rapidly evolved, necessitating a comprehensive analysis of existing methods, applications, and challenges. This review aims to explore the classification of drones based on size, shape, and functionality while providing an in-depth assessment of ML-driven advancements in navigation, energy management, mapping, control, and imaging applications.
To ensure a thorough and structured review, we conducted an extensive literature search across major academic databases:
IEEE Xplore – For peer-reviewed research on AI applications in UAVs.
SpringerLink – For studies on drone classification and ML techniques.
ScienceDirect – For drone-related technological advancements.
MDPI – For open-access studies on UAV and AI interactions.
arXiv – For preprint articles discussing the latest ML-driven UAV solutions.
A combination of keywords was utilized to retrieve relevant literature, including:
“Machine learning in drones” AND “autonomous UAVs”
“Drone classification by size and function”
“Deep learning in UAV navigation and obstacle avoidance”
“Energy-efficient drone operation using AI”
“Object detection and imaging using ML in UAVs”
“Drone swarms and reinforcement learning”.
The initial search resulted in 320 papers, from which 112 studies were shortlisted based on the following inclusion criteria:
Papers published in the last 15 years (2010–2025).
Studies focusing on ML-driven UAV applications rather than traditional methods.
Studies discussing real-world implementations of ML in UAVs.
Exclusion of studies focusing solely on hardware design without AI-based enhancements.
The scope of this review spans several domains where ML has significantly enhanced UAV functionality:
Aerial Surveillance & Object Recognition: ML models enabling drones to detect and track objects.
Navigation & Obstacle Avoidance: Reinforcement learning and deep learning in UAV path planning.
Energy Management in UAVs: AI-based optimization techniques for extended flight times.
Drone Imaging & Environmental Monitoring: AI-driven image classification for agricultural and environmental analysis.
Autonomous Drone Swarms: Deep reinforcement learning for multi-agent UAV coordination.
Security & Intrusion Detection: The role of ML in detecting unauthorized UAV activities.
This review differs from previous studies by:
Providing a structured classification of UAV types and their specific AI applications.
Highlighting emerging ML trends that impact UAV technology, such as transformer-based architectures and federated learning.
Discussing practical limitations such as onboard computation constraints and regulatory challenges.
By analyzing both theoretical advancements and practical implementations, this review offers a valuable resource for researchers, engineers, and industry professionals seeking insights into the intersection of ML and drone technology.
Unmanned Aerial Vehicles (UAVs) are broadly classified based on their size, shape, propulsion system, and functionality. The standard classification used by military, regulatory, and research institutions groups UAVs into five categories based on their weight, altitude, and operational range, as shown in Table 1.
Group | Weight (lbs) | Max Altitude (ft) | Example UAVs |
---|---|---|---|
Group 1 | 0–20 | <1,200 | DJI Mavic, Parrot Anafi |
Group 2 | 21–55 | <3,500 | RQ-11 Raven, ScanEagle |
Group 3 | 56–1,320 | <18,000 | RQ-7 Shadow, Boeing Insitu |
Group 4 | >1,320 | Any Altitude | MQ-1 Predator, MQ-9 Reaper |
Group 5 | >1,320 | Any Altitude | Global Hawk, MQ-4 Triton |
Industry standard UAV classification [5].
The U.S. Department of Defense (DoD) and the Federal Aviation Administration (FAA) classify UAVs into five groups based on their operational capabilities [5].
These groups provide a structured way to classify UAVs based on their size, endurance, and operational range. While this classification is widely accepted in military and regulatory contexts, UAVs are also categorized based on their shape and functionality, which we discuss in the following subsections.
Drone aircraft with a fixed, non-rotating wing structure are referred to as fixed-wing drones, aerial drones, and UAVs. Fixed-wing drones generate lift through the aerodynamics of their wings, whereas, rotary-wing drones rely on rotating blades to generate lift. They are great for surveillance, mapping, and delivery because of their ability to fly over great distances quickly and remain airborne for a long time.
Single or multiple engines are used to propel fixed-wing drones, which typically have a longer wingspan than rotary-wing drones. Due to their reliance on aerodynamic forces and lack of vertical takeoff and landing capabilities, they are typically more challenging to operate and control than rotary-wing drones. However, the advantages of longer flight times, faster speeds, and more effective power use outweigh this drawback [6].
Fixed-wing drones are further classified into:
Manned Fixed-Wing UAVs: Piloted onboard.
Unmanned Fixed-Wing UAVs: Remotely controlled or autonomous.
A prototype image of a fixed-wing drone is shown in Figure 1.
A fixed wing drone [6].
Multirotor drones, commonly known as quadcopters, hexacopters, or octocopters, use multiple rotors for both lift and propulsion. Their ability to hover and maneuver precisely makes them highly suitable for aerial photography, search-and-rescue, and inspection tasks [7]. An example of a multirotor drone is shown in Figure 2.
A multirotor drone [8].
Multirotor drones are known for:
Hovering Stability: Ideal for aerial photography and videography.
Maneuverability: Can fly in any direction and are easy to control.
Variety of Configurations: Common configurations include quadcopters, hexacopters, and octocopters.
A flapping-wing drone, also known as an ornithopter, mimics the flight of birds by flapping its wings, as shown in Figure 3. Unlike traditional drones, which rely on rotors, these UAVs generate lift and propulsion through wing motion.
A flapping-wing drone [9].
Advantages of Flapping-Wing UAVs:
Energy Efficiency – Uses less energy compared to conventional drones.
Low Noise – Quieter operation, ideal for wildlife monitoring.
Enhanced Stability – Allows for intricate flight patterns.
However, they also have limitations, such as small payload capacity and the need for complex control systems [10].
Tilt-rotor drones combine the capabilities of helicopters and fixed-wing aircraft by tilting their rotors to transition between hover and forward flight, as shown in Figure 4.
A tilt-rotor drone [11].
Key Advantages:
Vertical take-off and landing (VTOL) Capability – Can take off and land like a helicopter.
Longer Range – When in fixed-wing mode, operates efficiently over long distances.
Versatility – Suitable for both urban and remote environments.
Tilt-rotor UAVs require advanced control systems to stabilize their dual flight modes, making them more complex than traditional drones [12].
The classification of UAVs is essential to understanding their diverse applications and technological advancements. While industry-standard classification (Groups 1–5) helps categorize UAVs based on their operational capabilities, classification by design (fixed-wing, multirotor, flapping-wing, tilt-rotor) further refines their functional purposes. With ongoing advancements in AI and autonomous flight systems, UAVs are expected to evolve, improving efficiency and versatility in both civilian and military applications.
A deep reinforcement learning (DRL) strategy for coordinated multi-agent UAV systems has been presented in [13]. To achieve energy awareness in UAV-based big-data platforms, which are used for data collection and analysis in large-scale environments, the authors proposed a system that employs DRL. The system was intended to reduce the amount of energy used by multiple UAVs for data collection and analysis. To achieve energy efficiency, the actions of multiple UAVs are coordinated using the DRL algorithm. The findings demonstrate that the proposed system is capable of maintaining the performance of data collection and analysis tasks while simultaneously reducing energy consumption. The limitations of using DRL in UAV-based big-data platforms, as well as the need for additional research and development in this area have also been discussed by the authors. Overall, the study demonstrates the importance of continuing progress in this field and the potential of DRL to enhance energy efficiency in big-data platforms based on UAVs.
A multi-UAV charging system that makes use of orchestrated scheduling and multi-agent deep reinforcement learning (MARL) was presented in [14]. To provide UAVs with charging that is both efficient and effective, the authors suggest a system that combines charging stations, UAVs, and a cloud platform. The charging of the UAVs is orchestrated by the system using a scheduling algorithm that takes into account the battery levels, flight plans, and charging availability of the UAVs. After that, the scheduling results were used by the MARL algorithm to improve efficiency and shorten the charging time by optimizing the charging procedure. The study demonstrated the potential of orchestrated scheduling and MARL for improving the efficiency and effectiveness of UAV charging systems by demonstrating that the proposed system outperformed traditional charging methods in terms of charging time and energy consumption. The authors also discussed the problems with these kinds of systems and the future steps, like how much research and testing in real-world situations would be required. Overall, the study demonstrated the potential of cloud-assisted systems and the significance of integrated approaches for enhancing the performance of UAV charging systems.
A study [15] described a DRL strategy for allocating frequency bands and designing 3D UAV trajectories to ensure fair and energy-efficient communication. By optimizing the UAV’s flight path and frequency band allocation, the authors hoped to boost the efficiency of UAV communication systems. Based on real-time communication requirements and energy consumption, the proposed strategy made use of a DRL algorithm to learn and optimize the UAV’s trajectory and frequency band allocation. The study demonstrated that DRL has the potential to improve the performance of UAV communication systems by outperforming conventional methods in terms of energy efficiency and communication fairness. The authors also discussed problems with these kinds of systems and the future needs, like how much research and testing in real-world situations would be needed. Overall, the study highlighted the potential of DRL in this context and the significance of optimizing the trajectory and frequency band allocation of the UAV for energy-efficient and fair communication.
A deep learning (DL) approach to predict drone energy consumption was discussed in [16]. The authors modelled the drones’ energy consumption based on flight time, altitude, speed, payload, and weather conditions using a deep neural network. The findings demonstrated that the proposed method applied to real-world scenarios and could accurately predict drones’ energy consumption.
An investigation into the use of reinforcement learning (RL) for the autonomous navigation UAVs has been presented in [17–20]. For UAV navigation in unknown environments, where a UAV must learn to fly from one location to another while avoiding obstacles, the authors proposed an RL-based method. The authors compared their proposed method to more conventional approaches to navigation and evaluated its efficacy in a simulation environment. According to the findings, the RL-based approach performed better than conventional methods in terms of speed, accuracy, and efficiency. The authors concluded that RL can be a promising method for autonomous UAV navigation and recommended that future research concentrate on adapting the strategy to more realistic environments.
A study on using DRL for autonomous drone navigation has been presented in [21–24]. For drone navigation in complex environments where the drone must use sensor data to avoid obstacles and reach its destination, the authors proposed a DRL-based strategy. The authors compared their proposed method to more conventional approaches for navigation and evaluated its efficacy in a simulation environment. The findings demonstrated that DRL-based approach can deal with complex environments with multiple obstacles effectively and outperforms conventional methods in terms of efficiency, speed, and accuracy. The authors concluded that DRL is a promising method for autonomous drone navigation and recommended that additional sensor data and real-world testing be the primary focus of subsequent research.
Presently, several studies [25–30] provide an overview of the state of DL for drone navigation research. The authors, Thomas Lee, Susan Mckeever, and Jane Courtney, examined various methods needed for drones to navigate on their own and determine the problems that still need to be solved. They looked at how DL techniques like RL, CNN, and recurrent neural networks (RNNs) can be used in drone navigation. The authors also discussed how obstacle avoidance, flight planning, and landing are affected by DL in drone navigation. They concluded by highlighting the potential for profound comprehension to revolutionize drone navigation and provided research avenues for the future.
A novel method for using DRL for drone navigation and obstacle avoidance was proposed in several studies [31–37]. Using a deep neural network to make decisions about what to do based on observations of the environment, the authors created a system that lets a drone navigate through complex environments. The neural network was then trained by the system using RL to avoid obstacles and reach a desired location. In a simulation, the authors evaluated their strategy and demonstrated that the drone can successfully navigate and avoid obstacles in difficult environments. The findings indicate that real-world drone navigation and obstacle avoidance systems could benefit from the use of DRL.
In [38], a novel strategy for navigating UAVs was proposed using Deep Reinforcement Learning (DRL) and massive MIMO technology. These two methods have been combined by the authors to create a system that can avoid obstacles while navigating UAVs through complex environments. A deep neural network was used in the system, which was trained through RL to decide what needed to be done based on what it sees around it. The vast MIMO technology was utilized to enhance the quality of communication between the control system and the UAV. In simulations, the authors evaluated how the proposed method performed and demonstrated that, in terms of stability and accuracy for UAV navigation, it performed better than conventional approaches. The findings point to the possibility of putting massive MIMO technology and DRL together in UAV navigation systems.
One study [39] suggested a DRL strategy for navigating high-dynamic environments with UAVs. A system that uses a deep neural network that has been trained through RL to make decisions about what to do based on what it sees around it was created by the authors. The system was built to deal with the problems of highly dynamic environments, such as sudden changes in the environment or obstacles in the way of the UAV. In simulations, the authors evaluated how well their method performed and demonstrated that, in terms of UAV navigation accuracy and stability in highly dynamic environments, it performed better than conventional approaches. The outcomes demonstrated the usefulness of DRL for navigating UAVs in difficult environments.
The application of DL to autonomous indoor drone racing was investigated in a study [23]. Using perception, guidance, and navigation systems based on DL algorithms, the authors proposed a system that can effectively guide a drone through an indoor racecourse. A DL-based approach to overcome various obstacles encountered in indoor drone racing was presented and the results showed that the system could successfully navigate the racecourse and maintain a stable flight trajectory. The system was trained using a combination of visual and sensory data from the drone. The study’s results show that DL has the potential to improve the capabilities of autonomous indoor drone racing.
A method for simultaneous navigation and radio mapping for UAVs connected to cellular networks of UAVs was presented in [40]. To address the issue of radio mapping and navigation in a connected to cellular networks, the authors suggested using DRL. The UAVs can fly to areas with strong signals while avoiding areas with weak signals thanks to the DRL algorithm’s training to make decisions based on the quality of the cellular signal. In terms of radio mapping efficiency and navigation accuracy, the study’s findings demonstrated that the proposed DRL strategy performed better than conventional methods. The results show that cellular-connected UAVs can use DRL for simultaneous radio mapping and navigation.
An RL strategy for autonomous navigation of UAVs was presented in [41]. To boost both the speed and accuracy of the navigation process, the authors suggested incorporating function approximation into the RL algorithm. Based on the UAV’s current state, the function approximation was used to estimate the expected reward for each action. The study showed that, in terms of speed and accuracy for navigation, the proposed RL approach with function approximation performed better than conventional RL methods. The results show that autonomous UAV navigation with RL and function approximation is feasible.
A DRL strategy for distributed energy-efficient multi-UAV navigation for long-term communication coverage was presented in [42]. The authors attempted to reduce the amount of energy each UAV used while simultaneously increasing the network’s communication coverage. The proposed DRL algorithm was taught to decide how to navigate each UAV, taking into account the network’s energy efficiency and communication coverage. In terms of both communication coverage and energy efficiency, the study demonstrated that the proposed DRL approach performed better than conventional methods. The results show that DRL can be used to provide long-term communication coverage for distributed energy-efficient multi-UAV navigation.
In [43], a method for mapping soil pollution was presented using drone image recognition and machine learning. The authors suggested taking pictures of an arsenic-contaminated agricultural field with a camera on a drone. ML algorithms were used to analyze the images and figure out which parts of the field were contaminated with arsenic. A dataset of images that have been labeled as either contaminated or uncontaminated was used to train the ML algorithm. The findings demonstrated that the proposed method was successful in accurately locating arsenic-contaminated fields. The proposed method has the potential to boost soil pollution mapping’s efficiency and accuracy, which are crucial for ensuring the safety of agricultural products and human health.
A study [44] compared three approaches for using UAV images to map vegetation on small island ecosystems. To evaluate the precision and efficacy of each method, the authors compared ML, visual interpretation, and pixel classification approaches. The outcomes demonstrated that the ML approach performed better than the other two as it achieved high accuracy with lower effort. The dataset of labeled images used to train the ML algorithm could classify vegetation types based on patterns in the images. The proposed method has the potential to make vegetation mapping on small island ecosystems more effective and accurate, which is important for ecological and conservation research.
A new approach to design low-Reynolds-number airfoils using DL-based tailored airfoil modes was proposed [45]. The authors suggested using a DL algorithm to learn a low dimensional representation of the shape of an airfoil. This representation could then be used to create new airfoil shapes that are optimized for particular characteristics of aerodynamic performance.The low-Reynolds-number airfoils that are frequently found in small UAVs and micro air vehicles can be designed using the proposed approach. The findings demonstrated that the proposed approach can design airfoils with improved aerodynamic performance when compared to other approaches, achieving an improvement of up to 40% in the lift-to-drag ratio for certain design specifications. Small UAVs and MAVs, which are increasingly used in environmental monitoring, search and rescue, and precision agriculture, could benefit from the proposed method’s increased efficiency and performance.
A method for using DL to create a multi-fidelity surrogate model for aerodynamic design optimization was presented in [46]. A more accurate and effective surrogate model for the design optimization process was created by combining low-fidelity models with high-fidelity computational fluid dynamics (CFD) simulations in the proposed approach. From the combined data set, DL algorithm was used to learn the mapping between the input parameters and the output response. A UAV wing’s aerodynamic design was optimized using the proposed approach. The findings demonstrated that the proposed approach was effective in reducing the number of high-fidelity CFD simulations required by 25% and increasing overall design performance by 10%, both of which contributed to an increase in the process’s efficiency and accuracy in aerodynamic design optimization. Beyond UAVs, the proposed method has the potential to improve the design process for automotive and aerospace engineering, among other engineering fields.
A strategy for designing the trajectories and access control of aerial drones to optimize communication performance was proposed in [47]. The authors suggested coordinating the drones’ trajectories and access control by employing a MARL algorithm. By optimizing the drones’ position and movement at the ground base station, the proposed method aimed to improve communication performance. A scenario in which multiple drones are utilized to provide communication coverage in a disaster area was the subject of the proposed approach. The findings demonstrated that the proposed approach was successful in increasing communication coverage and decreasing communication delay, with an improvement in communication performance of up to 60% when compared to previous approaches. Communication systems for disaster response and other crucial applications could benefit from the proposed approach’s increased effectiveness and dependability.
The application of ML methods to aerodynamic shape optimization was discussed in [48]. Aerodynamic design optimization, which can be challenging due to the complexity of the underlying physics and the large design spaces involved, was highlighted by the authors as requiring precise and effective methods. Neural networks, genetic algorithms, and other optimization techniques were reviewed for the application of ML to aerodynamic shape optimization. The authors highlighted the potential of ML to enhance the accuracy and efficacy of aerodynamic design optimization and discussed the advantages and disadvantages of these methods. Several research directions for the future were identified, such as the creation of more sophisticated ML algorithms and its integration with other methods of optimization. The authors argued that the advancement of ML methods for aerodynamic shape optimization has the potential to revolutionize the field and make it possible to design aircraft that are effective and better for the environment.
The use of neural networks in the design of UAVs for the fitting and classification of CFD data was proposed in a study. The authors pointed out that the traditional method of designing UAVs required a lengthy process of trial and error, which can be significantly sped up with the help of ML methods. Neural networks were used in the proposed method to predict the CFD results for a specific UAV design and classify it according to various performance metrics. The collection of CFD data, the creation of a neural network model, and the utilization of the model to optimize the design of a UAV were the components of the method presented by the authors for putting the proposed strategy into practice. The authors presented the outcomes of several experiments, demonstrating that the proposed strategy was effective in reducing the amount of time required for design optimization and improving the performance of unmanned aerial vehicles. The use of RL to improve the design of UAVs in dynamic environments and the incorporation of additional performance metrics into the classification process are two potential areas for future research.
In [49], the authors presented a novel strategy for maximizing communication coverage while minimizing energy consumption by optimizing the flight path of UAVs. To accomplish this, the authors suggested using DRL, which involved training a neural network in a simulation environment to learn the best policy for UAV control. The authors proposed a novel reward function that takes into account the fairness and quality of the UAV’s communication coverage. To prevent an excessive concentration of resources in particular locations, the reward function was intended to encourage the UAV to prioritize areas with poor coverage while avoiding areas with good coverage. The authors presented the outcomes of experiments that were carried out in a simulated setting. These results showed that the proposed method was capable of increasing communication coverage while also ensuring that coverage was distributed fairly. The authors also determined how well the proposed method worked in different situations, such as when there are different demands on communication and the environment is different. The proposed method has applications in different areas, like disaster response, public safety, and military operations.
A system for social distancing monitoring in public places that makes use of a DL algorithm was proposed in [50]. In a public setting, the system detected individuals and their distances using a camera mounted on a drone. Two parts made up the proposed DL algorithm: estimation of distance and object detection. The You Only Look Once (YOLO) algorithm was used for object detection portion, and SD (Single Shot MultiBox Detector) algorithm was used for distance estimation. Multiple people could be tracked by the proposed system, which can be used in real time to make sure social distance rules are followed. The system can be used to effectively maintain social distance in public places and in identifying individuals and their distances.
A study [51] involved analyzing drone data using ML methods. The three stages of the proposed framework were: data processing, data analysis, and data acquisition. The collection of evidence from drones using a variety of tools and methods, including digital forensics tools and a physical examination of the drone, was a part of the data acquisition phase. Preparing and extracting data from the acquired evidence was a part of the data processing stage. ML methods were applied to the extracted data during the data analysis phase to find patterns, anomalies, and potential evidence related to the incident.
A case study is also included in the study to elucidate how well the proposed framework works. The crash of a drone that caused property damage is the subject of the case study. The collected evidence was analyzed using ML methods and the proposed framework was applied to it. The findings of the analysis provided valuable evidence for the investigation and revealed the incident’s likely cause.
In general, the proposed forensic framework and ML methods provide an efficient and effective method for investigating drone-related incidents.
For controlling UAVs [52] in a variety of tasks, including navigation, tracking, and obstacle avoidance, the various RL and deep learning algorithms were evaluated by comparing the algorithms based on accuracy, speed, and generalizability using a simulated environment and various metrics. Q-learning, SARSA, and ActorCritic are examples of RL algorithms tested, while CNNs and DBNs are examples of deep learning models.
While the speed of the RL algorithms was superior, the accuracy and generalization of the DL models were generally superior to those of the RL algorithms. However, each algorithm performed differently depending on the task and environment at hand. The authors also pointed out the possibility of enhancing the overall performance of UAV control by combining RL and DL.
Overall, the study demonstrated the control potential of DL and RL algorithms for UAVs and emphasized the need for additional research to tailor these methods to specific tasks and environments.
A method for controlling a drone in virtual reality using ML and MEMS sensor technology was proposed in [53]. To predict the drone’s motion, the authors used an IMU (Inertial Measurement Unit) to capture the drone’s motion in three-dimensional space. The authors discovered that the ML algorithm could accurately predict the motion of the drone, which could be utilized for virtual reality drone control. They posited how this technology could be used for training and simulation, as well as for making autonomous drones. Overall, the study showed that it is possible to control drones in virtual reality with MEMS sensor technology and ML, which could have an impact on the development of more advanced drone technologies.
Another study [54] utilized a deep DL-based approach to use RF signals in order to find and identify drones. To train and evaluate their DL model, the authors presented a dataset of RF signals gathered from various drone models. Two main steps in their proposed method were: classification and detection. A deep CNN was used in the detection step to find a drone signal in the RF data. An RNN was used in the classification step to divide the detected signals into various drone models. The collected dataset served as the basis for the authors’ evaluation of the proposed method, which yielded encouraging results with high detection and classification accuracies. Their method could be used to identify and detect drones in a variety of settings, like safety and security.
Using ML models, the authors [55] proposed a framework for the autonomous detection of malicious events in drone networks. Analyzing the network traffic data and identifying abnormal traffic patterns that could indicate an attack or malicious event, the authors made use of the power of AI, particularly DL techniques. The experimental results showed that the framework was effective at accurately detecting and classifying a variety of malicious events, including reconnaissance attacks, denial-of-service attacks, and malware infections, on a custom drone network dataset. By making it possible to detect malicious activities in a timely and accurate manner, the proposed framework can help to improve the security of drone networks, preventing potential damage and maintaining the network’s integrity.
A cooperative multi-agent DRL framework was proposed in [56] to achieve effective and reliable surveillance through autonomous multi-UAV control. By allowing each agent to learn from and work with other agents to improve the global objective function, the proposed framework addressed the difficulties of autonomous multi-UAV systems. The four main components of the proposed framework were: (1) observation extraction module; (2) communication module; (3) action selection module; and (4) learning and cooperation module. The experimental results of the proposed method tested in a simulated environment demonstrated that, while maintaining a robust and dependable surveillance system, it was capable of successfully detecting and tracking moving targets with a high success rate. The proposed method looks promising and could be used in real-world situations like disaster management, search and rescue, and surveillance.
The authors of the study [57] proposed a circulated calculation for improving a PID regulator for a heterogeneous group of UAVs utilizing AI. The calculation intended to limit the deviation of the genuine direction of the UAVs from the ideal direction. The proposed approach depended on the utilization of a versatile pundit plan strategy, which joins a pundit brain network that gauges the expense capability with an activity network that produces the control activities. The pundit network is based on internet utilizing the information produced by the UAVs during the flight, while the activity network is refreshed utilizing the approach inclination strategy. The creators showed that the proposed calculation outflanks conventional strategies and accomplished better intermingling.
For obstacle avoidance in UAVs with limited environmental knowledge, a study proposed a memory-based DRL strategy. The proposed approach [58] stored the current state of the UAV and its environment in memory by encoding it using an RNN. The decisions regarding the UAV’s trajectory to avoid obstacles were then made using the stored information.
A DRL algorithm known as proximal policy optimization (PPO), which optimizes the UAV’s policy for obstacle avoidance, was used to train the RNN. In a simulation, the proposed strategy was tested against other obstacle avoidance strategies that are based on rules and DRLs.
According to the findings, the success rate and path efficiency of the proposed memory-based DRL method were superior to those of the other approaches. The authors concluded that the memory-based strategy is a promising approach for UAVs with limited environmental knowledge to avoid obstacles.
One study [27] described a DRL for using UAVs to locate ground objects. An autonomous navigation and object detection system for a UAV was developed by the authors. A deep neural network trained through RL was used by the system to decide what to do based on what it sees around it. In a simulation, the authors evaluated their strategy and demonstrated that the UAV could precisely locate ground objects while avoiding obstacles in its path. In real-world applications, the findings demonstrated the potential of RL for autonomous UAV navigation and object localization.
A DL strategy for human target search and detection using UAVs was proposed in [28]. A system designed by the authors made it possible for UAVs to search for and identify human targets in challenging environments. A large dataset of human targets was used to train the system’s deep neural network, which was used to make decisions about what to do based on environmental observations. In simulations, the authors evaluated how well their method worked and demonstrated that the UAV can accurately identify human targets while avoiding obstacles in its path. In real-world UAV applications, the outcomes demonstrated the potential of DL for human target search and detection.
A DRL strategy for object tracking in drone images was proposed in [29]. A method devised by the authors enabled a drone to simultaneously take pictures and track objects in real time. Based on observations of the environment, including the captured images, the system made decisions about what to do with a deep neural network that has been trained through RL. In simulations, the authors evaluated their strategy and demonstrated that the system can accurately track objects in real time while avoiding obstacles in the path of the drone. The outcomes demonstrate the practical application potential of DRL for object tracking in drone images.
A study on the potential of DRL for vision-based autonomous drone navigation with a mission in real-world scenarios was presented in [30]. For drone navigation, the authors proposed a method based on DRL that decides what to do based on visual observations of the environment. The authors posited the difficulties and considerations when putting their method into practice in the real world and evaluated how well it worked in simulations. The findings emphasize the need for additional research and development to address the difficulties of putting DRL to use in the real world for mission-related vision-based autonomous drone navigation.
A DL strategy for human detection in real-time UAV applications was proposed in [31]. The authors detected humans in UAV images by employing YOLO-v2 object detection algorithm. The system’s speed and effectiveness make it ideal for use in real-time UAV applications. On a dataset of UAV images, the authors evaluated their method and demonstrated that the system could accurately identify human targets while maintaining a high processing speed. The outcomes demonstrate that real-time human detection in UAV applications using YOLO-v2 and deep learning is feasible.
The use of drones for search and rescue in natural disasters was suggested in [32]. The authors discussed the challenges and benefits of using drones in these kinds of situations, like being able to quickly survey large areas and get into hard-to-reach places. In addition, they suggested enhancing the capabilities of drones in search and rescue operations by utilizing cutting-edge technologies like computer vision and DL. The authors highlighted the significance of collaboration between stakeholders, such as government agencies, non-profit organizations, and technology companies, to effectively utilize drones for search and rescue operations and provided examples of successful drone surveillance during natural disasters. The findings of this study emphasize the significance of further research and development in this field as well as the potential of drones as a useful tool for responding to natural disasters.
A study on how to use DRL for the autonomous landing of UAVs was presented in [33]. An end-to-end DRL strategy for landing UAVs that can handle the dynamic and complex nature of landing tasks was proposed by the authors. In a simulation, they evaluated their strategy and demonstrated that the UAV was capable of landing accurately and successfully while avoiding obstacles in its path. The findings emphasize the significance of continuing research in this field and demonstrate the potential of DRL for autonomous UAV landing. The authors discussed the difficulties associated with putting the ideas into practice in the real world, the aspects to consider, and the need for more work to fix these problems. In general, the study presented a promising DRL-based strategy for autonomous UAV landing.
The use of DRL for autonomous UAV landing on a moving platform was suggested in [34]. A DRL algorithm that can successfully land the UAV and adapt to the shifting conditions of a moving platform was proposed in this study. To guarantee a secure and precise landing, the algorithm took into account wind, platform velocity, and platform orientation. The findings demonstrated that, in terms of stability and accuracy, the proposed algorithm performed better than conventional landing strategies. The authors also emphasized the need for additional research in this area and discussed the difficulties and limitations of using DRL for autonomous UAV landing on moving platforms. Overall, the study highlighted the significance of ongoing research and advancement in this field and demonstrated the potential of DRL for autonomous UAV landing on moving platforms.
The use of RL for the autonomous landing of a multi-copter UAV was the primary focus of [35]. To deal with the difficulties of landing a multicopter UAV, the authors proposed a vision-based RL strategy that took into account factors such as landing site detection, stability, and orientation. The algorithm improved its accuracy over time through trial-and-error learning. The findings demonstrated that the proposed strategy can carry out autonomous landings with high precision even in difficult circumstances like strong wind or uneven landing surfaces. The limitations of using RL for autonomous UAV landing were discussed, as well as the need for additional research and development in this area. Overall, the study demonstrated the importance of continuing progress in this field and the potential of RL for the vision-based autonomous landing of multi-copter UAVs.
A DRL strategy for the autonomous tracking and landing of UAVs was presented in [36]. To enable autonomous UAV tracking and landing, the authors proposed a RL algorithm that combined image recognition, control, and decision making. The UAV’s camera’s image data were analyzed by the algorithm, which used deep neural networks to make decisions based on that information. The findings demonstrated that the proposed algorithm was capable of autonomous tracking and landing with high stability and accuracy. The authors also offered suggestions for future research and discusses the difficulties and limitations of using DRL for autonomous UAV tracking and landing. Overall, the study demonstrated the importance of continuing progress in this field and the potential of DRL for autonomous UAV tracking and landing.
A DL-based strategy for landing drones was presented in [37]. To enhance the precision and stability of drone landings, the authors proposed a system that combined marker detection with super-resolution reconstruction. To enable more precise marker detection, the system used deep neural networks to perform super-resolution reconstruction on the images captured by the drone’s camera. The drone’s landing was then guided by the marker detection data. The findings demonstrated that the proposed system was capable of landing drones with high precision and stability even in challenging conditions. The limitations of using DL for drone landing were discussed, as well as the need for additional research and development in this area. Overall, the study emphasized the significance of continuing progress in this field and demonstrated the potential of DL to enhance the stability and accuracy of drone landings.
The authors in [38] proposed a novel coordinated multi-agent deep reinforcement learning (MADRL) algorithm for efficient energy sharing among multiple UAVs engaged in distributed big-data processing. Their approach enables UAVs to be autonomously charged via wireless power transfer from infrastructure-based charging towers, while minimizing energy purchases from external utility providers. The coordinated MADRL framework intelligently manages energy distribution among towers, aiming to reduce system-wide operational costs. Simulation results demonstrate that the proposed algorithm significantly enhances energy efficiency and system performance.
The objective of a study [39] was to use DL methods to classify the maturity of strawberries in images taken by near-ground imaging and UAVs. The goal of the authors was to create a classification model that can tell the difference between mature and immature strawberries in the images. This could be useful in agriculture for predicting yields and effectively harvesting fruits. The results demonstrated that the DL model can accurately classify strawberries’ maturity when trained on a dataset of UAV and near-ground images.
A method for estimating the damage caused by a disaster using aerial images captured by UAVs and DL algorithms was discussed in [40]. For disaster management, the purpose of the study was to provide an effective and precise method for assessing disaster damage. The authors suggested using DL algorithms to analyze UAV images and determine how much damage the disaster caused. The findings demonstrated that the proposed approach can provide useful information for disaster management systems and is both effective and efficient in estimating disaster damage.
The use of UAVs for remote sensing to detect citrus canker disease was investigated in [41]. The authors analyzed the data collected and used ML algorithms and hyperspectral imaging to find the diseased plants. The findings demonstrated that the proposed method was effective at detecting the disease, which can aid in early treatment and prevent further spread.
A method for detecting vine diseases in UAV multispectral images using optimized image registration and DL segmentation was proposed in [42]. A DL segmentation technique was used to identify the diseases, and an optimized image registration method was used to precisely align the multispectral images taken by the UAV. When compared to more conventional approaches, the outcomes of the proposed strategy are found to be promising for efficiently and accurately identifying vine diseases.
The creation of a real-time sound detection and analysis system for UAVs was the primary focus of [48]. This system aimed to use UAVs to detect and analyze sounds in real time, which can be useful for a wide range of applications, including search and rescue, environmental monitoring, and surveillance. The system’s design and implementation, as well as its hardware and software components, are described in detail, and the system’s performance was evaluated through experiments and analysis.
The development of a system for detecting and identifying drones based on their audio was the primary focus of [59]. The system processed the drones’ audio signals and used DL algorithms to identify them based on the distinctive sound patterns produced. The research aimed to develop an efficient and effective method for real-time drone detection and identification.
A novel method for detecting drones based on their sound was presented in [60]. To improve drone detection accuracy, the authors proposed a system that combines result- level fusion of multiple audio features with DL. To identify drone sounds, the DL model was trained on audio data from various drone types and ambient noise. The outputs of multiple audio feature extraction models were then combined with the result-level fusion to further enhance the drone detection system’s performance. The proposed system was tested and evaluated in a real-world setting, and the findings demonstrated that it was capable of accurately identifying drones in real time with low false alarm rates. The authors concluded that security and surveillance applications requiring real-time drone detection may benefit from the proposed method.
A drone detection system that uses multiple acoustic nodes and ML models to detect UAVs was presented in [61]. Multiple microphones were used by the system to collect ambient acoustic signals, which were then analyzed by ML algorithms to see if they came from a UAV. Additionally, a decision-level fusion mechanism was incorporated into the system to combine the results of multiple acoustic nodes to enhance overall detection accuracy. Using real-world data, the authors evaluated the system’s performance and demonstrated that, in terms of detection accuracy and response time, it performed better than conventional drone detection methods. The findings show that the proposed system could be used in a variety of security and surveillance scenarios where UAV detection is crucial.
A framework for using RF signals to identify UAVs was presented in [62]. To accurately detect and identify UAVs in real-time, the authors suggested applying a ML method to the RF signals gathered by multiple sensors. To classify the RF signals into UAV and non-UAV categories, the system employed a combination of feature extraction, dimension reduction, and ML algorithms. Using real-world RF data, the authors evaluated the effectiveness of their framework and demonstrated that it detects and identifies UAVs with high accuracy.
This study’s findings can be used in a variety of real-world situations, like security and surveillance, where accurate UAV detection is crucial.
A study [63] focused on using multiple-input multiple-output orthogonal frequency- division multiplexing (MIMO-OFDM) radar system for detecting drones. The proposed system was based on the MIMO-OFDM radar technology and aimed to provide an effective solution for drone detection. The design and implementation of the MIMO-OFDM radar system were presented and the authors analyzed its performance for drone detection. The results demonstrated the feasibility of using MIMO-OFDM radar technology for drone detection and the effectiveness of the proposed system.
Another study [64] presented a drone detection and tracking system based on phase-interferometric Doppler radar. The system utilized Doppler shift to detect the drone’s motion and determine its location. The authors described the design of the system and evaluated its performance through simulations and experiments. Results showed that the system was capable of accurately detecting and tracking drones in real time, making it a promising solution for drone detection and tracking in various applications.
Another study [65] investigated the feasibility of using passive radars to detect drones. Passive radar systems rely on reflected signals from existing transmitters instead of emitting their signals, making them stealthy and cost-effective compared to traditional active radars. The study explored the potential of passive radar systems to detect and track drones and analyzed the performance of different passive radar configurations. Results indicated that passive radar systems have the potential to detect and track drones in certain scenarios. However, the authors also highlighted the limitations of passive radar systems and suggested ways to improve their performance for drone detection.
A study [66] explored the use of CNNs to detect drones based on radio-frequency (RF) signals. The authors collected a dataset of drone and non-drone signals and used it to train and test the CNN. They evaluated the performance of the CNN-based drone detection system and compared it with a traditional machine-learning approach. Results show that the CNN-based system outperformed the traditional approach in terms of accuracy and computation time. The authors concluded that CNNs are a promising approach for drone detection based on RF signals.
This study [67] investigated the use of radar cross-section (RCS) signatures for the detection and classification of drones. The authors proposed a system that uses an RCS measurement to capture the reflection of a drone’s electromagnetic signal, which can be used to distinguish the drone from other objects in the environment. The system utilizes ML algorithms, specifically decision trees and support vector machines, to classify the RCS signatures into different drone categories based on their shape and size. The authors evaluated the performance of their proposed system through simulations and experiments, and showed that it can accurately detect and classify different types of drones. The results indicate that RCS signatures are a promising technique for drone detection and classification.
A DRL-based control strategy for UAV swarms was proposed in [68]. A control policy for coordinating the movements of the UAVs so that they exhibit flocking behavior, including maintaining formation and avoiding collisions, was learned by the authors using a deep neural network. Simulation was used to train the DRL algorithm, and the results showed that the proposed method worked well for flocking control of UAV swarms.
According to the authors, a deep neural network that maps the state of the UAVs to control actions could be trained using a digital twin, a simulated representation of the actual UAV system. The policy was optimized using the RL algorithm so the UAVs behave flockingly, staying in formation and avoiding collisions. The findings demonstrate that the suggested method was successful in achieving flocking control over multiple UAV systems [69].
An attention-based DRL strategy for collision-free flocking control of a scalable fixed-wing UAV swarm has been presented in [70]. A population-invariant DRL algorithm is proposed by the authors to deal with changes in the number of UAVs in the swarm while still maintaining efficient flocking behavior. An attention mechanism was used to train RL, allowing the UAVs to focus on relevant information in the swarm state. The outcomes demonstrate that the proposed method was successful in controlling a scalable UAV swarm without having to worry about collisions.
Another study [71] described a method for large-scale multi-UAV flocking and navigation that used DRL. The authors recommended using an oracle to steer the RL process and boost training efficiency. The oracle informs the RL algorithm of the best course of action, reducing the number of trial-and-error iterations and accelerating learning convergence. A policy that enables the UAVs to exhibit flocking behavior while avoiding collisions and successfully navigating to a target was optimized using the DRL algorithm. The findings demonstrate that the proposed oracle-guided strategy was successful in achieving multi- UAV flocking and navigation on a large scale.
A modified adaptive formation strategy for UAV swarms based on pigeon flock behavior in the local visual field is presented in [72]. Each UAV keeps a specific distance and orientation from its neighbors while avoiding collisions, according to the authors’ control algorithm, which is inspired by pigeon flocking. The control algorithm takes into account each UAV’s local visual field and is based on a modified version of the Reynolds flocking rules. The findings demonstrate that the proposed strategy was successful in achieving a stable adaptive formation for swarms of UAVs.
The integration of ML in the drone industry has brought numerous advantages, such as increased automation, enhanced data processing, and improved situational awareness. However, implementing ML in drones is not without its challenges. Addressing these challenges is critical for fully harnessing the potential of ML in this field. Key challenges include:
Data Quality: ML algorithms require substantial amounts of high-quality data for effective training. In drone applications, data collection can be hindered by sensor limitations, hardware malfunctions, and adverse environmental conditions, which may degrade the quality and reliability of the collected data.
Limited Onboard Computing Power: Drones often have restricted computing resources, making it difficult to execute computationally intensive ML algorithms. This limitation necessitates optimized models and, in some cases, reliance on cloud computing for processing.
Battery Life and Flight Time: The limited battery capacity of drones directly impacts their flight time. ML algorithms, particularly those requiring real-time data processing, can further strain power resources, reducing the operational duration of drones in the field.
Lack of Standardization: The drone industry lacks standardized data formats and protocols, complicating data sharing and integration across different systems and manufacturers. This lack of standardization creates barriers to collaboration and interoperability.
Regulatory Constraints: Regulations governing drone usage vary significantly across regions, limiting certain applications, such as beyond visual line of sight (BVLOS) operations. These restrictions can hinder the deployment of ML algorithms that rely on extended operational capabilities.
Safety and Reliability: Drones equipped with ML algorithms must operate reliably to ensure safety. Inaccuracies in obstacle detection or navigation algorithms can pose significant risks, emphasizing the need for robust and fail-safe ML systems.
To address these challenges, collaboration among drone manufacturers, software developers, and regulators is essential. Solutions include advancements in lightweight ML models, improved hardware, enhanced data acquisition techniques, and the establishment of global standards.
The adoption of ML in the drone industry has revolutionized various aspects of drone operations, enhancing their capabilities, safety, and efficiency. The most commonly used ML methods include:
Supervised Learning: This technique involves training algorithms on labeled datasets to identify patterns and make predictions. Applications in the drone industry include object detection, image and video classification, and obstacle avoidance.
Unsupervised Learning: Unsupervised learning enables algorithms to find patterns and structures in unlabeled data. It is particularly useful for anomaly detection and clustering tasks, such as identifying unusual behaviors in drone flight data.
Reinforcement Learning (RL): RL involves training drones to take actions in an environment to maximize a reward. Applications include autonomous navigation, where drones learn to navigate complex environments and avoid obstacles.
Deep Learning: Leveraging neural networks to model complex relationships in data, deep learning powers applications like advanced object recognition, autonomous flight, and high-accuracy video analysis.
Transfer Learning: This method involves adapting pre-trained models to new tasks or environments, significantly reducing the data and time required for training. Transfer learning is particularly effective in scenarios with limited labeled data.
Hybrid Methods: Combining multiple ML approaches often leads to superior results. For example, DRL integrates RL and neural networks to enable drones to navigate and avoid obstacles efficiently in dynamic environments.
Rapid adoption of these methods has made drones safer, more efficient, and more capable. Continued research and development in these areas will further enhance ML’s potential in the drone industry.
In recent years, the combination of drone technology with machine learning has become an important focus in research and development. This review provides an in-depth understanding of the current use of machine learning in drones and its potential applications. This study covers several machine learning methods, including reinforcement, supervised, and unsupervised learning, highlighting their strengths and weaknesses. For example, supervised learning methods like deep neural networks have significantly improved tasks such as object detection and tracking. In contrast, unsupervised learning has proven useful for identifying unusual patterns and clustering data. Reinforcement learning has enabled drones to perform tasks like autonomous navigation and mission planning more effectively. Machine learning plays a critical role in tasks like object detection, tracking, mapping, and classification. These capabilities have been applied across agriculture, wildlife management, and in search and rescue operations. Machine learning improves terrain and vegetation mapping, helping drones collect and analyze high quality data more efficiently. One of the biggest advantages of using machine learning in drones is its ability to handle complex tasks quickly and accurately. By processing large amount of data in real time, drones can make smarter decisions and adapt to environmental changes. This is particularly valuable for autonomous navigation, where drones need to respond instantly to new challenges. However, the use of machine learning in drones comes with challenges. High-quality training data is essential for accurate algorithms, but collecting such data is an arduous task. Drones also have limited processing power and memory, making it hard to run computationally demanding algorithms. Additionally, many machine learning models act like “black boxes,” making it unclear how they reach certain decisions, which can be a limitation for critical applications. Despite these obstacles, this study concludes that machine learning has great potential to improve drone capabilities, enabling them to perform a wide range of tasks with higher precision and efficiency. Industries like agriculture, forestry, and environmental monitoring benefit significantly from these advancements. In summary, this study is a valuable resource for researchers and practitioners. It highlights the challenges that need to be addressed and emphasizes the importance of continued research and development to unlock the full potential of machine learning in drones. With further innovations, machine learning will continue to enhance the abilities of drones and their applications across various industries.
Haque, Ahshanul: Conceptualization, Investigation, Formal analysis, Writing – original draft, Visualization; Chowdhury, Md Naseef Ur Rahman: Investigation, Formal analysis, Writing – review & editing; Hassanalian, Mostafa: Writing – review & editing, Supervision.
This research did not receive external funding from any agencies.
Not Applicable.
Source data is not available for this article.
The authors declare no conflict of interest.
Written by
Article Type: Review Paper
Date of acceptance: March 2025
Date of publication: March 2025
DOI: 10.5772/acrt.20240052
Copyright: The Author(s), Licensee IntechOpen, License: CC BY 4.0
© The Author(s) 2025. Licensee IntechOpen. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.
Impact of this article
374
Downloads
577
Views
1
Altmetric Score
Join us today!
Submit your Article