A route planning system based on deep reinforcement learning could, in the future, guide drones on the correct path when searching for missing persons in remote areas.
In the last decade, the use of drones in search and rescue operations across vast and complex terrains has increased, driven by advancements in technologies that enable Unmanned Aerial Systems (UAS). These include flight control systems, positioning and coordination mechanisms, onboard devices (such as cameras and remote sensing tools), and data analysis capabilities [source: “Key technologies for safe and autonomous drones” – Microprocessors and Microsystems, November 2021].
In 2017, at the twelfth edition of the International Scientific Conference of Young Scientists on Sustainable, Modern, and Safe Transport in Slovakia, a comprehensive study on the use of drones to assist public services in natural hazard interventions was presented for the first time. This study particularly focused on operations involving the «rescue of people and animals during emergencies triggered by fires, floods, and chemical, biological, radiological, and nuclear hazards» [source: “The Use of UAV’s for Search and Rescue Operations” – Procedia Engineering, 2017].
In February 2019, the first test using drones for a UNICEF search and rescue mission was conducted in a remote steppe area of Kazakhstan. This location was significant as, seven months prior, a child went missing there and was never found despite nearly two weeks of helicopter searches.
Another significant application of rescue drones is their use in mountain accidents in hard-to-reach areas, providing timely medical assistance for traumatic injuries, often delayed due to the challenging locations[source: “Drones reduce the treatment-free interval in search and rescue operations with telemedical support – A randomized controlled trial” – The American Journal of Emergency Medicine, April 2023].
Regarding the rapid transport of medical supplies via Unmanned Aerial Systems, Belgium conducted its first test in July 2023. A two-seater drone delivered blood bags for transfusions from Limbourg Airport over a one-kilometre dedicated route, paving the way for a potential new type of rapid, cost-effective, and environmentally friendly air ambulance service within the EU.
TAKEAWAYS
An overview of regulations on professional drone use
A group of scientists from the Aerospace Sciences Research Division at the University of Glasgow, authors of the study “Deep Reinforcement Learning for Time-Critical Wilderness Search And Rescue Using Drones” (arXiv, 22 May), acknowledge drones’ specific role in search and rescue operations. However, they note that this application is still relatively new and not fully mature, considering the necessary adherence to existing regulations for proper RPA use. Globally, there are numerous studies, tests, and experiments on the subject, but practical applications are less common.
Regarding regulations, the team cites the United States as an example, where professional drone use requires pilots to maintain «a constant visual line of sight between themselves and the aircraft». This presents challenges for drone operations in mountainous regions.
Such regulations are driven by safety concerns, aiming to prevent risks to people on the ground. However, in extensive and rugged terrains, these rules can limit drones’ search and rescue capabilities.
Specifically, within the European Union, Regulation (EU) 2019/947 outlines three categories for professional drone use: the “open” category, for Unmanned Aerial Systems up to 25 kg flying within visual line of sight below 120 metres; the “specific” category, for drones over 25 kg flying beyond visual line of sight and above 120 metres, requiring aviation authority permission; and the “certified” category, for drones performing high-safety standard flights, such as passenger transport.
The use of rescue drones in the wilderness: advantages and challenges
Researchers at the Scottish university highlight drones’ superior lightness, speed, and flexibility in Wilderness Search and Rescue (WiSAR) operations compared to traditional transport like helicopters.
«Deploying helicopters can be slow, especially on islands, where it may take hours for them to arrive. Additionally, helicopter operations are very costly. In contrast, drones offer an agile and cost-effective solution for aerial searches. Although they cannot currently replace helicopters due to a significant disparity in lifting capacity, drones can initiate searches quickly and work in synergy with helicopters».
In future scenarios, fleets of RPAs could complement helicopter fleets in searching for and rescuing people in remote and rugged areas. An illustrative case is Scotland, known for its often inaccessible terrain, where the study was conducted. Police Scotland Air Support Unit (PSASU) and Scottish Mountain Rescue (SMR) are working to position small drone fleets across the country «for rapid deployment in potential WiSAR scenarios».
A significant weakness of RPAs in WiSAR is the planning of aerial search routes, particularly due to remote piloting challenges. This necessitates the development of specific methods and technologies.
The current approach by Police Scotland Air Support Unit involves a «“pilot-observer model,” where two operators per drone are required: the observer maintains a constant visual line of sight with the drone, while the pilot flies it, focusing on the real-time video feed from the drone’s camera». Both operators are constantly moving.
It is important to note, as highlighted in “Sweep Width Estimation for Ground Search and Rescue”, that «a stationary operator has higher detection capabilities compared to one in motion, similar to helicopter pilots who stop and scan an area upon spotting something of interest».
This difference between cognitive processes in motion and at rest highlights the limitations of the pilot-observer model for drones in WiSAR scenarios. What solutions can be adopted then?
Air search pathways guided by deep reinforcement learning
The planning of search and rescue missions using drones in complex and rugged terrains involves collecting dataprimarily related to the area to be explored, and also considering the age, physical condition, and any potential health issues of the missing person. All this information, as the research team notes, contributes to generating a “Probability Distribution Map” (PDM) that indicates the likelihood of finding the missing person in a specific location. This map serves as a guiding compass throughout the various phases of the operations.
According to the authors, combining this map with a machine learning technique known as “deep reinforcement learning” enables the definition of autonomous and precise flight paths through a computational process, supporting ground operators and increasing the chances that a drone can more swiftly locate those missing in inaccessible territories.
The choice of deep reinforcement learning was driven by the need for an artificial intelligence technique capable of «learning to make a series of effective decisions aligned with the given objective within a very short timeframe».Compared to other AI techniques, «it was found that deep reinforcement learning outperforms other algorithms by over 160% in terms of precision and decision speed. This difference can mean the difference between life and death in real-world search and rescue operations».
Rescue drones: AI-optimised probability maps indicate priority pathways
The deep reinforcement learning model developed by the team of scientists has been trained using a vast amount of data from search and rescue missions conducted in wilderness areas worldwide over recent years.
This data includes the age of the missing persons, the activities they were engaged in when they lost their bearings (such as hunting, sports, horse riding, or hiking), and their health status. The training dataset also includes geographical video data from Scotland, particularly relating to its forests, steep mountains, slopes, and rock faces, which are considered challenging and often dangerous terrains, especially for solo adventurers.
Once trained, the AI model’s goal is to run millions of simulations aimed at identifying the paths that a missing person in a specific remote area is most likely to follow under certain circumstances. «The result is a probability distribution map that indicates priority search areas for the drones», emphasises the team. They tested their DRL algorithm against two commonly used search models for planning drone flight paths in difficult locations: one where the UAV flies over a target area in a series of “search strips” and another AI algorithm not designed to interact with probability distribution maps.
In virtual tests, the deep reinforcement learning model outperformed the two traditional approaches, particularly in two aspects: «the distance a drone needs to travel to locate the missing person and the probability of the drone finding the individual».
Specifically, the first comparison model managed to find a missing person in 8% of cases, the second in 12%, compared to 19% for the deep reinforcement learning-based approach.
Glimpses of Futures
If the methodology defined by the Scottish researchers is validated through field experiments in the future and proves effective in real-life rescue situations alongside helicopter fleets, it could one day speed up search and rescue operations for missing persons in critical areas, thereby saving more lives «in scenarios where every minute counts».
Aiming to anticipate potential future scenarios, we use the STEPS matrix to provide a vision of the impacts this methodology might have from social, technological, economic, political, and sustainability perspectives.
SOCIAL: scientists from the Aerospace Sciences Research Division at the University of Glasgow affirm that in the medium term, Remotely Piloted Aircraft (RPA), even with advanced artificial intelligence techniques, will not replace helicopters in the search for missing persons in hard-to-reach areas. They currently lack the lifting capacity and, being ground-controlled, must comply with strict safety regulations. In a future scenario, thanks to advancements in the developed method – which integrates probability distribution maps and deep reinforcement learning techniques for precise, autonomous flight path planning – their role could be to arrive at the scene before other means, ensuring that the search begins as quickly as possible without wasting time. Once helicopters arrive, rescue drones would support them, accelerating the location of the missing persons.
TECHNOLOGICAL: the performance level of the described aerial path planning model is proportional to the accuracy of the probability maps. This means that the maps must always be based on high-quality, well-organised, reliable, and especially up-to-date data. Therefore, the future development of the AI model created by the study’s authors will need continuous new training data, such as GPS data from recent rescue operations. The main area of focus for improving the new planning system will be highlighting correlations within the training data between where missing persons were last seen and the exact locations where they were found. If the AI system can connect these dots, it could enable rescue drones to achieve increasing levels of autonomy in the coming years.
ECONOMIC: economically, the comparison in search and rescue operations is between the cost of using drones (which have electric motors) and the cost of helicopters, which includes fuel, equipment, and onboard operators. To give an idea of the numbers, in Italy, the cost of helicopter rescue interventions, though varying by region, averages around 100-120 euros per minute, fully covered by the National Health Service as per the Decree of 27 March 1992. Considering that in 2023 alone, the Italian National Alpine and Speleological Rescue Corps recorded 12,349 activities, 5,845 of which involved helicopters, the overall expense borne by the system is significant. In a future scenario, thirty years from now, where drones alone could mark the new aerial rescue strategy, the savings factor would be substantial and noteworthy.
POLITICAL: when discussing drones surveying vast territories searching for missing persons, the immediate concern is the handling of the large amounts of video data showing third-party individuals in public places – captured in flight by the aircraft’s onboard cameras – and the privacy protection of those depicted. Imagining a long-term future scenario where RPAs might replace helicopters in search and rescue operations, reference within the European Union must be made to GDPR, which prohibits recording images concerning private spaces of non-involved individuals by the drone, often containing sensitive data such as addresses, vehicle license plates, and facial details.
SUSTAINABILITY: the future possibility, over an extended period, of electric motor drones replacing helicopters in rescue operations in difficult terrain would have a positive impact on environmental sustainability. Consider some numerical data: «… the quad-rotor drone currently used by alpine rescue services, with a payload of 0.5 kg, has an energy consumption of about 0.08 MJ/km, which corresponds to a carbon footprint of 12.1-23.5 g of CO2/km, depending on the electricity production method used by the motor. This translates to a carbon footprint of 250–500 g of CO2 for a 40-minute flight compared to 720 kg of CO2 for a typical 50-minute helicopter flight. If drones completely replaced helicopters, this could lead to a potential reduction in the carbon footprint by 1,500 to 3,000 times» [source: “Green HEMS in mountain and remote areas: reduction of carbon footprint through drones?” – Scandinavian Journal of Trauma, Resuscitation and Emergency Medicine, July 2023].