Comparative Analysis of Different UAV Swarm Control Methods on Unmanned Farms

Abstract

Unmanned farms employ a variety of sensors, automated systems, and data analysis techniques to enable fully automated and intelligent management. This not only heightens agricultural production efficiency but also reduces the costs associated with human resources. As integral components of unmanned farms’ automation systems, agricultural UAVs have been widely adopted across various operational stages due to their precision, high efficiency, environmental sustainability, and simplicity of operation. However, present-day technological advancement levels and relevant policy regulations pose significant restrictions on UAVs in terms of payload and endurance, leading to diminished task efficiency when a single UAV is deployed over large areas. Accordingly, this paper aggregates and analyzes research pertaining to UAV swarms from databases such as Google Scholar, ScienceDirect, Scopus, IEEE Xplorer, and Wiley over the past decade. An initial overview presents the current control methods for UAV swarms, incorporating a summary and analysis of the features, merits, and drawbacks of diverse control techniques. Subsequently, drawing from the four main stages of agricultural production (cultivation, planting, management, and harvesting), we evaluate the application of UAV swarms in each stage and provide an overview of the most advanced UAV swarm technologies utilized therein. Finally, we scrutinize and analyze the challenges and concerns associated with UAV swarm applications on unmanned farms and provide forward-looking insights into the future developmental trajectory of UAV swarm technology in unmanned farming, with the objective of bolstering swarm performance, scalability, and adoption rates in such settings.

1. Introduction

Agriculture holds a cornerstone position in safeguarding global population survival, effectively mitigating hunger crises through its accelerated evolution [1]. The 2018 report by the World Resources Institute underscores a key goal of sustainable agriculture: a significant augmentation of agricultural productivity by 2050 to accommodate the dietary needs of an expected global populace of 10 billion [2]. Unmanned farms, a fusion of biotechnology, intelligent agricultural machinery, and information technology, offer a potent solution to fulfill these goals [3]. Serving as the backbone of Farm 4.0, unmanned farms epitomize advanced agricultural development, embodying the amalgamation of next-generation information technology, smart machinery, and progressive planting and farming techniques [4,5]. Distinct from traditional farms, unmanned farms harness a gamut of sensors, automated systems, and data analytics to realize fully automated and intelligent management, consequently slashing labor costs significantly [6,7].

The overarching objective in the development of unmanned farms is to foster deep engagement with the entirety of the agricultural production process via robot-based intelligent agricultural machinery. In this approach three types of unmanned farms can be recognized: remote-control-based, remote-guarded, and those operating autonomously. Intelligent agricultural machinery, serving as the tangible backbone of unmanned farms, incorporates significant advances on traditional agricultural machinery in four distinct areas: intelligent perception, automated navigation, precision operation, and intelligent management. Appropriate utilization of this advanced agricultural machinery can effectively mitigate labor shortages while optimizing long-term costs [8,9].

The agricultural Unmanned Aerial Vehicle (UAV), a significant component of intelligent farming machinery, is distinguished by precise operation, high efficiency, environmental sustainability, simple operation, and a high degree of automation and intelligence [10]. At their current stage of development, UAVs have found extensive applications in various operational facets of unmanned farms. These applications include crop monitoring and management [11,12], precision spraying [13,14], automatic crop detection [15,16], as well as crop mapping and yield estimation [17,18]. However, the existing level of technological advancement, coupled with relevant policy regulations, imposes substantial limitations on UAVs in terms of payload and range [19]. Specifically, when UAVs are utilized in large-scale unmanned farms, challenges arise such as the frequent need to replace batteries and fluid refills during single UAV operations, thereby leading to reduced operational efficiency and increased labor intensity. Consequently, to enhance operational efficiency, strategies beyond simply increasing the payload capacity and endurance of single UAVs are required. A viable solution involves deploying UAV swarms to conduct cluster operations collaboratively, thereby expanding the operational area and consequently enhancing the operational efficiency of the UAV [20]. Recently, UAV swarm cooperative operation systems have emerged as a potential means to augment mission efficiency, warranting further investigation [21,22].

A UAV swarm employs predefined modes of operation to facilitate simultaneous execution of specific tasks. This strategy curtails individual UAV operation times, accelerates the overall task completion speed, and significantly enhances operational efficiency for similar tasks [23]. The contrasting operational modes of single UAVs and UAV swarms are depicted in Figure 1. Furthermore, compared with single UAV operation, UAV swarms offer superior system reliability. Uncertainties during operation can cause some UAVs to malfunction; however, in such instances, UAV swarms can complete the corresponding tasks using other operational UAVs. Additionally, during operation allocation, UAV swarms can flexibly distribute tasks to individual UAVs based on distinct areas and operational scenarios, thereby further improving operational efficiency.

Currently, both domestic and international researchers are enthusiastically exploring the application of UAV swarm cooperative operation systems in sustainable agriculture. Although there have been some preliminary achievements, the research is still in its nascent stage. The intricate and varied nature of the agricultural environment implies that a single UAV swarm control method may not be universally applicable across diverse scenarios [25]. This complexity and diversity encompasses a range of factors, including the variety of environments such as orchards and farmlands, diverse task types like pesticide spraying and seeding, and the intricacy of weather conditions, including light intensity, wind speed, and temperature, among others. Different tasks necessitate specific flight altitudes, speeds, payloads, and modes, while varying conditions influence UAV flight stability, sensor performance, energy consumption, and swarm reliability. To effectively integrate UAV swarm technology with sustainable unmanned agriculture and thus reduce production costs and improve operational efficiency, it is essential to comprehensively review and analyze the application scenarios of UAV swarms on unmanned farms.

This paper presents a comprehensive review of the current literature on UAV swarm technology and its applications in unmanned farming. The analysis encompasses the prevailing techniques, challenges, and future directions for agricultural UAV swarms. The objectives of this review are: (1) to assist researchers in identifying effective control methodologies and understanding the technologies that underpin the development of UAV swarm systems in agriculture; and (2) to highlight promising and notable methods for creating guidance maps by comprehensively reviewing the application of UAV swarms and technical characteristics of current unmanned farming.

The remainder of this paper is organized as follows. The second section primarily discusses the existing UAV swarm control modes, offering a comparative summary across four distinct aspects: task assignment mode, trajectory planning mode, communication navigation mode, and visual navigation mode. The third section provides a summary of the four stages of UAV swarm deployment in unmanned farms, namely, cultivation, planting, management, and harvesting. This section also presents an analysis to determine which collaborative approach is most suitable at various stages and under different operational tasks. The fourth section analyzes and discusses the application of existing UAV swarm techniques on unmanned farms. Finally, the fifth section discusses the conclusions and future prospects.

2. Materials and Methods

The sources used for the compilation of the present review were the databases of Google Scholar, ScienceDirect, Scopus, IEEE Xplorer, and Wiley. The inclusion criteria for studies in this review were: (1) A comprehensive review was undertaken of multi-UAV cooperative navigation techniques from the past decade. (2) All relevant research on UAV swarm technology across the four stages of agricultural production, namely “cultivation, planting, management, and harvesting”, was summarized. However, due to the limited application of UAV swarm technology in the agricultural domain at this stage, studies on individual UAV operations have been referenced as alternatives in areas lacking UAV swarm applications. (3) For articles by the same author(s) reporting incremental findings, only the most recent articles were incorporated into this review.

Information on the research objectives, type of UAV swarm control, agricultural stage in which applied, the technological pros and cons, and technical challenges was extracted from each selected paper. By juxtaposing various control paradigms against the distinct operational features inherent to each agricultural phase, a series of analyses was undertaken to discern which methods hold prominence in various agricultural production stages.

This review categorizes UAV swarm technologies into distinct types based on their salient features and pivotal technical nuances, and then evaluates the potential each offers within unmanned farming scenarios. In conclusion, by synthesizing the current technical impediments and emerging research avenues, a visionary perspective on the trajectory of UAV swarms in future automated agriculture is presented.

3. Existing UAV Swarm Control Methods

UAV swarm navigation methodologies can be bifurcated into two primary categories: absolute navigation and relative navigation. Absolute navigation entails the segmentation of a mission objective into several smaller tasks by a ground computer, either prior to or during the flight of the UAV swarm. These tasks are created using relevant UAV task or trajectory algorithms and transmitted to each UAV, which then performs the assigned tasks to execute the navigational flight of the UAV swarm [26,27]. This mode can be further subclassified into task assignment collaborative mode and trajectory planning collaborative mode.

Relative navigation, integral to enabling UAV swarm navigation, operates on the principle of onboard sensors capturing information such as relative positions, attitudes, and flight conditions of neighboring or primary UAVs during the flight of the UAV swarm. Subsequently, each UAV adjusts its flight status and position based on preset constraints, facilitating UAV swarm relative positional navigation [28,29,30]. This mode can be further subdivided into communication collaborative navigation mode and visual collaborative navigation mode.

In this section, a comprehensive review of contemporary collaborative navigation methodologies for UAV swarms is presented. These methodologies are summarized from the vantage points of task assignment collaborative mode, trajectory planning collaborative mode, communication collaborative navigation mode, and visual collaborative navigation mode.

3.1. Task Assignment Collaborative Mode

The task assignment mode of a UAV swarm operates on the principle that, prior to the UAVs’ takeoff, individual tasks are allocated to each UAV based on user task requirements. These tasks incorporate parameters such as task area, duration, altitude, and more for each UAV. Upon completion of task assignments, the parameters are uploaded to each UAV. Subsequently, UAVs compute their respective flight paths using their inherent path planning algorithms based on the task specifics, making necessary adjustments to the flight path as per real-time conditions during the flight, and eventually accomplishing the user’s comprehensive task [31,32]. This mode of UAV swarm task assignment can further be classified into two types depending on the task assignment methodology: the centralized task assignment method and the distributed task assignment typical method.

3.1.1. Centralized Task Assignment Method

The centralized task assignment typical method, being the most commonly employed navigation technique under the UAV swarm task navigation mode, is frequently utilized [33,34]. This method involves the decomposition of tasks into multiple subtasks based on the number of UAVs, types and weights of tasks, and flight environment constraints, all executed via a ground task assignment terminal. These subtasks are then uploaded to the flight control system of each UAV in the swarm, which executes its mission flight according to the received subtasks.

In one study [35], UAV target task planning is conceptualized as a UAV path problem incorporating a time window and priority requirements. The authors propose a genetic algorithm equipped with an integrated mixed-integer programming solver. Another study [36] introduces a distributed multi-hop relay path optimization algorithm based on a matching game. This algorithm establishes a priority function for source nodes and relay nodes, elucidating the task execution order. It represents a more typical list planning method for optimization.

Addressing the motion trajectory problem of fixed-wing UAVs in complex three-dimensional environments, ref. [37] employs genetic algorithms and particle swarm algorithms to compute feasible and optimal motion trajectories while taking into account the dynamic characteristics of the aircraft. Apart from utilizing particle algorithms, this study also incorporates genetic algorithms from intelligent algorithms. Genetic algorithms exhibit high scalability, can be amalgamated with numerous other algorithm types, and can initiate search from the population, which offers potential parallelism and facilitates multiple individual comparisons simultaneously. Figure 2 provides an example of a centralized task assignment method.

The distinguishing feature of the centralized task assignment method lies in the presence of a singular task assignment terminal that decomposes and assigns tasks, managing the operations of the entire UAV swarm system. Each UAV exclusively receives instructions from the task assignment terminal, devoid of any information sharing or interactions. The solutions derived through the centralized task assignment method represent either global or local optimal solutions, thereby achieving optimal system decision-making and control.

Nonetheless, the centralized task assignment method primarily resolves complex combinatorial optimization problems, and hence exhibits relatively high computational complexity. Moreover, due to the absence of information feedback between the UAVs and the task assignment terminal, the real-time performance of this method warrants further enhancement.

3.1.2. Distributed Task Assignment Method

Contrary to the centralized task assignment method, the distributed task assignment method allows for the transmission of individual UAV information back to the task assignment terminal in addition to task decomposition and distribution. Consequently, the task assignment terminal assumes the role of an inter-network connector [38]. This connector enables the UAV swarm system to form a networked structure, facilitating information sharing, task coordination, and conflict resolution among UAVs through the inter-network connector.

The predominant distributed task assignment methods include multi-Agent theoretical methods and market mechanism-like methods [39]. Multi-Agent theoretical methods predominantly consider each UAV as an autonomous intelligent entity, orchestrating the mutual actions of each agent, appropriately analyzing and interpreting the target task, segmenting the overarching task into multiple small tasks of comparable intensity, and allocating them to each agent. Subsequently, the agents collaborate to accomplish the target task [40,41,42]. Market mechanism-like methods involve the application of market models such as contract tendering and item auctioning for task decomposition. The decomposed tasks are made available for bidding via open tendering or auctioning, with UAVs bidding in accordance with their actual task situation and flight status. Eventually, tasks are assigned based on the actual auction outcome [43,44]. Figure 3 provides an illustration of the distributed task assignment method.

The distributed structure boasts high autonomy and robustness, adeptly adapting to dynamic and uncertain battlefield environments. Given the significant enhancements in UAV platform autonomy capabilities, the distributed control structure presents broad application prospects and has emerged as a vital research direction for UAV swarm control.

The UAV swarm task assignment mode offers notable benefits including a high success rate and robust resilience to unexpected events during task execution. Simultaneously, given the pre-established target tasks for each UAV, energy consumption and other costs can be kept relatively low during flight. In the course of task execution, each UAV is also equipped with the flexibility to adjust its target trajectory based on the actual conditions encountered during flight. Consequently, this method exhibits high robustness, substantially improving task completion rates.

3.2. Trajectory Planning Collaborative Mode

The operational principle of the UAV swarm trajectory planning mode entails determining the starting point of the operational task prior to the aircraft’s takeoff. This mode considers a variety of complex constraints such as the maximum operational radius constraint, minimum turning radius constraint, no-fly zone constraint, time-domain collaborative constraint, and airspace collaborative constraint for each aircraft. Collaborating with the methods between each aircraft, this mode plans a single or multiple sets of safe and reliable flight paths for UAV swarm with shared task objectives. This ensures optimal overall performance of the UAV swarm during operations, facilitating successful completion of the target task [46,47,48]. Research methods commonly employed for the UAV swarm trajectory planning mode primarily encompass the optimal path planning method, the planning method predicated on artificial potential fields, and the planning method based on swarm intelligence algorithms.

3.2.1. Optimal Path Planning Method

The A* algorithm, a typical method in optimal path planning, segments the surrounding environment into equally sized grids. It estimates the optimal positions in the surrounding grids using heuristic information and continually identifies the optimal nodes, thereby facilitating the search for an optimal and shortest path [49,50,51]. However, the A* algorithm has certain limitations. When there exist two or more optimal values in the path grid, the algorithm cannot guarantee the optimality of the searched path. Furthermore, the A* algorithm’s efficacy is heavily reliant on the environment and is only applicable under favorable conditions. In instances when the derived map is large, the computational load increases, resulting in higher latency and inferior parallelism.

The Probabilistic Graph Space method initially divides the environment into equally sized grids and subsequently excludes grids that fail to meet constraints such as mutual constraints within the UAV swarm, flight task requirements, etc. The remaining grids constitute the optimal flight path [52,53]. The Probabilistic Graph Space method exhibits benefits such as a low order of magnitude, swift construction time, and maximal distance from threats and obstacles. Nevertheless, during the modeling process, the finer the map grid division, the more accurate the map description and the higher the flight accuracy. However, this also results in the storage of more map information, slowing down computation speed and inducing higher latency during flight. Figure 4 provides an illustration of the optimal path planning method.

3.2.2. Planning Method Based on Artificial Potential Field

The planning method based on artificial potential fields is a trajectory planning method that emulates spatial force fields. During the execution of tasks by the UAV swarm, the method simulates the task target attracting the UAVs, while UAVs and obstacles within the flight trajectory repel each other. Due to this attractive force, the UAVs progressively converge towards the target, while the repulsive force ensures safe flight distances within the UAV swarm, thereby fulfilling the target task [56]. To address the limitations of traditional artificial potential field methods, such as unreachable targets, susceptibility to local minimum values, inability to circumnavigate obstacles, and the absence of trajectory optimization strategies, the authors of [57] proposed a hybrid method premised on a Lyapunov guidance vector field (LGVF) and an improved interference fluid dynamics system (IIFDS), integrating vertical components, steering control, and speed control. A UAV Swarm flight trajectory based on an artificial potential field is shown in Figure 5.

The planning method based on artificial potential fields enables the swarm to effectively circumvent obstacles in the flight path and is characterized by high safety, superior real-time performance, and rapid planning speed. However, in complex environments, it is prone to local optimal values. When the swarm state changes, the dynamic planning capability of each UAV is compromised, and when the repulsive force of obstacles equals the attractive force of the target point, the UAVs may oscillate near the obstacles.

3.2.3. Planning Method Based on Swarm Intelligence Algorithm

Planning methods premised on swarm intelligence algorithms draw parallels to the swarm intelligence methods used in task assignment modes. Such methods emulate natural phenomena like ant colonies and bee swarms, viewing the UAV swarm as a collective entity and individual UAVs as components within the collective. The collective perceives the task target as the ultimate state, with individual entities autonomously adjusting and formulating plans according to their specific environments. Through the influence of these individual entities, the collective progressively approaches the ultimate state, ultimately deriving a solution that approximates the global optimum [58,59].

Swarm intelligence-based planning methods currently rank among the most widely researched algorithms in the realm of UAV swarm trajectory planning. Prominent among these algorithms are the ant colony algorithms [60], bee swarm algorithms [61], wolf pack algorithms [59], fish swarm algorithms [62], and genetic algorithms [63], among others. Notably, traditional swarm intelligence algorithms are susceptible to missing the optimal solution during the search process within the solution space, due to the randomness inherent in initial value settings. This often results in settling for local optima, which substantially diminishes planning accuracy. Figure 6 illustrates an example of a planning method predicated on the swarm intelligence algorithm.

Swarm intelligence-based planning methods possess several merits including rapid planning speed, strong parallelism, easy coordination, and the propensity to converge towards optimal values. Utilizing swarm intelligence algorithms allows for the unification and processing of multi-machine collaborative target assignment and trajectory planning with similar algorithmic structures [65]. However, they also have certain drawbacks. For instance, intelligent algorithms require the adjustment of many parameters, which makes the selection of key parameters a challenging task. Additionally, computational loopholes within intelligent algorithms could cause the UAV swarm system to stagnate or enter an infinite loop, ultimately leading to failures in planning and coordination.

3.3. Communication Collaborative Navigation Mode

Communication cooperative navigation stands as the most fundamental and core navigation method for UAV swarms [66]. It refers to the process whereby a UAV swarm in flight does not rely exclusively on ground stations or satellites for control. Instead, each UAV, leveraging high-precision sensor-acquired information such as location and attitude, views itself as a communication node [67]. Via Bluetooth, ZigBee, or WiFi wireless transmission modules, a wireless communication network is established among the UAV swarm for information exchange, fostering cooperation among various UAVs, and forming a unified entity [68]. Communication cooperative navigation can effectively enhance the comprehensive operational ability and expandability of the UAV swarm system, and it effectively assists other communication methods.

Currently, as the number of UAVs in UAV swarm systems gradually increases, so does the volume of data generated. The existing UAV swarm architecture, characterized by MAC protocols and routing algorithms, is challenged in handling this large volume of data, impacting the overall operation quality of the UAV swarm system. Thus, the authors of [69] introduced a deep Q-learning (DQN) model designed to optimize the entire network performance in view of the dynamic swarm topology and time-varying link conditions. Another study [70] proposed a multi-method combined UAV communication strategy, which besides utilizing the traditional UAV-to-UAV (U2U) architecture to form a network, also integrated a cellular UAV communication strategy, treating each UAV as a relay node for data transmission. Figure 7 provides an illustration of the wireless router-based UAV swarm control method.

The communication cooperative navigation mode of UAV swarms allows for network construction independent of other pre-established network facilities. It is capable of automatically forming a network in any operational scenario and dynamically modifying the network structure in accordance with specific operational requirements, endowing UAV swarm communication cooperative navigation with robust anti-interference capability [72,73]. Furthermore, it possesses high intelligence and versatile functions. However, it also has noticeable drawbacks: (1) The network protocols need improvement. Currently, the speed of UAVs in communication cooperative navigation of UAV swarms primarily falls within the range of 5 m/s to 15 m/s. At higher speeds, rapid movement of UAVs leads to frequent changes in the communication network topology, escalating the overhead of the communication network and exacerbating data congestion, thereby affecting real-time data transmission. (2) The communication security is inadequate. The wireless links and mobile topology within a UAV swarm network present varying degrees of security weaknesses. Additionally, under UAV swarm cooperative navigation flight, due to the complexity of the operational environment, safety issues may arise from other radio frequency interferences, among other factors. (3) The substantial information payload results in compromised real-time performance. During flight, the UAV swarm system generates a considerable amount of redundant UAV information. Given the limited existing wireless transmission bandwidth of UAV swarm communication, it cannot process the information promptly. This directly impacts the efficiency of information transmission, leading to significant latency between UAV swarm systems.

3.4. Visual Collaborative Navigation Mode

Recent advancements in vision-related recognition algorithms have thrust visual navigation into the spotlight as a significant research topic in both domestic and international navigation fields [74]. Compared with conventional radio sensors, visual sensors have the benefit of compactness and cost-effectiveness. Notably, the images collected by visual sensors contain a wealth of information, which when coupled with increasingly sophisticated visual algorithms, makes visual navigation a highly efficient method. During visual cooperative navigation in UAV swarms, UAVs capture feature information of other UAVs using visual sensors. This information is then processed through modeling, filtering, and other techniques to track the targets that fulfill certain requirements [75]. Estimating their motion state based on the target’s position, UAVs adjust factors like flight attitude or speed to adhere to preset constraints and complete the UAV swarm’s navigational flight. Within the visual cooperative navigation mode for UAV swarm, the primary control information for the UAV is derived entirely from the feature information provided by the visual sensors. Hence, the results produced by visual analysis have a direct impact on the flight quality and safety of the UAV. The key technology in visual cooperative navigation for UAV swarms involves extracting feature information from UAVs based on the images gathered by visual sensors and fusing this extracted information with UAVs’ inertial navigation data to accomplish the tracking flight. Visual cooperative navigation mode for UAV swarms can be categorized into three types: visually guided navigation, visual multi-source information fusion navigation, and visual SLAM navigation.

3.4.1. Visually Guided Navigation

Visual guidance pertains to a technique where UAVs, during the flight process of a swarm system, obtain all feature information and position-related data through visual sensors. Utilizing this obtained data and preset constraints, UAVs adjust their flight statuses to track the flight of one UAV or a UAV swarm, thereby achieving coordinated UAV swarm flight. This method is primarily employed in areas where GNSS (Global Navigation Satellite System) signals are weak or absent. Given its reliance on direct visual tracking for flight, it effectively circumvents issues induced by GNSS signals.

A typical technique in visual guidance is the “Leader-Follower” approach [76]. In this methodology, certain UAVs assume the role of leaders while others act as followers within the swarm. Each follower has at least one designated leader, and within the swarm system, some UAVs may serve both as a leader and a follower. This hierarchical structure simplifies the swarm system into a singular motion planning problem. In a string or chain-like structure, each UAV merely needs to follow the preceding UAV. Figure 8 provides an illustration of the “Leader-Follower” visual guidance method.

The visual guidance approach allows UAV swarm systems to function effectively in areas where GPS signals are unavailable, broadening the potential application scenarios of such systems and introducing a novel methodology for their operation [77]. Nevertheless, given that visual guidance is entirely dependent on data provided by visual sensors, it places substantial demands on the reliability of both the visual sensors and visual recognition algorithms. Moreover, as visual sensors are sensitive to variations in light intensity, and visual recognition requirements can fluctuate under different sunlight conditions, enhancing the robustness of visual recognition algorithms remains a crucial research objective.

3.4.2. Vision-Based Multi-Source Information Fusion Method

Unlike the visual guidance method, which exclusively employs vision to guide the UAV swarm system, the vision-assisted method integrates information from GPS, Inertial Measurement Unit (IMU), and Magnetic (MAG) sensors in addition to using vision to extract and track feature information [78]. The master UAV is spatially localized by GPS and follows a predetermined route. Then, the follower UAV uses vision sensors to extract the master’s feature information. Upon obtaining the relevant positional information, it fuses the multi-source information with the GPS, IMU, and MAG data on the follower. This results in a flight path determined by the optimal method obtained from multi-source information fusion [79,80].

A typical vision-assisted method is the Parent-Child approach. In this approach, only one UAV is treated as the parent, and it does not require the installation of visual sensors. Instead, it directly uses GPS for path localization. There can be at least one or more UAVs acting as children, which are equipped with visual sensors along with other positional sensors such as GPS, IMU, and MAG. Concurrently, children UAVs can further function as parents for other UAVs.

The vision-assisted method can better realize UAV tracking and localization during the flight process through the fusion of multi-source sensor information, thereby improving the operational accuracy of the UAV swarm system. Figure 9 illustrates the vision-based multi-source information fusion method.

3.4.3. V-SLAM

Visual Simultaneous Localization and Mapping (V-SLAM) is a technology that utilizes visual sensors to simultaneously accomplish localization and mapping. While it can provide services for ground robots, it also has crucial applications in the localization and navigation of UAVs [82,83]. V-SLAM primarily consists of five components: sensor data acquisition, front-end visual odometry, back-end nonlinear optimization, loop closure detection, and map building [84]. Figure 10 illustrates the V-SLAM framework. In situations where the UAV lacks a priori information about the environment, it first acquires and preprocesses the image information through the vision sensor. The front-end visual odometry then utilizes the image information captured by the vision sensor and the related position information to reconstruct the 3D motion of the camera, resulting in a localized motion map.

However, the front-end visual odometry can only provide trajectories and maps for a short duration and cannot provide all maps. Therefore, it is necessary to employ back-end nonlinear optimization to match the optimal trajectory over an extended period with the large-scale optimization of the optimal mapping, ultimately leading to the construction of the map. The front-end visual odometry only considers the correlation of neighboring time frames and cannot build globally consistent trajectories and maps. Loop closure detection can effectively identify instances where the camera passes through the same location multiple times, providing longer constraints than neighboring frames.

UAV swarm visual co-navigation based on V-SLAM technology enhances the independence and anti-interference capabilities of UAVs [85]. However, the inherent limitations of V-SLAM frequently manifest as follows: Firstly, typical aerial images may exhibit local blurring phenomena, impacting the accuracy of UAV localization. Secondly, the extensive integration operation processes may lead to the emergence of cumulative error, which consequently influences the system’s accuracy. Additionally, SLAM itself suffers from issues of scale uncertainty and scale drift, which can severely affect system outcomes. Figure 11 depicts the UAV flight trajectory based on V-SLAM.

The UAV swarm visual co-navigation mode utilizes visual sensors to directly co-navigate UAV positions. By providing reliable relative position and attitude information to the UAV swarm system, it enables each UAV within the UAV swarm system to use visual sensors to estimate the relative positions and distances of other UAVs within its own coordinate system. In the presence of GPS, visual navigation can contribute to the navigation information, facilitating UAV swarm cooperative navigation. When GPS is denied, visual navigation can act as an independent navigation system to accomplish relative navigation between UAVs in a swarm. Moreover, visual navigation can achieve cooperative navigation of a UAV swarm system without requiring mutual communication between UAVs, making it an essential means of cooperative navigation in complex environments. However, the current UAV swarm visual cooperative navigation still has several shortcomings: (1) The complexity of the environment where the UAV swarm system operates necessitates the development of a robust recognition algorithm. This algorithm should be able to reduce interference of UAV recognition from obstructive objects and accurately identify and track UAVs in a complex environment. (2) In terms of data processing, while visual sensors provide rich information, they also generate a substantial amount of data. As the number of UAVs in the UAV swarm system increases, the amount of data produced by the visual sensor rises exponentially. Therefore, determining how to process and utilize the obtained visual information is a key issue that requires resolution. (3) After acquiring effective and reliable data, the formulation of a reasonable control logic and strategy to actualize the UAV swarm system is necessary. (4) Exploring strategies that ensure the UAV swarm system operates at maximum efficiency with minimal resource consumption during actual flight is also a direction worthy of investigation.

This section has comprehensively summarized and analyzed various implementations of the UAV swarm system at the current stage of development, and these are presented in

Table 1. Overview of the existing navigation systems.

Main Cooperation Strategy	Secondary Cooperation Strategy	Characteristics
UAV Swarm Collaborative Tasking Model	Centralized tasking model	(1) The solution obtained, whether it is the global optimal solution or a local optimal solution, facilitates the attainment of optimal decision-making and control within the system. (2) The process involves higher computational complexity due to the optimization of complex combinatorial optimization problems. (3) The lack of information feedback between the UAV and the task allocation endpoint leads to deficient real-time performance and poor robustness.
	Distributed tasking model	(1) UAV Swarm systems are characterized by high autonomy and robustness, enabling them to adapt to dynamic and uncertain battlefield environments effectively. (2) The systems can match tasks based on their own task situations, leading to high task operation efficiency.
UAV Swarm Collaborative Trajectory Planning Model	Optimal path planning method	The performance of the system is more dependent on the environment, and the amount of computation increases as the map size grows, resulting in higher latency when applied to larger maps.
	Planning method based on artificial potential field	(1) The system exhibits high security, good real-time performance, and fast planning speed. (2) In complex environments, it is prone to producing local optimal values. (3) The dynamic planning ability of each UAV is poor when the cluster state changes.
	Planning method based on swarm intelligence algorithm	(1) The planning process is fast, parallel, easy to collaborate, and capable of converging to the optimal value. (2) The algorithm requires a large number of parameters to be adjusted, making it challenging to select the key parameters effectively. (3) In the presence of computational loopholes in the intelligent algorithm, the UAV Swarm system may experience stagnation or dead loop phenomenon, resulting in planning failure and loss of synergy.
UAV Swarm Communications Co-Navigation Model	-	(1) The system exhibits the capability for automatic network composition, dynamically adjusts the network structure according to specific operational requirements, and possesses a strong anti-interference ability. (2) The system demonstrates a high degree of intelligence and offers diverse functionalities. (3) The system requires an upgrade in network bandwidth and network protocols, and currently exhibits weak real-time performance due to extensive information loads.
UAV Swarm Visual Co-Navigation Model	Visual guide	(1) The UAV swarm system can operate effectively in GPS-denied areas, thereby expanding its application range. (2) The visual sensor is characterized by its compact size and affordable cost. (3) The visual navigation system may experience tracking errors, necessitating high reliability in both the visual sensor and the visual recognition algorithm.
	Vision-based multi-source information fusion method	(1) The fusion of visual and other diverse information facilitates high-precision tracking of flight. (2) More robust on-board processors are required to process and fuse a greater volume of information.
	V-SLAM	(1) Greater independence and anti-interference capabilities. (2) Extensive integrated arithmetic processes can lead to an accumulation of errors, affecting system accuracy. (3) SLAM possesses issues with scale uncertainty and scale drift, which can impact the system’s results.

. Due to the complexity of the operational environment within farmlands, further summarization and analysis of the practical application scenarios of UAV swarms on unmanned farms at the current stage of development are required. Consequently, the following section will explore and analyze the application of UAV swarm on unmanned farms.

4. UAV Swarms for Unmanned Farms

The operational schema of autonomous farming primarily encompasses four stages: cultivation, planting, management, and harvesting. Each stage is integral to the overall farming process. UAVs, serving as a pivotal component of contemporary agricultural machinery, can significantly augment both the quality and efficiency of agricultural tasks across these stages. To establish a versatile control methodology for UAV swarm operations adaptable to diverse agricultural requirements, it is essential to conduct a comprehensive analysis of UAV swarm applications within the realms of cultivation, planting, management, and harvesting. Given the nascent stage of UAV swarm technology application in sustainable agriculture, this section seeks to elucidate the role of UAV and UAV swarm technology in sustainable agriculture, particularly within the context of the aforementioned four stages.

4.1. Application of UAV Swarms in Cultivation

The field cultivation stage primarily encompasses two facets. The first involves preparing the soil for crop planting through processes such as turning over, crushing, and smoothing. The second facet involves surveying and mapping the cultivated land to preemptively mitigate damage. At this stage, UAVs are predominantly utilized for the surveying and mapping of cultivated land.

Field cultivation, a critical component in human sustenance, is increasingly threatened due to economic development leading to the occupation and degradation of cultivated land. Consequently, safeguarding these lands has emerged as a global priority to ensure food security. As highly efficient low-altitude remote sensing equipment, UAVs, when equipped with the appropriate sensors, can automatically survey, map, and calculate the area of cultivated land. In [87], a method was proposed for extracting and mapping farmland plots in southern China using UAV multispectral images and deep learning. The application of a multispectral camera on a UAV enabled the capture of images while flying over cultivated land. Through deep learning technology, these captured images were stitched together, facilitating the mapping and calculation of cultivated land data. Figure 12a presents a surveying and mapping diagram of farmland plots.

The authors of [88] employed high-resolution images captured by UAVs, alongside multispectral images and elevation data from a digital surface model, to classify agricultural land use. The practical experiment demonstrated that the classification accuracy could reach up to 92%, attesting to the viability of UAVs in classifying and mapping agricultural land. Figure 12b illustrates the classification effect diagram. In another study [89], UAVs were used to detect and capture images of agricultural land areas. Through high-pass filters, wavelets, principal component transformations, and other methods, image fusion was performed to measure and calculate agricultural land areas. The experimental results indicated that the error of the obtained agricultural land area was less than 20 cm.

The application of UAV swarm technology in cultivated land mapping is relatively less explored. Research conducted by [90] proposed a method for mapping and measuring cultivated land information based on three UAVs. This approach leverages UAV swarm technology under the collaborative task allocation mode, in which an overarching flight task is divided among the UAVs. Post-division, the three UAVs proceed with autonomous flight operations as per their assigned tasks and conduct a collective analysis of the gathered image data. To validate the proposed method, a practical flight experiment was conducted in an area measuring 330 m × 200 m. The results indicated that each of the three UAVs accurately executed the mapping of the cultivated land in accordance with the assigned tasks, with the mapping accuracy observed to be high. Figure 13 shows the comparison between the theoretical flight trajectory and the actual flight trajectory of the three UAVs, providing evidence of the effective implementation of the task allocation strategy.

In conclusion, the measurement and mapping of cultivated land information require large-scale land coverage, whether it is cultivated or non-cultivated land. As a result, the information exchange requirement within the UAV swarm is relatively low, where the tasks only need to be sent to UAVs after decomposition. Each UAV can then conduct operational flights according to the scope of their respective tasks and collect relevant data. Therefore, the UAV swarm cooperative mode is well-suited to task assignment in the context of cultivated land information mapping and surveying. This approach allows for efficient and coordinated mapping operations, ensuring accurate data collection and analysis for agricultural land monitoring and management.

4.2. UAV Swarms in Planting

The planting phase is pivotal within autonomous farming, with the density and uniformity of planting directly influencing crop germination. Unmanned helicopters, given their compact size, flexible maneuverability, and ability to execute programmed flight paths, provide noteworthy advantages, particularly within hilly terrains typified by small plots and significant altitude variations, where conventional, larger ground seeding machinery may find navigation challenging. Currently, the implementation of seeding strategies that employ unmanned aircraft is a burgeoning area of research [91,92], with the potential to significantly enhance the precision and efficiency of crop planting operations, especially within challenging agricultural landscapes.

The authors of [93] delineate the design of a UAV-based device for rice seeding and sowing. As illustrated in Figure 14, the seeds stored in a seed box are funneled into a seed dispenser, from where they are distributed to various launching modules via a dispersing device. Subsequently, each launching module’s milling wheel accelerates the seeds prior to their deposit on the ground. Following field tests, this device exhibited an impressive performance under operational conditions of a 1.5 m working height and a sowing rate of 38.56 kg/hm². The results demonstrated a significant in-row sowing performance, exhibiting a sowing rate deviation of a mere 1.89%. Sixteen days post-sowing, the seed emergence rate stabilized at an average seedling rate of 82.63%, yielding 6775.50 kg/hm². This sowing approach has similar performance to the spot-shot sowing method [94,95], ensuring a relatively stable seed sowing position to effectively manage planting density and guarantee a high level of uniformity.

Contrasting with the methodologies previously described, in [96], the authors delineate the design of a UAV-based centrifugal disc seeding device. According to this study, seeds, after being discharged from the seed dispenser, are fed directly into a centrifugal disc. The high-speed rotation of this disc generates a centrifugal force that propels the seeds onto the ground, thereby completing the seeding process. The fundamental architecture of this system is depicted in Figure 15. However, this centrifugal disc seeding approach diverges from prior methods in its use of centrifugal force for seeding, rendering the control of seeding density and uniformity during the actual seeding operation unfeasible.

Currently, no research specifically focuses on UAV swarm navigation methodologies applied to planting and seeding. Hence, this section will undertake a relevant analysis using the aforementioned two seeding methods in conjunction with prevailing multi-agent cooperative control strategies. The point-and-shoot seeding method has the advantage of high precision, allowing for well-regulated seed placement. Consequently, during practical operations, it suffices to plan individual flight trajectories for each UAV, ensuring comprehensive coverage of the operation area and preventing instances of missed or redundant seeding. In contrast, the centrifugal disc method cannot precisely calculate the landing point of seeds on the ground due to the primary influence of the centrifugal force. Furthermore, the actual landing points of the seeds may not be evenly distributed, potentially leading to missed seeding when using a trajectory task coordination mode. Therefore, implementing a visual navigation mode during actual seeding operations allows for real-time analysis of ground seed distribution during cooperative flights. This approach facilitates timely reseeding for plots that are missed or under-seeded.

4.3. Application of UAV Swarms in Field Management

Field management encompasses an array of management measures executed throughout the cultivation process, from sowing to harvesting, within the realm of autonomous farming. It constitutes the core phase of the entire autonomous farming operation. Currently, the primary applications of UAV swarm systems in field management are irrigation control, weed identification, pest identification and control, and crop health surveys.

4.3.1. Agricultural Sprays

In Reference [97], a two-UAV field spraying device is proposed, with its structural design depicted in Figure 16. The working principle involves simultaneously connecting the chemical tanks of two UAVs to a single spray bar. The UAVs operate in synchrony, moving the spray bar in unison, thereby increasing the operation area and volume of spray per single application. Considering the payload capacity of the UAVs is limited, the distance between the two UAVs is not particularly large. Moreover, given the requirement for precise synchrony in the same direction, the implementation of visual collaborative navigation or communication navigation modes can fulfill the aforementioned requirements.

Reference [98] addresses several challenges pertaining to pesticide spraying operations conducted by UAV swarms. These challenges include hardware discrepancies, irregular field boundaries, spray drift, and varying operator skill levels. To address these issues, a mathematical model for a heterogeneous UAV swarm system for crop protection has been established. Building upon this model, an optimal spray task allocation method for a UAV swarm involved in crop protection was proposed. Following spray tests on a five-UAV system, the effectiveness of this optimal spray task allocation method was confirmed for actual pesticide spraying operations.

4.3.2. Field Monitoring

Field monitoring in autonomous farming comprises elements such as weed identification, pest detection, and remote sensing of crop growth trends. The authors of [99] propose a dynamic path planning coordination method for UAV swarms for field monitoring and fertilization tasks, based on an improved distributed ant colony algorithm. By comparing the improved algorithm with traditional zigzag methods in a simulated monitoring and fertilization environment, it was verified that the improved algorithm reduces the time required to complete tasks. The authors of [100] also conducted research on the control methods of UAV swarms for field environment monitoring, UAV flight trajectory transmission, storage of UAV ground part trajectories, and data transmission between the Wireless Sensor Network (WSN) and the UAVs. The reliability of the system was proven through experiments in both simulated environments and actual crop monitoring points.

Reference [101] introduces a cooperative remote sensing approach for real-time water management and irrigation control based on UAV swarms. It discusses the development of basic subsystems such as embedded multispectral imagers, image stitching, and registration, and provides the results of an actual UAV flight test for a typical flight mission. In [102], the authors propose a UAV swarm navigation method for crop detection based on wireless communication. UAVs share their position information through xBee, Wi-Fi, or other communication modules, and the on-board computer sends all this information to the ground station. The ground station then plans and adjusts the UAV flight missions collectively. Figure 17 depicts the high-level architecture of the prototype.

References [103,104] address the issue of weed coverage and mapping in farmland, proposing a UAV swarm flight system based on communication and visual cooperation. This system shares location information among UAVs through the XBee module and collects weed information through the visual collection module. The reliability of the system was validated through simulated experiments and actual flight tests. In [105], researchers also identify and detect field weeds through a UAV swarm system, adopting a coordinated mode of trajectory planning under an absolute navigation mode. Reference [106] presents a collaborative method based on trajectory planning for UAV swarm systems to collect low-altitude remote sensing data in farmland. By comparing four algorithms (Nearest Neighbour based on K-means clustering, Christofides, Ant Colony Optimisation and Bellman–Held–Karp), the feasibility of this research was validated. In [107], the authors propose an enhanced path planning technique for UAV swarm systems in areas of interest for the problem of visible light and multispectral collection path planning for assessing the health of field crops. By increasing the flight path and collecting more data in areas of interest, non-uniform sensor data collection is achieved. The feasibility of the theory was validated through simulation and actual field experiments.

In summary, UAV swarm involvement in the management process of autonomous farming mainly encompasses the two aspects mentioned above. In terms of pesticide spraying or crop fertilization, precision control is crucial, as the amount of spray directly influences crop growth. Therefore, during the operation process, it is necessary for each UAV to communicate and coordinate. In practical usage, communication collaboration methods or visual collaboration methods are recommended. During field inspections, the installation of relevant sensor equipment on the UAVs allows for the collection of field information. As this process does not demand high precision for minute details, UAVs can complete the corresponding tasks by flying according to pre-designed tasks or trajectories during the operation process.

4.4. UAV Swarms in Harvesting

The automation of the harvesting process on unmanned farms is indeed a critical component. It holds the potential to significantly enhance farm productivity, reduce labor costs, and ensure the quality and yield of crops [108]. UAV swarm technology can greatly contribute to the harvesting process on unmanned farms, enabling efficient and precise harvesting, while also effectively mitigating crop losses. This technology could revolutionize traditional harvesting techniques, paving the way for a new era of smart, automated farming.

The two predominant applications of UAV swarm technology in the harvesting process are yield estimation and automated picking. Yield prediction is of paramount importance for farmers as it facilitates informed decision-making related to crop insurance, storage requirements, cash flow budgeting, and allocation of inputs such as fertilizers and water [109]. Similar to the methods of farmland measurement, mapping, and field inspection previously discussed, yield estimation is conducted by installing various sensor devices on UAVs to perform low-altitude remote sensing of agricultural lands. The collected remote sensing data are then analyzed to estimate the yield. At present, there are limited studies applying UAV swarm technology to yield estimation. For instance, a study presented in [110] proposed a UAV-based citrus yield estimation approach. This approach employs automatic image processing methods for the detection, counting, and size estimation of citrus fruits on individual trees, utilizing deep learning techniques. The approximation error was a mere 4.53% with a standard deviation of 0.97 kg, as validated on 20 citrus trees. This validates the effectiveness of the proposed algorithm. Figure 18 illustrates this UAV-based citrus yield estimation approach. In another research study [111], a cotton yield prediction model was established. The model leveraged multi-temporal, high-resolution, visible and multi-spectral UAV remote sensing images. The Bayesian regularized BP neural network and the ENVINet-5 semantic segmentation model were used for the prediction process. The model was validated through actual field experiments, confirming its feasibility for yield prediction.

In [112], the authors propose a UAV swarm-based cooperative task assignment model specifically for orchard picking. This heuristic cooperative method, with the shortest travelling distance as a key indicator, swiftly addresses multiple tasks. The basic performance of the algorithm was tested through simulation experiments. In 2023, Tevel Aerobotics Technologies, based in Israel, developed an orchard picking system predicated on the UAV swarm system. A physical diagram of the system is displayed in Figure 19. The system comprises ground-based UAVs and a UAV swarm. These components share real-time location information with each other via a communication mode. Subsequently, they analyze and locate the fruits through vision. Upon successful localization, the UAV’s mechanical arm picks the apples and places them into the UAVs. This system significantly enhances the efficiency of orchard picking and has emerged as a new direction for development in the realm of unmanned farming [113].

5. Results and Discussion

Upon reviewing the literature discussed above, it is evident that UAV swarm technology has significant applicability in diverse aspects of unmanned farming. This includes efficient agricultural land management, crop fertilization, orchard harvesting, and pest and disease monitoring. Utilizing UAV swarm technology can effectively enhance agricultural efficiency, curtail costs, and concurrently improve the quality and yield of agricultural produce. Due to the varying technical requirements and operating environments of production processes on unmanned farms, it is essential to investigate the compatibility of UAV swarming technology with agricultural production stages. This paper examines the operational requirements of UAVs across the four stages: cultivation, planting, management, and harvesting. We then propose specific control methods for UAV swarms in each stage.

Table 2. Operational requirements and suggested control methods for UAV swarm on unmanned farms.

Operations Segment	Operational Characterization	Control Methods	Reasons
Cultivation	(1) The operational scope is expansive, with a significant UAV flight altitude. (2) The precision required for UAV flights is not stringent; the primary concern is to ensure that the targeted area falls within the visual capture range. (3) Inter-UAV communication is minimal.	Collaborative Tasking Model, Collaborative Trajectory Planning Model	The utilization of a UAV swarm in arable land predominantly focuses on large-scale measurement and mapping of agricultural land data. During operational procedures, it suffices to dispatch task decomposition or path planning to individual UAVs. Subsequently, each UAV conducts its flight operations within designated task boundaries and gathers pertinent data.
Planting	(1) Flight altitude is low, necessitating high precision in spreading. (2) For spot seeding devices, a precise flight trajectory is essential, accompanied by heightened accuracy in coverage. (3) Disc seeding devices mandate real-time monitoring of the seeding process to preclude redundancies or omissions.	Visual Co-Navigation Model, Collaborative Trajectory Planning Model	The benefit of point-and-shoot seeding lies in its exceptional precision, allowing for enhanced control over seed positioning. Consequently, during the operational phase, planning a distinct flight trajectory for each UAV suffices. In contrast, disc seeds do not possess a uniform distribution upon landing. Thus, employing a visual navigation mode enables analysis of seed distribution on the ground, facilitating timely replanting.
Field management	(1) The spraying operation requires a low flight altitude coupled with high spraying precision. (2) The flight altitude designated for field inspection is contingent upon the operational task, guaranteeing that the targeted plots fall within the visual capture range.	Collaborative Tasking Model, Collaborative Trajectory Planning Model, Communications Co-Navigation Model	During pesticide spraying or crop fertilization, the varying concentrations of liquid can significantly influence crop growth. Thus, adequate communication and coordination among the drones are imperative during the operation. Conversely, the fields do not necessitate meticulous detailing, allowing UAVs to operate based solely on pre-defined tasks or trajectories to fulfill their respective assignments.
Harvesting	(1) The yield measurement operation necessitates a higher flight altitude, with each UAV being responsible for image acquisition within its designated operational zone. (2) The fruit harvesting operation requires a lower flight altitude but demands greater flight precision. Additionally, UAVs must communicate interdependently for rapid information exchange.	Collaborative Tasking Model, Collaborative Trajectory Planning Mode, Communications Co-Navigation Model, Visual Co-Navigation Model	For yield measurement tasks, a UAV swarm predominantly focuses on capturing remote sensing images within the cultivation area. As such, UAVs merely need to operate based on predefined tasks or trajectories. In contrast, fruit harvesting operations necessitate intensive picking and harvesting activities within a confined space. Therefore, swift information exchange between UAVs is imperative to prevent collisions, overlooked harvests, and repeated harvests.

provides a detailed analysis of the operational requirements and suggested control methods for UAV swarm in these segments.

On unmanned farms, distinct operations necessitate unique flight and procedural requirements. By tailoring control methods to each operation’s specific traits, the UAV swarm optimizes efficiency and accuracy, thereby addressing the multifaceted demands of contemporary agriculture. As illustrated in Table 2, among the four cluster control models, the UAV Swarm Collaborative Tasking Model is the most frequently employed. This preference can be attributed to the versatility of remote sensing imagery technology in various agricultural domains. It offers specific functionalities without necessitating exceptionally high operational precision, ensuring effective data acquisition within the designated operational zone. Both the Communications Co-Navigation Model and the Visual Co-Navigation Model should be chosen and fine-tuned based on particular mission contexts. By assessing the diverse operational facets of unmanned farms and the corresponding UAV cluster methodologies, future unmanned farms can tailor UAV swarm co-navigation approaches based on the nature of their operation during the entire automation process. Such strategic alignment enhances the intelligence quotient of UAV swarms.

Nevertheless, despite the strides made in the application of UAV swarm technology within unmanned farming, several problems and challenges persist:

(1)

Limited agricultural application scenarios: It is evident from the current applications of UAV swarm technology in sustainable agriculture that the focus is primarily on acquiring low-altitude remote sensing images [87,88,102,110]. When UAVs execute low-altitude remote sensing operations, they navigate a relatively simple environment, primarily following pre-established tasks or routes; for example, Reference [114] uses simulated annealing algorithms to assign sub-tasks to each member of a swarm for operational flights. This process currently overlooks the dynamic complexities associated with other environments such as farmlands or orchards, such as the avoidance of poles or other obstacles, and the movement of plants and animals [115]. For other farm operations like crop sowing [93], pesticide spraying [96], and farmland harvesting [110], it is imperative to consider not only the operation’s completion efficiency but also the safety of the UAV and the farmland [116]. This makes the development of UAV clustering technology in these areas relatively slow. Therefore, further research is needed on UAV clustering technology in combination with the actual field operation environment, thereby extending its applicability to broader scenarios.

(2)

Insufficient levels of intelligence: The realization of unmanned farms requires advanced information perception, quantitative decision-making, intelligent control, precision inputs, and personalized services throughout agricultural production and management [7]. This facilitates the accomplishment of sustainable development goals such as intensive agricultural production, high yield, superior quality, high efficiency, ecological balance, and safety [117]. However, current UAV swarm systems still require manual operations for task planning and path scheduling, and cannot completely achieve autonomous decision-making and automatic task execution. For example, as reported in [21,35], it is necessary to manually plan the UAVs’ mission and adjust the flight trajectory before production operations can be carried out. Despite some studies exploring UAV control methods based on artificial intelligence algorithms, UAV swarm control algorithms are not yet fully capable of environment perception and autonomous decision-making. Moreover, with the increasing number and types of sensors carried by UAVs, there is an urgent need to solve the problem of how to effectively process and utilize the collected data to enhance the intelligence of the swarm [81].

(3)

Inadequate hardware and software support: The application of UAV swarms in agriculture is a long-term project, as UAV swarms can potentially contribute to all aspects of unmanned farm operations. Nevertheless, the majority of UAVs are currently designed to fly in optimal conditions, often falling short in responding effectively to various agricultural environment challenges, such as fluctuating weather conditions, physical shocks, and electromagnetic interference [118]. Additionally, current UAV operations predominantly rely on manual judgment to determine the appropriateness of operational timeframes, with the software support resources primarily serving to control UAV swarm flight. If there are any significant weather or scene changes during operation, manual adjustments or halting of the UAV swarm’s operational tasks may be necessary. Thus, beyond improving the operational perception of UAV swarms from an intelligence perspective, it is also vital to integrate other intelligent perception devices in unmanned farms to enhance the software decision-making capabilities of UAV swarms [119]. Consequently, to further the application of UAV swarms on unmanned farms, there is a pressing need to develop more comprehensive hardware and software facilities to support UAV swarm deployment.

6. Conclusions and Prospects

This paper delves into the application of UAV swarm technology in the context of unmanned farming. Initially, we provide a comprehensive summary and analysis of existing UAV swarm control methods, highlighting the strengths and weaknesses of various approaches. We proceed to evaluate the current applications of UAV swarm technology within the four primary processes inherent to unmanned farming: cultivation, planting, management, and harvesting. Moreover, we propose potential implementation strategies based on UAV swarm technology for each operational procedure.

Furthermore, this paper offers an in-depth discussion on the potential to leverage intelligent control aspects of UAV swarms in unmanned farm management. We have considered how to broaden the scope of UAV swarm technology applications within unmanned farms, improve the level of UAV swarm intelligence, and construct pertinent hardware and software infrastructures.

In recent years, the swift advancement of UAV swarm technology within sustainable agriculture and unmanned farms has created promising opportunities. However, it has also encountered several complex challenges. Addressing these practical problems necessitates consideration from multiple perspectives and dimensions. This paper aims to provide potential solutions for the robust development of UAV swarm technology in unmanned farms by exploring three specific aspects.

(1)

Accelerating the autonomy and viability of UAV swarms. The numerous components of unmanned farm production necessitate the use of diverse technologies and encounter a wide range of environmental variables. These variables may include differing crop types, growth conditions, pests and diseases, as well as shifting weather conditions. Therefore, to enhance the autonomy of UAV swarms, there is a need to improve various control algorithms embedded within the UAVs. These include environment perception algorithms, path planning algorithms, and cooperative control algorithms, among others. Such improvements can facilitate the capacity of UAV swarms to adapt their operational mode to cater to varying task information. They can accomplish operational tasks autonomously and make adjustments in real-time according to the prevailing environment and task conditions. This will essentially improve the level of intelligence of the UAV swarm, realizing the efficacy of single-button operation.

(2)

Building a comprehensive model for analysis and decision-making on unmanned farms. A comprehensive model enabled by complex multilayer network structures learns by abstracting high-level features from massive data sets to facilitate more accurate predictions or decisions. Therefore, by leveraging existing comprehensive model algorithms in combination with deep learning networks, data related to unmanned farm cultivation and production can be used as input parameters. These might include crop yield data, climate data, soil data, satellite images, and remote sensing data, among others. Such a comprehensive model could simulate and predict the entire agricultural production and planting process, allowing for analysis, summarization, and decision-making at each stage. Additionally, a comprehensive model could be integrated with an intelligent farm machine control system to realize autonomous understanding, prediction, and optimization of agricultural production. This integration would significantly enhance the intelligence level of agricultural production.

(3)

Promotion of an integrated system of smart agricultural technologies and the agricultural Internet of Things. UAV swarms are a critical component of unmanned farms, which also encompass other intelligent agricultural equipment such as sensor networks, automated agricultural machinery, data processing, and decision-support systems. Therefore, to further enhance the intelligence level of unmanned farms, it is not only essential to improve the autonomy and feasibility of UAV swarms but also to integrate existing smart agricultural equipment and technologies. This integration would facilitate the establishment of an intelligent unmanned farm system. Moreover, such a system would significantly boost the evolution and implementation of the Internet of Things (IoT) within the agricultural sector.

A schematic diagram of the system framework is depicted in Figure 19. Initially, a remote monitoring and control system is developed, through which agricultural producers can remotely supervise the status of the unmanned farm and operate various types of equipment. Moreover, this system is capable of summarizing the data from all the sensors within the unmanned farm and presenting it visually. Further, it can define the types of crops to be planted in the plots and the necessary field management requirements, for example, whether basic fertilizer application or pesticide spraying should be carried out.

Finally, we propose to develop an agricultural production model based on the unmanned farm. This model would disassemble the entire planting process for different crop types, analyzing and decomposing the operational tasks that need to be accomplished in each link from the four aspects of “cultivation, planting, management and harvesting”. At the same time, the model would determine the automated agricultural equipment necessary for each operational task. When the preset time for a certain operational task is reached, the remote monitoring and control system would deploy each piece of automated farming equipment to carry out the corresponding farm operation.

During operation, each piece of farming equipment would also feed back operational information to the system in real-time. Simultaneously, the system, taking into account the actual operational conditions, would make independent decisions, including considerations such as whether the current weather conditions are conducive to continued operation.

Apart from the preset farming operational tasks, the system could also analyze soil composition through the real-time data collected by the sensor network. This would allow it to make judgments about, for example, a certain plot’s fertilization needs, and to control the corresponding agricultural machinery to carry out the corresponding farm operation. The ultimate goal is to realize an entirely autonomous production process for unmanned farm planting, which requires no manual involvement.

By enhancing the intelligence and autonomy of UAV swarms, developing an integrated model for data analysis and decision-making in unmanned agriculture, and promoting an integrated system of smart agricultural technologies within the context of the IoT, we believe it is possible to significantly elevate the level of intelligence in unmanned agricultural operations, offering a fresh perspective for the future development of autonomous agriculture.