Mobile-Edge Computing (MEC) has been become the most promising framework for meeting Quality of Service (QoS) requests, taking into account the requirements of processing cycles and energy. This is essential for computationally intensive mobile applications and the Internet of Things (IoT) to cope with capability limitations of mobile and IoT devices [1–3]. A cloud-based layer supports an edge network of servers that handles computationally intensive tasks, allowing for a reduction in the required energy and corresponding turnaround time. The close proximity of the edge servers to IoT and mobile devices provides low latency and a high bandwidth for delay-sensitive IoT applications [4 –6]. However, in some locations, the deployment of edge servers may be difficult, if not impossible [7–11]. As a real-life example, monitoring and observing forests and woodlands is a difficult task because such monitoring is time-consuming and requires considerable effort along with suitable resources. Stopping dangerous undertakings that can cause great damage to nature is a major challenge and responsibility of monitoring forestlands. Because forests cover such a wide area, it is difficult for forest managers and workers to take immediate action in cases of problems such as forest fires and illegal tree felling. Communication with ground sensors, aerial surveillance, mapping, aerial photography, and thermal imaging are some approaches in which Unmanned Aerial Vehicles (UAVs) can carry out forest monitoring. Thus, monitoring by UAVs contributes to the conservation of wildlife, biodiversity, and vegetation, the balance of ecology, and other forestry-related issues.

Edge servers are always privileged by their high bandwidths and low latency. Therefore, the performance of offloaded tasks is enhanced by MEC for restricted systems in terms of energy and time [24,25]. When UAVs are equipped with MEC servers and communicating channels, UAVs contribute to computing services and instant communications. This yields a better energy consumption and rate of transmission for mobile devices. However, the UAV’s energy limitations have an adverse effect on ground mobile equipment with regard to communication and computation [26]. To address these key challenges, a variety of research studies have been conducted to explore different configurations, optimization objectives, and underlying constraints.

In [27], optimum planning of the UAV trajectory, ratio of offloading tasks, and user scheduling were studied together to minimize all users’ delays. In [28], UCMEC was proposed as a novel framework of agent-based task offloading. With the aim of obtaining the optimal offloading plan, the intelligence agent was led to obtain a plan that was as efficient as possible. However, this study did not address UAV trajectory optimization. In [29], the authors demonstrated that partial offloading can be optimized through the optimization of offloading ratios, local computing frequencies, transmission power, and edge server computing frequencies. However, their framework did not consider UAV mobility. In [30], the authors proposed an MEC system involving multi-UAVs and ground IoT nodes to offload computational tasks that they could not handle using their limited capabilities. To balance the computational load of the UAVs, they employed a deep reinforcement learning approach. However, stable MEC clustering with greater computing power does not guarantee generation by this method. In [31], using the Dinkelbath algorithm and “successive convex approximation” (SCA) techniques, the authors solved a nonconvex optimization problem formulated in the UAV scenario. In this system, neither the level of distribution of mobile devices on the ground level nor the timing of tasks that are to be offloaded were considered. As part of resource allocation and joint computation offloading, task assignment to MECs while reducing turnaround time and required energy is an important issue. Reference [32] focused on the number and location of UAVs in the development of algorithms for deploying multi-UAVs in a FANET. Based on whether tasks are offloaded or processed locally, a greedy algorithm was proposed to determine the optimal solution. However, it is important to determine when and how to offload tasks from IoT devices, particularly if the tasks are uncertain in order to improve accuracy while reducing cloud communications costs. Using the UAV computing platform, Chen et al. [33] proposed an intelligent task offloading algorithm (iTOA) that can solve the computationally intensive problem of task processing. The offloading operation is performed using the deep “Monte Carlo Tree Search” algorithm (MCTS), which is the principal algorithm of Alpha Go. To arrive at the optimal decision, MCTS simulates the future trajectory of the unloading decision and optimizes the rewards based on this simulation. To be effective, the proposed offloading strategy requires training data, a prediction model of the channel state, and a period of self-learning. In [34], the UAV served as a supporting unit for helping ground devices offload their tasks to access points for further computations. They employed a greedy algorithm to optimize the required energy based on a suitable selection of UAV paths and allocation of resources. However, only the UAV velocity is taken into account in the optimization process. In [35], the relation between energy consumption and latency is considered for task offloading and for evaluating MEC system performance. Their approach used iterative optimization to determine the path of the UAV. However, this optimization does not take into account how clustered fixed slots should be partitioned or consider time delays. In [36], the UAV energy consumption was minimized by optimizing its trajectory and task scheduling. This will further shorten the operational time if some memory-intensive and computation-intensive tasks are executed on the UAV. In [37], greedy search optimization based on the uplink and downlink communications between UAVs and devices was utilized as a means of offloading and receiving data while ensuring energy-efficient operation of the system. In [38], a new mobile edge system integrating UAVs and IoT devices was developed. By launching a UAV, the system enables the provisioning of services to IoT devices by using wireless communication. Additionally, their approach formulates the problem as a nonconvex optimization problem by optimizing UAV position, resource allocation decisions, and task splitting decisions jointly to minimize the total delay and consumption of energy. However, it does not take task scheduling into account. In summary, the approach based on exhaustive and greedy searches is not suitable for such optimization. In fact, the majority of problems in consideration have many factors that must be optimized, such as the shortest path planning, ground user number and locations, energy consumption, and time delay parameters. Thus, most researchers deployed heuristic settings such as fixed partitions. However, their frameworks are not extensible due to their high complexity. In addition, learning-based approaches [39,40] request a preprocessing training step. Such a preprocessing step makes these approaches inapplicable to probabilistic problems.

Thus, the metaheuristic optimization framework is a suitable candidate for dealing with such stochastic problems. In this study, we employ a combination of a computational geometry algorithm, namely, the VD construction algorithm, and a natural phenomena simulation, namely, PSO, to carry out the required minimization of the time delay and energy consumption of the task offloading problem using UCMEC in a reasonable timeframe.

Now, we describe the employed time and energy models in this section. Table 2 lists the set of symbols and factors employed in the suggested task offloading model.

Two layers comprise an offloading system based on UAVs. A ground layer consists primarily of IoT and mobile devices that have N fixed-position wireless devices on the ground. Let D={d1,d2,...,dN} denote the set of ground devices. To maximize efficiency and reduce overall energy consumption, wireless devices are encouraged to offload their computations to a flying MEC server for faster calculations and minimum time delay. A single UAV in conjunction with an MEC form the second layer, namely, a flying MEC. This layer provides cloud-based computations for the first layer devices while ensuring a minimal response time.

For the UAV, an altitude of H>0 is maintained continuously. In addition, let Rc be the radius of communication coverage of the UAV. Several K regions are covered by the UAV with respect to a predefined path. Without loss of generality, let the path start at R₀ and end at R K , i.e., following the sequence R 0 ⋯ R K . At the end of the path cycle, the UAV returns to the starting position after hovering over each region R_k to process the tasks that have been offloaded to this region. On the other hand, the position of each ground device d_i is specified in advance by the Cartesian coordinates (x_i, y_i). In addition, the stopping position of the UAV over the region R_k is specified by the Cartesian coordinates u_k=(x_k, y_k). Therefore, by calculating the Euclidean distance, we can evaluate the distance between the UAV and each d_i as follows: (1) dik=(xi−xk)2+(yi−yk)2+H2

Provided that the UAV with device d_i is in the covered region, i.e., the following condition is satisfied:

(2) dik<Rc

While the UAV hovers over a region R_k, device d_i requests to offload its tasks ti=(tic,tiτ,tid), where tid, tiτ, and tic correspond to the data size, task delay, and needed computing cycles, respectively. The presented UCMEC framework divides the set of devices D into K clusters on each UAV to offload the requested tasks. Let δik=1 indicate that the ground device d_i is permitted to offload task t_i by the UAV while flying over R_k. The following subsections describe the suggested optimization approach according to the energy consumption, computational, and communicating models.

In this section, we describe our proposed optimization approach. First, we propose a partitioning of the set of wireless devices, D, using the VDs with respect to the UAV communication range Rc. Second, optimization is performed using the GPUPSO algorithm to obtain better clustering and the shortest path for the UAV. We propose the implementation of a customized PSO on a GPU. The customized PSO is as simple as the classical PSO but shows better performance. Our aim is to speed up the computation and introduce a better local communication topology between the particles within the population as part of the global search. The parallelization of the PSO on a GPU enables a better exploration of search spaces with a high number of dimensions using a large population swarm.

The experiments are performed on an NVIDIA CUDA Maxwell architecture embedded in an Intel Core i7-8565U CPU running at 1.8 and 1.99 GHz and a GeForce MX130 GPU with 4 GB RAM under Windows 10. Here, we present some experimental results that demonstrate the effectiveness of the proposed approach. A set of wireless devices is randomly distributed over the square region of 200×200m². Many studies have been devoted to determining the flying and energy parameters for UAVs [51,52]. Similar simulation parameters are presented in Table 3. First, the partitioning of the devices is performed on the CPU, and then the proposed optimization using the PSO is applied on the GPU. The results concerning the optimization step are the average of 10 independent runs. The inertia ω, cognitive C₁, and social C₂ parts are set as 0.75, 2.5, and 1.5, respectively. The aging factor for the inertia ω is set to 0.99 and applied every 10 iterations. We have set the number of iterations to 1000.

Fig. 7 shows a graphical illustration of the proposed approach. First, a set of 40 wireless devices is randomly generated, as shown in Fig. 7a. The VDs of these devices are created using the sweep-line algorithm [53,54], as shown in Fig. 7b. For each device, the corresponding Voronoi circles are determined, as shown in Fig. 7c, to apply the merging step. In the merging step, the Voronoi vertices are merged together, as described in Algorithm1 with respect to the UAV covering radius Rc, as shown in Fig. 7d. Once the clusters of devices are obtained, the proposed GPU-based particle swarm algorithm is applied to obtain the shortest path for the UAV, as shown in Fig. 7e.

Moreover, we compare the proposed optimization of our system vs. a GA. In fact, the GA is a metaheuristic optimization approach based on natural reproduction phenomena. The GA simulates the three main Darwinian operators of natural evolution, namely, selection of the fittest, crossover, and mutation operators. The algorithm starts by randomly initializing an initial population and selecting the fittest individuals for reproduction (applying the crossover and mutation) of the next generation offspring. The individual structure is called a chromosome. In our experiments, we use a chromosome structure similar to the proposed particle structure, as described in Section 4.2.3, and apply the classical GA operators. Fig. 8 shows a comparison of the proposed optimization and a GA. The population size is 800 particles (or individuals for GA). Fig. 8 shows that the proposed GPU-based PSO has a higher convergence rate than the GA.

The analysis of the computational time shows that the GPU architecture speeds up calculation, as presented in Table 4. The columns from left to right are the number of wireless devices, the Voronoi construction time, the partitioning time, and the optimization time. The number of iterations and swarm size are set to 1000 and 800, respectively. The timings presented in Table 4 show that the optimization process incurs the main bulk of the computational cost of the proposed approach. In addition, when the number of wireless devices increases, the dimensionality of the problem also increases, and the optimization step is the most strongly affected portion of the proposed approach. However, due to the high performance of the GPU architecture, the performance difference is negligible.

This paper proposed a UCMEC system to improve the performance of offloading tasks from mobile devices with the goal of minimizing the task delay and the energy consumption of the system. The proposed system partitions the ground devices into regions where the UAV can hover over each region to process the offloaded tasks. A VD is used for the partitioning process. The UAV trajectory over the regions is optimized using a GPU-based PSO. The performance of the proposed system was validated by comparison with other algorithms in the literature.

The limitation of the proposed offloading process is that it is valid only for stationary devices. In other words, it does not take into account the dynamic positioning of the ground sensors. In addition, the performance of the PSO in defining the optimal clustering and the shortest path of the UAV is very sensitive to the intrinsic parameters of the PSO, namely, the inertia, cognitive, and social parameters. In fact, this is a general drawback for all metaheuristic approaches. Our future work will consider the mobility of ground devices as well as the adaptive setting of the algorithm parameters.