*4.1. Experimental Setting*

We generate four problem classes, varying the settings of the following two design factors: a geographical distribution of demands and a representative package delivery scenario.

For the geographical distribution of demands, we first create demand set U, where 100 demand nodes are uniformly distributed over a square-shaped 2D area. To see the impact of densely located demand nodes on the performance of the zoning approach, we also create demand set C, where demands are primarily located in the center of a service area. This demand set is created by sampling 100 demand nodes from the york dataset available from the r-package maxcovr. Figure 2 shows the two different demand node distributions.

**Figure 2.** The distributions of demands: (**a**) demand distribution U; (**b**) demand distribution C.

Next, we apply two different service modes to test the zoning approach under different service scenarios. For scenario NQ (no queue), we assume a situation where a service request, which cannot be served immediately by a responsible UAV, will leave the system without receiving service. This scenario represents an emergency situation, where a timely response to a demand is critical (e.g., visual surveillance of a traffic accident). For scenario FCFS, we assume a situation where demands wait until they receive service by UAVs. In this

scenario, demands in a queue are handled by the first come first serve (FCFS) policy. This scenario is implemented to represent a commercial package delivery scenario.

Following the setting, we generate the four different problem classes ({U, C}×{NQ, FCFS}) and test the performance of the zoning approach in a simulation environment. Note that the objective function of the zoning problem, that is, to minimize the within-zone variances, is a proxy for the actual UAV safety level. Therefore, a simulation model is implemented to investigate the performance of the proposed zoning approach close to its actual performance. In principle, the zoning approach can be applied to a UAV-based service system with any type of UAV. Since we examine the systemic performance of the zoning approach, the detailed dynamics of UAVs (e.g., the minimum turning radius of a unit, energy consumption as a function of weather conditions, and collision avoidance logic during operation) are simplified in the simulation. The simulation model generates demands following the demand rates of demand nodes and assigns available UAVs by the applied UAV deployment strategy, updating the operational status (busy/idle) of the UAVs.

For each problem class, we replicate a simulation run 20 times. The length of the simulation is four hours, and the status of UAVs and demands are updated every second in a simulation run. In the simulation, the demand rate of a demand node (the number of demand requests per minute) is drawn from a uniform distribution with the bounds [0.4, 0.6]. The number of UAVs is set by ten to provide a sufficient service capacity for the demands.

The output of a simulation run is the UAV flight trajectories for all time steps (every second), including the positions of UAVs and their status, and the records regarding how demands are served. Based on the output, the separation level between UAVs is analyzed to see how much the safety can be improved by the zoning approach, which is the main topic of this study (Section 4.2). The service quality with regard to the demands is also analyzed to evaluate the performance of the zoning approach from the customers' perspective (Section 4.3).
