arXiv CSopen access

A Vertical Look at UAV Connectivity in the Wild: Cellular vs. Starlink, 3D Characterization, and Performance Prediction

arXiv:2605.27755v2 [cs.NI] 28 May 2026

A Vertical Look at UAV Connectivity in the Wild: Cellular vs. Starlink, 3D Characterization, and Performance Prediction SRAVAN REDDY CHINTAREDDY, University of Kansas, USA SHERWAN JALAL ABDULLAH, University of Kansas, USA JUSTIN D. CLOUGH, University of Kansas, USA VICTOR S. FROST, University of Kansas, USA SHAWN KESHMIRI, University of Kansas, USA MORTEZA HASHEMI, University of Kansas, USA Reliable and resilient wireless connectivity is essential for unmanned aerial vehicles (UAVs) and advanced air mobility (AAM) applications in Beyond Visual Line of Sight (BVLOS) scenarios. Terrestrial cellular and emerging Low Earth Orbit (LEO) satellite networks promise enhanced performance for various use-cases. Yet, their comparative performance in the low-altitude airspace with complex 3D aerial propagation remains largely unexplored. In this paper, we present an open-source measurement platform designed to characterize the performance of commercial cellular (Verizon, a major US provider) and LEO satellite (Starlink) networks through real-world flight tests in rural environments. We implement a comprehensive multi-layer measurement approach spanning physical layer signal metrics, multi-cell network topology, and end-to-end (E2E) application performance. Through an extensive flight campaign with more than 10 flight tests, 4.5+ hours of flight time resulting in more than 18K samples, we present the first detailed, open-source dataset analyzing dual cellular and Starlink performance for low-altitude UAV operations. Our cellular-Starlink comparative results, which are collected simultaneously at the same time and location, demonstrate significant performance differences between the two technologies: the LEO satellite link achieves superior latency performance with 95% of Round-Trip Time (RTT) measurements below 50 ms compared to 80% under 150 ms for cellular, and exceptional downlink capacity with 95% exceeding 25 Mbps versus only 5 Mbps for cellular. Our analysis on cellular network performance demonstrates that while higher altitudes (e.g., 330+ m above the sea level) improve signal power by 15 − 20 dB via line-of-sight (LOS) propagation, it causes a 3 − 4 × increase in handover rates, which is due to excessive multi-cell visibility rather than signal degradation. Furthermore, we observe asymmetric impacts on the RTT performance due to handovers such that 53.5% of handovers improve RTT, but worst-case degradation (275 ms) is 2 × larger than best-case improvement (137 ms). Finally, to generalize our measurement capabilities, we develop a suite of machine learning (ML) models to extend the analysis beyond measured flight paths through spatial and altitude-based prediction of key performance indicators. The measurement platform combined with the ML-based predictions provides a comprehensive toolset for evaluating and optimizing BVLOS UAV connectivity in rural deployment scenarios. CCS Concepts: • Networks → Mobile networks; Network measurement; Network performance analysis; Wireless access points, base stations and infrastructure; • Computing methodologies → Machine learning.

1

Introduction

Unmanned aerial vehicles (UAVs) are rapidly transforming diverse sectors, including logistics, precision agriculture, infrastructure inspection, public safety, and real-time multimedia services [1, 2]. As these applications transition from isolated trials to large-scale deployments, they increasingly rely on reliable connectivity for telemetry, command-and-control (C2), and high-rate payload data. This evolution from standalone to connected UAVs elevates communication from a mere supporting role to a core enabler of safety, reliability, and economic viability. In particular, emerging BVLOS use Authors’ Contact Information: Sravan Reddy Chintareddy, [email protected], University of Kansas, Lawrence, Kansas, USA; Sherwan Jalal Abdullah, [email protected], University of Kansas, Lawrence, Kansas, USA; Justin D. Clough, [email protected], University of Kansas, Lawrence, Kansas, USA; Victor S. Frost, [email protected], University of Kansas, Lawrence, Kansas, USA; Shawn Keshmiri, [email protected], University of Kansas, Lawrence, Kansas, USA; Morteza Hashemi, [email protected], University of Kansas, Lawrence, Kansas, USA. , Vol. 1, No. 1, Article . Publication date: May 2026.

2

Sravan, et al.

cases require robust wide-area connectivity capable of supporting three-dimensional (3D) mobility, stringent latency and reliability demands, and extended mission durations [3–5]. Firstly, the integration of UAVs into existing cellular communication infrastructures has gained significant traction due to their wide-area coverage, mature deployment, and built-in mobility management [6–8]. The 3GPP Release 15 study on enhanced LTE support for aerial vehicles, along with subsequent work in 5G-NR [9, 10], has demonstrated that commercial networks can support UAV C2 and payload traffic. However, several studies (see, for example, [6, 11, 12]) have revealed significant challenges, including elevated interference, uplink congestion, and coverage variability at higher altitudes. Practical field evaluations further highlight the difficulty of meeting stringent reliability and latency requirements, such as 99.9% C2 reliability and sub-100 ms one-way delay under complex aerial propagation conditions and dynamic network loading [13]. These findings underscore the need for measurement-driven characterization of cellular network performance for aerial platforms and for data-driven models capable of predicting key radio access network (RAN) metrics, such as signal quality and throughput, across spatial and altitude domains. There have been numerous previous works on the topic of cellular-connected UAVs. Particularly, several experimental studies on cellular-connected UAVs consider mounting a cell phone on the UAV to measure the performance of the cellular network [12, 14–17]. However, these are limited in scope due to several reasons, such as lack of automatic flight control capabilities, limited radio access measurements, and time synchronization mismatch [14] across different subsystems of the measurement setup. Secondly, while cellular networks offer strong coverage in many populated and suburban regions, their availability and performance can degrade in remote, rural, or sparsely populated areas where many UAV missions, such as environmental monitoring, disaster assessment, and long-range logistics, are most compelling. In these settings, LEO satellite constellations (e.g., Starlink) provide an attractive complementary connectivity option, offering wide-area coverage and relatively low latency compared to traditional geostationary systems [18–20]. Integrating UAVs with satellite networks can enable resilient connectivity in areas with limited or no terrestrial infrastructure, extend BVLOS mission range, and provide redundancy for safety-critical operations [21–23]. There are several prior measurement campaigns to characterize the performance of Starlink (see, for example, [18, 24–28]) under different mobility patterns such as drive tests. However, the performance of satellite links in practical UAV scenarios, which are subject to antenna pointing constraints, beam misalignment, and fast mobility and maneuvers, remains largely unexplored. Finally, relying on a single connectivity substrate, either cellular or satellite, can leave UAV operations vulnerable to coverage holes, capacity limitations, or transient degradations [29, 30]. Dual connectivity, in which UAVs simultaneously utilize commercial LTE and LEO satellite links, offers a promising path toward resilient and flexible communication architectures, enabling failover, load balancing, and application-aware traffic steering across heterogeneous networks [31]. While a few prior studies have explored the integration of cellular and Starlink networks [23, 32], these efforts are predominantly restricted to ground-level scenarios. Currently, there is a lack of open, measurement-driven platforms that integrate both commercial cellular and Starlink connectivity on a UAV for simultaneous operation and monitoring. Comparative studies evaluating satellite versus cellular connectivity at matched spatial locations and altitudes remain limited. To address the aforementioned research gaps, the contributions of this paper are multi-fold. Firstly, we design, develop, and test a comprehensive UAV-based measurement platform with dual cellular and Starlink connectivity. This platform is integrated with our in-house automatic flight system (AFS), which enables extensive autonomous flight testing campaigns under realistic conditions. This platform simultaneously measures both cellular and Starlink networks performance, enabling detailed analysis of their performance and altitude-dependent trade-offs. To the best of our knowledge, no open-source UAV-based measurement platform currently provides synchronized, , Vol. 1, No. 1, Article . Publication date: May 2026.

A Vertical Look at UAV Connectivity in the Wild

3

Fig. 1. Overall system model and measurement scenario using UAV(s) that can hover above the ground level equipped with cellular and satellite modems and fly up to an altitude of 400 ft. UAV collects RAN metrics, E2E network measurements, and Starlink measurement data.

co-located Cellular and Starlink measurements during BVLOS flight. Secondly, we develop a comprehensive multi-layer measurement and data collection framework that operates across multiple layers of the protocol stack. At the physical layer, we collect spatiotemporal metrics such as RSRP, RSSI, RSRQ, and SINR, which collectively provide insight into cellular coverage, interference, and link quality as a function of altitude, position, and cell association. At the network layer, we capture information about serving and neighboring cells, handover behavior, and multi-cell network topology. Furthermore, we collect end-to-end (E2E) performance indicators, including latency, uplink/downlink throughput, and packet delivery statistics, from the UAV platform to our dedicated remote server. Our platform enables measuring and correlating all of these metrics for the cellular network and E2E performance metrics for the Starlink network at matched geolocations and altitudes. Our analysis results demonstrate several key findings on altitude-dependent LTE network performance, the relationship between multi-cell visibility and handover behavior as a function of altitude, and performance comparison of LTE vs. Starlink. Thirdly, to extend our analysis beyond measured trajectories, we leverage ML models to capture complex 3D propagation patterns and extrapolate performance to unmeasured altitudes. Therefore, our empirical performance measurement combined with ML-based prediction algorithms will ultimately provide a comprehensive framework for network performance characterization, coverage assessment, and safe BVLOS route/mission design. In summary, our contributions are as follows: • Dual-connectivity UAV measurement platform: We design and implement a UAV-based measurement platform that supports BVLOS operations and integrates both commercial LTE and Starlink connectivity, enabling simultaneous operation and coordinated measurement of heterogeneous wireless networks during real-world rural flights. To the best of our knowledge, this is the first UAV platform that integrates commercial cellular and Starlink connectivity for simultaneous operation, measurement, and comparative analysis of heterogeneous wireless networks during real-world flight. • Multi-layer cellular and Starlink measurement dataset: We develop custom software to capture multi-layer cellular metrics, including RAN-level parameters (RSRP, RSRQ, RSSI, SINR) for serving and neighboring cells, network-level topology information, and E2E performance , Vol. 1, No. 1, Article . Publication date: May 2026.

4

Sravan, et al.

metrics (latency, uplink, and downlink throughput), alongside Starlink performance measurements collected at the same geolocations and altitudes. This results in a comprehensive dataset collected from May 2024 to December 2025, comprising 10 flight tests with 18, 948 samples. The dataset includes 3.5 hours of LTE measurements across 7 flights and over 1 hour of Starlink E2E measurements across 4 flights, totaling approximately 4.5 hours of flight time. For application layer comparison, we have also collected synchronized LTE and Starlink E2E measurements across 2 flight tests (September 2025 and December 2025). • Comparative analysis of commercial cellular and Starlink connectivity for UAVs: By leveraging synchronized measurements at matched spatial locations, we provide a comparative evaluation of cellular and Starlink performance for UAVs, quantifying their respective coverage, latency, and throughput characteristics, and identifying regimes where one technology offers clear advantages or where dual connectivity can provide robustness benefits. • ML-based prediction of key RAN metrics in 3D space: To extend the analysis beyond the measured flight paths, we train ML models on the collected dataset to estimate key RAN metrics in both measured and unmeasured spatial and altitude regions, under both random-split and leave-one-altitude-out validation. These models enable the prediction of network performance in unobserved areas, supporting coverage planning, risk assessment, and route optimization for future UAV missions. • Open-source testbed design, software modules, and measurement data: We provide all platform designs, custom software implementations, and analysis tools as open-source resources to enable reproducible research, facilitate collaborative development, and provide broader access to advanced measurement capabilities for rural network evaluation. To maintain anonymity, references to our open-source repository (which contains accurate geographical information) are omitted. The complete dataset and codebase will be made publicly available in the final manuscript upon acceptance. The rest of this paper is organized as follows. Section 2 presents the platform architecture and measurement methodology, followed by spatial and temporal performance analysis of LTE network in Section 3. Section 4 presents our multi-cell measurement results and Section 5 provides comparison results between LTE and Starlink networks. In Section 6 we present our ML-based prediction results followed by concluding remarks and future works in Section 7. 2

Platform Architecture and Measurement Methodology

To bridge the gap between theoretical/simulation-based network performance characterization and real-world scenarios, we develop a modular and open-source measurement platform designed specifically for hybrid terrestrial-satellite network characterization. Commercial measurement tools such as Rohde & Schwarz’s Qualipoc [33], TEMS Pocket [34], and Keysight’s NEMO [35] (which are used for aerial radio measurements [15, 36, 37]) or software defined radio Fig. 2. A layered and modular design for (SDR) configurations (e.g., [38, 39]) are either proprietary integrating dual connectivity technolo(closed-source) or often lack full protocol stack measure- gies with onboard avionics and software ments capabilities. logging systems. To address this gap, our proposed UAV-based platform integrates automatic flight controller, onboard avionics and compute systems, commercial-grade cellular modem, and Starlink terminal with a synchronized data logging framework. Figure 2 shows the layered and modular design for our UAV platform, which consists of three distinct layers: (i) , Vol. 1, No. 1, Article . Publication date: May 2026.

A Vertical Look at UAV Connectivity in the Wild

5

radio layer and RF subsystems, (ii) onboard avionics, control, and computing subsystems, and (iii) software modules and data logging subsystems. In this section, we present the details of our platform architecture and measurement methodology. 2.1

Platform Design

2.1.1 Avionics Subsystems. For comprehensive and scalable network measurements, we designed and assembled large fixed-wing and vertical landing and take-off (VTOL) UAVs (> 12 lb) (shown in Figures 1, 3, and 4) for which the avionics suites are fully integrated with commercial LTE modem and Starlink Mini terminal. The VTOL platform combines fixed-wing and rotary-wing (e.g., hexacopters or quadcopters) configurations, which allow the UAV to switch from VTOL mode to fixed-wing mode during cruise flight. This dual-mode capability Fig. 3. UAV platform inteprovides enhanced payload capacity, extended range and endurance, grated with LTE modem. while allowing vertical takeoff and landing operations even from smaller fields. Our UAV platforms are equipped with an in-house Autonomous Flight System (AFS) that features a modular and adaptable architecture built on the Robot Operating System 2 (ROS 2) framework. The AFS has been successfully integrated into multiple UAV platforms with varying payload capacities and power constraints, enabling flexible mission configurations and scalable measurements. The AFS is specifically designed for advanced aerospace applications running multiple computationally intensive algorithms in real-time. As shown in Figure 1, the onboard avionics centers on the NVIDIA Jetson Orin 16GB [40] as the primary processing unit. In Subsection 2.2.2, we present further details on how the Jetson Orin interacts with other components, including the LTE modem and Starlink terminal. 2.1.2 Radio and RF Subsystem. To characterize the performance of terrestrial and satellite networks, we integrate a commercial-grade cellular modem and a Starlink Mini terminal with our UAV platform. Cellular Modem: For cellular network measurement, we use the Microhard pMLTE [41] modem, which provides complete access to serving and neighboring cell RAN parameters through AT command interfaces. The cellular modem provides compatibility with globally deployed LTE networks and 3G/HSPA fallback capability. The module supports LTE FDD with a maximum 150 Mbps downlink/50 Mbps uplink and LTE TDD with a maximum 130 Mbps downlink/30 Mbps uplink. Figure 3 depicts the pMLTE module, its custom 3D-printed enclosure for airframe integration, and the fully instrumented UAV in flight. Starlink Mini. The Starlink Mini is an integrated phased-array satellite terminal combining the antenna subsystem and IEEE 802.11ac WiFi router in a compact form factor. The terminal utilizes electronically steered beamforming that dynamically adjusts phase relationships to Fig. 4. Our UAV platform track multiple LEO satellites at approximately 550 km altitude. The satel- integrated with a Starlink lite interface operates in dual-band frequency allocation: Ku-band with Mini terminal, shown at 10.7-12.7 GHz downlink and 14.0-14.5 GHz uplink (primary), and Ka-band ground level and during at 17.8-20.2 GHz downlink and 27.5-30.0 GHz uplink. The system is ex- flight testing. pected to achieve throughput of 50 - 100 Mbps with 25 - 60 ms round-trip latency through the LEO constellation to terrestrial gateway infrastructure. Due to its lightweight , Vol. 1, No. 1, Article . Publication date: May 2026.

6

Sravan, et al.

(a) RTT. (b) Uplink and downlink throughput. Fig. 5. Ground-based orientation sensitivity analysis for Starlink end-to-end metrics vs time.

form factor and low power consumption, the terminal is well-suited for UAV platform integration and extended aerial operations. Figure 4 shows the integration of the Starlink Mini with the UAV airframe during (a) ground-based configuration and (b) active flight testing (discussed in Subsection 5.2). Ground-based Sensitivity Analysis. To evaluate the performance of the Starlink Mini terminal for UAV applications, we conducted a ground-based orientation sensitivity analysis. The groundbased analysis evaluated the sensitivity of satellite connectivity to terminal orientation, with the primary objective of identifying potential service degradation scenarios under various spatial configurations that simulate realistic UAV flight altitudes and determining the optimal installation configuration for airborne deployment. Understanding orientation sensitivity is critical for UAV applications, as aircraft undergo pitch, roll, and yaw maneuvers during flight operations that could affect antenna-satellite alignment. The experimental evaluation was performed at an outdoor location with an unobstructed view of the sky. We varied the Starlink Mini’s spatial position using a spherical coordinate system. The azimuthal angle 𝜙 represents rotational orientation around the vertical axis, with 𝜙 = 0◦ aligned with geographic north. Four azimuthal positions were tested: 𝜙 = 0◦ (north), 𝜙 = 90◦ (east), 𝜙 = 180◦ (south), and 𝜙 = 270◦ (west). The tilt angle 𝜃 simulates UAV pitch and roll maneuvers, with 𝜃 = 0◦ representing the terminal facing directly upward. We evaluated tilt angles of 30◦ and 60◦ relative to the vertical axis. The ground-based results, shown in Figure 5, demonstrate that most tested orientations provided acceptable performance, with RTT values ranging between 40 ms and 200 ms, and downlink throughput maintaining 20-100 Mbps. However, critical connection degradation occurred during specific misalignments, particularly at a 60◦ tilt facing south (𝜃 = 60◦, 𝜙 = 180◦ ), where RTT exceeded 800 ms and throughput dropped to near-zero levels due to severe antenna-satellite misalignment. Based on the ground-based findings, the Starlink Mini terminal was mounted on the UAV platform with its surface facing directly upward (𝜃 = 0◦ ) to minimize orientation-induced performance degradation during flight maneuvers. 2.2

System and Software Architecture

2.2.1 System Architecture. We implement a comprehensive measurement architecture that coordinates onboard data collection, real-time data transfer, and ground-based analysis framework. As shown in the system architecture in Figure 1, a remote application server (e.g., a remote pilot) communicates with the airborne UAV platform. In order to support both real-time monitoring and E2E network performance evaluation, the server performs multiple critical functions, including serving as a fixed measurement endpoint for latency and throughput testing, receiving real-time data transmissions during flight operations, and enabling preliminary analysis while measurements are being , Vol. 1, No. 1, Article . Publication date: May 2026.

A Vertical Look at UAV Connectivity in the Wild

7

collected. The fixed endpoint approach eliminates measurement variability introduced by automated server selection in consumer speed testing applications (e.g., Speedtest), enabling consistent comparative analysis across different flight tests and deployment scenarios. The synchronized data collection approach enables various analytical methods, including spatial interpolation to estimate coverage in unmeasured areas, time-series analysis to identify temporal patterns, and correlation studies to understand relationships between position, altitude, and network performance. 2.2.2 Communication Software Architecture. To define and create data packets, we use MAVLink [42], a “library for lightweight communication,” to encapsulate the data being transmitted. MAVLink is an open-source protocol used for communicating with UAVs, and between onboard UAV components (e.g., between flight controller and Orin in Figure 1). MAVLink is designed to be platform-independent and can be extended with new message types without requiring changes Fig. 6. Internal communication flow of to the current codebase. MAVLink supports a variety of mes- the NVIDIA Jetson Orin with the pMLTE sage types, including status, control, and mission messages. and Starlink Mini modules. Once the data is formatted into a MAVLink message, it is transmitted via a serial or Ethernet connection to the pMLTE module or Starlink Mini, as shown in Figure 6. The data is then transmitted over the wireless network and received by the remote server. In the opposite direction, upon successful arrival of a message, the MAVLink message is captured by onboard Intel’s MAVLink Router (IMR). The IMR scans the incoming data for a valid MAVLink message and once it finds one, it will transmit the packet via UDP to any endpoint specified. The only current endpoint given is a ROS node whose job is to unpack the MAVLink messages and extract the desired information. It then publishes that information to the rest of the ROS system via ROS messages. This ROS node also subscribes to ROS topics that give information about the current location and attitude of the UAV. These subscriptions are spawned in separate threads from the main execution loop of the node and will pull messages into a structure as soon as the data is available. The main execution thread is started in a separate thread within the ROS node and its job is to receive incoming messages and send outgoing messages. 2.2.3 Graphical User Interface. To facilitate efficient real-time monitoring and post-flight analysis, we developed a web-based Graphical User Interface (GUI) analysis tool, as shown in Figure 1. This tool streamlines the data analysis workflow by providing an integrated platform for log file processing, data extraction, and results visualization. The interface includes file browsing functionality that allows users to select measurement log files, data extraction routines that parse the collected metrics, and integrated visualization modules that generate both spatial coverage maps and statistical analysis charts. The results are displayed directly within the web interface, providing immediate feedback to the remote pilot during flight operations and enabling rapid iterative analysis during post-flight evaluation. 2.3

Multi-Layer Measurement Methodology

2.3.1 Metrics. We implement a multi-layer measurement approach capturing network performance metrics across three distinct layers: physical layer, network layer, and application layer. We have developed a custom Python-based data logger application deployed on the NVIDIA Jetson to enable automated, high-resolution data collection throughout flight operations. The application operates at a 1-second sampling interval, providing sufficient resolution to capture signal dynamics during aerial movement. At the physical layer, the data logger continuously retrieves RAN parameters for both the serving cell and up to three neighboring cells through telnet , Vol. 1, No. 1, Article . Publication date: May 2026.

8

Sravan, et al.

communication. Collected metrics include RSRP, RSRQ, RSSI, and SINR for each tracked cell. This comprehensive serving and neighboring cell monitoring capability operates continuously without manual intervention throughout flight operations, enabling interference analysis and handover assessment. At the network layer, the platform captures network topology identifiers, including Physical Cell ID, Cell ID, and Location Area Code. These identifiers enable spatial coverage mapping and cell-specific performance analysis, distinguishing the platform from single-cell measurement approaches. At the application layer, the platform integrates end-to-end performance evaluation capabilities. The data logger automatically executes Nping [43] for packet round-trip time (RTT) measurement and iPerf3 [44] for bidirectional throughput testing, both connecting to the dedicated remote server infrastructure. Additionally, packet delivery statistics are collected to assess link reliability. This automated testing framework runs periodically throughout flight operations, capturing latency, throughput, and packet delivery characteristics synchronized with RAN metrics and position data. The fixed remote server endpoint ensures consistent and comparable results across different measurements. Furthermore, we collect all UAV flight information (altitude, direction, speed). All measurement data are captured with GPS coordinates (latitude, longitude, altitude) and high-precision timestamps synchronized via Network Time Protocol, creating a spatiotemporal dataset where every network metric is linked to both physical location and time. 2.3.2 Flight Testing and Measurement Area. Platform validation was conducted in a semi-rural environment in the U.S., as shown in Figure 7. This test area was selected for its typical rural characteristics including limited cell tower installations with sparse infrastructure deployment and varied terrain with elevation changes affecting propagation. These conditions provide an appropriate environment to study real-world flight operations and validate the platform’s three-dimensional measurement capability in rural deployment scenarios. The platform operated on Verizon’s commercial LTE network (maximum 20 MHz Fig. 7. Spatial distribution of cell sites channel bandwidth) and Starlink LEO satellite, ensuring around the test area. Each site contains mulmeasurements reflect real-world operational conditions tiple sectors with unique Cell IDs. Red sites with realistic interference, load, and network manage- contribute to coverage in the test area, while ment characteristics absent from controlled testbed envi- black sites are non-contributing. ronments. Measurements were collected across altitudes from 240 to 400 meters above sea level (ASL) (corresponding to 0 − 160 meters above the ground level (AGL) in our test area), with data processed using 10-meter altitude bins to extract altitudedependent performance trends. This vertical sampling range, combined with horizontal spatial coverage, demonstrates the platform’s three-dimensional characterization capability while remaining within Class G airspace regulations for UAV operations [45]. Given the presented multi-layer measurement framework, we focus on three key sets of results: (1) three-dimensional performance analysis for cellular networks across various altitudes (Section 3), (2) multi-cell performance characterization via simultaneous tracking of serving and neighboring cells performance (Section 4), and (3) comparative end-to-end performance evaluation between terrestrial cellular (LTE) and Starlink satellite connectivity (Section 5). 3

Temporal and Spatial Characterization of Cellular Network Performance

In this section, we present our analysis results based on spatiotemporal dataset captured by the platform. First, we quantify the baseline service quality through statistical distribution analysis , Vol. 1, No. 1, Article . Publication date: May 2026.

A Vertical Look at UAV Connectivity in the Wild

(a) CDF of RSRP

9

(b) CDF of RSRQ

(c) CDF of RSSI (d) CDF of SINR Fig. 8. CDFs of LTE RAN metrics across five different flight tests denoted by “LTE-FLT-1” through “LTE-FLT-5”.

of key RAN metrics (RSRP, RSRQ, RSSI, and SINR). Second, we investigate three-dimensional performance to characterize altitude-dependent behaviors and empirically validate the trade-off between improved line-of-sight signal (LOS) strength and elevated interference at higher altitudes. 3.1

RAN Performance Analysis

In this part, we use the Cumulative Distribution Function (CDF) to characterize the distribution of RAN performance metrics across the flight testing area. As shown in Figure 8, CDF analysis was conducted for five flight tests (denoted by “FLT”) between May 2024 and December 2025, examining RSRP, RSRQ, RSSI, and SINR distributions. The plots present individual flight CDFs alongside mean trends with confidence bands. Since standardized thresholds for quality classification vary across manufacturers and application requirements, we employ the criteria proposed by Teltonika Networks [46, 47] for consistent performance categorization. The following thresholds define poor conditions: RSRP below -100 dBm for signal power, RSRQ below -20 dB for signal quality, RSSI below -95 dBm for channel power, and SINR below 0 dB for interference immunity. Figure 8a shows RSRP distributions across five flight tests, with the percentage of measurements below the -100 dBm threshold varying significantly from 20% to 67% across different flights. The mean CDF demonstrates consistent coverage trends despite these individual flight variations. Signal quality measurements in Figure 8b reveal excellent performance, with 0% of measurements falling below the -20 dB threshold across all five flights. Channel power conditions in Figure 8c show 100% of measurements well above the -95 dBm threshold, indicating uniformly excellent coverage. This high channel power suggests extensive signal overshooting from elevated cell sites. Figure 8d reveals that 100% of measurements exceed the 0 dB threshold, demonstrating good interference immunity across all flights. The shaded confidence regions indicate measurement variability across the flight tests, with RSRP showing the most variation while RSRQ, RSSI, and SINR remain consistently , Vol. 1, No. 1, Article . Publication date: May 2026.

10

Sravan, et al.

(a) 3D RSRP map.

(b) 3D RSSI map.

(c) 3D RSRQ map. (d) 3D SINR map. Fig. 9. LTE RAN metrics distribution in 3D maps.

strong. Overall performance comparison across all four metrics reveals that LTE-FLT-4 achieved the best results with the highest RSRQ (∼−14 dB) and SINR (∼ 9 dB) while maintaining competitive RSRP and RSSI values. LTE-FLT-3 exhibited the strongest raw signal metrics with the best RSRP (∼ -93 dBm) and RSSI (∼ -56 dBm), though with slightly lower SINR (∼ 7.5 dB). 3.2

3D Spatial Performance Analysis

The platform’s ability to collect GPS-synchronized measurements at multiple altitudes enables three-dimensional coverage mapping. The platform collected measurements from 240 to 400 meters ASL, with data processed using 10-meter altitude bins to extract statistical performance trends. This vertical sampling capability is impossible with ground-based drive testing and provides essential insights for understanding aerial cellular connectivity and coverage prediction for elevated devices. The system generates geospatial visualizations in 3D format, illustrating the distribution and behavior of key performance metrics across horizontal position and vertical altitude dimensions. Figure 9 demonstrates the platform’s 3D mapping capability through spatial visualization of RSRP, RSSI, RSRQ, and SINR distributions across the test environment. The RSRP and RSSI visualizations reveal coverage weak spots at lower altitudes corresponding to terrain obstruction, demonstrating acceptable signal strength at higher elevations but degradation near ground level in specific regions. The RSRQ distribution exhibits spatial variability with improved quality in certain areas, indicating varying levels of interference and channel quality across the coverage area. The SINR results show , Vol. 1, No. 1, Article . Publication date: May 2026.

A Vertical Look at UAV Connectivity in the Wild

(a) RSRP vs. altitude.

11

(b) RSRQ vs. altitude.

(c) RSSI vs. altitude. (d) SINR vs. altitude. Fig. 10. LTE RAN metrics as a function of altitude. As altitude increases, the RSRP values improve due to enhanced LOS conditions; however, this leads to increased interference from neighboring cells.

fair to poor values with altitude-dependent variations, reflecting the complex interplay between signal strength and interference in the rural environment. Furthermore, Figure 10 demonstrates the platform’s ability to reveal altitude-dependent signal strength trends. The RSRP measurements show progressive improvement with increasing altitude, attributed to enhanced LOS propagation conditions between the UAV and the base stations. At lower altitudes, terrestrial obstacles, including terrain features and vegetation, cause significant signal attenuation. As the platform ascends, the probability of achieving direct LOS increases, reducing path loss. This characterization capability enables network operators to understand vertical coverage profiles and predict performance for aerial users. The platform’s multi-altitude measurement capability reveals a critical trade-off for aerial cellular connectivity: while signal power improves with altitude, signal quality degrades due to increased multi-cell interference. At elevated positions, the UAV gains unobstructed visibility to multiple base stations simultaneously, including cells beyond the intended serving cell. This fundamental altitude-dependent behavior, where improved LOS conditions simultaneously enhance desired signal reception and interference from neighboring cells, can only be characterized through vertical measurements that the platform enables. These results provide comprehensive insights for identifying coverage gaps, interference patterns, and altitude-dependent signal quality variations in three-dimensional propagation environments. 4

Multi-Cell Performance Analysis

Aggregate signal strength is a primary indicator of connectivity quality in cellular networks. However, reliable BVLOS operations rely fundamentally on network redundancy and the ability to execute seamless handovers during flight time. In this section, we extend our analysis from the serving cell’s isolation to a broader multi-cell perspective. We first investigate the spatial distribution of serving cells in Section 4.1, and demonstrate that it is possible a “dominant” cell could provide the majority of aerial coverage despite its suboptimal radio performance. Next, in Section 4.2, we quantify the network performance by evaluating the availability and signal quality of neighboring cells to assess handover readiness. We present a detailed analysis on the handover behavior as a function of RAN metrics, flight altitude, and impacts on RTT performance. , Vol. 1, No. 1, Article . Publication date: May 2026.

12

Sravan, et al.

(a) Serving cells RSRP.

(b) Serving cells RSRQ.

(c) Serving cells RSSI.

(d) Serving cells SINR.

Fig. 11. LTE RAN metrics for different serving cells. In total, signals from 14 distinct cells were received in our testing area shown in Figure 7.

4.1

Multi-Cell Coverage Dominance and Spatial Distribution Analysis

While typical RAN metrics such as RSRP, RSRQ, RSSI, and SINR indicate signal quality, they alone cannot determine network performance without considering each cell’s spatial coverage contribution. A cell may exhibit excellent radio metrics but serve only a small fraction of the area, minimizing its impact on user experience. Conversely, a cell with suboptimal parameters that dominates the coverage area significantly affects overall service quality. To investigate these behaviors, we capture the normalized Cell ID occurrence frequency across all measurement samples collected from the measurement area (Figure 7). Results in Figure 12 show that Cell 8 dominates the coverage area with approximately 45% of measurement samples, followed by Cell 1 with approximately 30%, and Cell 2 with approximately 10%. The remaining cells (Cells 3, 7, 13 and 14) each contribute less than 5%. However, we observe interesting behaviors when contrasting this coverage dominance with the RAN performance metrics shown in Figure 11. In particular, Cell 8, serving approximately 45% of the test area, exhibits poor signal strength with mean RSRP around −100 dBm and mean RSSI around −65 dBm, but shows moderate RSRQ (∼ −16 dB) and fair SINR (∼ 8 dB). Cell 1, accounting for 30% of samples, demonstrates moderate RSRP (∼ −93 dBm) but maintains good RSRQ (∼ −14 dB) and SINR (∼ 7 dB). Together, these two cells serve approximately 75% of the test area that directly defines the network quality experienced by the majority of aerial users in the area. Cell Fig. 12. Normalized number of samples 2, contributing 10% of coverage, exhibits good performance received from each cell. across all metrics (RSRP ∼ −97 dBm, RSRQ ∼ −16 dB, SINR ∼ 9 dB) but due to limited spatial presence, does not contribute to overall service quality. The remaining cells, despite their individual metric variations, have minimal practical impact due to serving less than 5% of the coverage area each. , Vol. 1, No. 1, Article . Publication date: May 2026.

A Vertical Look at UAV Connectivity in the Wild

13

(a) Neighbor cells RSRP. (b) Neighbor cells RSRQ. Fig. 13. RSRP and RSRQ performance for the dominant cell (Cell 8) and three neighbor cells (denoted by “NB1” through “NB3”).

Spatial analysis using CellMapper [48], shown in Figure 7, confirms that most detected cells are appropriately positioned for test area coverage. Cells 13 and 14, installed on high-elevation terrain (confirmed via Google Earth analysis), reach the test area through signal overshooting despite contributing less than 5% coverage and exhibiting poor signal strength (average RSRP ∼-108 dBm). The site corresponding to Cells 1, 2, and 3 shows variable coverage contribution influenced by antenna orientation patterns, as verified through CellMapper directional data. For real-world rural UAV applications in this deployment area, network optimization efforts must prioritize Cells 8 and 1, as improvements to their performance yield the greatest impact on overall service quality by addressing the 75% of the coverage area they collectively serve. 4.2 Neighboring Cell Performance Analysis and Handover Assessment 4.2.1 Handover readiness. Maintaining service continuity during user mobility requires reliable handover mechanisms to neighboring cells when the serving cell signal degrades. The platform’s multi-cell monitoring capability enables simultaneous measurement of neighboring cell performance, which is essential for evaluating network reliability, coverage redundancy, and the system’s capability to maintain seamless connectivity across the deployment area. Our data logging platform continuously captures radio measurements from neighboring cells. For this analysis, we consider the three strongest neighboring cells, which constitute critical handover candidates during serving cell signal degradation events to prevent service interruption. Figure 13 presents the RSRP and RSRQ performance metrics for the dominant cell (index 8) along with three neighboring cells (denoted by “NB” in the figures). The experimental results confirm that good handover candidates are available, with at least one neighboring cell maintaining signal levels above threshold requirements (RSRP ≥ −100 dBm, RSRQ ≥ −20 dB) [49, 50]. This coverage overlap ensures seamless handover execution and mitigates the risk of service interruption during mobility events. 4.2.2 Handover Analysis. We use Cell ID information for handover detection, which provides globally unique cell identification compared to Physical Cell IDs (PCIs) that can be reused across the network. Our algorithm processes time-series LTE measurements to simultaneously detect handovers (via Cell ID changes) and classify their trigger causes based on signal quality at the handover instant. Based on 3GPP standard [49], Algorithm 1 categorizes each detected handover into one of four event types based on threshold-based decision logic: (i) Handover Event 1 (E1) happens when neighbor cell RSRP exceeds serving cell by ≥3 dB. This event corresponds to Event A3 (Strength-based) in the context of 3GPP. (ii) Handover Event 2 (E2) corresponds to Interference (Quality-based) when RSRP is strong (>-95 dBm) but RSRQ is poor (≤-18 dB). (iii) Handover Event 3 (E3) is triggered when serving cell RSRP falls below -110 dBm. This event corresponds to Event , Vol. 1, No. 1, Article . Publication date: May 2026.

14

Sravan, et al.

Algorithm 1 Combined Handover Detection and Classification Input: Time-series: {(CellID𝑖 , RSRP𝑖 , RSRQ𝑖 , RSRP𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑟,𝑖 , 𝑡𝑖 )}𝑛𝑖=1 Output: Set of classified handover events H 1: H ← ∅ 2: for 𝑖 = 2 to 𝑛 do 3: if CellID𝑖 ≠ CellID𝑖 −1 and CellID𝑖 ≠ null then 4: ΔRSRP ← RSRP𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑟,𝑖 −1 − RSRP𝑖 −1 5: if ΔRSRP ≥ 3 dB then 6: 𝑐𝑎𝑢𝑠𝑒 ← E1 ⊲ Corresponding to Event A3 in 3GPP. 7: else if RSRP𝑖 −1 > −95 dBm and RSRQ𝑖 −1 ≤ −18 dB then 8: 𝑐𝑎𝑢𝑠𝑒 ← E2 ⊲ Capturing increased interferences. 9: else if RSRP𝑖 −1 ≤ −110 dBm then 10: 𝑐𝑎𝑢𝑠𝑒 ← E3 ⊲ Corresponding to Event A2 in 3GPP. 11: else 12: 𝑐𝑎𝑢𝑠𝑒 ← E4 ⊲ Handover due to other events. 13: end if 14: H ← H ∪ {(𝑡𝑖 , CellID𝑖 −1, CellID𝑖 , 𝑐𝑎𝑢𝑠𝑒)} 15: end if 16: end for 17: return H

A2 (Coverage-based) in the context of 3GPP. (iv) Handover Event 4 (E4) captures other handovers from network optimization or load balancing. The primary limitation is potential missed handovers during brief radio link failures, though our dataset shows <0.1% null values ensuring high detection accuracy. This method detected and classified 217 handover events during a 60 minutes flight, which are distributed as follows: E1/Event A3 (121, 55.8%), E2/Interference (26, 12.0%), E3/Event A2 (66, 30.4%), and E4/Other (4, 1.8%).

Fig. 14. Handover rate per 10-minute blocks with normalized RSRP, RSRQ, and UAV altitude.

Temporal Handover Dynamics. Figure 14 presents the temporal evolution of handover events over the 60 minutes flight, with handovers aggregated into 10-minute bins and overlaid with normalized signal quality metrics (RSRP, RSRQ) as well as flight altitude. The first 30 minutes (0 30 min, 250 m ASL) show relatively low handover rates of 16 - 30 events per bin, with E3 (Event A2, coverage-based) handovers dominating due to weak RSRP (-95 to -100 dBm) characteristic of groundlevel operations at cell edge. A significant increase in the number of handovers begins around 30 minutes as the UAV ascends: handover rate surges to 65 - 75 events per 10-minute bin during the 30 - 50 minute period at 330 - 380 m ASL, coinciding with RSRP improvement to -81 to -85 dBm, the strongest signals recorded during the entire mission. This paradox reveals that stronger signal strength does not reduce handover frequency in aerial environments. Critically, E2 (Interference) , Vol. 1, No. 1, Article . Publication date: May 2026.

A Vertical Look at UAV Connectivity in the Wild

15

handovers emerge exclusively during this 30 - 50 minute high-altitude window, comprising 19 22% of handover events as RSRQ degrades to -16 to -17 dB despite strong RSRP, indicating high cochannel interference from multiple visible cells. E1 (Event A3) handovers dominate throughout (60 70% of events during the 30 - 50 min high-altitude phase), with the 3 dB threshold triggering frequent handover ping-pong as uniformly strong RSRP across multiple cells causes minor signal fluctuations to continuously meet the A3 criteria. The final 10-minute bin (50 - 60 min) shows handover rate returning to baseline (10 events) as altitude decreases to 250 m ASL, with E3 handovers re-emerging as the primary type. This temporal analysis demonstrates that altitude, rather than signal strength, is the primary driver of handover instability, with the 30 - 50 minute high-altitude period exhibiting 3 - 4× higher handover rate than ground operations despite 15 - 20 dB stronger RSRP. Cell Visibility and Network Reporting. Figure 15 examines the relationship between unique cell visibility and handover rate across four flight phases: Lift-off (0 30 min, ground level at 250 m ASL), Transition (30 - 35 min, rapid ascent from 250 - 330 m ASL), Ascent (35 - 50 min, sustained flight at 330 - 380 m ASL), and Descent (50 - 66 min, return to 250 m ASL). During the Lift-off phase, the UAV detects 7 unique Cell IDs with a handover rate of 2.2 HO/min, limited by terrain obstructions and typical ground-level propagation. The Transition phase (converting to fixed-wing), marking the beginning of vertical ascent, experiences a 3× surge in handover rate Fig. 15. Handover rate change and cell visibility across flight phases. to 6.6 HO/min with a slight increase in cell visibility (9 Cell IDs), demonstrating that rapid altitude change velocity alone drives handover instability through dramatic RSRP fluctuations. The Ascent phase reveals the most striking finding: cell visibility increases 33% to 12 unique Cell IDs due to LOS propagation, enabling detection of cells 5 - 10 km away, yet handover rate increases disproportionately by 223% to 7.1 HO/min, compared to ground level. The increased number of visible cells could have led to suboptimal and frequent handover decisions, which may not necessarily be due to performance degradation received from the serving cell [51–53]. The Descent phase shows handover rate dropping to 0.6 HO/min with 6 visible Cell IDs as the UAV returns to ground level, confirming that both altitude and vertical velocity are key factors in handover behavior. RTT Impact of Handover Events. Figure 16 overlays individual handover RTT impacts on the UAV altitude trajectory, revealing that handovers do not uniformly degrade latency as commonly assumed. Analysis of all 217 handovers shows that 116 events (53.5%) decreased RTT while 101 events (46.5%) increased RTT, with a maximum RTT increase of 275.4 ms and maximum RTT decrease of 136.6 ms. During the first 30 minutes at ground Fig. 16. Changes in RTT at Handover (HO) events overlaid on UAV altitude trajectory. level (250 m ASL), RTT changes are moderate and mixed, with roughly equal numbers of improvements and degradations typically ranging ±50 ms. The most significant RTT instability occurs during the 30 - 50 minute high-altitude phase (330 - 380 m ASL), where both the largest RTT increases (100 - 275 ms spikes) and substantial decreases (50 - 100 ms improvements) are concentrated. This period, which coincides with peak handover frequency (7.1 HO/min), strongest RSRP (-81 to -85 dBm), and degraded RSRQ (-16 to , Vol. 1, No. 1, Article . Publication date: May 2026.

16

Sravan, et al.

(a) RTT. (b) Downlink Throughput. (c) Uplink Throughput. Fig. 17. End-to-end network performance measurement. The LTE and Starlink (denoted by “SL”) performance were measured simultaneously at identical times and locations (between UAV and a dedicated remote server). The experiments were repeated two times in September and December 2025.

-17 dB), demonstrates that strong signal strength alone cannot prevent latency degradation. The RTT impact exhibits critical asymmetry: while 53.5% of handovers improve latency (often by moving the UE from congested or poorly-backhauled cells to better network conditions), worst-case degradation (275.4 ms) is 2.0 × larger than best-case improvement (136.6 ms), creating a high-risk profile where handovers either help modestly (typical decrease: 20 - 50 ms) or hurt severely (outliers exceeding 100 - 200 ms). The descent phase (50 - 60 min, returning to 250 m ASL) shows sparse RTT events reflecting the low handover rate (0.6 HO/min), with mixed but moderate impacts and no extreme outliers. At the sustained 7.1 HO/min rate observed during the high-altitude phase (one handover every 8.5 seconds), applications experience continuous latency disruptions with insufficient time for recovery between events, resulting in persistent performance degradation unsuitable for time-sensitive UAV operations such as real-time video streaming and command-and-control. Key Findings and Implications. Our handover analysis establishes the following key findings: • Altitude drives 3-4× handover rate increase despite 15 - 20 dB stronger RSRP, revealing that RSRP-only algorithms fail in aerial LOS conditions where uniformly strong signals from multiple cells trigger excessive E1 (Event A3) handovers. • Interference handovers appear exclusively during the high-altitude phase (30 - 50 min at 330 - 380 m ASL) where strong RSRP coexists with poor RSRQ (-16 to -17 dB), with E2 (Interference) events comprising 12% of total handovers. This demonstrates that quality metrics must supplement strength-based handover triggers for aerial platforms. • Handover impact is asymmetric: 53.5% of handovers improve RTT, but worst-case degradation (275 ms) is 2 × larger than best-case improvement (137 ms), creating high-risk latency profiles unsuitable for time-sensitive applications. 5

Cellular vs. Starlink: End-to-End Service Quality Assessment

Given the platform’s ability to simultaneously track metrics from both terrestrial cellular network and LEO satellite (Starlink Mini) connectivity, we perform dual connectivity analysis comparing end-to-end performance characteristics under operational flight conditions. The analysis focuses on three critical metrics: round-trip latency, bidirectional throughput, and packet delivery statistics. 5.1

Head-to-Head Comparisons

Latency Performance: We perform comparative analysis of LTE and Starlink performance measured simultaneously at identical times and locations (between UAV and a dedicated remote server). The comparison was repeated over two flight tests (September 2025 and December 2025), and the results are shown in Figure 17. The operational flight tests revealed significant performance , Vol. 1, No. 1, Article . Publication date: May 2026.

A Vertical Look at UAV Connectivity in the Wild

17

(a) Starlink RTT. (b) Starlink Downlink Throughput. (c) Starlink Uplink Throughput. Fig. 18. Temporal stability analysis of end-to-end metrics for Starlink across four distinct flight tests (denoted by “FLT-1” through “FLT-4”) conducted over a period of three months.

differences in real-time responsiveness between the two connectivity options. As shown in Figure 17a, the Starlink satellite link demonstrated exceptional latency stability, with 95% of RTT measurements remaining below 50 ms. In contrast, the cellular network exhibited higher latency, with approximately 80% of measurement data falling under the 150 ms threshold. This three-fold difference in latency distribution indicates that the LEO satellite constellation provides superior responsiveness for real-time applications despite the additional propagation distance to orbit. Throughput Performance: The throughput analysis reveals performance dichotomies between cellular and Starlink across downlink and uplink directions. For downlink performance, as shown in Figure 17b, Starlink demonstrates a significant capacity advantage with 95% of measurements exceeding 25 Mbps. The cellular network shows more modest performance, with the comparable 95% threshold at only 5 Mbps. This five-fold difference in downlink capacity highlights the bandwidth advantages of the LEO satellite system for data-intensive applications. However, uplink performance presents a contrasting picture. Approximately 65% of uplink measurements for both cellular and Starlink fall below 5 Mbps, indicating similar uplink constraints across both technologies. This symmetric uplink limitation suggests that uplink capacity represents a shared bottleneck regardless of the underlying connectivity technology. Packet Delivery Statistics: Both connectivity solutions demonstrated high transmission reliability during active flight operations. As shown in Table 1, the cellular connection achieved a packet delivery rate of 99.44% (5130 out of 5159 packets delivered), while the Starlink connection maintained a 99.31% delivery rate (3602 of 3627 packets delivered). These minimal packet loss rates (0.56% for cellular and 0.69% for Starlink) validate the robustness of both links for UAV operations and indicate sufficient reliability for data transfers. Connection LTE Starlink

Pkts Sent

Pkts Delivered

Pkts Dropped

Delivery Rate (%)

5159 3627

5130 3602

29 25

99.44 99.31

Table 1. Packet delivery performance comparison between LTE and Starlink connections.

Dual Connectivity Implications: The complementary performance characteristics of cellular and Starlink suggest opportunities for intelligent dual connectivity architectures. Starlink’s latency performance (95% under 50 ms) makes it well-suited for latency-sensitive control and telemetry traffic, while its high downlink capacity (95% exceeding 25 Mbps) supports bandwidth-intensive payload data transmission. The cellular network, while exhibiting higher latency, maintains competitive packet delivery reliability (99.66%) and provides consistent uplink performance comparable to Starlink. A dual connectivity approach could leverage Starlink for primary data transmission , Vol. 1, No. 1, Article . Publication date: May 2026.

18

Sravan, et al.

(a) RTT. (b) Uplink and downlink throughput. Fig. 19. Starlink E2E metrics conducted across 4 flight tests to identify any spikes similar to ground sensitivity analysis as shown in Figure 5.

and real-time operations while utilizing cellular for redundancy and failover, or alternatively, could intelligently distribute traffic based on latency requirements and bandwidth availability. The platform’s capability to simultaneously measure both links during flight operations enables real-time link quality assessment necessary for dynamic connectivity management in dual-link architectures. 5.2

Starlink Performance Deep Dive

5.2.1 Short-term Starlink Performance Evolution. In this part, we conduct more comprehensive performance analysis on Starlink network performance using our UAV measurement platform. We repeat the measurements across four different flight tests (denoted by “SL-FLT-1” through “SL-FLT4”) over a 3-month period (from September 2025 to December 2025). Figures 18 and 19 demonstrate the CDF and temporal performance across the four flight tests. From the results, we observe several interesting trends. As shown in Figure 19, the concatenated time series plots reveal a clear temporal performance evolution in Starlink connectivity from September to December 2025. SL-FLT4 conducted on Dec 17, 2025, demonstrates substantially superior performance across all metrics compared to earlier test days. Specifically, December shows ∼ 5× higher downlink throughput (consistently 60 - 100 Mbps vs 20 - 40 Mbps in September), ∼ 3× higher uplink throughput (10 - 20 Mbps vs 5 - 8 Mbps), and 40% lower RTT (20 - 30 ms vs 40 - 60 ms). Additionally, December data exhibits dramatically reduced variability as the performance curves are smoother and more stable, contrasting sharply with the erratic fluctuations visible in SL-FLT-1 and SL-FLT-2. This aligns perfectly with the CDF analysis (Figure 18), where we observed that the December distribution was better with mean improvements of +30% downlink, +160% uplink, and -39% latency. Furthermore, in contrast to the ground-based sensitivity analysis depicted in Figure 5, our observations revealed no unusual disruptions, with the exception of a handful of transient spikes. This configuration of Starlink antennas on the UAV likely enhanced LOS conditions, ensuring that connectivity remained stable even amid the UAV’s maneuvers. Based on the observed performance, we concur that Starlink’s LEO connectivity provides a reliable communication channel for real-time applications involving UAVs in rural settings. 5.2.2 Impacts of UAV Dynamics on Starlink performance. During flight test exercise, the UAV followed a race-track trajectory involving continuous banking, pitching, and turning. We observed that despite the frequent changes in the Starlink antenna’s physical orientation caused by aircraft banking and pitching, no significant impact on link stability was observed. Our statistical analysis confirms relatively small correlations between the aircraft’s altitude angles and network performance. The correlation coefficients between RTT and roll/pitch were calculated to be 0.0434 and −0.1122, respectively. Similarly, the correlation coefficients between throughput and roll/pitch were −0.0011 and 0.0544. These actual flight test results indicate that the Starlink Mini’s phased-array , Vol. 1, No. 1, Article . Publication date: May 2026.

A Vertical Look at UAV Connectivity in the Wild

19

beam-steering is sufficiently robust for standard fixed-wing UAV operations assuming the antenna maintained a general skyward orientation. 6

ML-based 3D Performance Prediction

As discussed in earlier sections, the platform captures measurements at various altitudes during flight tests, providing comprehensive spatial sampling across the deployment area. However, exhaustive measurement of every possible location and altitude combination is impractical. We leverage standard ML models to predict performance metrics at unmeasured altitudes and locations. Our approach involves training these models on GPS-tagged position data from flight tests, where each measurement is associated with a specific location defined by longitude, latitude, and altitude. We use position coordinates as input features (𝑥) to predict corresponding metric values as outputs ˆ to generate predicted values (𝑦). ˆ (𝑦). Once trained, the models accept new position coordinates (𝑥) We use two distinct evaluation paradigms: Leave-One-Altitude-Out (LOAO) cross-validation and the “standard” 80/20 Random Split approach. The LOAO strategy is specifically designed to test the models’ vertical extrapolation capabilities by removing all measurements from a target altitude during training and evaluating the model on that unseen vertical plane. Specifically, we create a truncated dataset by removing all measurements from one altitude level from the training set and train each model on the remaining altitude measurements. The trained model was then used to predict metric values for the excluded altitude. In contrast, the 80/20 Random Split serves as a benchmark for general spatial interpolation, where 80% of the shuffled dataset is used for training and the remaining 20% for testing across all spatial coordinates and altitudes simultaneously. This evaluation framework enables a comprehensive assessment of both altitude-specific extrapolation and general spatial prediction capabilities. Evaluation Metrics. To evaluate model performance, we used Mean Absolute Error (MAE) as the primary evaluation metric and Root Mean Square Error (RMSE) as a secondary metric. These metrics are computed as: v u t 𝑁 𝑁 1 ∑︁ 1 ∑︁ MAE = |𝑦𝑖 − 𝑦ˆ𝑖 |, and RMSE = (𝑦𝑖 − 𝑦ˆ𝑖 )2, (1) 𝑁 𝑖=1 𝑁 𝑖=1 where 𝑁 is the number of observations, and 𝑦𝑖 and 𝑦ˆ𝑖 are the actual (collected) and predicted values for the 𝑖 𝑡ℎ observation, respectively. Unlike MAE, which treats all errors equally, RMSE assigns greater importance to larger errors due to the squaring operation, making it particularly useful for identifying models that produce occasional large prediction errors. Results and Discussion. Experimental results in Table 2 demonstrate that while both training approaches achieve high fidelity, the models exhibit significantly higher precision when performing spatial interpolation compared to vertical extrapolation. As shown in Table 2, the Random Forest (RF) model achieved its peak accuracy under the 80/20 split with an RMSE of 1.764 dBm and an MAE of 1.258 dBm. However, when subjected to the more challenging LOAO task, the RMSE increased to 3.610 dBm. This “extrapolation penalty” quantifies the additional variance introduced when predicting signal metrics at entirely unobserved altitudes, which is a critical consideration for network operators managing 3D aerial coverage in rural terrain. The performance gap between the two methodologies highlights the complexity of the characterization of the 3D radio signal. For safety-critical UAV operations, the LOAO metrics (specifically the RMSE of about 3.6 dBm) provide the worst-case error margin for autonomous path planning at new flight levels. These findings confirm that while ML-based interpolation is highly precise, vertical extrapolation requires slightly higher safety buffers to account for altitude-dependent propagation phenomena like diffraction and ground reflections. , Vol. 1, No. 1, Article . Publication date: May 2026.

20

Sravan, et al.

RAN Metric RSRP RSRQ RSSI SINR

Machine Learning Model Random Forest (RF) Gradient Boosting (GB) Multi-Layer Perceptron (MLP) Random Forest (RF) Gradient Boosting (GB) Multi-Layer Perceptron (MLP) Random Forest (RF) Gradient Boosting (GB) Multi-Layer Perceptron (MLP) Random Forest (RF) Gradient Boosting (GB) Multi-Layer Perceptron (MLP)

RMSE (dBm) LOAO Generic 80-20 5.21 3.97 5.14 4.15 5.88 4.61 2.24 1.84 2.27 1.90 2.57 2.12 4.58 3.32 4.47 3.50 4.84 3.79 2.44 1.68 2.54 1.80 2.65 2.09

MAE (dBm) LOAO Generic 80-20 3.96 2.43 3.92 2.78 4.73 3.36 1.85 1.45 1.86 1.52 2.11 1.76 3.50 2.40 3.46 2.59 3.77 2.89 1.85 1.12 1.92 1.24 2.11 1.56

Table 2. Performance Comparison of Machine Learning Models for predicting the RAN metrics with 2 methods Leave-One-Altitude-Out (LOAO) and the generic 80/20% train/test split method.

These predictive capabilities have direct implications for the optimization of BVLOS UAV operations and cellular network planning. By enabling accurate 3D coverage extrapolation, the models allow network operators to estimate interference patterns and signal dead zones at altitudes where physical measurements are logistically infeasible. Furthermore, the integration of these models into autonomous flight control systems supports proactive path optimization. A UAV can leverage these signal predictions to dynamically adjust its trajectory, maintaining a reliable communication link by avoiding predicted zones of high multi-cell interference or weak signal power, thereby enhancing mission safety and data transmission reliability. 7

Conclusion

In this paper, we presented the first open-source, dual-connectivity UAV-based measurement platform to characterize the performance of terrestrial LTE and LEO satellite networks in rural environments. Our measurement platform integrates commercial-grade cellular modem and Starlink Mini terminal with onboard flight controller and computing systems to capture synchronized physical, network, and application-layer metrics across diverse flight altitudes. Experimental validation demonstrates the platform’s capability to reveal altitude-dependent performance tradeoffs: signal power improves 8 − 20 dB at higher altitudes due to enhanced LOS conditions, whereas signal quality degrades 3 − 5 dB because of increased interference from neighboring cells. This causes 3 − 4 × increase in handover rates due to excessive multi-cell visibility rather than signal degradation. Contrary to the general belief that satellite links are slower than terrestrial networks, our LTE-Starlink comparative analysis showed that Starlink maintained 95% of RTT samples below 50 ms, significantly outperforming the rural LTE network with 20% higher than 150 ms RTT. Finally, our ML-based prediction suite quantifies the “extrapolation penalty” inherent in 3D network measurements. To facilitate reproducible research, we will make the complete platform design, software suite, and spatiotemporal dataset publicly available. As the next steps, we plan to extend the platform capabilities to support 5G NR measurements and multi-UAV measurements across larger geographic areas. Furthermore, we will extend our measurement campaigns to other geographical environments and with multiple cellular network operators.

, Vol. 1, No. 1, Article . Publication date: May 2026.

A Vertical Look at UAV Connectivity in the Wild

21

References [1] A. V. Sheshashayee, M. Bordin, P. B. del Prever, D. Villa, H. Cheng, C. Petrioli, T. Melodia, and S. Basagni, “Experimental Evaluation of the Performance of UAV-assisted Data Collection for Wake-up Radio-enabled Wireless Networks,” in 2024 IEEE 99th Vehicular Technology Conference (VTC2024-Spring), 2024, pp. 01–06. [2] Geraci, Giovanni and Garcia-Rodriguez, Adrian and Azari, M Mahdi and Lozano, Angel and Mezzavilla, Marco and Chatzinotas, Symeon and Chen, Yun and Rangan, Sundeep and Di Renzo, Marco, “What will the future of UAV cellular communications be? A flight from 5G to 6G,” IEEE communications surveys & tutorials, vol. 24, no. 3, pp. 1304–1335, 2022. [3] S. Homayouni, M. Paier, T. Berisha, S. Woblistin, and J. Rehak, “3GPP-based verification of latency measurements in operational cellular networks with low-altitude drones,” in 2022 International Conference on Smart Applications, Communications and Networking (SmartNets). IEEE, 2022, pp. 01–05. [4] Alsabah, Muntadher and Naser, Marwah Abdulrazzaq and Mahmmod, Basheera M and Abdulhussain, Sadiq H and Eissa, Mohammad R and Al-Baidhani, Ahmed and Noordin, Nor K and Sait, Sadiq M and Al-Utaibi, Khaled A and Hashim, Fazirul, “6G wireless communications networks: A comprehensive survey,” IEEE Access, vol. 9, pp. 148 191–148 243, 2021. [5] J. Poorvi, A. Kalita, and M. Gurusamy, “Reliable and efficient data collection in uav based iot networks,” IEEE Communications Surveys & Tutorials, 2025. [6] Y. Zeng, Q. Wu, and R. Zhang, “Accessing from the sky: A tutorial on uav communications for 5g and beyond,” Proceedings of the IEEE, vol. 107, no. 12, pp. 2327–2375, 2019. [7] S. Zhang, Y. Zeng, and R. Zhang, “Cellular-Enabled UAV Communication: A Connectivity-Constrained Trajectory Optimization Perspective,” IEEE Transactions on Communications, vol. 67, no. 3, pp. 2580–2604, 2019. [8] Zeng, Yong and Guvenc, Ismail and Zhang, Rui and Geraci, Giovanni and Matolak, David W, UAV Communications for 5G and Beyond. John Wiley & Sons, 2020. [9] A. S. Abdalla and V. Marojevic, “Communications standards for unmanned aircraft systems: The 3gpp perspective and research drivers,” IEEE Communications Standards Magazine, vol. 5, no. 1, pp. 70–77, 2021. [10] S. D. Muruganathan, X. Lin, H.-L. Määttänen, J. Sedin, Z. Zou, W. A. Hapsari, and S. Yasukawa, “An overview of 3gpp release-15 study on enhanced lte support for connected drones,” IEEE Communications Standards Magazine, vol. 5, no. 4, pp. 140–146, 2022. [11] X. Lin, V. Yajnanarayana, S. D. Muruganathan, S. Gao, H. Asplund, H.-L. Maattanen, M. Bergstrom, S. Euler, and Y.-P. E. Wang, “The sky is not the limit: Lte for unmanned aerial vehicles,” IEEE Communications Magazine, vol. 56, no. 4, pp. 204–210, 2018. [12] R. Amorim, H. Nguyen, J. Wigard, I. Z. Kovács, T. B. Sørensen, D. Z. Biro, M. Sørensen, and P. Mogensen, “Measured uplink interference caused by aerial vehicles in lte cellular networks,” IEEE Wireless Communications Letters, vol. 7, no. 6, pp. 958–961, 2018. [13] Homayouni, Samira and Berisha, Taulant and Paier, Mario and Woblistin, Sebastian and Rehak, Johannes and Neubauer, Thomas, “Verification of Standardized Rel-15 Requirements for Drone’s Command-and-Control Link Reliability,” in 2023 IEEE 97th Vehicular Technology Conference (VTC2023-Spring). IEEE, 2023, pp. 1–5. [14] V. Platzgummer, V. Raida, G. Krainz, P. Svoboda, M. Lerch, and M. Rupp, “UAV-based coverage measurement method for 5G,” in 2019 IEEE 90th Vehicular Technology Conference (VTC2019-Fall). IEEE, 2019, pp. 1–6. [15] J. Sae, R. Wirén, J. Kauppi, H.-L. Maattanen, J. Torsner, and M. Valkama, “Public LTE network measurements with drones in rural environment,” in 2019 IEEE 89th Vehicular Technology Conference (VTC2019-Spring). IEEE, 2019, pp. 1–5. [16] R. Amorim, P. Mogensen, T. Sorensen, I. Z. Kovács, and J. Wigard, “Pathloss measurements and modeling for UAVs connected to cellular networks,” in 2017 IEEE 85th Vehicular Technology Conference (VTC Spring). IEEE, 2017, pp. 1–6. [17] X. Lin, R. Wiren, S. Euler, A. Sadam, H.-L. Määttänen, S. Muruganathan, S. Gao, Y.-P. E. Wang, J. Kauppi, Z. Zou et al., “Mobile network-connected drones: Field trials, simulations, and design insights,” IEEE Vehicular Technology Magazine, vol. 14, no. 3, pp. 115–125, 2019. [18] Wang, Sizhe and Ghoshal, Moinak and Feng, Yufei and Khan, Imran and Dinh, Phuc and Basit, Omar and Yu, Zhekun and Hu, Y Charlie and Koutsonikolas, Dimitrios, “Exploring the 5G Digital Divide in the Non-Contiguous US: LEO Satellites to the Rescue?” Proceedings of the ACM on Measurement and Analysis of Computing Systems, vol. 9, no. 3, pp. 1–25, 2025. [19] Ghafoori, Amirreza and Famili, Alireza and Stavrou, Angelos, “Stars and towers on the wheels: Global perspective on satellite networks vs. terrestrial 5G,” in GLOBECOM 2024-2024 IEEE Global Communications Conference. IEEE, 2024, pp. 301–306. [20] Ghoshal, Moinak and Basit, Omar and Khan, Imran and Kong, Z Jonny and Wang, Sizhe and Feng, Yufei and Dinh, Phuc and Hu, Y Charlie and Koutsonikolas, Dimitrios, “Replication: Performance of Cellular Networks on the Wheels,” in Proceedings of the 2025 ACM Internet Measurement Conference, 2025, pp. 381–396. , Vol. 1, No. 1, Article . Publication date: May 2026.

22

Sravan, et al.

[21] Baltaci, Aygün and Dinc, Ergin and Ozger, Mustafa and Alabbasi, Abdulrahman and Cavdar, Cicek and Schupke, Dominic, “A survey of wireless networks for future aerial communications (FACOM),” IEEE Communications Surveys & Tutorials, vol. 23, no. 4, pp. 2833–2884, 2021. [22] Zhou, Di and Sheng, Min and Li, Jiandong and Han, Zhu, “Aerospace integrated networks innovation for empowering 6G: A survey and future challenges,” IEEE Communications Surveys & Tutorials, vol. 25, no. 2, pp. 975–1019, 2023. [23] Hu, Bin and Zhang, Xumiao and Zhang, Qixin and Varyani, Nitin and Mao, Z Morley and Qian, Feng and Zhang, Zhi-Li, “LEO satellite vs. cellular networks: Exploring the potential for synergistic integration,” in Companion of the 19th International Conference on emerging Networking EXperiments and Technologies, 2023, pp. 45–51. [24] W. Qin, Z. M. Komodromos, and T. E. Humphreys, “An analysis of the short-term time stability of the starlink ku-band downlink frame clock,” in 2024 IEEE International Conference on Wireless for Space and Extreme Environments (WiSEE). IEEE, 2024, pp. 106–111. [25] L. Izhikevich, M. Tran, K. Izhikevich, G. Akiwate, and Z. Durumeric, “Democratizing leo satellite network measurement,” Proceedings of the ACM on Measurement and Analysis of Computing Systems, vol. 8, no. 1, pp. 1–26, 2024. [26] N. Mohan, A. E. Ferguson, H. Cech, R. Bose, P. R. Renatin, M. K. Marina, and J. Ott, “A multifaceted look at starlink performance,” in Proceedings of the ACM Web Conference 2024, 2024, pp. 2723–2734. [27] D. Laniewski, E. Lanfer, S. Beginn, J. Dunker, M. DÞckers, and N. Aschenbruck, “Starlink on the road: A first look at mobile starlink performance in central europe,” arXiv preprint arXiv:2403.13497, 2024. [28] D. Laniewski, E. Lanfer, and N. Aschenbruck, “Measuring mobile starlink performance: A comprehensive look,” IEEE Open Journal of the Communications Society, 2025. [29] X. Lin, S. Rommer, S. Euler, E. A. Yavuz, and R. S. Karlsson, “5g from space: An overview of 3gpp non-terrestrial networks,” IEEE Communications Standards Magazine, vol. 5, no. 4, pp. 147–153, 2021. [30] S. D. Muruganathan, X. Lin, H.-L. Maattanen, J. Sedin, Z. Zou, and W. A. Hapsari, “An overview of 3GPP release-15 study on enhanced LTE support for connected drones,” IEEE Communications Standards Magazine, vol. 5, no. 4, pp. 140–146, 2021. [31] M. Vaezi, A. Azari, S. R. Khosravirad, M. Shirvanimoghaddam, M. M. Azari, D. Chasaki, and P. Popovski, “Evolution of non-terrestrial networks from 5G to 6G: A survey,” IEEE Communications Surveys & Tutorials, vol. 24, no. 4, pp. 2634–2703, 2022. [32] B. Shang, X. Li, Z. Li, J. Ma, X. Chu, and P. Fan, “Multi-connectivity between terrestrial and non-terrestrial mimo systems,” IEEE Open Journal of the Communications Society, vol. 5, pp. 3245–3262, 2024. [33] “QualiPoc Tool for Networking Testing and Measurement.” [Online]. Available: https://www.rohde-schwarz.com/us/ products/test-and-measurement/network-data-collection/qualipoc-android_63493-55430.html [34] “TEMS Pocket – Portable Mobile Network Testing.” [Online]. Available: https://www.infovista.com/tems/pocket [35] K. Nemo, “Online,” Available https://www.keysight.com/zz/en/products/nemo-wireless-network-solutions.html. [36] R. Amorim, P. Mogensen, T. Sorensen, I. Z. Kovacs, and J. Wigard, “Pathloss Measurements and Modeling for UAVs Connected to Cellular Networks,” in 2017 IEEE 85th Vehicular Technology Conference (VTC Spring), 2017, pp. 1–6. [37] R. Amorim, H. Nguyen, J. Wigard, I. Z. Kovács, T. B. Sørensen, D. Z. Biro, M. Sørensen, and P. Mogensen, “Measured Uplink Interference Caused by Aerial Vehicles in LTE Cellular Networks,” IEEE Wireless Communications Letters, vol. 7, no. 6, pp. 958–961, 2018. [38] Maeng, Sung Joon and Ozdemir, O and Nandakumar, HN and Güvenç, İ and Sichitiu, Mihail L and Dutta, Rudra and Mushi, Magreth, “Spectrum activity monitoring and analysis for sub-6 GHz bands using a Helikite,” in 2023 15th International Conference on COMmunication Systems & NETworkS (COMSNETS). IEEE, 2023, pp. 857–862. [39] Maeng, S. J. and Ozdemir, O. and Güvenç, İ. and Sichitiu, M. L. and Dutta, R. and Mushi, M., “AERIQ: SDR-Based LTE I/Q Measurement and Analysis Framework for Air-to-Ground Propagation Modeling,” in 2023 IEEE Aerospace Conference, 2023, pp. 1–11. [40] Nvidia, “Online,” Available https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/. [41] pMLTE, “Online,” Available https://www.microhardcorp.com/pMLTE.php. [42] “Micro Air Vehicle Link (MAVLink).” [Online]. Available: https://mavlink.io/en/ [43] Nping, “Online,” Available https://nmap.org/nping/. [44] iPerf, “Online,” Available https://iperf.fr/. [45] P. Kopardekar, J. Rios, T. Prevot, M. Johnson, J. Jung, and J. E. Robinson, “Unmanned Aircraft System Traffic Management (UTM) concept of operations,” 2016. [46] T. Networks, “Online,” Available https://wiki.teltonika-networks.com/view/Mobile_Signal_Strength_ Recommendations. [47] Z. Shakir, A. Mjhool, A. Al-Thaedan, A. Al-Sabbagh, and R. Alsabah, “Key performance indicators analysis for 4 g-lte cellular networks based on real measurements,” International Journal of Information Technology, vol. 15, pp. 1–9, 03 2023. [48] CellMapper, “Online,” Available https://www.cellmapper.net. , Vol. 1, No. 1, Article . Publication date: May 2026.

A Vertical Look at UAV Connectivity in the Wild

23

[49] 3GPP, “Evolved Universal Terrestrial Radio Access (E-UTRA); Requirements for support of radio resource management,” 3rd Generation Partnership Project (3GPP), Technical Specification (TS) 36.133, 2025. [Online]. Available: https://www.3gpp.org/dynareport/36133.htm [50] Lai, Lifeng and Zheng, Fu-Chun and Luo, Jingjing, “Flight direction-based handover in cellular-connected UAV communications,” IEEE Transactions on Vehicular Technology, vol. 73, no. 11, pp. 17 771–17 775, 2024. [51] 3GPP, “Evolved Universal Terrestrial Radio Access (E-UTRA); Radio Resource Control (RRC); Protocol specification,” 3rd Generation Partnership Project, Tech. Rep. TS 36.331, Available at: https://www.3gpp.org/dynareport/36331.htm. [52] Tashan, Waheeb and Shayea, Ibraheem and Aldirmaz-Colak, Sultan and Ergen, Mustafa and Azmi, Marwan Hadri and Alhammadi, Abdulraqeb, “Mobility robustness optimization in future mobile heterogeneous networks: A survey,” IEEE Access, vol. 10, pp. 45 522–45 541, 2022. [53] Lobinger, Andreas and Stefanski, Szymon and Jansen, Thomas and Balan, Irina, “Load balancing in downlink LTE self-optimizing networks,” in 2010 IEEE 71st vehicular technology conference, 2010, pp. 1–5.

, Vol. 1, No. 1, Article . Publication date: May 2026.