High-quality wireless communication is paramount to coordinating flights and missions effectively on a micro-UAV (a.k.a. drone). However, most cellular antennas are optimized for users on the ground and have not been planned for aerial devices. Many have studied cellular communication quality, but few works explore the impact of altitude during a flight. Through real-world experiments using an actual drone, we demonstrate significant connectivity dynamics in the air for both 4G LTE and 5G NR, and reveal that fixed-coordinate flights cannot maintain high quality connectivity in response to those dynamics. To address this problem, we present <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ASCEND</i>, a reinforcement learning-based 3D altitude selection scheme that maintains high-quality connectivity during a flight over a planned 2D path without requiring prior training. We evaluate <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">ASCEND</i> at multiple real-world locations to demonstrate a notable increase in expected throughput and a reduction in the proportion of low-quality legs during a flight mission.