From Radiosonde Archives to Flight Commands: ML-Guided Adaptive Sampling for UAS Atmospheric Boundary-Layer Profiling

Fixed-rate drone soundings can miss thin inversions, humidity transitions and wind shifts important to severe-weather forecasting. This project tests adaptive sampling by matching partial atmospheric profiles to seasonally stratified historical IGRA2 radiosondes. K-medoids clustering and a two-of-three point-of-interest voting method are integrated with the BLISS CopterSonde and LAAIRS ground-control system. Correcting a temperature convention reduces matching errors from approximately 18°C to 0.4–1.6°C; gains level off beyond about 15–20 medoids. Field validation places transition zones at 800–1,100 metres above ground, and the matching method operates end to end in LAAIRS. The NOAA NERTO work, mentored by Elizabeth Smith at the National Severe Storms Laboratory, examines a hardware-independent route to more efficient atmospheric sampling.