Proof-of-Concept: Vertical Wind Profile Reconstruction from Ground-Based Optical Sensors Using Machine Learning

Wolfgang Schneider

Vertical wind profiles are critical for weather forecasting, aviation safety, and atmospheric research, yet remain sparsely observed due to the high cost of radiosondes (€100–200 per launch) and wind profiler radars (€100,000–1,000,000). We present a proof-of-concept demonstrating that ground-based optical measurements from low-cost amateur radio sensors can predict upper-air wind speeds across multiple atmospheric layers using machine learning. Two monitoring stations (DG2MCM-15 and DG2MCM-16) near Kempten, Germany (47.73°N, 10.32°E, 686 m ASL) measured infrared sky temperature (MLX90614, €15), multi-channel spectral radiance (AS7341, €8), and RGB photometry every 5 minutes during 15–21 February 2026. Random Forest regression models trained on 21 coincident samples with EUMETSAT Meteosat Third Generation (MTG) Atmospheric Motion Vectors achieved training R²=0.75–0.91 for wind speeds in the lower troposphere (1–3 km), middle troposphere (3–6 km), upper troposphere (6–9 km), and jetstream (9–12 km) layers. Feature importance analysis revealed that infrared-derived cloud base height was the strongest predictor for upper tropospheric winds (28% importance), while the spectral aerosol Ångström exponent ranked among the top features (10% overall), validating the hypothesis that atmospheric optical properties encode information about vertical structure and synoptic forcing. Surface wind speed contributed minimal predictive power (4%), indicating decoupling between 10 m winds and upper-air flows. Test set performance showed severe overfitting (R²_test = −2.59) due to small sample size, confirming that operational deployment requires larger training datasets. However, the proof-of-concept successfully demonstrates feasibility: total hardware cost of €250 per station enables dense observational networks impossible with traditional instrumentation. A follow-up 6-month validation study (Paper 2b) is planned to assess generalizability across seasons and weather regimes.