Event-Based Camera Modeling for Atmospheric Turbulence Prediction

Atmospheric turbulence degrades long-range imaging and free-space optical performance, yet conventional measurement systems such as large-aperture scintillometers require active transmitters, precise alignment, and dedicated deployment. This study investigates whether a passive neuromorphic event camera can provide reliable estimates of the refractive-index structure parameter Cn2 along a 300 m horizontal path. We conducted a week-long field experiment using a Prophesee EVK-4 HD event camera (Prophesee, Paris, France), a Basler acA2040-120um HD CMOS video camera (Basler AG, Ahrensburg, Germany), and a Scintec BLS900 scintillometer (Scintec AG, Rottenburg, Germany) as ground truth. A compact set of 19 statistical event-stream features was extracted over multiple integration times (2-50 s), and machine learning regression models were trained to predict the corresponding scintillometer-measured turbulence. Across the full turbulence range 10-14-10-12 m-2/3, the best-performing model (XGBoost) achieved a Pearson correlation of 0.93 and a mean absolute relative error of 35%, with longer integration times and higher-contrast regions yielding improved accuracy. The results also quantify, for the first time in field conditions, how integration time, target contrast, and feature stability influence event-based turbulence estimation. These findings demonstrate that passive event-driven sensing can approximate scintillometer-level turbulence measurements without active illumination, enabling compact, low-power alternatives for real-time atmospheric monitoring.