Horizontal wind estimation from consumer-drone hover telemetry: comparing machine-learning and calibration approaches

Abstract. Mounting external anemometers on small uncrewed aerial vehicles adds payload weight and complicates flight operations. During hover, horizontal wind can instead be derived from the drone tilt required to hold position. We compare two such tilt-based approaches using flight telemetry from an unmodified DJI Air 3S. Ridge regression fits a linear model to pitch- and roll-derived features and requires reference measurements for training, while a calibration baseline converts mean tilt to wind speed using quadratic curves fitted during separate near-calm flights and requires no reference anemometer during deployment. Both methods were evaluated across 50 hover sessions in Bremen, Germany, against a reference anemometer 1.6 m above ground. The sessions yielded 71,749 overlapping 30 s windows. Session-mean wind speeds ranged from 0.8 to 7.6 m s⁻¹, and the highest 30 s vector-mean speed was 11.5 m s⁻¹. In leave-one-flight-out cross-validation, Ridge had a speed root-mean-square error (RMSE) of 0.32 m s⁻¹ and a direction mean absolute error (MAE) of 7.0°, compared with 0.39 m s⁻¹ and 8.4° for the baseline. For session-mean winds above 2 m s⁻¹, direction MAE was 6.0° for Ridge and 7.0° for the baseline, close to the estimated 5–7° directional uncertainty of the reference setup. Across the six sessions above 6 m s⁻¹, Ridge and the baseline had similar speed RMSEs of 0.49 and 0.48 m s⁻¹. Tree-based models (XGBoost and LightGBM) did not improve upon Ridge. Hover telemetry can therefore provide accurate horizontal wind estimates without dedicated payload sensors. Further validation is needed for other airframes, greater heights, and sustained stronger winds.