• Wind estimation from UAV telemetry only, no external sensors required • 1,750 real free-flight missions enable realistic evaluation • Robust predictions demonstrated across multiple learning architectures • Wind-tunnel features improve learning but remain optional • Errors mainly due to ground-reference limits, not model capability Reliable local wind estimation using onboard telemetry alone enables autonomous unmanned aerial vehicle (UAV) operation in environments where external sensing is unavailable or impractical. This work proposes a data-driven framework for estimating wind velocity from multirotor UAV telemetry, without additional instrumentation. A dataset comprising 1,750 flights under diverse trajectories and atmospheric conditions was analyzed. Four machine learning architectures were evaluated: a multilayer perceptron (MLP) with an attention mechanism, a gated recurrent unit (GRU), a Bayesian neural network (BNN), and TabNet. Aerodynamic and thrust-related features were combined with predictions from a neural network trained on wind-tunnel data rated to enrich model inputs. A fixed ground-based anemometer positioned 5 m above ground level served as the reference for supervision and validation under real operational conditions. All models achieved comparable predictive accuracy and successfully captured both the temporal evolution and magnitude of wind dynamics across a wide range of atmospheric conditions, including low, moderate, and high wind regimes. The analysis indicates that estimation errors are influenced by the inherent spatial variability of the wind field, particularly as the UAV operated at increasing distances from the reference anemometer. Overall, the results demonstrate the feasibility of learning-based wind sensing under free-flight conditions using only standard onboard telemetry. Among the evaluated approaches, TabNet exhibited strong and consistent performance, highlighting its suitability for future real-time onboard deployment.
García-Gascón et al. (Sun,) studied this question.