Drone Flight Dataset and Lightweight LSTM-Based Wind Estimation for Semi-Autonomous Quadcopter Control

Environmental forces like wind present a persistent challenge to drone stability, particularly during autonomous missions where precise controller performance is critical. Traditional systems often neglect wind effects or depend on costly external sensors, limiting scalability and adaptability. This study introduces a novel, lightweight LSTM-based model for effective wind speed estimation directly from onboard multivariate time-series data, eliminating the need for external hardware. A major contribution is the development of a comprehensive, publicly available drone flight dataset captured across varied wind conditions and terrains. The dataset includes 17 synchronized features—such as IMU readings, GPS velocity, motor outputs, and orientation—providing rich context for wind–drone interactions. To integrate predictions into real-time operations, a custom semi-autonomous control interface is designed, visualizing RC commands, system feedback, and model-inferred wind estimates for operator transparency and intervention. The proposed model achieves high predictive accuracy (RMSE = 0.0245, R² = 0.9945) while remaining computationally efficient (1,281 parameters). Under controlled outdoor experiments, the interface reduced spatial drift by 61–74%, roll/pitch RMS error by 41–44%, and improved landing accuracy by 83% (reducing average drift from 2.4 m to 0.4 m). Together, the dataset, model, and control interface establish a complete pipeline toward wind-aware, adaptive drone navigation. The proposed methodology is the subject of Indian Provisional Patent Applications Nos. 202541069264, 202541068445, and 202541070585, filed in 2025.