MACHINE LEARNING AND DEEP TEMPORAL NETWORKS FOR UAV VELOCITY CLASSIFICATION USING HIGH-DIMENSIONAL CSI AND RSSI DATA

With increasing utilization of Unmanned Aerial Vehicles in civilian and industrial scenarios, reliable velocity estimation is imperative for safe tracking and airspace security. This work presents a 5G-centric RF sensing framework that jointly leverages Received Signal Strength Indicator and Channel State Information using a USRP-based SDR testbed under controlled line-of-sight flights of drones. More than 19 million RSSI samples and high-dimensional CSI measurements were collected over multiple velocities. For RSSI-based inference, the proposed custom 1D-CNN+BiGRU model significantly outperforms classical machine learning baselines. For CSI-based inference, CNN-LSTM and ensemble models-like Random Forest and XGBoost-powerfully capture velocity-dependent channel variations. Results show that RSSI provides coarse temporal cues while CSI encodes fine spatial-multipath structure, and hybrid deep temporal models generalize best. The proposed dual-modality framework is scalable and easily extensible to UAV detection, localization, and intent-aware surveillance.