MeL-RF: Meta-Learning Assisted Hierarchical Ensemble Framework for RF-Based Drone Classification

In this paper, we propose aMeta-Learning assisted hierarchical ensemble framework for radio frequency (RF)-based drone detection referred to asMeL-RF. The proposed MeL-RF framework adopted long-short term memory (LSTM)-based deep neural network (DNN) as a base models, independently trained for each drone class (Background, Bebop, AR Drone, and Phantom) to extract distinct temporal features from drone RF signatures. Then these base models predictions are fused through a lightweight meta-learner, which intelligently weighs their outputs to produce the final classification. Unlike prior stacking and averaging ensembles, MeL-RF introduces aclass-specialized hierarchicaltraining mechanism in which each LSTM-DNN is optimized as a binary detector for a single drone type, and a shallow meta-learnerlearns inter-class reliabilityto fuse their outputs. This design results in improve accuracy. Moreover, it also balanced recall across confusable classes (e.g., AR Drone vs. Phantom) while keeping the meta stage extremely lightweight, which facilitates real-time and edge deployment. The framework is evaluated using a public RF drone dataset comprising drone activity across various flight modes, including off, connected, armed, and hovering. Simulation results reveal that the proposed MeL-RF framework achieves consistent gains across all performance metrics, including precision, recall, and F1-scores, with substantial improvements for difficult and easily confusable classes such as AR Drone and Phantom. Unlike the conventional single-model classifier, which suffers from severe recall degradation on these classes, the ensemble delivers balanced and reliable detection performance across all drone types, thereby demonstrating a clear system-level advantage beyond overall accuracy improvement.