Electromagnetic Environment Classification for Robust UAV and Robotic Operation Using Ensemble and Deep Learning Models

Abstract

This paper proposes an electromagnetic (EM) environment classification framework that enables unmanned aerial vehicles (UAVs) and robotic platforms to perceive EM conditions as structured environmental states rather than unmodeled noise or isolated sensor faults. A physics-inspired, imbalanced multi-class dataset with nine EM-related features and six representative EM environment classes is generated to emulate realistic operating conditions spanning clean, disturbed, and hostile regimes. Heterogeneous classifiers—including Random Forest, Gradient Boosting, support vector machine (SVM) with radial basis function kernel, and a cost-sensitive feature-wise long short-term memory (LSTM) network—are trained using standardized features, stratified splits, and class-weighted learning to address minority yet safety-critical EM states. A stacked ensemble that fuses the probabilistic outputs of all base learners through a logistic regression meta-classifier achieves the best overall performance, with macro-averaged precision, recall, and F1-score of 0.9642, 0.9062, and 0.9319, respectively, outperforming individual models under statistically validated comparisons. The results demonstrate that multi-sensor EM awareness, implemented via ensemble learning, can provide robust and balanced classification of degraded and hostile EM environments, establishing EM perception as a key capability for resilient autonomous operation; future work will transition from synthetic to real robotic and flight data and integrate EM awareness into higher-level planning and control.