<p>Accurate and timely classification of near-Earth asteroids is fundamental to planetary defense, as it directly supports threat assessment, observation prioritization, and mission planning. Despite recent advances, existing classification methods remain limited by severe class imbalance, insufficient robustness, and a lack of physical interpretability. This study introduces a comprehensive framework for predicting asteroid orbit classes using attention-driven Deep Learning (DL) and explainable techniques. These architectures are systematically benchmarked using orbital elements and physical parameters from a publicly available dataset for orbit class prediction. To ensure robustness under real-world data skew, multiple imbalance-handling strategies are incorporated into the evaluation pipeline. Experimental results demonstrate that the Tabular Network consistently outperforms competing models, achieving over 99.9% accuracy, precision, recall, and F1-score across the diverse balancing settings. Beyond predictive performance, this work advances scientific transparency by incorporating comprehensive global and local explainability analyses. The results indicate that absolute magnitude, semi-major axis, perihelion distance and aphelion distance dominate model decisions, closely aligning with established principles of orbital mechanics. This study establishes a robust, interpretable, and physically consistent DL framework for reliable asteroid orbit classification in planetary defense applications.</p>