Accurate classification of drone targets using radar systems is becoming increasingly important for surveillance, airspace monitoring, and security applications. This paper presents a machine learning-based framework for radar drone detection and classification using a synthetically generated radar dataset under both ideal and realistic operating conditions. The dataset incorporates key radar features, including radar cross section (RCS), Doppler shift, range, range resolution, aspect angle, and micro-Doppler signatures generated by rotating propellers, along with realistic clutter and noise associated with ground, weather, and urban environments to emulate practical radar scenarios. Three machine learning classifiers, Support Vector Machine (SVM), Random Forest (RF), and Boosted Trees (BT), were trained and evaluated. Under ideal conditions, all three models achieved classification accuracies close to 99%. When realistic clutter and noise were introduced, performance degraded substantially, with accuracy dropping to 78.8% for both SVM and RF and 78.5% for BT. Feature selection and hyperparameter tuning via Bayesian Optimization (BO) recovered comparable overall accuracies of 86.8–87.2% across all three classifiers. However, the optimized models exhibited a consistent trade-off rather than a single dominant classifier: SVM achieved the highest recall (90.4%), RF achieved the highest Area Under the Receiver Operating Characteristic Curve (AUC) (0.953) and the lowest false-positive rate (3.4%), and BT achieved the highest overall accuracy (87.2%) and F1-score (87.15%). No single classifier dominates across all metrics, so identifying the most suitable model requires an explicit evaluation priority. This study adopts the position that, in surveillance and security contexts, missed detections typically carry a greater operational cost than false alarms; under this recall-weighted priority, SVM is identified as the preferred classifier, while RF is preferable where minimizing false alarms is instead the priority. These results demonstrate that optimized machine learning techniques, combined with radar-derived target features, can substantially improve drone classification robustness in cluttered environments, while highlighting that classifier selection should be guided by the relative cost of missed detections versus false alarms specific to the deployment context.