A Comparative Analysis of Machine Learning Models for Robust UAV-Bird Classification in Aerial Surveillance

The proliferation of Unmanned Aerial Vehicles (UAVs) necessitates advanced surveillance systems to distinguish them from birds, a critical challenge for airspace security. This study addresses the problem of high false alarm rates in traditional systems by evaluating the efficacy of various machine learning models for accurate, real-time classification. The research contribution is a comprehensive benchmarking of six machine learning algorithms—Logistic Regression, Neural Networks, Stochastic Gradient Descent, CN2 Rule Induction, Naive Bayes, and Support Vector Machines—trained on a curated dataset of bird and drone images. The methodology involved rigorous preprocessing, including resizing, normalization, and augmentation, followed by stratified 10-fold cross-validation. Results demonstrated that Neural Networks, Support Vector Machines, and Logistic Regression were the top performers. The Neural Network model achieved the highest accuracy (98.6%) and AUC (0.998), with the lowest LogLoss (0.056), significantly minimizing false positives and negatives. In contrast, Naive Bayes underperformed substantially (accuracy 82.2%, LogLoss 5.528). The discussion contextualizes these findings within existing literature, highlighting the superiority of complex models capable of capturing nonlinear patterns in image data. This study concludes that advanced machine learning models, particularly deep learning architectures, are highly effective for UAV-bird discrimination, thereby enhancing real-time surveillance capabilities. Future work will focus on integrating these models with radar data and testing them in dynamic operational environments.