Robust UAV vs Bird Classification using Ensemble Deep Learning with Confidence-Based Decision and Explainable AI

Abstract

Unmanned Aerial Vehicle (UAV) detection has become increasingly critical in surveillance, airspace management, and security systems, where distinguishing UAVs from visually similar objects such as birds remains a significant challenge. This work presents a robust ensemble deep learning framework for UAV versus bird classification, integrating multiple convolutional neural network architectures with confidence-based decision strategies and explainable artificial intelligence. The methodology incorporates comprehensive image preprocessing techniques including resizing, augmentation, and normalization to enhance generalization. Multiple deep learning models—Custom CNN, ResNet50, MobileNet, EfficientNet, and ConvNeXt—are employed to extract discriminative features. A confidence-based ensemble mechanism is utilized to improve classification reliability, while Gradient-weighted Class Activation Mapping (Grad-CAM) is integrated to provide visual explanations of model predictions, enhancing interpretability. Performance evaluation demonstrates that the ResNet50 model achieves superior results with an accuracy of 99.8%, precision of 99.8%, recall of 99.8%, and F1-score of 99.8%, indicating its effectiveness in handling complex visual patterns. The system is deployed using a Flaskbased web interface with SQLite integration, enabling user authentication, image upload, and real-time prediction.