An Explainable Deep Transfer Learning for Multi-Class Aerial Vehicle Classification Using Fine-Tuned ResNet50 with Grad-CAM

Abstract

Aerial object classification for intelligent surveillance, airspace monitoring and autonomous security systems is essential but challenging because of the diversity in scale, background complexity and visual similarities among aerial objects. In this paper, a fine-tuned ResNet50 based deep learning system is proposed which can be used to classify the aerial vehicles into four classes (AirPlane, Bird, Drone, Helicopter) from a publicly available Kaggle dataset. The dataset consists of 11,998 RGB images that have been preprocessed before training by cropping the objects, resizing the images, normalizing, and encoding the labels. Transfer learning is used to use ImageNet pre-trained weights for the ResNet50 backbone, then fine-tune with a customized classification head, improving feature extraction and classification ability. In order to make the model more transparent, Gradient-weighted Class Activation Mapping (Grad-CAM) is incorporated to get visual explanations showing the parts of the image that are important for the classes predicted by the model. The proposed framework shows its effectiveness through the experimental evaluation, and the results show that the classification accuracy is 98.8%, and the precision, recall, F1-score, and ROC-AUC value are excellent in all classes. Comparative analysis shows that the proposed model is able to achieve better performance than several baseline models without compromising the good generalization ability. This work brings together the two main pillars of the field of artificial intelligence - transfer learning and explainable artificial intelligence (XAI) - in order to present an accurate, interpretable, and reliable aerial vehicle classification solution suitable for intelligent aerial surveillance applications.