Deep Learning Based Micro-Drone Detection forReliable Counter-Drone Systems Using MultipleSensors

Micro-drones are easily attainable due to their affordability and ease of operation. However, their malicious use has led to increasing security concerns, necessitating the development of reliable counter-drone systems. In the literature, numerous deep learning models have been proposed to achieve high detection accuracy using image-based approaches. However, these methods tend to be unreliable under low-light or dark conditions. Hence, to ensure robust and reliable detection, a multimodal approach using both audio and image data is proposed. Datasets containing audio and aerial images of drones and non-drones are utilized, and data augmentation is applied to increase variability and improve generalization. PANNs-CNN10, YAMNet, and ResNet50 CNN models are explored for acoustic detection using micro-drone sounds, while InceptionV3, ResNet50, and DenseNet121 are investigated for optical detection using images. The optimal models from each category are selected for fusion. The performance of the proposed method is evaluated using standard metrics such as accuracy, precision, recall, F1-score, and ROC curves. Experimental results show that PANNs-CNN10 is lighweight and achieved near-perfect accuracy (99.8%), while DenseNet121 achieved 98.9%. The combined probability outputs of these optimal CNN models are fed into a logistic regression (meta-classifier) to exploit the complementary strengths of both modalities. The fusion model achieved an accuracy of 99.95%, representing a slightly higher improvement compared to the individual models. The ensemble approach effectively leverages the strengths of acoustic and optical modalities, addressing the limitations of standalone techniques and reducing false alarms to enhance reliability. Overall, this study contributes to the development of intelligent detection mechanisms that can strengthen counter-drone defense systems.