Cognitive MobileNetV2-Based Micro-Doppler Analysis for Small UAV Detection and Classification

The growing use of small unmanned aerial vehicles (UAVs) in civilian and defense airspace are increased the need for reliable detection method which can operate under clutter and low altitude condition. Radar based sensing using micro-Doppler signatures provide valuable motion related information however, effective use of these signatures still remain challenging because of background interference and limited signal strength. In this study, a lightweight deep learning framework based on MobileNetV2 is investigate for detection and classification of small UAV using micro-Doppler spectrogram derived from DIAT-uSAT dataset. Raw radar signal is processed to obtain time frequency representation which then analyzed using multiple convolution neural network architecture to evaluate the classification performance. Comparative experiment involving EfficientNetB0, DenseNet121, Xception, ResNet50 and VGG16 shows that MobileNetV2 achieve faster convergence and more stable generalization on compact spectrogram representation. The proposed model achieved training accuracy of 96.1% and validation accuracy of 95% with minimum validation loss. These result suggest that properly designed lightweight convolution network can effectively capture discriminative micro Doppler pattern and suitable for real time UAV surveillance application in complex environment.