CNN Performance Analysis for SAR Object Classification

Advancing Aerial View Object Classification (AVOC) methods benefit many applications, such as surveillance and geospatial analysis. Synthetic Aperture Radar (SAR) imaging can operate at night and in various weather conditions, giving an advantage over Electro-optical (EO) sensors. However, the unique properties of radar imaging pose a challenge for human interpretation. Hence, automatic methods such as a Convolutional Neural Network (CNN) have gained popularity in recent years. However, deep learning approaches are considered black box models due to the difficulty of understanding their predictions. We trained a CNN model based on the DenseNet architecture for vehicle classification in SAR images. The classification performance was investigated using the model’s learned latent space, and important features considered for predictions were visualized. Our proposed method achieved 24.88% accuracy in the testing phase of the PBVS 2022 Multi-modal Aerial View Object Classification (MAVOC) Challenge.