Generalization in deep learning-based aircraft classification for SAR imagery

Automatic Target Recognition (ATR) from Synthetic Aperture Radar (SAR) data covers a wide range of applications. SAR ATR helps to detect and track vehicles and other objects, e.g. in disaster relief and surveillance operations. Aircraft classification covers a significant part of this research area, which differs from other SAR-based ATR tasks, such as ship and ground vehicle detection and classification, in that aircrafts are usually a static target, often remaining at the same location and in a given orientation for longer time frames. Today, there is a significant mismatch between the abundance of deep learning-based aircraft classification models and the availability of corresponding datasets. This mismatch has led to models with improved classification performance on specific datasets, but the challenge of generalizing to conditions not present in the training data (which are expected to occur in operational conditions) has not yet been satisfactorily analyzed. This paper aims to evaluate how classification performance and generalization capabilities of deep learning models are influenced by the diversity of the training dataset. Our goal is to understand the model's competence and the conditions under which it can achieve proficiency in aircraft classification tasks for high-resolution SAR images while demonstrating generalization capabilities when confronted with novel data that include different geographic locations, environmental conditions, and geometric variations. We address this gap by using manually annotated high-resolution SAR data from TerraSAR-X and TanDEM-X and show how the classification performance changes for different application scenarios requiring different training and evaluation setups. We find that, as expected, the type of aircraft plays a crucial role in the classification problem, since it will vary in shape and dimension. However, these aspects are secondary to how the SAR image is acquired, with the acquisition geometry playing the primary role. Therefore, we find that the characteristics of the acquisition are much more relevant for generalization than the complex geometry of the target. We show this for various models selected among the standard classification algorithms.