Assessment of AI classifier robustness under atmospheric effects
Alexander Schwegmann, Claudia S. Huebner · 2025
Atmospheric effects such as haze, turbulence, and scattering significantly degrade the visual quality of images captured and thus the performance of computer vision systems, especially in long-range surveillance and defense application with AI-based detection and classification systems, which are often trained on high-quality or digitally augmented data. While human observers can often still recognize distorted objects, convolutional neural networks (CNNs) may fail unpredictably. In this study, we systematically evaluate 18 different CNN classifiers ranging from early architectures (AlexNet, VGG) to state-of-the-art models (NASNet) on eight video sequences of a military jet recorded under different atmospheric conditions. Each sequence contains 2500 frames at four distances (800m, 1300m, 1800m, 2300m) and two times of day (morning, noon), resulting in varying turbulence levels. We find that turbulence strongly affects classifier confidence and stability. Networks can be grouped into four categories: (1) models that fluctuate heavily (confidence oscillating between 10–80%), (2) models resilient to mild turbulence, (3) models failing already at weak turbulence, and (4) one model (NASNet) that remains surprisingly stable across conditions. These results highlight the vulnerability of CNNs to turbulence-induced distortions and motivate future research on algorithmic turbulence correction versus turbulence-aware training strategies to enhance operational robustness.