Noise-Aware Data Augmentation for Robust Road Detection in Small Satellite Imagery

Road extraction from small satellite imagery is challenging because raw images often suffer from low signal-to-noise ratio (SNR), high radiometric variability, and reduced sharpness. In this work, we investigate whether noise and blur data augmentation during pretraining can improve robustness in such conditions. We use a two-stage transfer-learning framework in which a U-Net with a ResNet-50 encoder is pretrained on PlanetScope RGB imagery and fine-tuned on NEMO-HD imagery. During pretraining, we evaluate Gaussian, ISO-like, and Perlin noise, as well as Gaussian and motion blur, each at three severity levels. On the internal held-out test split, augmentation effects were modest, with the best strict IoU improving from 26.7% for the geometric-only baseline to 27.2%. However, evaluation on external full-scene NEMO-HD images showed clearer benefits. Augmentation-based models consistently improved road detection in raw imagery, mainly by increasing completeness and recall, while there was little or no systematic benefit in stacked imagery. No clear trend was observed across augmentation severity levels, indicating that performance depended more on scene conditions than on perturbation strength. The results show that augmentation is most useful for single-acquisition small satellite imagery, where it improves robustness to lower image quality.