Abstract. 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.
Nina Krašovec, Aleš Marsetič· The International Archives o...· 0 citations
Abstract. In recent decades, many different satellites for Earth observation have been launched. They produce large amounts of data that, if properly preprocessed, can be used in many applications. A rapidly growing portion of these data comes from small satellites, which remain underused by scientists and entrepreneurs. However, greater utilisation can be ensured by producing images with positional accuracy of at least two pixels, which is necessary for their reliable use. This paper presents the upgraded version of the geometric correction module of the STORM processing chain. The module can automatically orthorectify images from the NEMO-HD small satellite, which, like other small satellites, in principle has a lower signal-to-noise ratio (SNR) and higher radiometric variability. It automatically extracts ground control points (GCPs) by matching freely available reference vector roads and reference images to roads extracted from the satellite image using a deep learning method trained on PlanetScope and NEMO-HD imagery. The performance of the geometric correction module was evaluated using three images acquired over Slovenia. The road extraction method can achieve an F1-score of approximately 60%. The tests demonstrated that automatic GCP extraction based on roads detected by a deep learning method is a viable approach for achieving geometric model accuracies of two pixels or less at independent check points when using small satellite imagery.
Aleš Marsetič, P. Pehani, Nina Krašovec· The International Archives o...· 0 citations