Quantifying the Impact of Data Augmentation on Road Segmentation Performance in Deep Learning Models Julio Jeremia Sinabutar (a*), Jiann-Yeou Rau (a)
a) Department of Geomatics, National Cheng Kung University
No.1, Daxue Rd, East Dist, Tainan City 701, Taiwan
*juliojeremia46[at]gmail.com
Abstract
This study investigates the impact of data augmentation on road segmentation performance using deep learning models trained on UAV imagery. The dataset consisted of 894 training images and 100 validation images, sourced from 3 geographically diverse locations, i.e., Tainan City (Taiwan), Kendari City (Indonesia), and Yogyakarta City (Indonesia). Model performance was evaluated on a high-resolution true orthophoto from Pematangsiantar City (Indonesia). A total of 40 model configurations were combined from various segmentation architectures, encoder backbones, and two probability thresholds for evaluation (0.05 and 0.5). Data augmentation covered 100% of the training data, distributed across saturation (1/3), hue (1/3), darkening (1/6), and brightening (1/6) transformations. Based on the Intersection over Union (IoU) metric, 26 out of 40 models show improved performance after augmentation, while the remaining 14 experienced a decrease. On average, data augmentation results in a 4.185% increase in IoU. A one-tailed paired t-test confirmed that this improvement was statistically significant for a 95% confidence interval (p = 0.02056), and Cohen^s d of 0.33 indicated a small effective size. These results suggest that even modest augmentation strategies can enhance overall segmentation performance in UAV-based road extraction tasks, although the degree of benefit may vary across model configurations.
Keywords: Cohen^s d- Data augmentation- Deep learning- Paired t-test- Road segmentation
