Assessing the transferability of road network extraction methods to Slovenian orthophotos for power distribution network planning

english

2026

volume 93, issue 4

benchmarking   domain shift   power distribution   road extraction   satellite imagery   spatial planning

Elektrotehniška zveza Slovenije
(Obvezni izvod spletne publikacije)

In spatial planning of power distribution networks, road corridors provide an important basis for routing medium- and low-voltage power cables, as they often simplify permitting procedures, construction, and maintenance access. Automatic road extraction from aerial and satellite imagery is therefore not only a segmentation task, but also an input to subsequent optimization procedures, where connected and navigable road networks are required. The paper compares six recent methods for road network extraction: SAM-Road, D-LinkNet, CRESI, CU-dGCN, Sat2Graph, and U-Net-ResNet18. They are first evaluated on the SpaceNet dataset and then tested for transfer to Slovenian GURS orthophoto imagery. The results show that methods with an explicit graph output generally preserve road network connectivity better in the experiments. When transferred to Slovenian GURS orthophotos, their relative performance is largely retained. The methods that perform well on the SpaceNet also tend to perform well in this setting. A decline in their performance is observed mainly in rural areas, where roads are often narrow, partly occluded by vegetation or shadows, and visually less distinguishable from the surrounding terrain. Besides a comparative evaluation, the paper therefore provides practical guidance for selecting methods when constructing road networks for the optimization-supported power distribution network planning.

20.10.2026