A Comprehensive Evaluation of Deep Vision Transformers for Road Extraction from Very-high-resolution Satellite Data
The result's identifiers
Result code in IS VaVaI
<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216305%3A26220%2F25%3APU154965" target="_blank" >RIV/00216305:26220/25:PU154965 - isvavai.cz</a>
Result on the web
<a href="https://www.sciencedirect.com/science/article/pii/S2666017224000749" target="_blank" >https://www.sciencedirect.com/science/article/pii/S2666017224000749</a>
DOI - Digital Object Identifier
<a href="http://dx.doi.org/10.1016/j.srs.2024.100190" target="_blank" >10.1016/j.srs.2024.100190</a>
Alternative languages
Result language
angličtina
Original language name
A Comprehensive Evaluation of Deep Vision Transformers for Road Extraction from Very-high-resolution Satellite Data
Original language description
Transformer-based semantic segmentation architectures excel in extracting road networks from very-high-resolution (VHR) satellite images due to their ability to capture global contextual information. Nonetheless, there is a gap in research regarding their comparative effectiveness, efficiency, and performance in extracting road networks from multicity VHR data. This study evaluates 11 transformer-based models on three publicly available datasets (DeepGlobe Road Extraction Dataset, SpaceNet-3 Road Network Detection Dataset, and Massachusetts Road Dataset) to assess their performance, efficiency, and complexity in mapping road networks from multicity VHR satellite images. The evaluated models include Unified Perceptual Parsing for Scene Understanding (UperNet) based on the Swin transformer (UperNet-SwinT), and Multi-path Vision Transformer (UperNet-MpViT), Twins transformer, Segmenter, SegFormer, K-Net based on SwinT, Mask2Former based on SwinT (Mask2Former-SwinT), TopFormer, UniFormer, and PoolFormer. Results showed that the models recorded mean F-scores (mF-score) ranging from 82.22% to 90.70% for the DeepGlobe dataset, 58.98% to 86.95% for the Massachusetts dataset, and 69.02% to 86.14% for the SpaceNet-3 dataset. Mask2Former-SwinT, UperNet-MpViT, and SegFormer were the top performers among the evaluated models. The Mask2Former, based on the SwinT, demonstrated a strong balance of high performance across different satellite image datasets and moderate computational efficiency. This investigation aids in selecting the most suitable model for extracting road networks from remote sensing data.
Czech name
—
Czech description
—
Classification
Type
J<sub>imp</sub> - Article in a specialist periodical, which is included in the Web of Science database
CEP classification
—
OECD FORD branch
20202 - Communication engineering and systems
Result continuities
Project
—
Continuities
S - Specificky vyzkum na vysokych skolach
Others
Publication year
2025
Confidentiality
S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů
Data specific for result type
Name of the periodical
Science of Remote Sensing
ISSN
2666-0172
e-ISSN
—
Volume of the periodical
11
Issue of the periodical within the volume
9
Country of publishing house
NL - THE KINGDOM OF THE NETHERLANDS
Number of pages
19
Pages from-to
„“-„“
UT code for WoS article
001399325200001
EID of the result in the Scopus database
2-s2.0-85214196971