What question did this study set out to answer?

This review aims to evaluate and compare various techniques for road segmentation, extraction, and classification from satellite imagery.

May 26, 2026Open Access

A comprehensive review of road segmentation, extraction, and classification from satellite imagery

Key Points

This review aims to evaluate and compare various techniques for road segmentation, extraction, and classification from satellite imagery.
Comparative analysis of conventional and contemporary road segmentation techniques.
Examination of datasets like SpaceNet and performance metrics including F1-score and IoU.
Discussion on challenges such as occlusion and the scalability of vector-based methods.
Deep learning approaches achieved F1-scores ranging from 0.78 to 0.94 and IoU values between 0.65 and 0.83.
Vector-based methods exhibited precision and recall frequently surpassing 0.94, yet faced scalability issues.
Road classification F1-scores varied from 0.71 to 0.97, emphasizing the effectiveness of CNN and transformer models in well-annotated conditions.

Abstract

Road segmentation, extraction, and classification from satellite imagery are integral parts of geospatial analysis, enabling applications in urban planning, navigation, disaster management, and autonomous systems. This review provides a comparative view of the latest and conventional techniques in road segmentation, extraction, and classification, emphasizing the transition from the initial morphological and ruled based techniques to the latest machine learning and deep learning techniques. We review methods such as edge detection, object-based image analysis, Markov Random Fields, and Support Vector Machines, and then recent developments with convolutional neural networks and attention mechanisms. The most important datasets, metrics, and challenges, such as occlusion, road textures, and multi scale appearance, are also presented. Results from major datasets including SpaceNet, Massachusets Roads, OpenStreetMap derived data, and DeepGlobe, show that deep learning based road extraction approach achieve F1-score ranging from 0.78 to 0.94, with IoU value between 0.65 and 0.83; multispectral based models outperform RGB. Vector-based methods show high levels of precision and recall, frequently surpassing 0.94, yet they face challenges related to scalability and automation. In road classification, a F1-scores variation ranging from 0.71 to 0.97, demonstrate the efficacy of CNN and transformer-based models when conditions are well-annotated. In addition, road segmentation research shows IoU values reaching as high as 0.79 and Kappa coefficients exceeding 0.80, proving strong pixel-level consistency in controlled environments. Even with these encouraging outcomes, obstacles remain in maintaining network structure, addressing variations between different areas and sensors, and reducing inaccuracies in user-generated annotations. The review ends with an outlook for current research trends and future work for strong and scalable road extraction systems.

Ask AI

Helpful

Bookmark

View Full Paper