A Systematic Review of GeoAI and Deep Learning for Automated Road Infrastructure Damage Detection Using Satellite Imagery

Abstract

This study addresses the problem that research on automated road infrastructure damage detection using satellite imagery remains fragmented across damage categories, data sources, model architectures, and evaluation practices, making it difficult to form unified evidence base for infrastructure monitoring and decision support. The purpose of the study was to systematically synthesize how GeoAI and deep learning have been applied to road damage detection between 2018 and 2026, identify dominant data and model trends, compare case-based findings, and evaluate recurring methodological challenges. Methodologically, the study used a cross-sectional, case-based systematic literature review design with light quantitative synthesis rather than primary field experimentation. The sample consisted of published multi-country, region-specific, and disaster-context studies treated as analytical cases. Key variables included road-damage detection outcomes, satellite imagery characteristics, preprocessing and data quality conditions, model architecture, spatial context, and damage-mapping outputs. The analysis plan combined structured screening, eligibility assessment, extraction and coding of study characteristics, and thematic narrative synthesis supported by descriptive quantitative summaries. Headline findings showed that deep learning-based GeoAI has become the dominant paradigm, with very strong support for H1 and H3 and strong support for H2 and H4. Quantitatively, the RDD2020 benchmark included 26,336 road images and more than 31,000 damage instances across India, Japan, and the Czech Republic; a pixel-level satellite road-damage model achieved an F1 score of 76.09%; a transfer-learning road-quality study using 53,686 images covering 2,400 km reported 80.0% accuracy and 99.4% predictions within the true or adjacent class, rising to 94.0% after adaptation; post-earthquake road-damage detection over more than 530 km of roads reported 87.1% accuracy; and a semi-supervised GAN approach achieved 81.540% mean IoU and 79.228% F1. Overall, the findings imply that high-resolution imagery, strong labels, and context-aware architectures improve performance, but standardization, cross-region transferability, and deployment readiness remain major limitations for real-world geospatial infrastructure intelligence.