Artificial intelligence for resilient transportation infrastructure: the first Ecuadorian pavement damage dataset and an assessment of domain transfer across international road networks
Journal
Frontiers in Built Environment
ISSN
22973362
Date Issued
2026-08-26
Author(s)
Type
Article
Abstract
Background
Pavement-condition monitoring is a cornerstone of resilient and sustainable transportation infrastructure, yet existing public datasets for automatic damage detection are dominated by data from high-income countries. Deep-learning models trained on those data often transfer poorly to Latin American road networks, where deterioration patterns, materials, and maintenance regimes differ. No publicly available pavement damage dataset for Ecuador has been reported to date, to the best of our knowledge.
Methods
We present the Ecuadorian Pavement Distress Dataset (EcuPDD), 1,295 images annotated under ASTM D6433 into six damage classes (4,559 annotated instances), collected across seven cities in the three geographic regions of Ecuador (Sierra, Coast, and Amazon). The dataset is released under CC BY 4.0 on Zenodo (DOI:
10.5281/zenodo.20454801
). On this base we trained five YOLOv8n variants progressively fused with three open international datasets—RDD2022 Japan, RDD2022 China_MotorBike, and McMaster Pavement Crack Detection—for a total corpus of up to 15,941 images (of which roughly 80% were used for training and 20% for validation, following the splits described in Section 3). Each version was trained five times with independent random seeds (25 runs in total), and results are reported as mean ± standard deviation. We evaluated each version on its mixed validation set and, separately, on a fixed 259-image Ecuadorian held-out subset. We complemented the standard mAP50/ mAP50-95 metrics with Wilson 95% confidence intervals on per-class detection rates, a Gini coefficient for class imbalance, and a Domain Gap Index (DGI) that we define as 1 − (mAP50Ecuador/ mAP50general), summarising the divergence between the general and local performance of each model.
Results
Adding international data lifted the mixed-validation mAP50 from 14.65% (V1, Ecuador only) to 57.66% (V5, Ecuador + McMaster + China_MotorBike), an absolute gain of 43.01 percentage points. However, when evaluated only on the 259 Ecuadorian held-out images, mAP50 stayed between 14.65% and 15.90% across all five versions—a swing of 1.25 pp despite the inclusion of up to 14,646 additional images. The Wilson 95% confidence intervals on those Ecuadorian-only mAP50 values overlap heavily for all versions, indicating substantial overlap and no clear evidence of separation among versions at this sample size. The Domain Gap Index reaches 0.733 for V5 and 0.713 for V2, while remaining ∼0 for V1 and V3. Patching, despite contributing only 2.3% of the instances, reached a Wilson-CI-supported detection rate of 0.78 [0.73, 0.82] thanks to its strong visual distinctiveness when paired with the McMaster dataset.
Conclusion
International datasets improve a YOLOv8n detector globally but do not, by themselves, close the domain gap to Ecuadorian roads. The bottleneck is the volume and representativeness of local annotated data. EcuPDD provides the first openly licensed Ecuadorian benchmark for this task and an empirical calibration point that researchers in Latin America and other under-represented regions can use to plan dataset-construction and model-development efforts aligned with the resilience and sustainability goals of the AI-for-infrastructure agenda.
Pavement-condition monitoring is a cornerstone of resilient and sustainable transportation infrastructure, yet existing public datasets for automatic damage detection are dominated by data from high-income countries. Deep-learning models trained on those data often transfer poorly to Latin American road networks, where deterioration patterns, materials, and maintenance regimes differ. No publicly available pavement damage dataset for Ecuador has been reported to date, to the best of our knowledge.
Methods
We present the Ecuadorian Pavement Distress Dataset (EcuPDD), 1,295 images annotated under ASTM D6433 into six damage classes (4,559 annotated instances), collected across seven cities in the three geographic regions of Ecuador (Sierra, Coast, and Amazon). The dataset is released under CC BY 4.0 on Zenodo (DOI:
10.5281/zenodo.20454801
). On this base we trained five YOLOv8n variants progressively fused with three open international datasets—RDD2022 Japan, RDD2022 China_MotorBike, and McMaster Pavement Crack Detection—for a total corpus of up to 15,941 images (of which roughly 80% were used for training and 20% for validation, following the splits described in Section 3). Each version was trained five times with independent random seeds (25 runs in total), and results are reported as mean ± standard deviation. We evaluated each version on its mixed validation set and, separately, on a fixed 259-image Ecuadorian held-out subset. We complemented the standard mAP50/ mAP50-95 metrics with Wilson 95% confidence intervals on per-class detection rates, a Gini coefficient for class imbalance, and a Domain Gap Index (DGI) that we define as 1 − (mAP50Ecuador/ mAP50general), summarising the divergence between the general and local performance of each model.
Results
Adding international data lifted the mixed-validation mAP50 from 14.65% (V1, Ecuador only) to 57.66% (V5, Ecuador + McMaster + China_MotorBike), an absolute gain of 43.01 percentage points. However, when evaluated only on the 259 Ecuadorian held-out images, mAP50 stayed between 14.65% and 15.90% across all five versions—a swing of 1.25 pp despite the inclusion of up to 14,646 additional images. The Wilson 95% confidence intervals on those Ecuadorian-only mAP50 values overlap heavily for all versions, indicating substantial overlap and no clear evidence of separation among versions at this sample size. The Domain Gap Index reaches 0.733 for V5 and 0.713 for V2, while remaining ∼0 for V1 and V3. Patching, despite contributing only 2.3% of the instances, reached a Wilson-CI-supported detection rate of 0.78 [0.73, 0.82] thanks to its strong visual distinctiveness when paired with the McMaster dataset.
Conclusion
International datasets improve a YOLOv8n detector globally but do not, by themselves, close the domain gap to Ecuadorian roads. The bottleneck is the volume and representativeness of local annotated data. EcuPDD provides the first openly licensed Ecuadorian benchmark for this task and an empirical calibration point that researchers in Latin America and other under-represented regions can use to plan dataset-construction and model-development efforts aligned with the resilience and sustainability goals of the AI-for-infrastructure agenda.