Scaling limits of federated learning for vision-based road damage detection across countries

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Federated learning enables cross-country collaboration for training vision-based road damage detection models without sharing raw data. However, empirical guidance on federation configuration and client composition remains limited. Specifically, the effects of increasing federation size and cross-country heterogeneity, including differences in imaging platforms and pavement and climate conditions, on detection accuracy, cross-test stability, and minority-class performance (scarcity of some damage types) are not well quantified. In this context, we evaluate progressive multi-country federated YOLOv8 by incrementally integrating clients across road damage datasets from six countries and multiple sensing modalities. We observe diminishing returns as federation scale increases: a four-country configuration matched the mean mAP@50 of a five-country setup (0.41 vs. 0.40) while exhibiting lower variability across test sets (0.329 vs. 0.334), indicating more stable aggregation. To characterize the underlying heterogeneity, we quantify geographic and acquisition-related domain shifts using cosine similarity of intermediate feature representations. To mitigate minority-class scarcity, we propose and evaluate synthetic augmentation, which improves detection under severe imbalance but yields negative gains when inter-domain feature alignment is limited. The results characterize scaling limits in federated road inspection systems and support compatibility-aware client selection and domain-aware imbalance mitigation as actionable deployment strategies. Observations are specific to the evaluated datasets, model architecture, and aggregation configuration. Beyond road damage detection, the findings inform engineering-informatics practice for configuring distributed, data-driven infrastructure monitoring systems.