This paper presents a federated learning (FL) approach for global road damage detection, combining FL with YOLOv5l to develop models for single- and multi-country applications. The study addresses the challenges of centralized model training, including data privacy concerns, large data transfers, and computational resource requirements. FL allows models to be trained without sharing data, only exchanging model parameters between clients and a central server. The FL-based models achieved 21%–25% lower mean average precision (mAP) than centralized models but outperformed local client models by 1.33%–163% on global test data. The research uses the Road Damage Dataset 2022 (RDD2022) to create data sets with different data distributions for single- and multi-country applications. The dataset includes road damage images from six countries, with annotations for four types of road damage. For single-country detection, the Japan road damage dataset (JRDD) was used, while for multi-country detection, the multicountry road damage dataset (MRDD) was used, which includes data from Japan, India, and the United States. The study compares the performance of centralized and FL models using mAP@IoU = 0.5 as the main metric. FL models were trained using the Flower framework with FedAvg as the server aggregation method. The results showed that FL models trained on non-IID data had lower mAP than those trained on IID data. However, when evaluated on global test data, FL models outperformed local country models, demonstrating better generalization. The performance of FL models was influenced by factors such as data distribution, data quality, and the number of clients. The study also found that FL models trained on non-IID data performed better than local country models when evaluated on global test data. Additionally, FL models trained on multiple countries can be used to develop more robust and generalizable road damage detection models without accessing data from individual countries. The research highlights the potential of FL for global road damage detection, particularly in scenarios where data privacy is a concern or data sharing is not feasible. The results suggest that FL can provide a more efficient and privacy-preserving solution for road damage detection across multiple countries. Future work could explore different server aggregation methods, object detection algorithms, and advanced supervised learning techniques to further improve FL performance for road damage detection.This paper presents a federated learning (FL) approach for global road damage detection, combining FL with YOLOv5l to develop models for single- and multi-country applications. The study addresses the challenges of centralized model training, including data privacy concerns, large data transfers, and computational resource requirements. FL allows models to be trained without sharing data, only exchanging model parameters between clients and a central server. The FL-based models achieved 21%–25% lower mean average precision (mAP) than centralized models but outperformed local client models by 1.33%–163% on global test data. The research uses the Road Damage Dataset 2022 (RDD2022) to create data sets with different data distributions for single- and multi-country applications. The dataset includes road damage images from six countries, with annotations for four types of road damage. For single-country detection, the Japan road damage dataset (JRDD) was used, while for multi-country detection, the multicountry road damage dataset (MRDD) was used, which includes data from Japan, India, and the United States. The study compares the performance of centralized and FL models using mAP@IoU = 0.5 as the main metric. FL models were trained using the Flower framework with FedAvg as the server aggregation method. The results showed that FL models trained on non-IID data had lower mAP than those trained on IID data. However, when evaluated on global test data, FL models outperformed local country models, demonstrating better generalization. The performance of FL models was influenced by factors such as data distribution, data quality, and the number of clients. The study also found that FL models trained on non-IID data performed better than local country models when evaluated on global test data. Additionally, FL models trained on multiple countries can be used to develop more robust and generalizable road damage detection models without accessing data from individual countries. The research highlights the potential of FL for global road damage detection, particularly in scenarios where data privacy is a concern or data sharing is not feasible. The results suggest that FL can provide a more efficient and privacy-preserving solution for road damage detection across multiple countries. Future work could explore different server aggregation methods, object detection algorithms, and advanced supervised learning techniques to further improve FL performance for road damage detection.