Towards Smarter Cities: Federated Learning and CNN for Pothole Detection

One specific need often associated with the maintenance of urban public assets, including roads, is the identification of potholes as part of maintaining the structures’ reliable, safe condition. This research presents an FL-CNN model that eliminates the limitations of the traditional approaches used for pothole detection. Our technique leverages the feature extraction strength of CNN and the privacy-preserving advantage of FL to enhance detection accuracy without compromising the patients’ data. The CNN-FL model was employed with test results and data collected from various cities containing a broad spectrum of road conditions and circumstances of the environment. The model got an accuracy of 91. 01%, exactness of 92. 1%, while the recall value registered an average of 87. 5%, F1-score of 88. 3%, which also outperforms the original centralized CNN model’s performance. The latter reached an accuracy of 88.5 % and a precision of 85. 3% and the recall capacity of 82. While the Ridge classifier attained the highest accuracy level of 76. 9%, precision of 7%, and an F1-score of 84. 0%. Hence, Federated Learning (FL) enables distributed training models while ensuring the original data does not leave the client’s device. Thus, it guarantees privacy and security. It is also ideal for broad applications that can be carried out in various large metropolitan areas. Additionally, the high evaluation of accuracy and recall shows that the model efficiently identifies potholes with fewer false positives and negatives.