An Explainable Smartphone-Based Deep Learning and Geospatial Framework for Road-Defect Surveillance and Maintenance Prioritisation in Rwanda: A Synthetic Proof of Concept

Source abstract

Road infrastructure is central to mobility, public safety, access to services and local economic development. Nevertheless, recurrent pavement defects, drainage obstruction and rain-related surface deterioration can outpace periodic manual inspection, particularly where inspection resources are constrained. This paper proposes a smartphone-based deep-learning and geospatial framework for detecting, classifying and prioritising road defects in Rwanda. The framework combines vehicle-mounted smartphone imagery, global positioning data, lightweight object detection and segmentation models, civil engineering severity rules and a geospatial maintenance dashboard. A reproducible synthetic proof of concept was implemented using 600 training images and 200 test images across normal-road, pothole, crack and standing-water classes. A two-hidden-layer neural net-work achieved 79.0% test accuracy and a 79.3% macro F1-score. Potholes obtained the strongest F1-score (88.4%), whereas crack recall was limited to 66.0%, exposing a mate-rial false-negative risk. These results establish software functionality but do not estimate performance on Rwandan roads. The paper synthesises computer-vision research with infrastructure-management and development perspectives and specifies a locally grounded data-collection, annotation, validation, explainability and governance protocol. Its principal contribution is not merely automated defect recognition, but a decision-support architecture linking low-cost sensing to accountable public maintenance. The framework and accompanying programme offer a feasible basis for field research, municipal piloting and scalable digital road stewardship in Rwanda and comparable low-resource settings.