Automated Railway Sleeper Inspection Integrating UAV Photogrammetry and Deep Learning: Towards an Intelligent Geospatial Monitoring System.

Railway sleeper inspection traditionally relies on manual visual surveys, which are time-consuming, hazardous, and subjective, particularly along extensive railway corridors. This study presents a preliminary integrated workflow combining UAV imagery, deep learning, and geospatial reporting for automated sleeper detection and visual condition assessment. High-resolution nadir imagery was acquired using a DJI Matrice 4E RTK-enabled UAV along a railway corridor in Lima, Peru. The current pilot dataset comprised 77 annotated UAV images, divided into 54 training images (70.1%), 9 validation images (11.7%), and 14 test images (18.2%). The test subset was obtained from a spatially distinct railway section not used during training or validation. A pretrained YOLOv8n model was fine-tuned to detect sleepers and classify their visually observable condition as good or damaged. On the independent test set, the model achieved an overall precision of 50.23%, recall of 72.86%, F1-score of 59.46%, mAP@0.50 of 52.88%, and mAP@0.50:0.95 of 32.20%. Performance was substantially higher for good-condition sleepers (mAP@0.50 = 87.59%) than for damaged sleepers (mAP@0.50 = 18.16%). The lower performance of the damaged class appears to be associated with greater visual variability of deterioration and frequent partial occlusion by ballast, indicating the need for additional and more diverse damaged-sleeper samples and improved annotation consistency. Potential missing sleepers are identified through a post-processing procedure in which distances between consecutive sleeper centroids are compared with local reference spacing based on median normal spacing, allowing anomalous gaps to be flagged as potential missing-sleeper candidates. Detection results are subsequently associated with UAV spatial metadata to support approximate geolocation and georeferenced reporting. These preliminary results demonstrate the feasibility of the proposed UAV–deep learning workflow while identifying damaged-sleeper recognition and missing-sleeper threshold calibration as the main areas for further improvement and validation.