Remote Sensing, Vol. 15, Pages 2099: State-Level Mapping of the Road Transport Network from Aerial Orthophotography: An End-to-End Road Extraction Solution Based on Deep Learning Models Trained for Recognition, Semantic Segmentation and Post-Processing with Conditional Generative Learning
Remote Sensing doi: 10.3390/rs15082099
Authors: Calimanut-Ionut Cira Miguel-Ángel Manso-Callejo Ramón Alcarria Borja Bordel Bordel Sánchez Javier González González Matesanz
Most existing road extraction approaches apply learning models based on semantic segmentation networks and consider reduced study areas, featuring favorable scenarios. In this work, an end-to-end processing strategy to extract the road surface areas from aerial orthoimages at the scale of the national territory is proposed. The road mapping solution is based on the consecutive execution of deep learning (DL) models trained for ① road recognition, ② semantic segmentation of road surface areas, and ③ post-processing of the initial predictions with conditional generative learning, within the same processing environment. The workflow also involves steps such as checking if the aerial image is found within the country’s borders, performing the three mentioned DL operations, applying a decision limit to the class predictions, or considering only the central 75% of the image to reduce prediction errors near the image boundaries. Applying the proposed road mapping solution translates to operations aimed at checking if the latest existing cartographic support (aerial orthophotos divided into tiles of 256 × 256 pixels) contains the continuous geospatial element, to obtain a linear approximation of its geometry using supervised learning, and to improve the initial semantic segmentation results with post-processing based on image-to-image translation. The proposed approach was implemented and tested on the openly available benchmarking SROADEX dataset (containing more than 527,000 tiles covering approximately 8650 km2 of the Spanish territory) and delivered a maximum increase in performance metrics of 10.6% on unseen, testing data. The predictions on new areas displayed clearly higher quality when compared to existing state-of-the-art implementations trained for the same task.