Road network extraction from airborne digital camera images: A multi-resolution comparison

Abstract

As image resolution increases from 10-30 m to 0.5-2 m, road networks will appear to be narrow areas rather than thin lines. This becomes a challenge for traditional linear analysis methods based on mask operations but creates an opportunity for classification based methods. We experimented with an advanced linear analysis, gradient direction profile analysis, and a few classification algorithms including a maximum classification, clustering and a contextual classifier for road network extraction using airborne digital camera data acquired over Livermore, California with approximately 1.6 m spatial resolution. Results indicate that both the linear extraction and image clustering algorithms worked reasonably well. Best road network results have been obtained by applying the linear extraction algorithm to a morphologically filtered image that was generated by combining the near infrared (NIR) and red (R) image bands through NIR/R+NIR. With this method, the correctly extracted road pixels account for 78.7% of the total road pixels obtained from image interpretation with field verification. The image clustering method resulted in 74.5% correctly extracted road pixels. When experimenting with the images resampled at approximately 3 m and 5 m resolution, the best overall accuracies for road extraction decreased to 74.6% and 61.6%, respectively.