Multi-spectral and Topographic Fusion for Automated Road Extraction

Abstract

AbstractRoad geometry is pertinent information in various GIS studies. Reliable and updated road information thus calls for conventional on-site survey being replaced by more accurate and efficient remote sensing technology. Generally, this approach involves image enhancement and extraction of relevant features, such as elongate gradient and intersecting corners. Thus far, its implication is often impeded by wrongly extraction of other urban peripherals with similar pixel characteristics. This paper therefore proposes the fusion of THEOS satellite image and topographic derivatives, obtained from underlying Digital Surface Models (DSM). Multi-spectral indices in thematic layers and surface properties of designated roads were both fed into state-of-the-art machine learning algorithms. The results were later fused, taken into account consistently leveled road surface. The proposed technique was thus able to eliminate irrelevant urban structures such as buildings and other constructions, otherwise left by conventional index based extraction. The numerical assessment indicates recall of 84.64%, precision of 97.40% and overall accuracy of 97.78%, with 0.89 Kappa statistics. Visual inspection reported herewith also confirms consistency with ground truth reference.