Unsupervised Multispectral Gaussian Mixture Model-Based Framework for Road Extraction

Abstract

The inherent composition of roads and buildings project spectral and hierarchically similar characteristics in remote-sensing images. Gray values of both background pixels and roads overlap when a large area of a remote-sensing image is considered. As a consequence, segmenting road networks and buildings in an urban environment presents critical challenges. So far, the literature suggests that supervised algorithms outperform their unsupervised counterparts when it comes to segmenting roads and buildings. However, supervised algorithms require a massive database in the training stage. This can cause a bottleneck as the percentage of pixels in urban remote sensing images depicting roads is very low when compared to the background. Index integrated spatially constrained Gaussian Mixture model (IISC-GMM), a novel unsupervised algorithm that overcomes the aforementioned constraints by integrating a Morphological Building Index (MBI) mask with a novel Gaussian mixture model (GMM) is proposed. To better distinguish foreground from background pixels, this novel algorithm blends localized spatial smoothness of neighboring pixels with spectral information. The gaps in the road network are eliminated by applying path morphology. The algorithm generates a Dice coefficient of 80.00%, a Completeness of 77.41%, a Correctness of 82.75%, a Quality of 73.80%, and a Misclassification rate (MCR) of 11.36% when validated on the Massachusetts Road dataset. In addition to being faster and less computationally intensive, the results obtained by IISC-GMM are comparable to those obtained by the computationally intensive Deep Learning methods.