Urban Terrain Segmentation Using Multispectral Satellite Imagery

Urban terrain classification is an important yet challenging problem that lies at the crossroads of remote sensing, computer vision and machine learning. The rapid growth of urban areas and urban population over the past century has made urban terrain studies an indispensable part of climate change research. Terrain classification has been utilized in many climate change-related studies, but only few attempts have been made to implement terrain classification on complex urban landscapes. The lack of labelled high-resolution satellite datasets proportional to the scale of urban terrain and artefacts is one of the major challenges in urban terrain classification. This project exploits multi-spectral Google Earth Engine satellite images paired with three customized spectral bands to build simple machine learning models to identify and segment various types of urban terrains including buildings, trees, roads, grass, and greenery and water bodies. The random forest classifier was able to achieve the highest accuracy of 80.35% on the unseen test set.