Automatic Building Footprint Extraction using Random Forest Algorithm from High Resolution Google Earth Images: A Feature-Based Approach

Abstract

Mapping building structures in urban areas are fundamental for many urban studies. High-resolution satellite images and machine learning algorithms are often used for semi-automatic and automatic building detection. This study explores the utility of classification algorithm and high-resolution Google Earth images to detect buildings from a complex urban area having heterogeneous building structures of various shapes, sizes and construction materials. The proposed model converts the RGB images to grayscale followed by derivation of several filtered profiles to capture spatial and spectral feature vectors. Differential morphological profiles (DMPs) were constructed from consecutive morphological profiles to identify the structural information of probable buildings present in the image. Consequently, morphological building index was computed by averaging the DMPs that included mislabeled rooftops due to the presence of shadows. Global Otsu thresholding was used to reduce the number of mislabeled buildings. Finally, Random Forest algorithm was used for the classification of buildings based on the computed feature vectors. After noise removal, the quality of final building maps were assessed using evaluation metrics with reference to the ground truth which showed 4.4% improvement (from 81.2 to 85.62 %) over RF model while also revealing the detailed structural information of buildings.