Abstract

Building extraction from aerial imagery facilitates many geo-specialized tasks like Urban Planning, Map Generation and Disaster Management. Well planned cities ensure good sanitation, lesser pollution and hence, a better standard of living for its citizens. This is essential for developing countries which face a major crisis of urban migration and space crunch, and where planned cities would be a move towards smart living. The objective of this work is to segment building footprints from aerial images. Traditional pixel clustering algorithms like K-means, Color Quantization (CQ) and Gaussian Mixture Model (GMM) are implemented with inclusion of preprocessing steps for improved performance. These techniques are compared based on performance and time taken. The number of clusters/components are selected on the basis of Silhouette Score and Akaike Information Criterion/ Bayesian Information Criterion (AIC/BIC). A commonly encountered problem in building segmentation is misclassification of pixels due to shadows. This challenge is dealt by masking shadows using morphological operations as a part of preprocessing.