Urban classification from optical satellite images: a comparison between conventional machine learning and deep learning approaches

Kurzfassung

Urban classification is a challenging problem for many reasons; the diverse types of urban areas with different appearance in remotely sensed image data (residential areas, industrial infrastructures, sport facilities, etc.), the different look of urban areas from one country to another, or even within the same country. There have been many attempts to develop a global classifier for urban areas producing urban maps with various accuracy and diverse resolutions. To answer the question of whether it is possible to develop a single high accuracy global classifier which can classify urban areas worldwide, the first step would be to extract a good set of global features, the second would be to design a good generic classifier. In this study, both aspects of the problem were investigated. To this end, we followed two alternative approaches. The first alternative consists of conventional feature extraction (e.g., Gabor filtering, NDVI, local variance computation, etc.) followed by classification (e.g., Support Vector Machine). The second alternative is an end-to-end Deep Learning solution with no feature engineering. In our study, we explored and compared the performance of different feature extractors and different classifiers. As a first step towards solving the problem in a global scale, we will demonstrate typical cases of urban classification using LANDSAT ETM+ data of a large area in the eastern part of the USA, covering few cities and their surroundings in four side-by-side LANDSAT scenes. For ground truth we used settlement maps from national agency, in addition to the freely available data from OpenStreetMap. Our initial results show that Support Vector Machines (SVM) provide good results, while the deep learning model using Convolutional Neural Networks (CNNs) performs similarly well in local areas and generalizes better.