Machine Learning-based Regression for Characterization of Urban Environments with Sentinel-2

Abstract

Knowing the characteristics of urban environments is crucial for managing cities and planning an infrastructure which satisfies the needs of its inhabitants. This knowledge is also essential to assess the effects of natural hazards and, amongst others, it allows estimations about the amount of people that could be affected. For that reason up-to-date information about urban environments is of high interest. The average built-up height and the share of built-up area are two basic parameters which allow conclusions to be drawn about existing buildings and about the amount of people living in an area. However, a Manual recording of these parameters is often not possible due to too fast and uncontrolled urban growth, and many surveying techniques which provide a high accuracy, e.g. LIDAR, are too expensive for a comprehensive application. Thus, the aim of this master thesis was to develop a method which provides exhaustive, up-to-date and accurate information about both parameters mainly based on freely available data. In this study, satellite imagery recorded by ESA’s Sentinel-2 satellite was chosen as data basis, because it is freely accessible, globally available, has a high resolution of 10 meters and with a revisit time of five days it is always up-to-date. In the first step of the method presented several sets of features, like mathematical morphologies, textures and statistical features, were derived from Sentinel-2 scenes showing the urban areas of interest. Subsequently, based on these features the average built-up height and the share of built-up area were predicted with four different regression algorithms: random forest, gaussian process regression, neural network and support vector regression. Afterwards, the single predictions were combined into a final result via an ensemble learning technique. Within this study stacked generalization and local selection were applied as ensemble learning approaches and their performance was compared. Before the prediction the four regressors had been trained with the features calculated from the Sentinel-2 imagery and with reference data derived from TandDEM-X data, which is quite costly. However, since the aim of this study was to develop a low-priced method, the use of expensive training data for every prediction was contrary to the goal set. In order to overcome this drawback and reduce the necessary usage of TanDEM-X data, a domain adaptation procedure was integrated. In the domain adaptation process the four regressors were trained on a Sentinel-2 scene, the so-called source domain, where reference data were available. Afterwards, the regressors predicted the average built-up height and the share of built-up area for another scene, the target domain, where only Sentinel-2 imagery was present. In the end, the single predictions of the four different regressors were again combined via an ensemble learning procedure. Experimental results were obtained for the cities of Berlin, Cologne, Hamburg and Munich for which reference data were available. Each city was separately used as source domain for the other three cities so that the accuracy of the presented method could be assessed via the available reference data. Finally, the domain adaptation approach developed in this study had a mean absolute error (MAE) of 3.97 meters on average regarding the average built-up height and a MAE of 9.05 % on average regarding the share of built-up area. However, depending on the combination of source and target Domain the MAE can vary a lot and under optimal conditions a MAE of 1.23 meters or 3.73 % can be achieved.