Data-driven Feature Learning for High Resolution Urban Land-Cover Classification

Abstract

Automated classification of aerial and satellite images is one of the fundamental challenges in remote sensing research. Over the last 30 years, researchers have tried to overcome the tedious and time consuming manual interpretation of images. With the advent of digital technologies, classification approaches facilitating image interpretation have emerged. They were quickly embraced, and nowadays classification of remote sensing imagery is a mature field with many well-established methods. However, a major yet largely unsolved problem is the design and selection of features, that would be appropriate for a specific classification task. Usually, it is not known in advance which image features would help separating object classes in an optimal way and manual feature by trial and error is still a common practice. In the last decade rapid development of remote sensing sensors gave the end-user access to very high resolution imagery. At a ground sampling distance below a meter, small objects and fine-grained texture of larger objects emerge. Thus, to properly exploit the information that these images contain, additional contextual and textural properties of objects should be extracted. Unfortunately, classification of such images is often performed using features tailored to lowand medium resolution sensors: raw pixel values, usually augmented with either simple band ratios (e.g. in form of vegetation indices), or specific texture filter banks (e.g. Gabor filters). In this thesis we consider the problem of feature design and selection for classification of urban land-cover from very high resolution (VHR) remote sensing images. To appropriately capture characteristic object patterns, we propose a set of simple and efficient features, called random quasi-exhaustive (RQE) feature bank. It consists of a multitude of multiscale texture features computed efficiently via integral images inside a sliding window. At the same time, we propose to sidestep manual feature selection, and let a boosting classifier choose only those features from a RQE feature bank that are able to efficiently discriminate between different object classes in a specific classification task. We believe that the proposed feature set is fairly generic to many urban remote sensing datasets, such that the features selected by the classifier can be adapted to the characteristics of a certain image: different lighting or different scene structures. We start with presenting the developed framework for supervised classification of land-cover in urban environments. We demonstrate the efficiency of a boosting classifier used in conjunction with the RQE feature databank on five different very high resolution remote sensing datasets. Next, we move from supervised feature learning to unsupervised methods. Using random forest classifier, we investigate the performance of features extracted using data-driven methods, such as principal component analysis (PCA) or Deep Belief Networks (DBN). We show that, at least in our study, complex unsupervised and non-linear feature learning did not improve classification accuracy over standard linear baseline methods. Finally, we use the developed supervised classification framework for an application in the field of urban hydrology. We produce imperviousness maps, which are then used to model rainfall-runoff processes in urban catchments. We show that the proposed method yields results superior over state-of-the-art methods in the field of urban hydrology. Furthermore, we perform an end-to-end comparison, in which different image data sources produced using different classification methods are used as an input for a hydraulic sewer model.