Construction of a new similarity metric for domain adaptation-based characterization of built-up areas

Abstract

A more and more frequently encountered problem in Machine learning is the comparison of similarity between the feature spaces of different data sources. This is needed for example to evaluate the suitability of a data set for training a model which would then be applied to the second data set for predicting one or more target variables. In urban Remote Sensing, researchers encounter this problem when it comes to pursuing a Domain Adaptation (DA) classification approach, e.g. by investigating multiple cities. Normally, one would have to compare all combinations of data sets and models in order to select the optimal combination of source and target domain. This can serve e.g. to predict variables for which training data is only available for a different location. However, this requires large computational resources and long calculation times, in order to train different models, and apply various methods such as Features Selection, Boosting and Ensemble Learning. The goal of this thesis was therefore to develop a framework for comparing the feature spaces before training the models. This would allow to reduce the number of models to train to a single, optimized model in order to reach optimal performance. In this thesis, which is based on an existing master thesis, it was tried to predict the average building height and share of built-up area for 4 German cities (Berlin, Cologne, Hamburg, Munich) with data derived from TanDEM-X and validated with LOD-1 cadastral data on 3 different aggregation levels. For each of these combinations, we compared the performance of 2 models (RandomForest, SVM) with several metrics in form of combinations of existing statistical similarity and distance measures (e.g. χ2) between the two feature spaces. This is done via comparing the ranked rank sums of different metrics formed by measure grouping. In order to perform the total number of experiments without redundancies, an algorithm has been developed and implemented in R. Finally, the method was validated by comparing the numerical similarity with the similarity of the actual built-up structures in each city by investigating the cadastral data and identifying common and individual types of built-up structures. Despite the fact that the exponential runtime of our method prevented us from processing the entire amount of available data, it was possible to identify several metrics with good performances.