Multi-Output Regression: On the Impact of Individual Model Parameters for Built-Up Height and Density Prediction

Kurzfassung

Urbanisation is an ongoing process and will gain importance in the future. It comes with multiple challenges as inhabitants are dependent on water and energy supply, a functioning street network and health care system | all these require a deliberate management. However, this is not an easy demand as administrative areas can cover several thousands of square kilometers. Therewith, remote sensing methods constitute a reliable source to observe large areas as cities. To observe the growth of cities, variables as built-up height and built-up density have emerged as reliable attributes that characterize well the urban morphology. They can be obtained by integrating remote sensing data from optical and other sensors such as synthetic aperture radar (SAR). The application of machine learning algorithms makes it feasible to interpret the large amount of data generated in remote sensing. This study focuses on the optimization of machine learning algorithms for predicting builtup height and built-up density in four German major cities based on remote sensing data, by integrating so-called multi-output regression (MOR) methods. Instead of processing and predicting each target variable independently, MOR methods incorporate all target variables into one process which, in the best case, increases the accuracy of predictions. Recent literature highlights the benefit of exploiting possible correlations between target variables. In this work, four methods are applied and modified according to state-of-the-art models: multi-target stacking (MTS), multi-target regressor chains (MTRC), multi-target regressor chains without repetitive permutation (MTRC-nrp) and single-target stacking (STS). Each method is used with four different regression models, namely random forest (RF), Gaussian process (GP), support vector regression (SVR) and neural networks (NN). Additionally, the impact of different stacking options as well as the impact of the feature space is evaluated. The extensive and systematic evaluation of the aforementioned parameters provides several insights. It shows, that all models (MTS, MTRC, MTRC-nrp, STS) outperform models that do not use multi-target stacking or chaining or single-target stacking. Furthermore, it shows that MOR models behave differently depending on which regression model is used for the prediction. Finally, it gives recommendations on which MOR methods and which additional parameters are suitable for particular use cases similar to those evaluated in this study and discusses possibilities for future research.