Analysing Urban Dynamics

Projekt: Der zeitliche Wandel von Geodaten (BKG Gauss-Zentrum)

Forschende:  Mireille Fangueng, Frank Thiemann, Philipp Otto, Monika Sester

Projektidee: This project is part of an effort to better understand the spatial logic of urbanisation in relation to landscape features, in a context of rapid territorial transformation. By combining traditional statistical approaches and deep learning techniques, the aim is to identify the determining factors of residential development and assess their evolution between 2004 and 2023.

Why do people build houses in certain places? How do cities spread out? What role do certain landscape features like forests or rivers play? To answer such questions, we can look into the past and use historical maps to determine which factors may have been decisive for settlement. In addition to statistical methods, AI methods such as deep learning could also help to improve the resulting predictions for future development.

This project focuses on the Donauwörth region in Bavaria, a rural area characterised by diffuse urban growth and a highly heterogeneous landscape. The area was chosen because accurate spatial data was available and because the area is relevant for the analysis of urban dynamics.  To find out which factors are decisive for the development of residential areas and to evaluate their development between 2004 and 2023, the probability of the presence of residential areas at different points in time was modelled - based on their distance from certain landscape features (farms, forests, rivers).

The first step was the pre-processing of the data: Points with and without residential areas were identified on historical maps.  The distances to farms, forests and rivers were calculated for each point. Surrounding image patches centred on each pixel were also generated to feed SDDR models (semi structured deep distributional regression model) integrating the visual context.

For the analyses several statistical approaches like GLM (generalised linear model) and GAM (generalised additive model) and hybrid approach (SDDR models, combining deep neural network and GAM) were used. Together, they provided a better understanding of the spatial dynamics of urbanisation between 2004 and 2023. The results showed that distances to farms, forests and rivers have different effects on the probability of residential areas, with notable changes over time. People seem to prefer open areas near agricultural infrastructure for housing and to avoid the immediate vicinity of forest areas. GLM models show a stable and significant linear influence of distances to farms (positive effect) and forests (also positive).  On the other hand, proximity to rivers reveals a more variable influence, ranging from a moderately positive to a strongly negative effect depending on the year. The reasons for this could be that in certain places the risk of flooding is high and in others access to water is more of a priority.

The GAM model refines this interpretation by capturing non-linear relationships and complex spatial effects.  For example, at the beginning of the period, there is a strong attraction for areas less than 2,000 meters away from farms. However, this effect later diminishes. In addition, areas directly adjacent to forests are clearly rejected. Areas that are 4 kilometres or more away from forests, on the other hand, have a strong attraction. Rivers appear to be ambivalent elements, attractive at moderate distances but dissuasive at very short or long distances.

SDDR models, particularly those incorporating a spatial image context, provide a clear predictive gain. This confirms the assumption that it is favourable to combine continuous geographical data with contextual information from images. These results suggest that, while traditional models (GLM, GAM) remain relevant for interpreting underlying mechanisms and formulating hypotheses, deep learning approaches such as SDDR are more suitable for operational prediction, particularly in complex and heterogeneous spatial contexts. Together, these approaches highlight a trend towards greater complexity in urban location logic, marked by increasing consideration of environmental risks, resource accessibility and landscape preferences.