Characterization of urban land use and urban land dynamics in data scarce regions

The aim of this thesis is to advance the characterization of urban land and urban land dynamics to further the understanding of urban areas in data scarce regions. In order to achieve this, this thesis will answer the following research questions: RQ1: How can the integration of satellite imagery and socioeconomic data contribute to mapping urban land use (at a large spatial scale)? RQ2: How can we further the understanding of urban structure in cities located in data scarce regions? RQ3: What insights do building-level changes provide in urban dynamics? Chapter 2 uses a combination of open-source satellite imagery and socioeconomic data to classify urban land use at a national scale, using a deep learning approach. Combining Sentinel-2 and Sentinel-1 imagery with statistics from POIs and road networks increases the overall accuracy of the classification for the Netherlands by 3 percentage points. The Netherlands was divided into four regions, to test whether the combination of data types increased the transferability of the approach. When trained on three regions and tested on the independent fourth one the results showed a clear increase in classification accuracy between 3 and 5 percentage points relative to only using satellite data. However, when trained on one region and testing on another the results varied more strongly with differenced between 0 and 9 percentage points. Chapter 3 produces urban land use maps of three East African cities from satellite imagery and building footprint data. This chapters shows that the required amount of reference data needed when classifying new cities, using a combination of data sources, can be reduced by an order of magnitude by using a transfer learning approach. The combination of freely available PlanetScope satellite data, and publicly available Google building footprint data means that this approach is more cost effective compared to using VHR imagery. Despite using lower resolution imagery, the achieved classification accuracy was comparable to studies using VHR imagery. Chapter 4 develops new land use maps for multiple cities in East Africa, and uses these to compare the spatial structure of cities in this region with cities in Europe and the USA. By using urban land use maps, building footprint, and population data, several metrics were calculated in order to quantify the differences between cities of across three regions. Cities in East Africa, on average, have smaller building footprints compared to cities in the USA and Europe, but regarding to land use clustering, built-up density, and population distribution the variations within regions are higher than between them. These results indicate that the city structure of East African cities is not consistently different than cities from other continents, yet the variability found between cities of the same region challenge the idea that there is such a thing as a relevant city model. Chapter 5 uses VHR imagery to analyse different types of building-level changes in Nairobi, Kenya, between 2010 and 2021. Buildings were manually mapped to investigate whether they newly appeared, persisted, changed, were replaced, or removed. Over this period, removal, replaced, and renewal combined made up as much as 29% of the total mapped building area in 2010. In addition, the majority of newly build structures and buildings replacing other buildings are significantly larger than removed and replaced buildings. The observed changes, for example, relate to the removal of informal settlements and urban renewal policies. Due to the relative share of buildings affected, as well as the importance of the related change processes for sustainability, the study argues for the need to further develop automated methods that are able to detect multiple types of urban change on a large scale.