Validation of global urban density products with very high resolution imagery

Abstract

Remotely sensed human settlement datasets are of high importance to monitor human activity and assess relevant developments for human presence on earth, i.e. risk of urban floods, the urban heat island phenomenon or population density on a global scale. In this context, the German Aerospace Centre (DLR) has developed a new set of high resolution global layers, which describes the built-up environment and its spatiotemporal characteristics: (I) the World Settlement Footprint 2015 layer (WSF-2015) and (II) the World Settlement Footprint Evolution Density layer (WSF-Evo-Density). The WSF-2015 layer is a binary mask outperforming all other similar human settlement layers, especially in rural areas. To enhance the thematic scope of the WSF-2015 layer the WSF-Evo-Density layer was developed which shows the percentage of impervious (PIS) within the WSF-2015. This work focuses on the validation of the WSF-Evo-Density layer. A best practice sampling design, a comprehensive pool of ~32,800 very high resolution (VHR) imagery, the Settlement Size Complexity (SSC) index and the Köppen-Geiger climate classification ensured an independent and systematic validation of the WSF-Evo-Density layer. For the first time, a global human settlement dataset is validated with a reference dataset created via object-, expert- and rule-based image classification and VHR imagery. The results show a low underestimation of the WSF-Evo Density layer in urban and peri-urban areas, especially in areas with a high proportion of vegetation. Furthermore, the results show low overestimations for rural areas and strong overestimation for arid areas. In general, the standard deviation of the WSF-Evo-Density layer is ~28% and stable for the investigated parameters. The VHR impervious binary masks can be used to refine the classification of the WSF-Evo-Density layer. Excepted certain VHR imagery, the data and method in this work are freely available and hence, the reference dataset is extendable.