Miroslav Despotovic (p), David Koch , Gunther Maier

Assessing the quality of real estate requires an understanding of the objective and subjective sub-criteria such as the quality of the architecture, interior or location. Especially the quality of location is treated as one of the most important aspects in this respect. In order to understand related spatial dependences, regional scientists and economists apply various methods investigating e.g. neighborhood features, socio-demographics, proximity, and accessibility to spatial externalities. In order to explain the quality of the real estate location, distances to different areas such as schools, but also socio-demographic aspects are often used. The problem with this approach is that the features and major parameters of the assessment need to be chosen and defined by the analyst. The analyst needs to select from the theoretically infinite list of features, define distance decays and cut-off distances. Therefore, this approach depends upon a substantial amount of subjective and implicit choices made by the analyst. Moreover, the features used in the assessment have to be extracted from maps, feature lists, or site visits.
Our approach, which is presented in this paper, is to use satellite images instead and to test, whether the quality of the location can appropriately be determined only from satellite images. For that purpose, we have developed a survey-based model for evaluating the real estate location. Students and real estate experts were shown 600 orthophotos of inhabited areas in five Austrian cities (Innsbruck, Kufstein, Kitzbühel, Salzburg, and Vienna) with a pseudo Mercator scale of 300x300m. The respondents' task was to classify the images into different location quality classes and to define the location as suitable for apartments, single-family homes or both. We have standardized and aggregated the estimations of test persons and included them as a dummy variable in the regression model. The property price was modeled as a function of the building characteristics, location, and estimations. The results show a significant influence of the estimates on the property price. We conclude from this that a satellite image contains valuable information, which is not included in the "numerically measurable" variables. The results show the clear potential of image recognition for real estate assessments.

Ricardo Ribeiro Barranco (p)

The spatial allocation of services, either public like schools and health care centres, or private as commercial activities, is driven by certain determinants. These include population density, accessibility, public transportation, distance to city centres, road infrastructure and gross value added (GVA). By creating 1 km2 density grids for a set of 654 European cities, these determinants and services datasets were quantified spatially in standardized grids. Using the determinants as explanatory variables and services as dependent, for each city a Gradient Boosting machine learning regression model was trained on predicting the dependent grid. Once trained, these models provided the importance of each explanatory variable per city. A K-Means clustering algorithm grouped cities with similar profiles. These groups when plotted on a map helped understand the possible presence of regional resemblances. Further, the statistical comparison between each cluster and the entire aggregated results, enables measuring how diverging or similar these are from the overall European allocation profile. In general the location of consumption services, as shops, department stores and markets, is favoured by higher population, accessibility and GVA values. Due to the spatial nature of the datasets used, it is possible to detect areas within cities, were the most important allocation determinants have lower values. Such mapping exercise permits delimiting possible policy intervention areas and which particular determinants to address in order to boost a certain service. This approach offers both detailed insight of the determinants driving the allocation of a particular service within a city and also comparison between cities at European level.

Alan Smith (p)

Human populations are not static in space or time. However, traditional standalone census typically only provide a static day or nighttime population count every 5-10 years. Despite this they are still widely used for applications that would benefit greater from a dynamic, accurate, population estimate. For many applications, a level of spatiotemporal accuracy and resolution is required beyond datasets that are currently available. The method developed in this paper produces detailed, time-specific, dynamic population estimates. The Population 24/7 time-specific modelling framework has been applied and evaluated within the context of the United Kingdom (UK) and Australia, using Perth as an example. This paper sets the agenda for the international applicability of this framework but also the broader application requirement for spatiotemporal population estimates. This paper concludes that the spatiotemporal population datasets produced have come a long away from earlier predecessors. The growing trend in alternative data sources such as the a future administrative data census for the UK means that these developments will have a principal role in population mapping. Traditional census commuter flows have already been re-created using mobile telephone tracking data to a high level of resemblance by the UK's Office for National Statistics. There is a growing need to address that is termed as the spatiotemporal modifiable areal unit problem (STMAUP). This paper shows that population data can be delivered to a high spaiotemporal resolution using open source data and software. In doing so this paper makes recommendations for greater consideration in future work within the research community.