Estimation of missing building height in OpenStreetMap data: a French case study using GeoClimate 0.0.1

Abstract

Abstract. Information describing the elements of urban landscapes is required as input data to study numerous physical processes (e.g., climate, noise, air pollution). However, the accessibility and quality of urban data is heterogeneous across the world. As an example, a major open-source geographical data project (OpenStreetMap) demonstrates incomplete data regarding key urban properties such as building height. The present study implements and evaluates a statistical approach that models the missing values of building height in OpenStreetMap. A random forest method is applied to estimate building height based on a building’s closest environment. A total of 62 geographical indicators are calculated with the GeoClimate tool and used as independent variables. A training dataset of 14 French communes is selected, and the reference building height is provided by the BDTopo IGN. An optimized random forest algorithm is proposed, and outputs are compared with an evaluation dataset. At building scale for all cities, at least 50 % of the buildings have their height estimated with an error of less than 4 m (the cities' median building heights range from 4.5 to 18 m). Two communes (Paris and Meudon) demonstrate building height results that deviate from the main trend due to their specific urban fabrics. Putting aside these two communes, when building height is averaged at a regular grid scale (100 m×100 m), the median absolute error is 1.6 m, and at least 75 % of the cells of any city have an error lower than 3.2 m. This level of magnitude is quite reasonable when compared to the accuracy of the reference data (at least 50 % of the buildings have a height uncertainty equal to 5 m). This work offers insights about the estimation of missing urban data using statistical methods and contributes to the use of open-source datasets based on open-source software. The software used to produce the data is freely available at https://doi.org/10.5281/zenodo.6372337 (Bocher et al., 2021b), and the dataset can be freely accessed at https://doi.org/10.5281/zenodo.6855063 (Bernard et al., 2021).

A generic algorithm to automatically classify urban fabric according to the Local Climate Zone system: implementation in GeoClimate 0.0.1 and application to French cities

“…Concerning building height lack, it is a major concern for urban climate studies. For this reason, the missing heights have been estimated using a Random Forest method based on geographical indicators characterizing the building environment (Bernard et al, 2022). Although the RSU mean building height could be quite far from the truth for certain areas, other indicators such as the ground sky view factor are less impacted by the quality of the building height estimation.…”

“…French communes that have previously been used in Bernard et al (2022). More information about these territories is given Tab.…”

“…The built-up zone is visually more homogeneous in OSM than in BDT. The first reason is that the model used to predict OSM building height smooths its spatial variations (Bernard et al, 2022). The second reason is that the territory is more fragmented using BDT than OSM data (the number of RSU is 4,548 and 2,819 respectively for the whole Dijon commune).…”

“…For such territory having very low LCZ built type, the vegetation type has a considerable impact since once the 265 vegetation is merged, the general agreement fraction is very high (94%). This outcome is probably due to the model used to estimate OSM building height which often produces underestimations of 275 mid and high-rise buildings in open areas (Bernard et al, 2022).…”

“…As described by Bernard et al (2022), Annecy and Meudon are two cities where the model used to estimate OSM building height fails quite a lot. This is clearly highlighted by their very low agreement for compact mid-rise LCZ (47% for Annecy and 40% for Meudon while the average is 81%).…”