Urban form metrics for building occupancy modelling: applications and impacts

3.1 LITERATURE RESEARCH

For this study, the guidelines of the PRISMA2020 statement (Page et al. 2021) are applied where applicable to ensure methodological reproducibility. The scientific databases to be queried are Scopus and WoS. Based on first impressions, Scopus has higher relevancy with regard to the research questions posed, therefore the initial query construction and refining is done for Scopus and the finalised query adapted to WoS.

Initial searches in Scopus based only on keyword combinations yield too many results (over 20,000 publications). Therefore, additional criteria are applied. Based on the number of publications by year (see item 1 in the supplemental data online), only publications from 2014 and after are selected. With this limitation, the upward trend that began in 2017 is captured, possibly due to the rising popularity of ML algorithms at that time, plus a buffer of three years. Additionally, the search is limited to journal articles written in English from the categories ‘computer science’, ‘Earth and planetary science’, ‘engineering’ and ‘environmental science’.

Experiments with different queries show that using only the word stem (classif* versus classification) results in a superior selection of certain publications, since the latter would exclude any publication that uses the term ‘classifying buildings’, for example (for the full queries, see items 2 and 3 supplemental data online).

3.2 SEARCH RESULTS

The search was conducted in September–October 2024 and yielded 896 (Scopus) and 873 (WoS) publications as a first result. Building upon this, multiple steps of filtering are applied, where all filtering steps are saved as separate bibliographies. No automation tools are used in the screening process. The first filtering is by title only, whereas all publications are removed from the selection where the title did not fit the research goals. In unclear cases, a publication is not removed. The remaining 137 publications are then filtered by analysing the abstract, which reduced the number to 56 publications. For the final selection, only full-text accessible publications are considered, assessing whether any UFM are used to derive values related to building occupancy, resulting in 24 remaining publications that fit the criteria (see items 4–6 in the supplemental data online).

All UFM from the 24 selected publications are extracted and compiled into a comprehensive list. This list is then scanned to identify and consolidate semantically close/identical metrics, which are referred to by different names by various authors.

3.3 CATEGORISATION

Next, the individual metrics in the catalogue are categorised according to the spatial scale and semantic category they belong to. For the spatial scale classification, this paper follows what Biljecki & Chow (2022) propose: (1) the independent building level; (2) the contextual building level; and (3) aggregated indicators. For the purpose of this study, these scales correspond to the level of data aggregation required for calculating specific metrics. At the independent building level, each building is considered separately, yielding an individual value per building for a given metric. At the contextual building level, the influence of neighbouring buildings is incorporated, yielding an individual value per building for a given metric, where the value is dependent on multiple buildings. Finally, aggregated indicators consolidate data from multiple buildings into a single representative value, where multiple buildings are assigned the same value for a given metric.

For semantic classification, this paper combines and modifies classification schemes of several other authors (Fleischmann et al. 2021; Hecht 2014; Zhang et al. 2023) with additional criteria so that they fit the presented metrics catalogue. The categories proposed are shown in Table 1. Every individual metric in the catalogue can only belong to a single scale and category, but categories can be present in multiple spatial scales. While some of the categories are not closely linked with building occupancy at first glance (e.g. Road Network), they are included because some of the collected metrics belong to these domains. The goal is to include all used metrics from all selected studies, which is why these categories are used instead of ‘Other’. This way, it can be ensured that supposedly unrelated metrics are not overlooked in further steps.

The category Density & Coverage contains metrics that represent counts of buildings within an area (number of buildings in a block) or the area covered by neighbouring buildings within a defined radius, for example. Environmental consists of metrics such as land surface temperature or night-time light (NTL) intensity. Metrics in the category Geometry & Shape are derived values that describe buildings beyond their physical appearance (e.g. compactness or shape index), while Physical Appearance & Size consists of all directly measurable or visible metrics (e.g. area and height). Neighbourhood & Proximity contains all those metrics that describe distances to neighbouring buildings in some form. Position & Orientation metrics describe the positioning of buildings against each other or to a reference object. Road Network consists of all metrics that involve street networks (e.g. distance to next street). Semantics are mainly building type, land use and age metrics. Lastly, the Socio-economic category contains all those metrics that deal with point-of-interest (POI) data, closeness to amenities (e.g. closeness to healthcare facilities), land/building value and population.

Regarding the grouping of UFM, this paper intentionally does not group related metrics such as height and number of floors. This allows for an assessment of each individual UFM in terms of their influence on the results as the authors in the selected studies originally report, while grouping may lessen or even hide effects of specific UFM. While there are arguments in favour of grouping such as reducing collinearity effects and highlighting combined influence of related UFM, analysing them separately provides more granular insights, allowing one to identify specific metrics that drive observed effects in line with the core intention of this paper. Nonetheless, handling correlation issues between UFM should be addressed when using the UFM catalogue in practice through appropriate statistical techniques.

Additionally, the countries the studies are applied in and the data sources that have been used to extract or generate the metrics are analysed. Here, the following categories are defined: Census data, LiDAR, NTL data, Official data, OSM, POI data, Remote sensing or Other. While there are also other important data sources for UFM (e.g. cadastral data) available, only such sources that are used in the 24 selected studies are included in this paper.

3.4 METRIC IMPORTANCE RATING

A key aspect of this paper is to assess the importance the authors of the reviewed studies saw in a metric. It was crucial that not only the frequency with which UFM are used in publications but also their individual impact on the results of a study is considered, if possible. To assess the impact on the results, it was specifically analysed whether a feature importance ranking (e.g. ablation study, permutation importance or Shapley additive explanations (SHAP) values) or similar measure is provided in a study. If yes, the presented ranking of the metrics is extracted and used to calculate a scoring termed the importance score (I):

where P is the number of publications in which a metric appears; and n is the number of metrics in a specific publication. If no rating is presented, each metric is assigned the rank n. After calculating, I is then normalised to a range of 0–1 for each metric individually:

This normalised importance (nI) is closer to 1 for metrics that are highly ranked and/or frequently used in the publications found through literature research, and therefore deemed more important.

4.1 METRIC FREQUENCIES

Figure 1 shows the distribution of metrics by scales and categories. For the Independent building-level, Geometry & Shape is the most frequent category (81 metrics). For Contextual building level it is Neighbourhood & Proximity (15 metrics) and for Aggregated indicators Socio-economic (36 metrics) is the most frequent category.

Figure 1

In total, only 12 metrics appeared in five or more publications (3.8%). A total of 61 metrics appeared between four and two times (19.5%). The remaining 240 metrics appeared only in a single publication (76.7%).

4.2 METRIC IMPORTANCE

Table 2 shows the five most important UFM for each scale level, as well as the respective category to which they belong, their nI and frequency of appearance. Of the 24 publications analysed, 18 provide a measure of UFM importance. The overall most important metric is the area of a building, the average nI is 0.061 and the median nI is 0.0167.

4.3 DATA SOURCES AND GEOGRAPHICAL DISTRIBUTION

Certain UFM are only usable if specific data sources are available (Table 3). The important metrics are mostly provided by either official datasets or OSM, with remote sensing following close behind. Highly specific metrics such as NTL intensity or POI kernel density are only provided by a specific data source.

The total number of different target countries in which the studies were applied according to the source papers is 11 (Figure 2). The most studies were applied in China (n = 12), followed by Germany (7) and the US (3).

Figure 2

4.4 METRIC IMPLEMENTATION AND SELECTION

The application strategy of UFM within a study is another aspect to examine. Figure 3 gives an overview of the algorithms used as well as the metric importance ranking and selection strategies used in the publications.

Figure 3

It can be noted that the application of UFM in all contexts is dominated by ML algorithms, where UFM are used as input features in the respective model. The most frequent type of algorithm are tree-based classifiers (random forest, XGBoost, LightGBM) followed by neural networks (artificial neural network—ANN, DeepFM, dual environment feature fusion, HexaGAN) and graph-based models (GraphSAGE). For those algorithms that appeared multiple times, the ranges of reported metrics can give an initial indication of their respective performance. Relevant in this regard are GraphSAGE (88–91% F1-score), random forest (74–89% accuracy) and XGBoost (80–93% accuracy).

Looking at metric importance ranking, the most common strategy found is an ablation study (which also doubles as a method of variable selection) and subsequently frequency-based or permutation importance ranking. A total of 25% of selected studies did not analyse metric importance in any form.

Concerning variable selection, it stands out that over half of the selected publications did not assess this topic at all. Those that did relied mostly on the results of ablation studies used also for importance ranking, followed by backwards elimination strategy. Only two publications applied extensive statistical testing in the form of analysis of variance (ANOVA) and principal component analysis (PCA).

5.1 METHODOLOGICAL ASPECTS

The literature review resulted in a catalogue of 313 UFM that are used in the context of building occupancy modelling. This aligns well with the magnitude that can be found in related studies on UFM (Biljecki & Chow 2022; Fleischmann et al. 2021; Zhang et al. 2023), which identify between about 300 and about 750 metrics (including ranked variants). The lower end number in this study can be attributed to its focus on a specific subdomain and a more targeted query, which increases thematic relevance compared with broader reviews. A notable difference lies in the number of variables used per study. The median number of UFM in this paper is 18.5, which is substantially higher than in previous work. For example, Biljecki & Chow (2022) report a median of four. This may indicate that building occupancy modelling requires more information, potentially due to the complexity of the target variable. Another reason might be the increasing availability of detailed datasets in recent studies.

Terminological inconsistency remains a critical issue. As noted by Fleischmann et al. (2021), identical metrics are often labelled differently, or identical labels refer to different definitions (an exemplary case being ‘shape index’). To avoid false aggregation where information on niche, but important, UFM might get lost, metrics were kept separate if definitions were unclear.

A key contribution of this paper is the integration of an importance measure, extending prior work (Biljecki & Chow 2022; Fleischmann et al. 2021; Zhang et al. 2023) beyond descriptive compilations. For that, information on metric importance if provided was compiled and an index from these data was calculated. The derived index, however, is context specific and only comparable within the analysed sample and it should be noted that the robustness of the importance ranking is limited by the relatively small sample size. Additionally, importance inflation or selection bias may occur depending on research questions or modelling strategies used in the analysed studies.

5.2 METRICS EXAMINATION

A key observation is that more than 75% of UFM appear only once in the analysed studies. This likely reflects both context-specific data usage and regional differences in data availability. At the same time, approximately 30% of these single-appearance UFM show an above-average nI, indicating that infrequent use does not imply low importance.

The main findings of this paper suggest that frequently used metrics are not necessarily the most impactful ones, highlighting the need to critically assess UFM selection rather than adopting established variable sets without evaluation. For instance, ‘rectangularity’ appears in seven studies, but ranks relatively low (only at 56/313) in importance, whereas ‘number of holes’ ranks highly at 7/313 due to its relatively high impact the two times it was used. Another important consideration is whether high-ranking UFM reflect true explanatory power or simply ease of availability. Easily accessible variables may be overrepresented and repeatedly reused, resulting in an increasing bias towards these metrics over time. The rank-based scoring approach applied here aims to mitigate this effect, but cannot fully eliminate it.

Interdependencies between metrics present an additional challenge. Common variables such as area, height and volume are strongly correlated, yet more than half of the analysed studies do not address collinearity at all, and instead rely on the capabilities of the algorithms used to handle such issues automatically. This raises concerns about potential bias in reported importance rankings. A similar issue applies to shape-related metrics (e.g. ‘perimeter’ or ‘roundness’), which are sensitive to data resolution and geometric representation and therefore may be inconsistent across different datasets (Basaraner & Cetinkaya 2017; Labetski et al. 2023), possibly altering their modelling importance. Some metrics are also dependent on the definition of a building, whether they are seen as singular units or multi-object clusters has significant influence on topological metrics (such as ‘number of neighbours’). Although noteworthy, all these technical issues are considered beyond the scope of this paper.

A notable gap across the analysed studies is the limited discussion of metric appropriateness. A majority of authors report which UFM are impactful but provide little explanation of why this is the case. While some studies briefly address this issue (Chen et al. 2021; Du et al. 2015; Wang et al. 2021), systematic reasoning remains largely absent. Several reasons may explain the limited focus on such mechanisms. In many studies, the primary objective is model performance rather than interpretability. In addition, the internal behaviour of ML models is often difficult to disentangle, which limits insight into how specific UFM influence occupancy. Moreover, relevant assumptions are frequently embedded implicitly in the selection of variables, e.g. that larger buildings accommodate more people (Schug et al. 2021) or that urban centres exhibit higher population densities (Boo et al. 2022), leading to a focus on related indicators such as building height. As mentioned above, authors should critically reflect on such assumptions when applying UFM in their research. Overall, the appropriateness of UFM remains underexplored in the current debate and represents a key avenue for future research, particularly in application-oriented contexts such as building occupancy modelling.

This gap also defines a boundary of the present paper. While linking specific UFM to building occupancy would require a more in-depth conceptual and empirical analysis of causal relationships. Such an investigation was not the primary objective of this review. Establishing robust explanations for why and how individual metrics influence occupancy would necessitate controlled comparative studies, detailed contextual data and the application of explicit statistical modelling frameworks. Nevertheless, the literature suggests some indicative patterns. For instance, variables such as building area or number of floors are commonly interpreted as proxies for physical capacity and thus relate directly to potential occupancy. In contrast, accessibility-related indicators (such as proximity to public transport or points of interest) tend to influence occupancy indirectly by affecting residential attractiveness and demand. However, these relationships are highly context dependent and require further systematic validation. Future research should therefore aim to explicitly investigate these mechanisms and strengthen the conceptual linkage between UFM and building occupancy.

5.3 PRACTICAL APPLICATION

The implementation procedures in the selected studies mostly follow the same basic scheme: UFM are used as a proxy to infer knowledge on different aspects of building occupancy. The selection process is often based on literature research, using or recombining sets of UFM already proposed by other authors. Another option used in multiple papers is the adaption of predefined UFM frameworks such as the momepy toolkit (Fleischmann 2019), which offer a simple solution to selecting and calculating UFM.

Judging which of the algorithms/strategies is best suited for each task is difficult, since every analysed study uses a different classification scheme and reports different metrics. Also, a lot of the algorithms only appeared in a single study. From the results presented, there is not a great difference between the three most prominent algorithms, but for definitive recommendations the scope of this paper is not sufficient and further research is needed.

5.4 GEOGRAPHICAL DISTRIBUTION AND DATA SOURCES

When looking at the geographical distribution, it can clearly be noted that no studies target countries in the Global South. This could be explained by the lack of data, especially of high-quality official data, on buildings available for these countries (Young et al. 2021). Another potential factor could be the language of some relevant papers, as this study was limited to papers written completely in English. It also does not include conference proceedings or grey literature within the search criteria.

Adding to this topic is the general availability and completeness of the data sources. Some metrics are quite impactful, but can only be provided by very specific sources. On the other hand, sources that are broadly available, such as OSM, cannot provide the full range of metrics on their own and also suffer from data quality issues (Biljecki et al. 2023; Herfort et al. 2023).