Our study focuses on the Ahrweiler District (“Landkreis Ahrweiler”) in western Germany (Fig. ). In July 2021, the region experienced a severe flood event that caused significant economic losses and fatalities (see Fig. ). Ahrweiler’s history of flooding spans centuries, with over 70 floods recorded in the past 500 years , the first one being reported in 1348 .

The service sector dominates Ahrweiler's economy, accounting for 70.9 % of gross domestic product (GDP), followed by manufacturing industry (27.8 %) and agriculture, forestry, and fishing (1.2 %) . Land use is 53.3 % forest and vegetation, 31.1 % agriculture, 7.2 % settlements (3.3 % residential, 1.0 % industrial, 1.6 % recreational), 6.8 % infrastructure, and the remainder water bodies and other uses . Public records show over 40 000 residential buildings with 88.8 % being detached and semi-detached houses (as of 2023) and roughly 6000 registered companies .

To develop and process our candidate exposure model workflows, we began by searching for freely available building-exposure and asset-value datasets, with a focus on building fixed asset values (i.e., values for structural and non-structural architectural components of the buildings rather than movable contents/equipment). A systematic search using Google  and Google Scholar  yielded the datasets summarised in Table S2, from which we then selected those datasets listed in Table  for further analysis (see Sect. S2 for details).

The three asset value candidate datasets considered here are the Basic European Asset Map (BEAM), EHRE, and data derived from the European Statistical Office (Eurostat). While the raw EHRE data is object-level and model-ready, the BEAM and Eurostat data require additional processing with ancillary data as described in the following sections.

In natural hazards research, the variables selected for exposure models are typically guided by vulnerability research and the availability of data. This commonly leads building-related models to include categorical dimensions such as use, construction type, and economic sector. These categories provide informative predictors of vulnerability behaviour (e.g., via stratified modelling) and, in the case of economic sector classification, useful dimensions for communicating results to stakeholders. In practice, sector classification (e.g., residential, industrial, service) is frequently applied deterministically to each asset by transferring labels from a related, more granular categorical source (e.g., building function or land use) with a simple lookup table or reclassification schema. While there is no standard sector classification label set for building use, common frameworks include NACE 2.0 with its 21 economic areas (see Table S3) or a more basic framework like the one employed by BEAM (Table ).

Here, we adapt or develop the workflows for assigning sector labels to building assets shown in Table , and evaluate these against our hand-labelled benchmark dataset.

From these, the three best-performing workflows are further developed into asset value workflows formulated on BEAM, Eurostat, and EHRE datasets as described below. All resulting derived datasets are provided in .

To develop our benchmark dataset, we selected four representative study plots based on knowledge of the region (Fig. ). From these, a dataset of 844 sample buildings focusing on the four study plots and the village of Bad Neuenahr-Ahrweiler was created by hand (based on LoD1 geometries) representing typical residential and non-residential buildings in Ahrweiler .

Assets were classified into 24 building use types from , which we further categorised into the three basic sectors residential, industrial, and service and a fourth ambiguous category as a catch-all for building use types (e.g., “commercial buildings, with apartments”, “Warehouse/storage building, single-use”) whose sector assignment does not neatly fit into the three basic sectors (see Supplement Data Table 1 for complete specification of the sector label mapping). Building classifications were determined through detailed visual inspections using Google Earth , Google Street View , Mapillary , and ImmoScout24 in March 2024, comparing 3D representations to descriptions and images in and site visits (see Sect. S4 for details). Unit construction costs from ranged from EUR 370–495 m⁻³ for residential, EUR 210–290 m⁻³ for industrial, EUR 215–545 m⁻³ for service, and EUR 160–495 m⁻³ for ambiguous building use types (Table S9). New construction costs per-asset were then calculated by multiplying the relevant unit construction cost, the regional cost factor for Ahrweiler county (0.986; also provided in ), and individual building volumes from LoD1.

The benchmark dataset creation is summarised in Fig. S1.

To quantify the classification performance, we first calculated one-vs.-rest confusion counts and derived precision, recall, and F1 for each benchmark dataset sector and each workflow (Sect. S5). The differences in classification results between the four workflows against the four sectors of the benchmark dataset for the 844 sample buildings are summarised in Fig.  with per-label performance metrics provided in Table S11 and mean values for each workflow in Table S12.

Across the four workflows, precision ranges from 0.50–0.66, recall from 0.57–0.78, and the F1 score (F1) from 0.45–0.71 (Table S12), with LoD1 + Eurostat (Fig. a) and LoD1 + BEAM (Fig. b) performing best (F1 = 0.71).

Both these workflows are in perfect agreement with each other (assuming equivalence of all Eurostat services labels per Table S10) as both share a LoD1 “function” basis and vary only in their terminal category assignments (see Supplement Data Table 1). This LoD1 data assigns roughly 75 % of the non-residential buildings in our benchmark dataset to “Building for business or trade” which we map to market services and service in the Eurostat and BEAM workflows respectively. However, our manual classification of these buildings in the benchmark dataset describes many as industrial (n=66) or ambiguous (n=68). In other words, while some rare buildings with specific uses are well classified in LoD1 (e.g., schools, churches, fire departments), there is insufficient resolution (i.e., label groups are too broad) to differentiate between the more numerous industrial and service buildings. Beyond this, the two additional categories in the Eurostat workflow compared to the BEAM workflow (six vs. four) give Eurostat slightly more resolution. The utility of this can be seen in the non-market services category (Fig. a), 97 % of which match the benchmark dataset service classification. However, because we opted for a simple four-category benchmark dataset, this advantage is not reflected in the performance metrics of Table S12.

The OSM land-use workflow (Fig. d) performs the worst of all workflows (F1 = 0.45), chiefly by over-assigning buildings to the residential sector. For example, about 90 % of buildings labelled as service in the benchmark dataset are misclassified as residential by this workflow. This bias arises from the application of coarse land-use polygons downscaled to individual buildings in dense, predominantly residential areas that contain a small number of service buildings. In such settings, buildings belonging to a minority-sector (e.g., service) are misclassified to the dominant residential land use during the polygon-to-object translation.

To compare against EHRE (Fig. c), which is indexed to OSM geometry, it was necessary to spatially join the EHRE records to the LoD1-based benchmark dataset (see Sec. S5).

Because LoD1 is designed for 3D building representation, some objects that are single-part in EHRE are multi-part in LoD1; to provide a fair comparison, we merged these LoD1 building parts (see label “c” in Fig. S2). We also manually shifted <1 % of LoD1 features in obvious near-match cases, but left benchmark dataset records without a plausible EHRE counterpart unmatched. This yielded 617 record matches and 82 orphaned benchmark dataset records, which are likely attributable to differences in data vintage because they exist in LoD1 but not in OSM (see label “b” in Fig. S2). Conversely, EHRE records not present in the benchmark dataset were discarded (see label “d” in Fig. S2).

Considering these disparate geometries, EHRE had comparable performance to the other non-land-use-based workflows in the residential sector (F1 = 0.89).

However, EHRE generally assigned the service sector benchmark dataset labels to the other category, which EHRE uses as a catch-all for buildings whose occupancy type is unknown (because of a lack of OSM tags, or lack of OBM rules to translate those tags into a final occupancy type) and buildings whose OBM occupancy type does not correspond to any of the three ESRM20 occupancy cases (residential, commercial, industrial). We treat EHRE's application of these other labels as omission errors, because replacement values are omitted from these assets, rather than as commission errors, which would imply an active assertion of the wrong label (see label “e” in Fig. S2).

If we were to instead map this other category to service, EHRE's weighted score for this sector would increase enough to surpass the LoD1-based workflows (F1 = 0.85).

Further, as our evaluation only considers asset-to-asset comparisons, we do not consider any of the residential or commercial remainder buildings imputed by the EHRE model (Table ).

The three classification sectors of the benchmark dataset (industrial, residential, and service) show variable performance across the four workflows.

The highest classification agreement was observed for residential buildings, with all workflows classifying nearly all sample buildings consistently (recall ≥0.98); over-assignment of residential labels was more challenging, especially for OSM land-use (Fig. d; precision = 0.57) as discussed above.

For the service sector, the LoD1 function-based workflows (Fig. a, b) had reasonable true positive capture (recall = 0.83), but over-classified industrial and ambiguous buildings as service (precision = 0.65).

For industrial assets, EHRE performed best (F1 = 0.72), with the OSM land-use workflow close behind (F1 = 0.69), suggesting industrial land use is relatively distinct and well mapped in OSM for our study area. On the other hand, the LoD1-based workflows failed to map any industrial assets owing to the lack of any such labels within our sample building set (see Fig. ).

The final benchmark dataset category ambiguous was used as a catch-all for unclear or mixed building use in our hand labelling.

As the two LoD1-based workflows (Fig. a, b) do not have any mixed or ambiguous categories, these benchmark dataset buildings were forced into standard LoD1 function classes. For example, the benchmark dataset ambiguous class includes labels such as “Commercial buildings, with apartments”, “Warehouse/storage building, single-use”, and “Single, multiple and multi-storey garages”, whereas LoD1 represents buildings through ordinary function labels such as “Building for business or trade”, “Garage”, or “Parking garage” rather than a mixed or ambiguous class.

Similarly, EHRE does not have an ambiguous category because it deterministically maps MIX occupancy types to residential, industrial, or other categories (see Supplement Data Table 1). If these EHRE mixed (MIX) assets were instead mapped to the benchmark ambiguous category, EHRE would yield fewer false positives (precision increasing from 0.56 to 0.60), but with a small loss in recall (0.57 to 0.55) and therefore a negligible improvement overall (F1 increasing from 0.53 to 0.54).

In summary, the LoD1 function-based workflows (Fig. a, b), with their government-collected per-asset data (and unfortunate use restrictions) unsurprisingly provided the closest match to our hand-labelled benchmark dataset, despite being unable to identify any industrial or ambiguous buildings. The open-source object-based EHRE also performed well (Fig. c), especially if the other category were considered service and the MIX category were considered ambiguous. Lastly, the open-source hierarchical imputation OSM land-use-based workflow (Fig. d) performed the worst against our benchmark dataset, primarily through the over-assignment of residential labels; therefore, we exclude this workflow from the subsequent sections. Including it in the regional-total and per-asset value comparisons would require joining the OSM labels to an asset-value source such as Eurostat, producing an OSM + Eurostat hybrid that would mix label performance with the same Eurostat value basis already represented by LoD1 + Eurostat. This would add redundancy and confusion rather than new asset-value information, while EHRE already provides an OSM-based asset estimate.

We next extended three of the sector-classification schemes above with asset values from regional datasets. Table  summarises the resulting district-scale asset values for Ahrweiler.

Looking first at the building counts, the two LoD1-based workflows share the same inventory of 61 756 buildings, whereas EHRE includes 54 490 explicit buildings and 65 056.9 buildings when its remainder layers are included (Table ). Next, this table highlights the different cost bases used by the workflows: LoD1 + Eurostat and EHRE report RC, while LoD1 + BEAM reports DC for residential buildings and NAV for non-residential buildings. Comparing across workflows, we find large variations in estimated asset value district-wide aggregates.

This echoes the findings of , who showed that model estimates can diverge by factors up to 50–200 at 10 km² aggregation in Switzerland.

Comparing our findings for the residential sector for the two LoD1 function-based workflows with identical assets (LoD1 + BEAM and LoD1 + Eurostat), we find the Eurostat-based RC estimate is nearly five times BEAM's DC. Computationally, this arises from disaggregating with LoD1 + BEAM's EUR 118 m⁻³ DC versus 's EUR 2478 m⁻² RC applied to estimated usable floor area (Table ), and is unsurprising considering BEAM includes depreciation whereas does not.

Although LoD1 + Eurostat estimates lower non-residential totals than LoD1 + BEAM, its larger residential RC dominates the aggregate result, making its all-sector total about 2.1 times larger.

Looking at the unit costs, both LoD1-based workflows produced some implausible values as a result of disaggregating regional values onto sectors poorly represented by LoD1. For example, the 19 production/industrial buildings (which are exclusively LoD1 “substations”): LoD1 + Eurostat's production and LoD1 + BEAM's industrial unit costs are both roughly two orders of magnitude beyond comparable BKI unit costs (Table S9). Conversely, LoD1 + Eurostat's market services unit cost is roughly an order of magnitude lower than the BKI service unit costs because the EUR 489.5 million sector total is distributed across about 23 million m³ of mapped LoD1 building volume for this sector, yielding only EUR 21 m⁻³. Together, these extreme values suggest our sector mapping (Supplement Data Table 1) and LoD1 function data are not well matched to the Eurostat reporting structure in our region.

EHRE estimates a lower residential RC than LoD1 + Eurostat, with EUR 12.3 billion in explicit buildings and EUR 14.7 billion when the residential remainder is included, despite its higher building count (Table ). By contrast, EHRE and LoD1 + Eurostat estimate similar non-residential totals (EUR 3.0 billion and EUR 2.9 billion, respectively), both less than half the LoD1 + BEAM estimate (EUR 8.1 billion).

While these total asset value figures are not directly comparable to loss model estimates, the large disparities found here suggest exposure model asset values contribute a similar level of uncertainty to that of other components in a risk model chain. For example, BN-FLEMOps has a mean absolute error of around 15 % and FLEMOps has a mean absolute error of 24 % when comparing modelled to observed losses.

Comparing asset-to-asset, Fig.  presents the two LoD1-geometry distributed asset value model workflows against the construction-cost benchmark dataset (also LoD1-geometry distributed). Because this benchmark dataset is drawn from a single, predominantly residential region, it is likely not representative of areas where industrial and service buildings are more prevalent.

Figure  shows a near-linear relationship between modelled and benchmark dataset asset values for all sectors, which is intuitive if we consider that both the workflows and the benchmark dataset are geometry-based; i.e., they differ mainly in the unit cost multipliers applied to the building area or volume (see Tables  and S9 for unit costs).

Setting aside the sectors with implausible disaggregated unit rates (e.g., market services and corporate services), most non-residential assets in the LoD1 + Eurostat workflow underestimate the benchmark dataset. This can also be seen by comparing the non-market services unit cost (EUR 251 m⁻³) to the BKI service unit costs (Table S9), nearly all of which are higher. This non-residential underestimation likely arises from LoD1 + Eurostat's reliance on stock-average values , in contrast to the new-build construction cost averages reported by BKI.

Surprisingly, we find that the Eurostat residential asset values do not follow the same pattern and are generally higher than the benchmark dataset. This overestimate is likely influenced by the method we used to calculate the usable floor area, which multiplies the LoD1 footprint area by a height-derived storey count and a 0.7 usable-area coefficient after . Figure S5 shows the lack of precision of the storey counts estimated in this way, which generally exceed the observed benchmark dataset counts, in some cases by three storeys, inflating usable floor area and therefore residential RC at the object level. Further, sensitivity to the area multiplier is substantial. For example, using LoD1 footprint area directly gives a residential RC roughly EUR 14.5 billion (56 %) lower than the usable floor area estimates used here.

Conversely, all residential BEAM assets underestimate their BKI counterparts, with a unit cost of EUR 118 m⁻³, lower than any reported by BKI. While this is unsurprising considering BEAM includes a depreciation factor, the modest residential over-assignment (62 false positives; Table S11) could also influence the discrepancy. In contrast, the LoD1 + BEAM service assets are less biased, with the EUR 295 m⁻³ falling within BKI's wide range (Table S9); however, this results in higher variance than the other sectors. Thus, these differences should be read primarily as conceptual alignment with stock-average or depreciated values rather than inherent model deficiency; with the available data, we cannot isolate the specific mechanisms driving each discrepancy.

Unlike the LoD1-based workflows and the benchmark dataset, EHRE is object- rather than geometry-distributed, assigning asset values using probabilistic construction types from ESRM20's regional totals. In other words, EHRE preserves regional average replacement costs (from ESRM20) whereas the LoD1-based workflows considered here preserve asset-level unit costs. Further, EHRE's OSM geometry basis provides a simpler, more consolidated footprint-based representation of buildings, whereas LoD1 – designed for 3D buildings – represents complex buildings with multiple features; consequently, building-by-building value comparison is subjective and challenging (unlike categorical comparison). Even the relatively similar aggregate inventories in Table  (LoD1: 61 756 buildings; EHRE with remainder: 65 056.9, 5.3 % higher) mask important differences in object definition and sector composition, so total count agreement should not be interpreted as asset-level comparability across LoD1 and EHRE. Our early attempts to address this evaluation challenge failed to find a satisfactory definition of a building object between the two datasets (esp. for large complex structures), but in general we found EHRE strongly under-estimates residential replacement costs, while commercial and industrial comparisons are more difficult to interpret because of small sample sizes and object-matching ambiguity. The open-source nature of EHRE and the ESRM20 exposure model it builds upon mean, however, that the performance of the EHRE model in this regard could be improved, firstly, by changing the average replacement cost per area and, secondly, by using the area of the OSM building footprints to calculate the replacement costs, instead of retaining ESRM20's average per-asset-type replacement costs (which are based on average building areas). Future work should explore frameworks for comparing complex and disparate asset value models like these; ideally in a way that balances complexity with interpretability.

To provide a more robust basis for model selection we use a semi-quantitative Weighted Scoring Model (WSM) to communicate the relative strengths and weaknesses of the three exposure model workflows presented above. Derived from the multiple-criteria decision-making mathematical framework, this technique is widely used to prioritise options by assigning weights to various criteria, which are then scored and ranked . For this, we developed the criteria and weights shown in Table S13 to inform decisions on model selection based on previous experience. As presented in Table , each workflow is scored on a scale of 0–10 for each criterion before scores are multiplied by their respective weights, then summed. The benchmark dataset classification score was calculated from Table S12, while the criterion weights and remaining workflow scores were assigned through iterative discussion among the authors until consensus was reached on a final set of values.

In the WSM, LoD1 + Eurostat slightly outperforms the other workflows owing mainly to its benchmark dataset performance, sustainability, and adaptability.

Eurostat data are updated annually, offering a clear advantage, whereas the update cycles of BEAM are unclear and EHRE currently has no planned updates. Like BEAM, Eurostat data tables are web-hosted with open access; however, the LoD1 data requires a licence (commonly provided for free to public bodies and research institutes under specific agreements), while commercial users typically obtain fee-based licences.

Of the asset value model workflows considered, LoD1 + Eurostat is more onerous than LoD1 + BEAM, requiring the downloading and disaggregation of regional and categorical estimates (see Eq. ). EHRE, on the other hand, is already disaggregated to per-asset stochastic classes; therefore pre-processing is only required if the user desires deterministic values (like our study).

Considering the transparency of the underlying data, BEAM is substantially more opaque than the other workflows, with no source code or parameter values provided and documentation only a short paragraph, in some cases without references. In contrast, we find LoD1 + Eurostat and EHRE similarly transparent, both relying on large-scale opaque source data (e.g., OSM, LoD1) but with subsequent calculations performed transparently. For example, the EHRE source code is published while the simpler LoD1 + Eurostat workflow is described by Fig.  and Eq. ().

EHRE is the only workflow to explicitly consider uncertainty, both through the incorporation of probabilistic building types and the use of remainder layers and the other building sector.

Finally, while our study focuses on building structural asset values, numerous other asset types are often incorporated into exposure models (e.g., population, building contents). In this regard, LoD1 + Eurostat provides the narrowest set of asset types as it is the least pre-processed for exposure modelling.

However, this minimal processing affords the greatest flexibility for LoD1 + Eurostat, which provides data for any year across the Eurozone, allowing advanced modellers to tune the workflow for their context.

With this study, we provide the first per-asset evaluation of asset value model workflows for exposure modelling.

Although extending these workflows to other regions in Germany would be straightforward, the need for a labour-intensive hand-labelled benchmark dataset restricted our evaluation to a single region, meaning the findings are only transferable to regions with construction practices, economic conditions, and exposure distributions comparable to Ahrweiler. Future studies with more resources should consider extending the breadth and coverage of our benchmark dataset.

Similarly, we chose to focus on asset values and sector classification as the simplest means for comparing diverse workflows; however, exposure models typically include additional variables like building size, construction type, construction materials, etc. Including these, and non-building assets, in the benchmark dataset and evaluation would provide a more complete comparison. However, as the significance of each of these variables will depend on the particulars of the hazard in question and its associated vulnerability model, the usefulness of such a general study (rather than one specific to a particular case) should be carefully considered. Further, these additional variables are rarely ingested by loss models, and therefore their inclusion in our evaluation would be of limited use to most readers.

The dependence of our benchmark dataset on the underlying data, which only provides average new construction costs in Germany, makes the comparison less useful for certain disaster modelling applications like estimating insurance claims or disaster reconstruction for a heterogeneous set of buildings. More useful would be observed replacement costs from disaster-affected buildings; however, this information is notoriously difficult to collect and plagued with uncertainty when self-reported .

For the LoD1 + Eurostat residential estimate in particular, the reliance on inferred rather than observed usable floor area introduces uncertainty and complexity: future work should improve this conversion.

Similarly, because we opted to provide a broad evaluation, we include asset value model workflows with different cost bases (RC, DC, NAV) and years (2018, 2020, and 2022), making our findings less conclusive.

Aligning cost bases or developing standardised conversion approaches could improve the comparability of the evaluation, but at the expense of attribution and interpretability.

While our work only considers classical disaggregation workflows built on standard accounting and cadastral datasets, the challenges reported here suggest opportunities for employing machine-learning and remote-sensing-based methods, especially for asset classification.

For example, and trained algorithms on labelled images of buildings with promising results.

Lastly, during data collection and publication we struggled to obtain and release asset-level data, a challenge commonly faced by European exposure modellers. This type of building-level economic data is generally considered to be privacy-protected under the EU's General Data Protection Regulation . While many benefits arise from such protections, policymakers should consider the burden this places on asset-level disaster modelling .