BMAT: a footprint-level building facade material dataset for 73 major cities worldwide

Building facade materials are closely associated with urban microclimates, energy consumption, and carbon emissions, yet existing urban building datasets typically capture only building footprints and heights, leaving semantic material information unavailable at the individual-building level. Here we present BMAT, a footprint-level dataset of dominant visible street-facing facade materials covering 22.09 million buildings across 73 major cities worldwide. BMAT links individual building footprints with facade material labels inferred from over 147 million temporally stamped street-view images collected between 2007 and 2025, where historical imagery is available. We developed an automated inference pipeline based on a fine-tuned vision-language model trained on 39 405 manually annotated samples, achieving an F₁ score of 0.91 against held-out test data. Dataset reliability was further assessed through comparison with 50 823 Overture Maps facade-material records, yielding agreements of 0.72 for the latest BMAT records and 0.79 under a historical matching strategy. BMAT reveals spatially distinct material signatures associated with climate, geography, and regional building traditions. In cities with sufficient repeated imagery, the temporally stamped records further provide exploratory evidence of facade material transitions, including increasing glass facade adoption in selected rapidly urbanizing regions. By linking building geometry with facade-material information, BMAT provides facade material as an important building-attribute input for studies of building energy, embodied carbon, urban microclimates, and hazard risk. The dataset is openly available at https://doi.org/10.6084/m9.figshare.31569370 (Yin, 2026a).

1 Introduction

Global urbanization is accelerating, with nearly 68 % of the world's population projected to reside in urban areas by 2050 (United Nations Population Division, 2014). This unprecedented expansion intensifies environmental challenges, including urban heat islands (Manoli et al., 2019; Hsu et al., 2021), escalating building energy consumption (Güneralp et al., 2017; Wang et al., 2023), and rising carbon emissions (Yap et al., 2025; Gurney et al., 2021). Facade materials form the critical interface between buildings and their surrounding environment. Their thermal and radiative properties directly affect building envelope performance and urban microclimates (Wang et al., 2024a; Liu and Lei, 2025), with implications for building energy consumption, embodied carbon estimation (Liang et al., 2025a), and disaster risk assessment (Zamanialaei et al., 2025). Beyond environmental performance, the choice of facade materials also reflects the construction technologies (Xie et al., 2025; Huang and Hsu, 2003), aesthetic preferences (Xue et al., 2025), and urban development processes of different eras (Naik et al., 2017), making them a tangible record of how cities have changed over time. However, the lack of large-scale, multi-temporal facade material datasets with fine-grained classification leaves a critical gap for data-driven urban sustainability and resilience research.

Despite rapid progress in large-scale building mapping, no existing dataset provides systematic facade material information at the individual building level. Recent advances in remote sensing and deep learning have enabled the extraction of building footprints and heights at global scale, producing widely used datasets such as 3D-GloBFP (Che et al., 2024), GlobalBuildingAtlas (Zhu et al., 2025), Google Open Buildings (Sirko et al., 2023), and Microsoft Building Footprints (Microsoft, 2024). While these resources have greatly advanced urban morphology research, they are limited to geometric attributes – an inherent constraint of the nadir satellite perspective, which captures rooftops but cannot observe vertical facades. Open geospatial platforms such as OpenStreetMap and Overture Maps do include building material tags, but coverage is extremely sparse, with the vast majority of buildings worldwide lacking any material annotation (Herfort et al., 2023; Biljecki et al., 2023). Moreover, these records are typically static snapshots lacking temporal metadata, making them unable to reflect material changes from renovations or redevelopment (Zhang et al., 2025). As a result, no existing dataset provides ready-to-use, footprint-level facade material attributes across a broad set of major cities, while also preserving temporal references for multi-year analysis.

Street-view imagery has emerged as a promising alternative for facade material identification. Unlike nadir satellite observations, street-view images capture the urban environment from a horizontal perspective, clearly revealing facade materials, textures, and architectural details (Wang et al., 2024b; Chen et al., 2025). Importantly, major platforms such as Google Street View (GSV) and Baidu Street View (BSV) now provide dense, repeated coverage across cities worldwide, offering both broad spatial extent and a multi-temporal archive (Biljecki and Ito, 2021). Recent studies have demonstrated the feasibility of combining street-view images with deep learning for building material classification, progressing from end-to-end CNN and Transformer classifiers applied to whole images (Raghu et al., 2023; Sun et al., 2025), to patch-based strategies that better handle mixed-material facades (Chen et al., 2025), and more recently to fine-tuned vision-language models that achieve superior accuracy (Liang et al., 2025b). However, existing efforts have focused primarily on methodological development, releasing annotated image samples (Raghu et al., 2023; Wang et al., 2024b) or inference toolkits (Liang et al., 2025b) rather than producing a ready-to-use, building-level facade material dataset across major cities.

In summary, existing building datasets derived from remote sensing capture only geometric attributes, while street-view-based material studies remain at the methodological stage. No publicly available dataset yet links facade material attributes to individual building footprints across a broad set of major cities. To address these gaps, this study presents BMAT, a footprint-level dataset of dominant visible street-facing facade materials derived from over 147 million multi-temporal street-view images covering 73 major cities worldwide – 37 in mainland China (via Baidu Street View, 2013–2023) and 36 in other countries and regions (via Google Street View, 2007–2025). These cities encompass national capitals, provincial centers, and economically significant metropolises with large populations, providing broad geographic coverage across diverse urban contexts worldwide. By linking material attributes to individual building footprints and retaining the acquisition year of each street-view observation, BMAT moves urban material analysis beyond assumption-based or material-blind approaches. It supports applications in building energy simulation, embodied carbon accounting, urban microclimate analysis, and urban hazard risk assessment, while also providing a basis for exploratory longitudinal analysis in cities and periods with sufficient records. The main contributions are:

• (1)

  Broad multi-city coverage: the dataset spans 73 major cities across diverse climate zones and development contexts, providing broad geographic coverage for building-level facade material mapping and enabling cross-regional comparative analysis;

• (2)

  Building-level granularity: material attributes are mapped to 22.09 million individual building footprints, providing the spatial precision required for downstream urban analytics that aggregate-level data cannot support;

• (3)

  Temporally stamped records: the dataset preserves acquisition-year information from available street-view imagery between 2007 and 2025, supporting exploratory analysis of facade material transitions in cities and periods with sufficient repeated observations.

2 Data sources

2.1 City administrative boundaries

This study uses the GADM database v4.1 to obtain administrative boundaries of major cities worldwide. GADM provides multi-level administrative division vector data for countries and regions worldwide (https://gadm.org/data.html, last access: 9 June 2025). It should be noted that for mainland Chinese cities, the official administrative boundaries are typically much larger than the actual built-up areas. Therefore, this study focuses on the main urban areas where street-view image coverage is available, rather than the entire administrative region, to ensure targeted and efficient data collection.

2.2 Building footprint data

Building footprint data were acquired from the Overture Maps Foundation global dataset (schema version: v1.13.0). This repository integrates multiple authoritative sources – including OpenStreetMap, Esri Community Maps, Microsoft Buildings, and Google Open Buildings, among others. It comprises over 2.6 billion individual building vector features worldwide, referenced to the WGS84 coordinate system and attributed with geometric boundaries, heights, and classifications. The dataset's multi-source fusion strategy significantly improves coverage in developing regions and emerging cities compared to single-source alternatives. For buildings lacking height records in Overture Maps, missing values were supplemented using the GlobalBuildingAtlas (Zhu et al., 2025) dataset (Fig. S1 in the Supplement). We also collected building-level construction-year data for four cities to explore the relationship between building age and facade material types (Sect. S10).

2.3 Road network data

Road network data were retrieved from OpenStreetMap (OSM) using the OSMnx Python library (version: v2.0.4). Using the city administrative boundaries defined previously as spatial filters, we queried the network with the type parameter set to “drive service”. This configuration captures not only public drivable streets but also service roads (e.g., alleyways, access aisles), ensuring comprehensive coverage of potential street-view sampling locations.

2.4 Street-view images

Street-view images were collected from two platforms. Google Street View (GSV), launched in 2007, provides regularly updated panoramic imagery with historical archives dating back over a decade in many cities; GSV images were collected for 36 cities outside mainland China, covering available observations between 2007 and 2025. As GSV is not available in mainland China, we used Baidu Street View (BSV), which has been operational since 2013; BSV images were collected for 37 cities, covering the period 2013–2023. The 73 cities were selected to cover major urban centers of administrative, demographic, or economic importance, including national capitals, first-order administrative centers, and major population and economic hubs across diverse geographic regions worldwide (Sect. S2). In total, approximately 133 million GSV images and 14 million BSV images were acquired, yielding around 147 million street-view images across all available timestamps. After filtering out images in which the target buildings are not visible or are obstructed by intervening buildings, approximately 96 million (65.3 %) images remained as valid samples for facade material inference. The city-level numbers of total, not-visible, obstructed, and valid images are reported in Table S6.

3 Methods

The overall workflow comprises three main stages (Fig. 1). In the first stage, street-view sampling points are determined based on the spatial relationship between building footprints and the road network, and multi-temporal building facade images are retrieved with calculated camera parameters. In the second stage, a lightweight classifier first filters out outlier images; the remaining valid images are then passed to a fine-tuned vision-language model for material type inference. In the third stage, the trained model is applied to all valid building images across 73 cities to produce the final multi-temporal facade material dataset.

Figure 1 Overview of the dataset production workflow: (1) multi-temporal street-view image collection, (2) model training and material inference, and (3) multi-city dataset generation. Aerial images are from Google Earth Pro. Imagery © 2025 Landsat / Copernicus, Map data © 2025 Google.

3.1 Data collection

3.1.1 Retrieval of time-series building images

This study aims to acquire time-series street-view imagery at the individual-building scale. We developed a geometric workflow that links building footprints to the road network. For each building, we first identified a theoretical sampling location by projecting its geometric centroid onto the nearest road segment. Because this location may not coincide with an available street-view panorama, the nearest recorded panorama was used as the actual sampling point. Camera parameters, including heading, pitch, and field of view (FOV), were then determined from the spatial relationship between the actual sampling point and the target building (Fig. 2a). This parameterized workflow enabled automated, large-scale collection of building-centered imagery across entire cities.

Figure 2 Workflow for facade image acquisition. (a) Sampling location and camera parameter determination; (b) Line-of-sight obstruction detection.

The heading was defined as the clockwise angle from true north to the line connecting the actual sampling point and the building centroid, thereby directing the camera toward the target building. The pitch represented the vertical angle toward the midpoint of the building facade, estimated from half of the building height. For buildings without valid height information in Overture Maps, a default pitch of 30° was used. The FOV was adjusted according to building height and the distance from the panorama to the building (Sect. S3). Taller or closer buildings were assigned a wider FOV to reduce the risk of cropping, whereas shorter or more distant buildings were assigned a narrower FOV to preserve facade detail. All images were generated at a fixed resolution of 640 pixels×640 pixels.

3.1.2 Line-of-sight obstruction detection

In densely built-up areas, intervening structures may obstruct the view between street-view sampling points and target buildings. To systematically identify building visibility, we implemented a two-stage line-of-sight verification strategy (Fig. 2b). First, a sight line is cast from the sampling point to the centroid of the target building footprint and tested for intersection with surrounding building polygons. If the centroid sight line is obstructed, a multi-point verification is triggered, in which sight lines are cast to all boundary vertices of the target footprint and each is tested independently. Buildings are classified as fully visible if the centroid sight line is unobstructed, or partially visible if at least one vertex sight line remains clear. Completely occluded buildings, where all sight lines are blocked, are excluded from further processing. Both fully and partially visible buildings are retained for material inference, as even partial facade exposure provides sufficient visual information for classification. This approach maximizes valid sample retention while ensuring that the target building is genuinely captured in the imagery.

Figure 3 Example images of the 9 human-annotated material types.

3.2 Model training and inference

3.2.1 Filtering out low-quality building images

Street-view images frequently contain outlier samples unsuitable for material classification, such as overexposed scenes, vehicle-occluded views, indoor captures, or heavily mosaicked areas. To exclude these, we trained a binary classifier based on the MobileNetV2 architecture. A labeled dataset of 8934 images (5903 valid and 3031 invalid) was constructed and split into training (80 %) and validation (20 %) sets. The model was fine-tuned from pre-trained ImageNet weights using the Adam optimizer (learning $\text{rate}=\mathrm{1}\times {\mathrm{10}}^{-\mathrm{4}}$) for 100 epochs, achieving a validation accuracy of 88.75 %. This classifier was then applied to the full image collection to retain only valid building facade images for subsequent inference.

Table 1 Comprehensive performance comparison of different models. The proposed model and the best performance values are highlighted in bold.

3.2.2 Manual annotation of facade images

We assembled 39 405 manually annotated building facade images. The collection comprised 38 873 images sampled from the street-view imagery acquired in this study and 532 images obtained from previously published annotated resources (Sect. S5). To ensure consistent application of the material definitions, all images were annotated over one month by a doctoral researcher in urban planning with research experience in urban built environments and street-view image analysis. Following previous studies, facade materials were classified into nine categories (Fig. 3): brick (5502 images), concrete (5490), glass (4804), metal (2831), stone (4773), stucco (7359), tile (2067), wood (5521), and other (1058). Common materials were represented by larger sample sizes, while all nine categories retained sufficient observations for model training. The other category mainly included buildings under construction or renovation, as well as facades that could not be reliably assigned to the main categories. For mixed-material facades, the label was assigned according to the material occupying the largest visible proportion of the main facade. For example, a building with a concrete ground floor but tile cladding across most of its upper floors was labeled as tile.

3.2.3 Material inference using a fine-tuned vision-language model

The 39 405 annotated facade images described above were divided into training and test sets at a 7:3 ratio. We adopted Qwen2.5-VL-7B-Instruct as the backbone model and implemented supervised fine-tuning (SFT) using the LLaMA-Factory framework. To optimize computational efficiency, we applied LoRA (low-rank adaptation) with 4 bit QLoRA quantization in bfloat16 precision. Training was conducted for 3 epochs with an initial learning rate of $\mathrm{5}\times {\mathrm{10}}^{-\mathrm{5}}$. The prompt employed during fine-tuning was: “What is the facade material of the central building in this image? Please choose one of the following: [tile, brick, concrete, glass, metal, other, stone, stucco, wood] Only answer with the exact word from the list.”

Figure 4 Comparison of classification performance between the fine-tuned Qwen2.5-VL-7B-Instruct model and traditional vision backbones on the test set.

3.3 Development of a multi-city building facade material dataset

The fine-tuned vision-language model was applied to infer facade materials for all buildings with valid street-view images across 73 cities. Each valid image was linked to its target building footprint and acquisition year. For buildings captured at multiple timestamps, the year-specific material labels were retained as a temporal sequence in the “materials” field, while the latest available label was recorded in the “material” field. The full material inference over the filtered street-view images required approximately 6353 GPU-hours on NVIDIA RTX 4090 GPUs. By applying this approach across major cities worldwide, we finally produced a multi-temporal, building-level building material dataset covering 73 cities (36 cities based on GSV and 37 cities based on BSV).

4 Results

4.1 Model performance

We evaluated several vision-language models on the manually labeled test set (11 827 images) to select the most suitable backbone (Table 1). Among the models tested without fine-tuning, Qwen2.5-VL-7B-Instruct achieved the highest F₁ score while maintaining moderate inference time. To further validate this selection, we fine-tuned representative VLMs using the same training set (27 578 images), test set, and evaluation metrics. The fine-tuned Qwen2.5-VL-7B-Instruct achieved an F₁ score of 0.91 with a favorable accuracy-efficiency trade-off, substantially outperforming the non-fine-tuned VLMs and traditional vision backbones (e.g., ResNet, ViT) trained on the same data (Fig. 4, Table S2).

Figure 5 Comparison of model predictions with Overture Maps records for (a) latest and (b) historical material attributes, including (c) visual examples with green for correct and red for error.

Figure 6 Building material dataset overview: (a) geographic distribution of 73 cities; (b) detailed comparison of Google Earth imagery and recorded material classifications for local regions. Aerial images are from Google Earth Pro. Imagery © 2025 Landsat / Copernicus, Map data © 2025 Google.

To further validate model performance in real-world settings, we compared our inference results against Overture Maps building records that contain “facade material” attributes (Fig. 5). Because Overture Maps material records do not contain acquisition dates, two comparison strategies were used. First, matching the latest BMAT material record with the Overture Maps record yielded an agreement of 0.72 (Fig. 5a, Table S4). This comparison is transparent and conservative, but it may underestimate agreement if the undated Overture record corresponds to an earlier facade state rather than the latest BMAT observation. Second, we used a historical matching strategy in which a building was considered matched if any available annual BMAT material record agreed with the Overture Maps record, yielding an agreement of 0.79 (Fig. 5b). This value should be interpreted as an upper-bound sensitivity estimate rather than a conventional accuracy estimate, because buildings with more annual observations have more opportunities to produce a match by chance. Therefore, the Overture comparison suggests that the external agreement of BMAT is likely between the conservative latest-record estimate of 0.72 and the historical-match upper-bound estimate of 0.79.

It should be noted that both our model predictions and the Overture Maps records may contain errors. Figure 5c illustrates three representative scenarios: agreement between both sources, only our model correct, and only Overture Maps correct. Our model performs reliably in the majority of cases, with misclassifications concentrated in visually ambiguous scenarios, such as brick with indistinct textures being classified as concrete, stucco facades dominated by large glass windows being identified as glass, and wood and metal surfaces with similar visual appearance.

4.2 Dataset overview

Figure 6a presents the geographic distribution of the 73 cities covered in this dataset, including national capitals, first-order administrative centers, and major population and economic hubs across diverse geographic regions. The dataset classifies building facade materials into nine categories: brick, concrete, glass, metal, stone, stucco, tile, wood, and other. Material attributes are assigned to all buildings for which valid street-view images are available; buildings without street-view coverage do not have material records. In total, the dataset encompasses 73 cities with 40.99 million buildings, of which 22.09 million buildings (54 %) have material attributes assigned. Figure 6b illustrates the dataset at both city and local scales, showing the spatial distribution of facade material classifications across entire cities alongside zoomed-in views of representative neighborhoods in New York and San Francisco. A qualitative comparison with contemporaneous Google Earth Pro aerial imagery indicates that the dataset effectively distinguishes material types among adjacent buildings – for example, correctly differentiating glass-clad high-rises from brick-faced low-rise buildings in dense urban cores – demonstrating both the spatial resolution and high accuracy of the material classifications.

Table 2 Statistics of buildings and material information coverage across different cities in the dataset.

¹ “w/ Mat.” refers to the number of buildings with identified material information. Region abbreviations: UAE = United Arab Emirates; UK = United Kingdom; US = United States.

Figure 7 Distribution of building facade materials across the 73 studied cities.

Figure 7 illustrates the city-level composition of BMAT material records by building count, with buildings without valid material records included as a separate “No material record” category. Cities outside mainland China generally show higher proportions of valid material records, reflecting the broader spatial coverage and more frequent update cycles of GSV compared with BSV (Fig. S2). Among buildings with valid material records, distinct material patterns emerge across cities. Coastal cities (e.g., Singapore, Los Angeles, San Francisco, Tokyo, Mumbai) show a high proportion of stucco facades, consistent with its suitability for humid maritime environments due to moisture resistance. Cities in mid-to-high latitude cold climate regions (e.g., Ottawa, Toronto, Amsterdam, Chicago, London) are characterized by a high proportion of brick, a material valued for its thermal mass and heat retention. Chinese cities exhibit considerably higher proportions of tile facades than other regions, likely reflecting local ceramic manufacturing traditions and architectural preferences. Glass facades are also more prevalent in Chinese cities, corresponding to rapid urbanization and the widespread adoption of curtain wall construction in recent decades.

Table 2 summarizes the number of buildings with material information, the total building count, and the material coverage rate for each city. Coverage varies substantially across cities. Economically developed cities with relatively low building density – such as Sydney, Toronto, London, Boston, Chicago, New York, and San Francisco – exhibit notably higher coverage rates. In contrast, cities with high building density – such as Mumbai, Delhi, Nairobi, and many cities in mainland China – show considerably lower rates. This disparity may reflect limited street-view accessibility in dense residential areas, narrow streets, private or restricted roads, and platform-specific coverage gaps (Fig. S3).

4.3 Spatial patterns of building facade materials across cities

Figure 8 illustrates the spatial coverage of building material data inferred from GSV images. Figure 8a shows building counts and material coverage proportions at 1 km grid resolution across representative cities. In these developed cities, central areas generally exhibit higher proportions of buildings with material information than suburban zones, with most regions maintaining consistently high material data visibility. Figure 8b presents the proportion of buildings with material information (by count and by area) at the city level in this dataset. Most cities achieve count-based proportions exceeding 60 %, while cities characterized by large areas of high building density (e.g., Delhi, Dubai, Hong Kong, Mumbai, and Nairobi) show lower count-based proportions. These lower proportions primarily reflect limited street-view accessibility in dense neighborhoods with narrow streets, where numerous buildings lack corresponding images (Fig. S3). Since buildings in these inaccessible areas tend to be smaller, area-based proportions are typically higher than count-based metrics in these cities.

Figure 8 Spatial coverage of building material data inferred from GSV images: (a) Building counts and material coverage proportions at 1 km grid resolution in representative cities; (b) Proportion of buildings with material information by count (blue bars) and by area (red scatter) at the city level.

Figure 9 illustrates the spatial coverage of building material data inferred from BSV images. Figure 9a shows building counts and material coverage proportions at 2 km grid resolution in mainland Chinese cities. Central urban areas demonstrate notably higher material data visibility, while suburban zones show progressively lower proportions of buildings with material information, revealing distinct spatial patterns. Figure 9b and Table S5 present the proportion of buildings with material information across buffer zones at increasing distances from city centers. These proportions generally decline with distance from urban centers. Some exceptions exist, such as Beijing, where the proportion within the 5 km buffer is lower than within the 10 km buffer, primarily due to street-view image gaps in restricted areas like the Forbidden City. Overall, while BSV-derived material data exhibit lower completeness than GSV-derived data, average proportions of buildings with material information within 10 km buffers reach approximately 40 %.

Figure 9 Spatial coverage of building material data inferred from BSV images: (a) Building counts and material coverage proportions at 2 km grid resolution in representative cities; (b) Proportions of buildings with material records at varying distances from city centers.

To assess the spatial coverage quality of the dataset, we quantified building material diversity and street-view visibility across concentric annular buffer zones at 3 km intervals extending to 30 km from each city center (Figs. 10a, b, S4 and S5). Visibility rates generally decline with distance from city centers, reflecting sparser street-view sampling in peripheral areas. Material diversity patterns are spatially coherent and geographically interpretable (Figs. 10c, S6 and S7). Chinese cities sustain consistently high diversity across all buffer rings, reflecting rapid and spatially diffuse urbanization since the 1990s in which construction occurred simultaneously across all urban zones. Cities such as Chicago, Sydney, and Toronto show monotonically declining diversity outward, consistent with the concentric zone model where centers accumulate historically diverse building stock while suburbs are dominated by uniform residential materials. More complex profiles are equally interpretable: Singapore's U-shaped curve reflects standardized public housing in inner zones and industrial facilities at the fringe, while Mexico City's outward-increasing diversity captures informal settlements built with locally variable materials at the urban edge. These geographically consistent patterns across diverse global cities confirm that the dataset captures meaningful material variation across urban gradients.

Figure 10 Spatial gradients of building material diversity and composition across different cities. (a) Illustration of equal-interval concentric buffers. (b) Shannon diversity trends and visibility rates across buffers. (c) Material proportions across buffer zones for some cities.

4.4 Multi-year records of building facade materials

This dataset collects all available historical street-view images at the individual building level to derive year-specific material records. The temporal completeness of these records depends on street-view acquisition frequency, with GSV providing more frequent updates and more complete spatial coverage per cycle than BSV (Fig. S2).

GSV coverage spans 2007 to 2025 with frequent acquisition cycles, and the proportion of buildings with material records is relatively high across most cities, with many exceeding 40 % coverage in multiple years (Fig. 11). Cities ranked by overall coverage show that top-ranked cities such as Paris, San Francisco, and Boston maintain consistent records across nearly all years, while lower-ranked cities exhibit more intermittent patterns. BSV coverage starts from 2013 with less frequent acquisition intervals, and building coverage rates are generally lower than GSV (Fig. 12). For most Chinese cities, material records are concentrated in several acquisition years rather than evenly distributed over time. For buildings with material records from multiple years, the latest available street-view observation was used to define the latest material label. Overall, the latest material labels are mostly derived from observations after 2021, indicating that BMAT provides relatively recent material information for most buildings with material records (Fig. S8).

Figure 11 Annual proportion of buildings with material records derived from Google Street View (GSV) images across cities. Red crosses indicate years with no valid records.

Figure 12 Annual proportion of buildings with material records derived from Baidu Street View (BSV) images across cities. Red crosses indicate years with no valid records.

We further summarized the temporal records of BMAT, including the number of annual material records per building, the temporal span of material records, and the temporal sequence types of material labels by city. These statistics show that the frequency and duration of temporal records vary across cities: many GSV cities contain a large share of buildings with repeated annual records, whereas several BSV cities are dominated by single-year records (Figs. S9 and S10). For buildings with multiple annual records (Figs. S11 and S12), the sequence-type statistics show that most records are either stable across years or exhibit high-confidence material changes, suggesting that BMAT provides reliable temporal evidence for many repeatedly observed buildings. Complex fluctuations account for a smaller subset and are treated as uncertain cases, potentially reflecting mixed-material facades or model uncertainty.

To examine the potential use of BMAT's multi-year records, we tracked material transitions from 2015 to the latest available observation year across five Asian cities (Fig. 13a). The recorded transitions are broadly consistent with expected urban development patterns. Glass coverage increases across many cities, reflecting the broad trend toward modern cladding materials, further supported by building age analysis showing that glass and metal facades are concentrated in buildings constructed after 2000 (Figs. S13 and S14). Mainland Chinese cities show greater material turnover than Singapore and Tokyo, consistent with their faster pace of ongoing urbanization. At the individual-building level, paired street-view images show visually verifiable examples of apparent facade transitions, including brick-to-glass and concrete-to-tile changes on the same buildings across years (Fig. 13b). These temporally coherent patterns suggest that the dataset's multi-temporal records can provide exploratory evidence of facade-material transitions across cities.

Figure 13 Building material transitions over time: (a) Flows and statistical shifts of material composition in case study cities; (b) Visual comparison of building facade before and after material changes.

5 Discussion

To our knowledge, BMAT is the first large-scale dataset that links dominant visible street-facing facade material labels to individual building footprints (Table S3). In addition to these spatial features, this dataset incorporates a temporal dimension, explicitly mapping time-series material attributes to specific building footprints. This dataset is derived from over 147 million street-view building facade images. We utilized a fine-tuned vision-language model (Qwen2.5-VL-7B-Instruct), which was trained on 39 405 human-labeled facade images – the largest manually annotated dataset for building material classification to date, notably exceeding the second-largest benchmark of 17 914 images (Chen et al., 2025). As a result, the model achieves an overall F₁ score of 0.91, surpassing traditional deep learning models and ensuring high data reliability for downstream analysis.

This dataset enables diverse applications in built environment research and urban analytics. The building-level material information supports refined building stock estimation (Liang et al., 2025a; Fishman et al., 2024), facilitating more accurate assessments of urban material composition and construction patterns (Dai et al., 2024; Rajaratnam et al., 2025). Material attributes can be integrated into urban heat environment simulations to improve thermal modeling accuracy (Tarkhan et al., 2025; Wonorahardjo et al., 2022), and incorporated into building energy consumption estimation by accounting for envelope material properties (Liu and Abbasabadi, 2025; Liu et al., 2025). The dataset also enables specialized analyses such as photovoltaic potential assessment of glass facades (Dong et al., 2025; Zhao and Gou, 2023) and material-specific building vulnerability studies for hazards including floods (Balasbaneh et al., 2019; Xing et al., 2023) and fires (Zamanialaei et al., 2025; Kee et al., 2025). In addition, the year-specific material records can provide exploratory evidence of temporal material changes associated with urban renewal processes (Naik et al., 2017; Xue et al., 2025) and life-cycle assessment of building materials (Sharma et al., 2011; Zabalza Bribián et al., 2011).

Beyond applications in the built environment, the dataset's metadata structure offers unique analytical opportunities. As material records derive from temporally stamped street-view images, the data inherently capture spatiotemporal sampling patterns of street-view platforms (Biljecki and Ito, 2021; Kim and Jang, 2023). Temporal analysis can reveal geographic disparities in image update frequency (Wang et al., 2024c), identifying areas with outdated building records (Fig. S2) or systematic differences between platforms (e.g., Google Street View vs. Baidu Street View) (Wang et al., 2025). Spatially, missing material records may also contain useful spatial information. In some areas, large clusters of missing records may result from narrow roads, private or restricted access that make it difficult for street-view vehicles to enter or capture usable facade images (Kamalipour and Dovey, 2019). In other cases, missing records may reflect platform-specific coverage gaps (Fan et al., 2025).

However, this dataset also has some limitations. First, because the dataset relies on street-view images, it cannot capture every building in a city (Fan et al., 2025). Buildings located in gated communities or far away from main roads may be missing. Since building materials often show strong similarities within the same neighborhood (Fleischmann et al., 2020), future work could potentially address this by using data from nearby buildings to estimate these missing attributes. Second, the dataset currently records only the dominant facade material, which may simplify the representation of buildings composed of mixed materials (Wang et al., 2024b; Chen et al., 2025). It is worth noting that the time-series material information recorded in this study offers a potential way to address this. Frequent changes in predicted materials over time, for example, switching between brick and glass, might not be model classification errors. Instead, these variations likely reflect the fact that the building is composed of multiple materials, causing the model to generate inconsistent predictions across different years. Third, BMAT records the dominant material visible from street-facing views. For a small proportion of buildings whose materials differ across facades, the street-facing facade material may not fully represent the material composition of the entire building.

6 Data availability

BMAT is openly available at https://doi.org/10.6084/m9.figshare.31569370 (Yin, 2026a).

7 Code availability

The code developed for this study is publicly available on GitHub (https://github.com/yhyJoy/BMAT, last access: 10 November 2026). The archived version (v1.0.0) is available on Zenodo at https://doi.org/10.5281/zenodo.23940973 (Yin, 2026b).

8 Conclusions

This study presents the first large-scale, footprint-level dataset of dominant visible street-facing facade materials, effectively bridging the gap between geometric building footprint data and facade material attributes. Derived from over 147 million street-view images via a fine-tuned vision-language model, the dataset covers 73 major cities with high classification accuracy. The dataset records material attributes for 22.09 million individual buildings across major cities worldwide. In addition, unlike conventional datasets that capture material information at a single time point, this dataset includes temporal records spanning multiple years, enabling the analysis of material changes and building renovation patterns over time.

BMAT adds building-level facade-material information to large-scale urban building datasets. These records provide an important material input for studies of building energy, embodied carbon, and disaster vulnerability, reducing reliance on uniform or assumed material assignments. The temporally stamped observations can also support exploratory identification and comparison of persistent facade-material changes across buildings and cities. BMAT therefore provides an empirical basis for material-sensitive research on the built environment and urban sustainability.