Reconstructing building genealogy with visual intelligence

Abstract

The study of vernacular buildings involves recording, classifying, and analyzing large building populations to probe their environmental and cultural explanations. Traditionally, this process has relied on manual observation and researcher intuition, which can lead to selective and imprecise interpretations. Here we propose a computational framework that integrates visual intelligence to support the reconstruction of building genealogy. The framework is based on the premise that interpreting historical artefacts involves identifying a function from discrete data. We apply it to the stylistic classification of Singapore shophouses using frontage images of 1,276 historic shophouses in Singapore’s Chinatown. Findings move beyond the chronological order established by the Urban Redevelopment Authority of Singapore in the 1990s and offer a critical perspective on the multi-ethnic character of Singapore shophouses. The study serves as a model for collaboration between studies in human settlements and computer science, demonstrating how leveraging the strengths of both fields can yield remarkable insights.

This study uses visual intelligence to uncover a hidden genealogy of Singapore shophouses, showing that vernacular architecture evolved through dynamic cultural interaction rather than through a simple linear history.

Introduction

Vernacular buildings are one of the most ubiquitous forms of material culture, providing a robust empirical record to understand people and their place^1–3. Vernacular buildings are also one of the most underrated cultural heritages, often overshadowed by monumental or iconic structures and constantly threatened by abandonment and destruction under relentless waves of rapid urbanization and modernization^4–6. Despite international charters and agendas being proposed to advocate their conservation^7–9, our understanding of these vernacular built remains insufficient, which in turn hinders practical judgment of what and how to preserve such buildings in practice. The study of vernacular buildings has been a common ground for antiquarians, historians, geographers, folklorists, archaeologists, anthropologists, architects and more^10,11. For years, they have made great strides to record, classify, and order buildings, in order to probe environmental and cultural explanations for their form, which in turn help inform future conservation and development practices¹². Traditionally, this work is conducted manually and intuitively by researchers, entailing extensive labor that often overlooks intricate building types or variants¹³. Such a situation might be tenable from an art historical perspective because its interpretation of vernacular buildings is intrinsically selective, depending on the compatibility of form with contemporary aesthetics^10,14,15. However, from a social scientific perspective, “the attempt to account for all buildings is fundamental”¹⁴, and the conditions noted above may impede the credibility of building genealogy, undermining efforts to understand its connotative socio-cultural change¹.

In recent years, the rapid development of various digital technologies has showcased enormous potential for moving the research paradigm of vernacular buildings towards greater efficiency, objectivity, and profundity. Firstly, the sophistication of remote sensing tools, such as cameras¹⁶, LiDAR (Light Detection and Ranging)¹⁷, satellites¹⁸, UAV (Unmanned Aerial Vehicle)¹⁹, and more, has produced vast amounts of data relating to vernacular buildings. Platforms like Google Street View have provided online access to images that cover most parts of the world and anything photographable from a public street. This democratic rather than discriminative record of extant buildings ensures greater credibility^20,21. They are also publicly accessible resources, showcasing the great potential of liberating researchers from the intense fieldwork of data collection^22,23.

Secondly, technological advancements in computer vision have developed varied image-processing algorithms that can be used to solve the problems of architectural classification²⁴. Early works focused on combating repetitive labour in architectural classification tasks. They used simpler machine-learning algorithms to classify images with manually extracted features (such as edges, textures, corners, etc.). Shalunts et al²⁵. proposed classifying a small group of images of building façade windows with three pre-determined architectural styles: Romanesque, Gothic and Baroque. By using Scale Invariant Feature Transform (SIFT) to extract the local features, the study completed the task with a high accuracy of 95.16%. The rise of deep learning in the 2010s, specifically convolutional neural networks (CNNs), has allowed for more sophisticated and automated image classification tasks. Works that probed the architectural typological systems started to emerge. Based on the deep convolutional neural network (DCNN), Sun et al²⁶. re-examined the stylistic system of residential buildings in Cambridge, formulated by Lindenthal and Johnson²⁷. The study demonstrates the correlation between style and age, which manifests most prominently in the architectural feature of the window. This discussion on AI-augmented formal taxonomy is preliminary but already shows great potential for deeper investigation.

Lastly, scientific data analysis packages have developed many mathematical models to describe and interpret the relationship among taxa^28,29. These models have been widely applied in archaeology and linguistics, but rarely in the study of vernacular architecture. A significant gap is that the theoretical hypothesis that underpins much of the study of vernacular buildings fails to be elaborated explicitly^10,30. In contrast, other domains within cultural studies have evolved to establish mathematical models based on certain propositions about culture³¹. For example, archaeologists employ phylogenetic approaches to model change in artefact forms across time and space, like stone tools³², pottery³³, carpets³⁴ and more. This is based on the proposition that “cultural change is systematic rather than capricious and on the auxiliary proposition that an important basis for the systematic behaviour of culture is its continuous transmission through the agency of person-to-person contact.”³⁵ Transferring these propositions and numerical methods to vernacular buildings studies is promising. On the one hand, it contributes to validating existing theories about vernacular buildings; on the other hand, it facilitates the results being theoretically justifiable, reproducible, and quantitatively defensible.

Drawing on previous scholarly endeavours, this study aims to advance computational vernacular studies by formulating the building genealogy through guiding human intuition with vision intelligent technologies assemblage. By incorporating collection, classification and interpretation, our proposed framework provides researchers with a systematic visual intelligence toolkit to examine built forms in dispersed geographical contexts, in order to derive justifiable maps of relatedness. The Singapore shophouse, a quintessential urban vernacular building^36–38, is selected as the subject of our scholarly investigation. The study creates and analyzes a georeferenced image set of surviving 19th and 20th-century shophouses in downtown Singapore. The computational modeling results indicate that the formal change of shophouses across time and space is best captured by a phylogenetic network, which organizes the shophouse types into nine distinct clusters. In turn, we also show how this genealogy reveals concurrent signals of cultural evolution and diffusion across a range of building data. Here, cultural evolution manifests as a transition from traditional to modern architectural styles, while cultural diffusion, particularly during periods of economic prosperity, is characterized by the emergence of diverse ethnic styles, indicating cultural exchange among different ethnicities, alongside competition and integration. The paper begins with some conceptual background and literature review on how we might “compute the vernacular”. The goal is to develop a computational framework to facilitate visual intelligence in interpreting vernacular built forms. This is followed by some contextual information on Singapore and its shophouse vernacular. The data is then presented in a three-fold discussion highlighting the processes of collection, classification and interpretation. The study ends by verifying the hypothesis and thinking through the wider applications of our work to other architectural contexts, practices and empirical settings.

Results

Computing the vernacular

While it is tempting to imagine vernacular buildings as changeless, we must acknowledge that change is inherent in the nature of the vernacular^10,39–41, reflecting the complexity of the socio-cultural system. Genealogy serves as an indispensable means of describing the change in vernacular buildings and landscapes across time and space.

Traditional approaches to genealogy rely on researchers’ intuition, established through the constant re-experiencing of identical or similar artifacts. Fred Kniffen’s pioneering work on settlement geography in the U.S. involves the recording and classification of folk housing, in which he lamented, “There are times in the field when a deluge of variant or new types and combinations leads to a state of utter confusion on the part of the observer. There is then no alternative to stopping for as long as is necessary to identify the pieces and fit them into place”¹³. Because human perception is selective and partial, the resulting interpretation of buildings is necessarily broad and loose, often lingering on form descriptions that follow a linear historical progression or a crude regionalist model. The Encyclopedia of Vernacular Architecture of the World, edited by Oliver Paul (1997), is often regarded as one of the most comprehensive reference works for international vernacular buildings studies till today, bringing together the vernacular traditions of over 2000 different cultures from around the world⁴². Its categorization of vernacular buildings is based on cultural regions delineated by existing global mappings. While reflecting the nature of vernacular buildings generally, scholars admit that this system is unable to capture the cross-cultural or geographical particularities of vernacular traditions⁴³.

Researchers have acknowledged the gravity of the situation, and efforts to deepen the study are underway. Henry Glassie is one such vanguard scholar who advocated precise definitions and systematic coding of artefacts under investigation. Despite no computational technique being involved, his work Folk Housing in Middle Virginia: A Structural Analysis of Historic Artifacts adopts a rigorous approach to formal taxonomy to investigate 338 real houses. Each building is closely examined based on a pre-defined artefactual grammar through wide fieldwork and manual coding, which describes the recurring building elements and their syntactic arrangement to generate a genealogy of building types (See Supplementary Fig. 1). Informed by structuralism, his study aims to move beyond the building forms and uncover the human ideas behind them. He explained the architectural change in Middle Virginia through the makers’ design competence: “he is reliant not on one original, but on a competence constructed out of numerous originals. He labors within tight lines of correctness, but his design competence allows for a variable range of actions”¹.

Drawing on Glassie’s work, we develop a computational framework in which researchers can use visual intelligence tools to guide their intuition concerning vernacular buildings, helping them categorize buildings while also unpacking deeper structures that govern formal changes. Sitting at the core of this framework is our proposal that the interpretation of historical artefacts is a process of finding a Function from a discrete data set. The research framework begins with establishing a hypothesis, followed by three phases of data collection, classification, and interpretation, concluding with the verification or revision of the hypothesis and the provision of final explanations (Fig. 1).

Fig. 1. The visual intelligence framework of vernacular buildings.

It begins with establishing a hypothesis, followed by three phases of data collection, classification, and interpretation, concluding with the verification or revision of the hypothesis and the provision of final explanations.

Formulating a genealogical hypothesis means identifying a principle that could plausibly organize how one group of forms gives rise to another. In its initial stage, the hypothesis does not yet specify a detailed evolutionary sequence; instead, it primarily serves to delimit the scope of observation and to identify which aspects of form are treated as potentially meaningful. This hypothesis is therefore broad and mainly achieved by considering the many factors that shape the evolution of vernacular buildings, such as geographical environment⁴⁴, ethnicity⁴⁵, craft tradition⁴⁶, and others.

In the data collection phase, researchers prepare building image sets and formulate the feature rules based on the established hypothesis. The image set, which has benefited from various remote sensing technologies, can perform with high integrity and accuracy. Concurrently, the manually formulated feature rules aim to translate each building image into a structured description, forming discrete data points in a multi-feature space.

In the classification phase, these discrete data points will be clustered into a finite set of formal types. This can be achieved by annotating each building with feature rules and identifying building groups that share similar assemblages of features. However, direct manual annotation is both inefficient and prone to missing or misreading subtle elements. To address this, the framework incorporates supervised learning. By labelling a certain number of building images with key feature rules, we can train the supervised model to automatically detect all formal features of each building image and represent these features using strings. Building images with similar strings will be automatically grouped into the same cluster. The classification phase concludes with a set of equivalent clusters X(y), each characterized by a particular feature vector y.

In the interpretation phase, these equivalent clusters will be connected by structural lines informed by Function f. Here, the Function f is instantiated by computational models that infer the evolutionary relationship among clusters from their feature strings. Taking X(y) as input, these models produce a candidate genealogy f′(X(y)). This candidate structure is then confronted with independent historical evidence (such as dated photographs, building regulations, and archival descriptions). If f′(X(y)) is broadly consistent with these sources, the original general hypothesis can be sharpened into a more specific explanation of how particular building types emerge, transform, and coexist, and the function f is approximated by an acceptable ƒ′.

We suggest integrating these well-established techniques from data acquisition, deep learning and computational modelling, which offer a rigorous and productive method to support the researcher’s work. This visual intelligence framework, we argue, establishes a rationalist epistemological stance on vernacular buildings studies. The next section provides a brief background on Singapore’s shophouse vernacular before we demonstrate in the subsequent section how the framework has been successfully used to document and explain formal changes in Singapore shophouses across time and space.

Mapping the genealogy of Singapore shophouses

Built between the 1840s and 1960s, the Singapore shophouse is “a vernacular building of two to five-storey height; following a basic floor plan of a narrow frontage and common party-walls; with a versatility of being able to serve many functions such as economic, residential and recreational purposes”⁴⁷. Shophouses constituted much of the urban fabric in pre-World War II Singapore⁴⁸, but rapid urban renewal since the 1960s has resulted in the demolition of many in the city. It was not until the late 1980s that the Urban Redevelopment Authority recognized shophouses as artefacts of a nation’s cultural identity and proposed to preserve shophouses. Currently, there are approximately 6500 shophouses included in the conservation scheme, and those in the four historic districts of Boat Quay, Chinatown, Kampong Glam, and Little India are regulated under the strictest conservation rules, which require the entire building envelope to be retained and restored.

The front façade assumes a pivotal role in expressing the cultural meaning of shophouses and is frequently singled out for structuring formal taxonomies of shophouses. Most prominently are the six styles proposed by the Urban Redevelopment Authority (URA) of Singapore via extensive fieldwork in the 1980s and 1990s, namely early shophouse style (1840–1900), first transitional shophouse style (1900–1910), late shophouse style (1900–1940), second transitional shophouse style (late 1930s), art deco shophouse style (1930–1960), and modern shophouse style (1950–1960)⁴⁹ (See Supplementary Fig. 2).

Upon detailed scrutiny, however, we contend that this taxonomy raises concerns that are at odds with the perspectives of various other studies^47,50,51. Jon Lim’s research shows that architects like Wan Mohamed Kassim and Yeo Hock Siang had already submitted building drawings with spectacular elevations before 1900 for approval by government authorities⁵⁰. These cases do not fit well with the current stylistic classification. A similar situation is observed in Emerald Hill, where, according to architectural historian Lee Kip Lin “the age of the older houses, which span between 1901 and 1918, the façade treatment var(ies) considerably in detail.”⁵¹ Lee, in his article “Identifying the shophouse”⁵¹, opposes the categorization of shophouses by time period. An inclusive formal taxonomy of Singapore shophouses may not necessarily adhere to a chronological principle. Identifying “key styles” serves, therefore, as a heuristic device for understanding Singapore shophouses but can hardly cover their myriad variations¹⁵.

This study employs the visual intelligence framework to inform the formal taxonomy of Singapore shophouses, thereby providing a deeper understanding of Singapore’s society. We hypothesize that there is an undiscovered relationship between the shophouse facade form and Singapore’s societal culture (h1). As suggested in the framework, this hypothesis is preliminary and will be gradually refined and further developed in subsequent steps.

Given that the act of naming buildings as vernacular marks a transition from the unknown to the known, the process of studying vernacular buildings is easily involved in the biased narrative caused by intentional building filtering^20,21. The visual intelligence framework underlines a non-discriminative data collection, which could be achieved by using varied remote sensing technologies.

Applying this approach to Singapore shophouses, we assembled two image sets, each with 3361 façade images, drawn from four conservation districts. Within this, Chinatown has all 1276 listed buildings documented exhaustively and constitutes the primary analytical focus. The remaining images, collected across Boat Quay, Kampong Glam, and Little India, were incorporated to reflect a broader spectrum of architectural variation and thereby strengthen model robustness in downstream classification. Together, this multi-district documentation spans the full range of vernacular conditions, from highly ornamented façades to heavily weathered, altered, or partially obscured structures, thereby mitigating the selection bias typically associated with conventional sampling.

Massive amounts of building data need to be characterized for the following systematic classification. The framework suggests formulating the feature rules at this stage so that varied building forms can be rationally represented by the assemblage of features. The feature rules are formulated manually by researchers based on comprehensive literature review and systematic field observations. Ensuring that each of the feature units has a relatively stable form and meaning is very important.

In the case of Singapore shophouses, the feature rules are set on the scale of façade elements. Each façade element on the shophouse has a relatively constant formal character and an explicit similarity with European classical, Chinese traditional, Malay traditional, and modern architectural vocabularies. Supplementary Note 1 provides a glossary of the Singapore shophouse façade.

The classification stage aims to identify groups with similar assemblages of features from massive amounts of building data, within which the key is to tally the features on each of the building units. However, manual annotation is very inefficient and prone to missing or misreading certain featured elements¹³. To address this, the visual-intelligent framework incorporates supervised learning to automatically detect and record^33–36, enabling a more detailed and comprehensive classification system. Notably, not all feature elements are to be covered, and only the core elements that may best capture the cultural meaning of vernacular forms will be distilled.

Initially, we selected 16 façade elements from the glossary of Singapore shophouse façades as label classes and trained a supervised learning model to automatically detect their presence or absence on shophouse façades. The workflow is illustrated in Fig. 2 and is explained in detail in the Method section.

Fig. 2. Collection and classification.

a The workflow from collection to classification. b The 16 featured façade elements. c The performance of the YOLOv5-based detection model in identifying the 16 label classes.

Upon reviewing the presence or absence of 16 elements on the façades of 1276 shophouses in Singapore’s Chinatown, we removed 2 façade elements, green glazed tiles canopy and casement windows, because of their low discriminability in differentiating different building types. The results produced a 14-digit binary string for each building, where 0 indicates absence and 1 indicates presence. By counting duplicates of their 14-digit strings and filtering out those with duplicate values of less than 8, we identified 55 shophouse variants, representing 55 distinct building types, exemplified by 122 individual buildings. See Supplementary Data 1 for the list of featured elements in all shophouse facades and Supplementary Table 1 for the list of shophouse variants. Based on frequency analysis of the remaining 1154 stable shophouses, we identified 25 string labels, each representing a distinct façade type with specific element composition (Fig. 3a).

Fig. 3. The classification results of Singapore shophouse façades.

a The 25 types of Singapore shophouse facades in Chinatown. b The correlation index of façade elements in composing main types. c The frequency and percentage of shophouses containing each façade element.

Figure 3b shows the correlation index of façade elements in composing the main types. Majolica tiles and Malay transoms are absent from the 25 main types of compositions and are only occasionally shown on the remaining variants. The other 12 elements, however, automatically aggregated into several clusters. For example, three elements of stepping parapet, modern windows and shades correlated frequently. However, their co-emergence with other elements is low. Figure 3c shows the frequency and percentage of shophouses containing each façade element. Long windows and main pilasters are the most prevalent façade components, found in 81% and 82% of the buildings, respectively. In historical archives, the two elements were also noted as the earliest found in shophouses, a prominent example being Baba Yeo Kim Swee’s godowns, constructed in the 1840s⁵². They were designed by George Coleman, Singapore’s first professionally trained architect with classical architecture education. Although these buildings are devoid of ornamentation, the modular repetition of their two principal elements along the row produces a sense of classical order, which appealed to the colonial authorities of the day.

Fanlights and secondary pilasters are found in 52% and 28% of the buildings, respectively. Chinese calligraphy and Chinese decorative panels are found in 15% and 12% of the buildings, respectively. Majolica tiles, Malay transom, and fretwork fascia registered the lowest percentages, at 0.1%, 1.6%, and 0.18%, respectively. The results indicate that the influence of the Malay is weak in an ethnic district predominated by the Chinese. However, being a British colony, the influence of European classical architecture dominates the urban vernacular landscape. The analysis of building elements can therefore help suggest the factors driving building and architectural changes, but this perspective remains very preliminary.

Vernacular buildings are changing products and are successively produced in one place. Therefore, they are all related. The genealogy of vernacular buildings aims to discern a pattern of relationships among building types. The visual-intelligence framework suggests introducing statistical models with cultural hypotheses to model the change in building forms, thereby deriving theoretically justifiable maps of relatedness. This study employs a phylogenetic method to relate 25 Singapore shophouse types, each represented by a 12-digit string. This 12-digit string is derived from the previous 14-digit string by excluding Majolica tiles and Malay transoms, which are absent from the 25 main types. Cultural phylogenies are built on the proposition that artefacts are phenotypic features of culture and that variations in these artefacts result from cultural transmission. We retrieved three hypothetical graphic patterns from existing phylogenetic models to map the relatedness of Singapore shophouse façade types: line (h1-1), branch (h1-2), and network (h1-3). These three patterns embody alternative interpretations derived from the preliminary hypothesis (h1) (Supplementary Fig. 5).

The linear model, commonly known as seriation, suggests “arrang(ing) comparable units in a single dimension (that is, along a line) such that the position of each unit reflects its similarity to other units.”⁵³ Seriation assumes cultural transmission is one-way. This one-way transmission can be further incorporated with temporal dimension to describe a cultural evolutionary mode characterized by heritable continuity. An essential characteristic of this heritable evolution is that “the parental taxon goes extinct when the daughter taxon appears”⁵⁴.

Figure 4 shows that 25 types have been seriated based on their composed façade elements in a non-constrained environment. Results specify a signal of heritable evolution emerging between the former nine elements of Chinese calligraphy—Chinese decorative panel, long window, main pilaster, fanlight, secondary pilaster, modillion, festoon, fretwork fascia, and the latter three elements of stepping parapet, modern windows and shades—resulting in three distinct groups: Group traditional (24, 22, 19, 23, 13, 20, 14, 09, 15, 17, 06, 16, 21, 18), Group Transitional (11, 10, 07, 08), and Group Modern (4, 2, 5, 3, 1). Historical archives support that evolutionary signal. Traditional shophouses built before and during the early 20th century feature ornaments with ethnocultural traditions (e.g., see drawings of shophouses in 1899^55,56, 1904^57–60, and 1909^61,62). However, the development of new materials and technologies in the 20th century has led to a rethinking of tradition and a preference for streamlined ornamentation (e.g., see drawings of shophouses in 1925⁶³). Entering the 1930s, modernism took hold amidst an economic recession and the influence of global architectural trends (e.g., see drawings of shophouses in 1934⁶⁴). Modernist architecture rebelled against tradition, viewing ornamentation as a crime. The principle of “form follows function” dominated, incorporating geometric forms with functional elements, presenting a transformative look. Seriation analysis reveals a chronological narrative hidden in the 25 main types by concluding on the three groups as noted above. However, a closer look at the neighbor relationship between two of the seriated types reveals a problem. For example, Type 24, composed of four elements of Chinese calligraphy, Chinese decorative panels, main Pilasters, and long windows, is designated as the earliest type, but this does not align with historical archives (e.g., see drawings of shophouses from 1884 to 1886⁶⁵, the earliest drawings on shophouses submitted to Municipal Architect & Building Surveyor’s Department). This may be explained by the fact that the linear mode has greatly simplified the cultural narrative of Singapore shophouses.

Fig. 4. The seriation of Singapore shophouse facades.

Results reveal a chronological narrative hidden in the 25 main types by concluding on the three groups: Group traditional in red, Group Transitional in gradient, and Group Modern in blue.

The tree model, developed upon the neighbor-joining algorithm, suggests a phylogenetic tree to depict the evolutionary relationship among the taxas. This tree assumes a branching pattern of lineage splitting to model cultural transmission. When a lineage splits, it divides into two distinct groups. Like the linear mode, the tree mode also adheres to the principle of heritable continuity from ancestors to descendants.

Figure 5 illustrates the phylogenetic tree of the 25 shophouse facade types. According to the hypothesis of cultural evolution, each internal node of the tree represents a heritable evolution event that “the parental taxon goes extinct when the daughter taxon appears”⁵⁴. When viewed as a whole, the tree has three prominent clades, containing types with predominantly modern elements (Clade Blue), European elements (Clade Orange), and Chinese elements (Clade Red), respectively. The phylogenetic tree indicates that Clades Blue and Orange diverged from the same node, suggesting they share a common ancestor (A_bo), whereas Clade Red and ancestor A_bo diverged from a different node, indicating they evolved from a distinct lineage. Each clade is composed of many subclades. The smallest clade always includes two types split from the same node. They are considered sister groups, indicating the highest cultural proximity (3 and 5, 8 and 10, 1 and 6, 18 and 21, 24 and 25, 22 and 23, 19 and 20).

Fig. 5. The phylogenetic tree of Singapore shophouse facades.

The tree has three prominent clades, containing types with predominantly modern elements (Clade Blue), European elements (Clade Orange), and Chinese elements (Clade Red), respectively. Clades Blue and Orange are diverged from the same node, suggesting they share a common ancestor (A_bo), whereas Clade Red and ancestor A_bo diverged from a different node, indicating they evolved from a distinct lineage.

Based on our direct observations and historical archives, the smaller clades in the tree offer greater explanatory power for types in which most compositional elements are repetitive. For example, types 3 and 5 exhibit very similar visual language and construction techniques, indicating that both were formulated by architects based on a particular prototype, with minor revisions. However, larger clades with multiple layers of evolution in the tree are less explanatory. This is because phylogenetic trees follow a strict bifurcating model. In referring to historical archives, many of the ancestor-descendant relationships indicated by the phylogenetic tree are invalid. Most prominently, no evidence shows that the Clade Red featuring predominantly Chinese elements emerged earlier than the Clade Orange featuring predominantly European elements (e.g., see drawings of shophouses in 1899^55,56, 1904^57–60 and 1909^61,62). We conclude that the tree model cannot fully address the evolution of Singapore shophouses.

The network model, developed upon the neighbor-net algorithm, suggests a split graph to depict the relatedness among the cultural taxes. Unlike the tree model, which assumes binary evolution via hierarchical ancestor-to-descendant inheritance, the network model posits reticulate transmission, allowing multiple connections. These better capture non-hierarchical signals in cultural transmission, like cultural hybridization, borrowing, and parallel evolution. In this sense, the network model can best represent the hypothesis of cultural diffusion that “the patterns of similarity and difference among cultural assemblages are a consequence primarily of individuals in different groups copying each other’s practices, exchanging ideas and objects, and marrying one another”⁶⁶. Building on preliminary studies of façade elements that suggest frequent intra-ethnic exchange in Singapore’s society, we speculate that a network model more effectively captures the relationships among the 25 types.

Figure 6 illustrates the phylogenetic network of 25 facade types. When viewed as a whole, the network exhibits a branching growth pattern extending outward from the center toward distinct directions; the types in the center have fewer featured façade elements. In contrast, those in the peripheral area exhibit more facade elements, suggesting cultural evolution. We identified four clusters in the network: the Group Plain in grey, the Group Chinese in red, the Group Euro in yellow, and the Group Modern in blue. The Group Plain comprises Types 12 and 13, characterized by basic elements such as long windows and main pilasters. Historical archives also suggest that these two types were the first to emerge in Singapore, dominating the city’s urban landscape in the early 19th century⁶⁷. Viewed in conjunction with the branching growth pattern, Group Plain can also be seen as the earliest-evolved entity. Group Chinese in red comprises six types with Chinese elements, including Chinese calligraphy and decorative panels. Historical archives suggest that these types in Group Chinese emerged later than Group Plain and were initially constructed for affluent Chinese owners (e.g., see the shophouse drawings at no. 68 and 69 Upper Hokien Street for Loh Seng Tai in 1904⁶⁰). Group European comprises six types with elaborate European elements, like fanlight, secondary pilaster, modillion, and festoon. Historical archives suggest that these types also emerged later than Group Plain but proximately in the same period as Group Chinese (e.g., see the shophouse drawings from 1889 to 1891⁶⁸). Since the late 19th century, the number of Group Yellow buildings in Chinatown has increased, even exceeding that of Group Red. They were constructed under the commission of not only European owners but also other ethnic groups, reflecting a broader influence of colonial power in the society. Group Modern in blue comprised 11 types with modern elements, like stepping parapets, modern windows and shades. Historical archives suggest Group Modern emerging as the latest in the 20th century (e.g., see the drawings of shophouses in 1934⁶⁴). Despite comprising 11 types, the total number of constructions in Group Blue in Chinatown is low. The economic downturn and disruptions of World War II significantly slowed new construction. Moreover, there is a dearth of heritable continuity between Group Modern and earlier Group Chinese and European. Viewed in conjunction with the branching growth pattern, Group Modern, like Group Chinese and European, can all be seen as having evolved from Group Plain. Comparatively, the evolutionary directions of Group Chinese and European are closer, while that of Group Modern is distinctively far away.

Fig. 6. The phylogenetic network of Singapore shophouse facades.

We identified four clusters in the network: the Group Plain in grey, the Group Chinese in red, the Group Euro in yellow, and the Group Modern in blue. Nine distinct subgroups were further identified based on the four primary groups, which are also the nine styles that best capture the character of Singapore shophouses: Plain Style (13, 12), Plain Chinese Style (19), Chinese Style (22, 24), Chinese-European Style (20, 25, 23), Plain European style (14), Elaborated-European Style, Ethnic-hybrid Style (21), Stripped Ethnic Style, and Modern Style. Horizontal cultural transmission exists among the styles.

In addition to its tree-like structure, the phylogenetic network contains numerous loops, indicating conflicting transmission signals. Each group and type is influenced by multiple traditions, which complicates classification based on the tree. Nine distinct subgroups were further identified based on the four primary groups, which are also the nine styles that best capture the character of Singapore shophouses: Plain Style (13, 12), Plain Chinese Style (19), Chinese Style (22, 24), Chinese-European Style (20, 25, 23), Plain European style (14), Elaborated-European Style (15,17,16,18), Ethnic-hybrid Style (21), Stripped Ethnic Style (9, 11, 6, 10, 7, 8), and Modern Style (1, 2, 3, 4, 5) (Fig. 6). Their spatial distribution was shown in Supplementary Fig. 7. Plain Style overlaps with Group Plain in grey. It is primarily function-oriented, with minimal ornamentation: a modest amount of moulding serves as capitals for the columns and pilasters, while the windows are square-headed and fitted with jalousie shutters.

Plain Chinese Style, Chinese Style and Chinese-European Style are all split from Group Chinese in red. Chinese Style is defined by traditional Chinese decorative elements, either wall plaques with auspicious characters, plaster relief panels with Chinese motifs, or a combination of both. Chinese-European Style, however, incorporates these traditional Chinese decorative elements with classical European elements, creating a fusion look. Historical studies indicate these Chinese-European Style shophouses were primarily built by Chinese merchants who wanted to preserve their cultural roots while aligning with colonial elites to enhance their social and economic status⁶⁹. The style exemplifies the cultural diffusion from Group European to Group Chinese.

The Plain European Style, Elaborated European Style and Ethnic-hybrid Style are all split from Group European in yellow. The Plain European Style is similar to Plain Style but with added European fanlights. The Elaborated European Style, developed from the Plain European Style, features more decorative European elements such as secondary pilasters, modillions, and festoons. The Ethnic-Hybrid Style combines traditional Chinese, European and Malay elements, presenting a peculiarly vernacular mix. This explains Type 21, which has an exceptionally long line extending from the rest of the European Group. The Ethnic-Hybrid Style shares many similarities with the Chinese European Style within Group Chinese, as both represent forms of cultural hybridity. However, the Ethnic-Hybrid Style also displays a formal language dominated by a decidedly European classical language.

Historical archives reveal a more nuanced understanding of the cultural transmission of ethnic traditions as reflected by the Ethnic-Hybrid Style. Statistically, there are eight Ethnic-Hybrid Style shophouses in Chinatown, all located on Bukit Pasoh Road. Historical architectural drawings suggest that they were all designed by Swan & Maclaren, one of the most renowned architectural firms in Singapore since the colonial period, employing professional architects trained in European classical architecture. The owner of this row of eight shophouses is Seow Chwee, a Straits-Chinese merchant. In fact, the initial design by Swan & Maclaren showcased an orthodox classical architectural image. Alternating triangular and segment pediments were employed to crown the window, while the decorative panels beneath them featured festoon ornamentation (Supplementary Fig. 6). However, their design was not fully reflected in the final façade and could have been modified by the Chinese artisan under the owner’s commission, which was a common practice during the time⁷⁰. The festoon on the second floor was replaced by eight Chinese symbols with auspicious meanings, the Covert Eight Immortals. They also added the fretwork fascia of the traditional Malay house under the roof edge. The Ethnic-Hybrid Style is the most elaborate among the 25 types. It is also the only case that incorporates Malay elements in Chinese-owned buildings. Despite Type 21 exemplifying multi-ethnic integration, historical archives reveal that ethnic competition also played a significant role in shaping its hybrid style⁷⁰.

The Stripped Ethnic Style and Modern Style are both split from Group Modern in blue. While the Stripped Ethnic Style reduces traditional ornamentation but incorporates modern geometrical elements, leading to some innovative forms, the Modern Style features only modern geometrical elements like stepping parapets, modern windows, and shades. The Stripped Ethnic Style occupies a large area of the phylogenetic network, indicating that the six types are relatively distant. This is because, driven by the manifesto of anti-tradition, architects in the early 20^th century were keen on exploring a new visual language that could better embody the spirit of the industrial era. The Stripped Ethnic Style was born out of this context, showcasing great diversity and a spirit of experimentation. As the Modern Style emerged, experimental exploration stabilized.

The phylogenetic network offers a critical perspective for examining Singapore shophouses, which feature multi-ethnic characteristics. The exchange of ethnocultural traditions occurred during a period of economic prosperity from the late 19th century onwards. However, ethnic power dynamics remained unbalanced, as reflected in the dominance of European classicism in architectural styles. Statistically, European influences prevailed, with the Elaborated-European and Plain European Styles accounting for 23.57% and 17.23% of the total, respectively. Chinese influence was comparatively lower, with the Chinese Style comprising 13.63%. Fusion styles were the least common, with the Chinese-European, Ethnic Hybrid, and Stripped Ethnic Styles representing 4.54%, 3.52%, and 4.70% of the total, respectively. This distribution suggests that different ethnic groups remained in competition, displaying a pattern of parallel evolution rather than direct convergence. While some degree of architectural fusion did occur, it remained limited, reinforcing the prevailing cultural distinctions in Singapore.

Discussion

In this study, we propose a research framework for using visual intelligence to elucidate the genealogy of vernacular buildings. We suggest viewing the interpretation of historical artefacts as a process of finding a Function from a discrete data set. With this, visual intelligence technologies can be assembled to support researchers in four key aspects: efficiently mining building data; converting data into feature clusters; identifying the mathematical function that best relates these clusters; and converting that function into an architectural narrative. By recasting an architectural issue as a mathematical question, a computational framework for vernacular buildings can move beyond the reductive narrative models of traditional studies. It probes more deeply into the synchronic threads of development embedded within diachronic processes, promising a breadth and depth that conventional methods have yet to achieve.

Grounded in an epistemological stance on solving functions, we formulate a systematic workflow that integrates data collection, building classification, and relatedness interpretation using remote sensing, computer vision, and computational modelling technologies. First, spatial data acquisition technologies have significantly improved the quality of primary data for vernacular buildings studies. Unlike traditional methods that rely on piecing together fragmented materials to understand buildings, spatial data acquisition techniques enable large-scale, multi-temporal building representation with systematic data.

Second, the incorporation of machine learning, especially computer vision and deep learning, has ushered vernacular building research into a new stage of automation in feature extraction and feature clustering. Traditional studies have conventionally depended on experts manually identifying features based on their experience. In contrast, machine learning algorithms can automatically learn morphological characteristics from vast amounts of architectural images and structural data. This shift from individual expertise to a data-driven approach can help to uncover deeper, more nuanced patterns. In the case of Singapore shophouses, we identify 25 building types, which corrects the six types formulated by the URA in chronological order.

Finally, developing mathematical functions that capture the relatedness among feature clusters is key to unveiling the genealogy of vernacular buildings. A common pitfall in vernacular building studies is overlooking the need to set cultural hypotheses at the outset. As a result, research often becomes a linear narrative of progress or a fragmented functionalist framework, lacking depth. A vision-intelligent approach to vernacular buildings demands that scholars begin with clear hypotheses, aiming for both comprehensive, nuanced observations and straightforward, impactful concepts. The hypotheses also need to be constantly revisited and updated in view of empirical findings as they are uncovered, as our study shows.

We have demonstrated the effectiveness of the research framework through a case study of Singapore shophouses. Findings suggest that the formal change of shophouses across time and space is best captured by the phylogenetic network, which organizes the types into nine distinct clusters and reveals concurrent signals of cultural evolution and diffusion. Here, cultural evolution manifests as a transition in styles from the traditional to the modern, while cultural diffusion, particularly during periods of economic prosperity, is characterized by the emergence of diverse ethnic styles. However, we also observe that different ethnic groups remained in competition in colonial Singapore, displaying a pattern of parallel evolution rather than direct convergence. While some degree of architectural fusion did occur, it remained limited, reinforcing the prevailing cultural distinctions. As noted in our prior review, existing manually constructed genealogies are predominantly chronological and can overlook the simultaneous presence of markedly different types. Our computed approach, by comparison, integrates diachronic and synchronic dimensions and can accommodate higher complexity.

Rather than directly using digital tools to generate cultural narratives, we focus on how these tools can be assembled more effectively to guide scholars’ intuition and yield insightful findings. In this sense, expert knowledge does not disappear; it is redistributed. Experts remain responsible for formulating the feature rules, framing the cultural hypotheses, and judging which inferred genealogies are historically and theoretically persuasive. The automated components of the framework take over the exhaustive work of collecting and grouping thousands of façades, and of computing alternative classifications. The expert then reads these outputs not as ready-made explanations but as prompts for further interpretation: asking why particular buildings cluster, how loops in the network relate to known episodes of contact or competition, and which outliers warrant closer archival or field investigation. Our case study suggests that this complementarity opens lines of enquiry that would be prohibitively time-consuming by manual methods alone, such as systematically comparing alternative style groupings, probing the extension and transformation of types across districts and periods, or testing how different feature rules reshape genealogy. Despite this, our study is constrained by 2D street view imagery and 14-digit binary feature encoding. Future endeavors should focus on optimizing the relevant technologies for data collection and feature rule formulation (see Supplementary Note 2). By doing so, we hope this framework serves as a valuable mechanism for incorporating digital tools into vernacular buildings studies, fostering more fruitful collaborations between architectural and computer science scholarship in the future.

Methods

Collection

We collected building frontage images from two sources: Google Street View (GSV) and on-site photography. This strategy is not pre-designed but rather emerges from iterative experiments; a 100% reliance on the public geospatial database is impractical in our case. This is because the currently achievable maximum resolution of GSV images is relatively low, which limits the ability to complete the entire data analysis procedure^71,72. As such, we used the GSV image set to identify the featured domain on the building façade⁷³. Concurrently, we introduced on-site photography at a higher image resolution to identify the featured element on each building façade. GSV imagery was acquired using the Google Street View Static API with standardized parameters: 640 × 640-pixel resolution, 90-degree field of view, and systematic coordinate-based sampling across all target districts. Images were retrieved with consistent viewing angles positioned perpendicular to building frontages, as determined from vector-based building footprints by projecting a normal line to the nearest street segment, to minimize perspective distortion^74,75. High-resolution field photography (3024 × 4032 pixels) was conducted using standardized protocols with cameras positioned 15–20 m from façades, maintaining perpendicular orientation to building centroids⁷⁶.

We developed a comprehensive vocabulary of Singapore shophouse façades through an extensive literature review and field study, which includes many featured façade elements that have long been overlooked in existing studies. These façade elements are either decorative, functional, or a combination of the two. The frequency of these façade elements emerging on the shophouse frontage also varies; some are frequently found, like festoon, bian-e (Chinese plaque/caligraphy), and kepala tingkap (Malay transom); some others are less seen, like the Diocletian window, ge-shan (Chinese partition screen), and tangga batu (stone staircase).

Classification

We began by setting up label classes for façade elements in the model for training. These label classes were extracted from the glossary of the Singapore shophouse façade based on two actions. First, we trained a ResNet-50 classifier⁷⁷ on the GSV image set across four historical districts using ImageNet pre-trained weights, with a 80:20 train-test split, batch size 32, learning rate 0.001, and 50 epochs. Grad-CAM⁷⁸ analysis identifies façade elements above the ground floor contributing most to cultural classification (see Supplementary Fig. 3). Later, we applied frequency-based filtering, excluding elements that appeared fewer than 40 times to ensure adequate representation for model training. In the end, we defined 16 label classes of façade elements: main pilaster, fanlight, secondary pilaster, festoon, modillion, Chinese calligraphy, Chinese decorative panel, Green glazed tiles canopy, Malay transom, fretwork fascia, Majolica tiles, long window, modern window, casement window, shades, and stepping parapet. We also introduced the building label to focus the learning area within the upper level of the shophouse façade.

We annotated 708 of 3361 on-site photographs with 17 label classes (16 architectural elements plus building class) and partitioned them into training, validation, and testing sets. A frequency-stratified splitting strategy was employed: elements with fewer than 200 instances followed a 70:15:15 (train: validation: test) ratio, while more frequent elements maintained an 80:10:10 split ratio. Additionally, data augmentation techniques and class-weight adjustments were implemented to improve the model’s performance on less frequent classes.

Several data augmentation strategies, including mosaic composition, mix-up blending, geometric transformations, and photometric variations, are employed to enhance model robustness under varying façade conditions (weathered surfaces, varied lighting, partial occlusions) and across scale differences among elements (from small modillions to large pilasters). We used 3024 × 4032-pixel on-site photography (resized to 640 × 640 pixels) for model training, aiming to preserve fine decorative details. Its performance achieved a mean Average Precision of 0.936 (mAP@0.5) across all 17 classes. Complete model performance information is provided in Supplementary Methods, Section 3.

When using the model for feature detection, we assigned class-specific confidence thresholds to account for varying visual characteristics across feature elements. We also introduced a conflict-resolution mechanism to address view-dependent classification ambiguities in which characteristic features of certain elements were obscured by viewing angles or open positions. For example, casement windows photographed when open often expose only glass panes, potentially leading to misclassification as modern windows. When the spatial overlap between conflicting detections exceeded IoU > 0.5, priority rules were applied to retain the more architecturally characteristic classification. This process generated binary character states (0 = absence, 1 = presence) for each of the 14 architectural elements, producing 14-dimensional binary feature vectors for each shophouse façade. Complete threshold specifications are provided in Supplementary Methods, Section 3.

We manually checked the 16-dimensional character vectors generated for all shophouse façades in Singapore’s Chinatown and corrected all errors in the detection results. At this stage, elements obscured by vegetation or structurally damaged through weathering were reinterpreted by experts using fragmentary evidence. Expert validation was used to assess whether each element possessed sufficient typological discriminative power to justify its inclusion in the classification framework. Results show that two featured elements, green glazed tile roofs and casement windows, should be excluded because of their low discriminability in differentiating different building types.

The remaining 14 elements are identified as the primary features that most effectively represent the formal character and cultural significance of Singapore shophouses (See Supplementary Data 1). We generated a 14-dimensional character vector for each shophouse façade. After filtering out character string label 00000000000000 and other categories with fewer than 8 instances, we yielded 25 distinct character string labels.

Interpretation

To map the relatedness among 25 shophouse types, we tested three distinct phylogenetic functions, line, tree, and network, each corresponding to a different mode of cultural transmission. The linear model was built using PAleontological STatistics (Version 4.17)^29,79, while the tree and network models were built with SplitsTree6 (Version 6.4.7)²⁸.

To begin, we prepared a data matrix that tabulates the presence and absence of m facade features across n shophouse types. Presences are denoted by 1, and absences are denoted by 0. The façade features are on columns, whereas the shophouse types are on rows. Two featured façade elements, Majolica tiles and Malay transom, were omitted at this stage because they are absent from all 25 shophouse types. The n × m binary matrix has n = 25 shophouse types (rows) and m = 12 facade features (columns).

The goal is to reorder the data matrix so that feature occurrences are aligned as closely as possible along the diagonal, thereby minimizing the span of each feature across the reordered sample set. We applied unconstrained seriation following the iterative computational method introduced by Wilkinson^80,81. At each iteration, we calculated the mean column position of presences for each shophouse type and reordered rows accordingly; we then calculated the mean row position of presences for each facade feature and reordered columns. The cycle was repeated until row and column orders stabilized. The mean column positions of presences (row centroids) and the mean row positions of presences (column centroids) are calculated as follows:

$${\mu }_i^{\left(t\right)}=\frac{1}{k_i}\sum _{j\in J_i^{\left(t-1\right)}}{\mathrm{pos}}_C^{\left(t-1\right)}\left(j\right)$$ 1

$${\nu }_j^{\left(t\right)}=\frac{1}{k_j}\sum _{i\in I_j^{\left(t\right)}}{\mathrm{pos}}_R^{\left(t\right)}\left(i\right),$$ 2

where ${\mathrm{pos}}_C^{\left(t-1\right)}\left(j\right)$ is the column position of feature j at iteration $t-1$, and ${\mathrm{pos}}_R^{\left(t\right)}\left(i\right)$ is the row position of facade type i after the current row reordering. Sorting by ${\mu }_i^{\left(t\right)}$ and ${\nu }_j^{\left(t\right)}$ yields $R^{\left(t\right)}$ and $C^{\left(t\right)}$, respectively.

To construct the following tree and network model, the prepared data matrix must first be converted into a distance matrix. Pairwise morphological distances between all 25 architectural facade types were calculated using proportional Hamming distances (p-distances)^82,83, generating a symmetric distance matrix D of dimensions 25 × 25, providing standardized similarity measures between building types suitable for phylogenetic reconstruction. Each matrix element $d_{ij}$ represents the proportion of differing architectural elements between facade types i and j, calculated as:

$$d_{ij}=\frac{{\sum }_{k=1}^n∣X_{ik}-X_{jk}∣}{n},$$ 3

where n represents the total number of morphological features encoded in the facade feature matrix ($n=14$ in this study), matrix $X_{ik}$ denotes the binary state (0 or 1) of the k-th architectural element in facade type (binary encoding type) $i$, and $d_{ij}$ represents the proportional distance between facade types i and j. The resulting symmetric distance matrix D serves as the basis for subsequent distance-based analyses.

$$D=\left(\begin{array}{ccccc}\hfill 0\hfill & \hfill d_{12}\hfill & \hfill d_{13}\hfill & \hfill \cdots \hfill & \hfill d_{1,25}\hfill \\ \hfill d_{21}\hfill & \hfill 0\hfill & \hfill d_{23}\hfill & \hfill \cdots \hfill & \hfill d_{2,25}\hfill \\ \hfill ⋮\hfill & \hfill ⋮\hfill & \hfill \ddots \hfill & \hfill \ddots \hfill & \hfill ⋮\hfill \\ \hfill d_{25,1}\hfill & \hfill d_{25,2}\hfill & \hfill \cdots \hfill & \hfill \cdots \hfill & \hfill 0\hfill \end{array}\right).$$ 4

The Tree function employs the Neighbor-joining algorithm⁸⁴ to construct a hierarchical tree that reveals evolutionary relationships among facade types. Initially, each facade type represents an independent unit. Grounded in minimum evolution principles, the algorithm iteratively identifies the most evolutionarily related pairs using the Q-criterion⁸⁵, which selects pairs that are close to each other while both being distant from all other types, thereby distinguishing genuine evolutionary relationships from coincidental proximity. The two facade types with the minimum Q-value are then merged into a new internal node representing their hypothetical common ancestor, and this process continues until all facade types are integrated into a complete binary tree, revealing the comprehensive evolutionary hierarchy of architectural facade typologies.

At each iteration, the selection criterion Q-value for facade type pairs $\left(i,j\right)$ is computed by applying p-distances (calculated according to Eq. 3) in the following formula:

$$Q\left(i,j\right)=\left(r-2\right)d_{ij}-\sum _{k\ne i}^r{d}_{ik}-\sum _{k\ne j}^r{d}_{jk},$$ 5

where r denotes the number of remaining taxonomic units in the current iteration, $d_{ij}$ represents the pairwise distance between types i and j, and the summation terms ${\sum }_{k\ne i}^r{d}_{ik}$ and ${\sum }_{k\ne j}^r{d}_{jk}$ represent the total distances from types i and j to all other remaining types, respectively. The formula implements this principle by subtracting the total distances (summation terms) from the scaled direct distance, ensuring that pairs exhibiting the desired pattern yield minimal Q-values as optimal merging candidates. Once identified, the pair with the minimum Q-value is selected, and their branch lengths are recorded as:

$$L_i=\frac{1}{2}\left[d_{ij}+r_i-r_j\right],$$ 6

where $r_i$ and $r_j$ represent the average distances from facade types i and j to all other remaining types. Branch lengths quantify the evolutionary divergence from the hypothetical common ancestor to each descendant type, providing a measure of morphological change along each lineage. The algorithm then updates distances by treating the newly formed internal node as a single unit in subsequent iterations, repeating this process until all facade types form a complete binary tree.

To capture more complex, non-hierarchical relationships reflecting architectural style hybridization and convergent design evolution, we applied the Neighbor-Net algorithm^84,86, which generates circular split collections represented as planar networks. Unlike NJ’s immediate merging approach, Neighbor-Net introduces clusters to facilitate delayed grouping and selection. Delayed agglomeration aims to preserve multiple relationship hypotheses, generating complex split collections that cannot be represented by a single tree. The algorithm first computes the Q-criterion to identify optimal cluster pairs at the cluster level, then selects specific facade types from these clusters at the node level to establish neighborhood relationships. When a node is paired for the second time, a three-node agglomeration occurs, replacing the three connected nodes with two new internal nodes and updating the distance matrix. As the algorithm progresses, the cluster structure updates continuously. This iterative process continues until few nodes remain, then reverses the agglomeration sequence to construct split collections, optimizing split weights through constrained least squares to generate the final circular split network characterizing complex architectural relationships.

Based on the p-distance matrix (Eqs. 3 and 4), the algorithm established a clustering framework in which each facade type initially formed an independent single-node cluster. Inter-cluster distances were calculated as:

$$d\left(C_p,C_q\right)=\frac{1}{∣C_p∣∣C_q∣}\sum _{i\in C_p}\sum _{j\in C_q}{d}_{ij},$$ 7

where$C_p$ and $C_q$ represent the clusters (groups) of facade types, $d\left(C_p,C_q\right)$ represents the distance between clusters $C_p$ and $C_q$, $∣C_p∣$ and $∣C_q∣$ denote the number of façade types in each cluster, and $d_{ij}$ is the proportional distance between facade types i and j.

Cluster pair selection employs a modified Q-criterion operating at the cluster level:

$$Q\left(C_p,C_q\right)=\left(m-2\right)d\left(C_p,C_q\right)-\sum _{\begin{array}{c}\hfill \ell =1\hfill \\ \hfill \ell \ne p\hfill \end{array}}^{m}d\left(C_p,C_{\ell }\right)-\sum _{\begin{array}{c}\hfill \ell =1\hfill \\ \hfill \ell \ne q\hfill \end{array}}^{m}d\left(C_q,C_{\ell }\right),$$ 8

where $Q\left(C_p,C_q\right)$ represent the cluster selection criterion, m denotes the number of active clusters, and $C_{\ell }$ represents any other cluster ($\ell \mathcal{\{\{\backslash relax\; \backslash special\; \{t4ht̂3)\}\backslash o:mathrel:\; \{\backslash unhbox\; \backslash voidb@x\; \backslash special\; \{t4ht@+\{38\}\{35\}x2260;\}x\}\}\}}p,q$) in the summation terms.

When neighbor marking results in a facade type having two neighbors, three-node agglomeration triggers, replacing the three connected facade types $i,j,h$ with two new internal nodes u and v. Distance recalculation follows the reduction formulae:

$$d\left(u,a\right)\&=\alpha d\left(i,a\right)+\beta d\left(j,a\right)$$ 9

$$d\left(v,a\right)\&=\beta d\left(j,a\right)+\gamma d\left(h,a\right)$$ 10

$$d\left(u,v\right)\&=\alpha d\left(i,j\right)+\beta d\left(i,h\right)+\gamma d\left(j,h\right),$$ 11

where $\alpha$, $\beta$, $\gamma$ are non-negative real numbers with $\alpha +\beta +\gamma =1$, $a$ represents any other node (facade type) in the network.

Upon convergence, split weights are optimized using constrained least squares:

$$\mathbf{w}=\arg \underset{\mathbf{w}\ge 0}{min}\mid \mid \mathbf{Aw}-\mathbf{d}{\mid \mid }^2,$$ 12

where w represents the vector of split weights, A is the split-distance incidence matrix that encodes which splits separate each facade pair, d contains the vectorized observed proportional distances, and the constraint $\mathit{w}\ge 0$ ensures all split weights are non-negative. This optimization seeks to minimize the discrepancy between network-predicted distances and actually measured architectural distances, ensuring that the phylogenetic network accurately reflects observed formal relationships among facade types.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

X.X., Y.Y. and Z.T. contribute equally to this work. X.X. conceived the project ideas and designed the research framework. X.X. and Z.T. collected the image and textual data. Y.Y., Z.T. and X.X. implemented the computational analysis. X.X. implemented the non-computational archival studies. X.X. and Z.T. interpreted the results and co-wrote the manuscript. T.-C.C. and C.-K.H. were involved in the discussion and manuscript revisions.

Peer review

Peer review information

Nature Communications thanks Divyarajsinh Rana, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

X.X. discloses support for the research and publication of this work from the Jiangsu Education Department, China through the Jiangsu Specially-Appointed Professor Research Gran [grant number 492511301]. C.-K. H. and T.-C. C. disclose support for publication of this work from the National Heritage Board, Singapore through the Heritage Research Grant [grant number HRG-057].

Data availability

The list of featured elements in this study is provided in Supplementary Data 1. The raw facade image data generated in this study have been deposited in the Zenodo repository under accession link 10.5281/zenodo.17009527⁸⁷ and are publicly available. Google Street View images cannot be redistributed due to Google’s Terms of Service; the sampling points and coordinates used to retrieve the Street View imagery are provided in Supplementary Data 2. The model training code and JSON annotation files used in this study are available in the GitHub repository https://zenodo.org/records/19110655⁸⁸.

Code availability

The code used in this study is available at https://github.com/ZITA98/Decoding-the-Singapore-Shophouse.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-73376-7.