A Systematic Review of Ontology–AI Integration for Construction Image Recognition

Abstract

This study presents a systematic review of ontology–AI integration for construction image understanding, aiming to clarify how ontologies enhance semantic consistency, interpretability, and reasoning in AI-based visual analysis. Construction sites involve highly dynamic and unstructured conditions, making image-based hazard detection and situation assessment both essential and challenging. Ontology-based frameworks offer a structured semantic layer that can complement deep learning models; however, most existing studies adopt ontologies only as post-processing mechanisms rather than embedding them within model training or inference workflows. Following PRISMA 2020 guidelines, a comprehensive search of the Web of Science Core Collection (2014–2025) identified 587 publications, of which 152 met the eligibility criteria, and 16 explicitly addressed construction image data. Topic modeling revealed five functional objectives—regulatory compliance, hazard reasoning, decision support, knowledge reuse, and sustainability—and four primary data modalities: BIM, text, image, and sensor data. Two dominant integration patterns were observed: training-stage and output-stage enhancement. While quantitative performance improvements were modest, qualitative gains were consistent across studies, including reduced false positives, improved interpretability, and enhanced situational understanding. Persistent gaps were identified in standardization, scalability, and real-world validation. This review provides the first structured synthesis of ontology–AI research for construction image understanding and offers an evidence-based research agenda that links observed limitations to actionable directions for semantic AI in construction.

1. Introduction

Construction projects are complex socio-technical systems involving interdependent processes, diverse stakeholders, and heterogeneous data across all lifecycle stages—from design and execution to operation and maintenance [1]. These characteristics demand continuous decision-making in process management, quality control, and safety monitoring [2]. On construction sites, rapidly changing environmental and operational conditions, together with the interaction of workers and heavy machinery, further complicate real-time risk assessment and hazard response.

Safety, in particular, is marked by high uncertainty. The International Labour Organization (ILO) reports that 30–40% of global occupational fatalities occur in construction [3], while Korean national statistics indicate that 53% of occupational fatalities in 2023 occurred in this industry [4]. Such figures underscore the need for intelligent systems capable of supporting hazard detection, situational awareness, and informed decision-making.

Recent advances in deep learning–based image recognition have enabled notable progress in construction safety monitoring. YOLO-based detectors have demonstrated strong performance in identifying workers, personal protective equipment (PPE), and machinery under complex visual conditions [5]. Similarly, CNN-based intrusion detection systems have shown near–real-time performance with high accuracy [6]. However, these models depend on large-scale curated datasets and often struggle to generalize in highly variable construction environments [7]. More critically, deep learning models exhibit limited capability in representing contextual relationships or reasoning about site semantics, hindering interpretability and adaptability.

Ontologies have therefore emerged as a promising knowledge-based structure for supporting semantic interpretation and logical reasoning. Ontologies formalize domain concepts and relationships, enabling transparent inference pathways, improved explainability, and knowledge-based validation of AI outputs. For example, Zhong et al. [8] used ontology-driven semantic analysis to address fragmented safety knowledge, demonstrating how integration with BIM can support automated hazard detection. Ontologies further contribute to robust reasoning with relatively small datasets and facilitate human–AI semantic alignment, thereby improving interpretability and trustworthiness [9,10].

Despite these advantages, the construction domain remains challenging for AI-driven interpretation due to its unstructured environments, dynamic activities, and heterogeneous data streams. Ontology techniques—leveraging predefined domain concepts and inference rules such as SPARQL and SWRL—offer practical and reliable mechanisms for structuring domain semantics and supporting rule-based reasoning [11]. Existing literature shows a steady increase in research combining ontology with AI, particularly in computer vision and machine learning. Oluborode et al. [12] provided a broad review of ontology-based computer vision and machine learning research from 2011 to 2024, and Ghidalia et al. [13] examined integration strategies across 128 studies, including learning-enhanced ontologies, semantic data mining, and learning–reasoning hybrids.

However, these reviews remain general and do not provide a systematic understanding of ontology–AI integration specifically for construction image analysis. Prior studies have focused primarily on BIM- and text-based applications, while image-based ontology–AI research—critical for hazard recognition and spatial understanding—has received limited attention. Moreover, differences in ontology design choices, integration strategies, and AI pipeline configurations remain insufficiently explored.

This study addresses this gap by systematically reviewing ontology–AI integration approaches for construction image understanding. The objectives are threefold: (1) to identify and classify ontology–AI integration methods applied in construction image analysis; (2) to examine the functional roles of ontologies across these approaches; and (3) to assess reported impacts and limitations. Rather than claiming universal performance gains, this review emphasizes ontology’s role as a semantic framework that enhances interpretability and enables context-dependent improvements.

To achieve these aims, the study conducts a systematic literature review (SLR) that classifies ontology–AI integration according to application purpose, ontology construction method, AI model type, integration strategy, and evaluation practice. Following PRISMA 2020 guidelines (see Supplementary Materials), Section 2 presents the scope, search strategies, and eligibility criteria. Section 3 provides classification results, including research trends, integration types, and a detailed review of image-based approaches. Section 4 synthesizes research gaps and proposes evidence-based directions for future work. Section 5 concludes by highlighting the unique contributions and implications of this study.

2. Materials and Methods

2.1. Ontological Perspective

Ontologies can be interpreted through two philosophical perspectives: subjectivism and objectivism. A subjectivist view treats ontologies as consensus-based representations shaped by social agreements and contextual meanings [14,15]. In contrast, an objectivist perspective regards ontologies as formal semantic structures that encode factual relationships among entities, attributes, and processes [16,17].

This study adopts the objectivist perspective, viewing ontologies as explicit semantic frameworks that structure relationships—such as worker–equipment–location or hazard–condition–response—and support logical interpretation of construction data. Under this stance, ontologies are not merely taxonomic hierarchies but knowledge models capable of enabling inference, semantic consistency, and transparent reasoning. This perspective guided the analytical framework in two ways: (1) classifying ontology–AI integration by how ontologies structure information flow across preprocessing, model learning, and output interpretation and (2) analyzing functional roles such as hazard identification, workflow monitoring, and knowledge transfer.

2.2. PRISMA Process

This systematic review followed the PRISMA 2020 guidelines to ensure transparency, methodological rigor, and reproducibility. A ten-step procedure was applied in accordance with the official PRISMA flowchart instructions.

Step 1. Preparation: The official PRISMA flow diagram template was obtained from the PRISMA website and used to document each phase of the review. The search strategy, screening criteria, and exclusion rules were predefined to minimize bias.

Step 2. Database Search: The Web of Science (WoS) Core Collection was selected as the sole database because it indexes high-impact journals and major conference proceedings in construction, engineering, computer science, and AI. The search was conducted in April 2025, covering publications from January 2014 to March 2025. The query was formulated as: (“Ontology” OR “Knowledge Graph”) AND “Construction” AND “AI”. This search yielded 587 records.

Step 3. Removal of Duplicates: Because only a single database was used, no duplicates were found (n = 0). All retrieved records proceeded to screening.

Step 4. Title/Abstract Screening: All 587 records were screened to assess relevance to the construction domain based on title and abstract.

Step 5. Records Excluded (Title/Abstract Screening): A total of 240 papers were excluded because they were unrelated to construction or lacked construction-specific context.

Step 6. Reports Sought for Retrieval: The remaining 347 papers were selected for full-text retrieval.

Step 7. Reports Not Retrieved: All full texts were successfully obtained (n = 0).

Step 8. Eligibility Assessment (Full-Text Screening): The same 347 papers were assessed for conceptual and methodological eligibility.

Step 9. Reports Excluded After Full-Text Screening: Two exclusion categories were applied: 165 studies used ontologies or knowledge graphs only conceptually (e.g., background explanation or terminology) without operational integration into methods, models, or analysis, and 30 studies were surveys or reviews without empirical implementation. In total, 195 papers were excluded.

Step 10. Studies Included in the Review: Ultimately, 152 studies met all inclusion criteria. These studies explicitly integrated ontologies or knowledge graphs with AI techniques—such as machine learning, deep learning, NLP, embeddings, or GNNs—within a unified research workflow.

A summary of the screening procedure is illustrated in Figure 1. The final corpus of 152 publications forms the basis of the thematic and analytical framework presented in Section 3 [9,11,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94,95,96,97,98,99,100,101,102,103,104,105,106,107,108,109,110,111,112,113,114,115,116,117,118,119,120,121,122,123,124,125,126,127,128,129,130,131,132,133,134,135,136,137,138,139,140,141,142,143,144,145,146,147,148,149,150,151,152,153,154,155,156,157,158,159,160,161,162,163,164,165,166,167].

2.3. Data Collection and Extraction

For this review, ontology–AI integration was defined as any study in which ontologies or knowledge graphs were used in conjunction with AI techniques within a unified analytical workflow. Importantly, studies were considered eligible only when ontologies or knowledge graphs played an explicit and functional role in the analytical pipeline, rather than being mentioned superficially or conceptually. AI was interpreted broadly to include deep learning, machine learning, natural language processing (NLP), embedding-based learning, computer vision, transfer and few-shot learning, and graph neural networks (GNNs).

A study was classified as ontology–AI integration if it satisfied at least one of the following conditions: (1) AI contributed to ontology construction (e.g., concept or relation extraction, preprocessing, automated classification, topic modeling, or embedding generation); (2) ontologies supported AI training or inference (e.g., semantic constraints, semantic feature generation, or interpretability enhancement); (3) ontologies and AI operated in complementary roles, regardless of direct data exchange; or (4) AI outputs were incorporated into an ontology or ontology inference results were used as inputs to an AI model. Studies that merely referenced ontologies, knowledge graphs, or semantic concepts without demonstrating their operational use within the AI workflow were excluded during screening.

To extract analytical information from the 152 included studies, a predefined data-collection framework was developed and applied. One reviewer manually extracted all relevant information using a structured template aligned with the analytical dimensions presented in Section 3. During screening and extraction, particular attention was paid to distinguishing substantive ontology–AI integration from nominal references, ensuring consistency with the inclusion criteria. Extracted items were cross-checked during the synthesis stage to ensure consistency and accuracy, and no automation tools were employed.

Data extraction was guided by three research questions that operationalize the definition of ontology–AI integration and determine the specific information required for classification and thematic synthesis.

First, RQ1 (“What are the key trends and functional objectives of ontology–AI integration in the construction domain?”) informed the extraction of publication metadata, including year, venue, and country, as well as the identification of high-level functional objectives reported across the literature.

Second, RQ2 (“How are ontologies and knowledge graphs integrated with AI techniques across different data modalities?”) guided the collection of ontology construction approaches, data modalities such as image, BIM, text, and sensor data, the types of AI models employed, and the integration strategies used within AI pipelines. For data modalities involving structured building information, studies were included only when semantic structures (e.g., object attributes, relationships, or rule-based reasoning) were explicitly utilized, rather than purely geometric or non-semantic representations.

Third, RQ3 (“In ontology-integrated image recognition systems, how is ontology applied, what functional roles does it serve, and what impacts are reported?”) directed the extraction of functional ontology roles—including semantic reasoning support, interpretability enhancement, and hazard inference—as well as implementation practices and reported performance outcomes.

All extracted information was organized into summary tables and thematic matrices, which directly correspond to the analytical dimensions described in Section 3. When studies lacked detail—for example, when ontology structures or model specifications were incomplete—classification was based solely on explicit statements within the article, and unresolved cases were coded as “not specified.”

A formal risk-of-bias assessment was not conducted because this review synthesizes methodological characteristics and qualitative patterns rather than statistical effect sizes. Studies were grouped for synthesis based on the analytical structure of Section 3. Extracted information was tabulated in summary tables and thematic matrices, and a qualitative thematic synthesis supported by topic modeling was used to integrate findings. Heterogeneity was examined by comparing ontology construction approaches, integration types, and data modalities across studies.

3. Results

3.1. Bibliometric Overview

This section provides a comprehensive overview of the bibliometric characteristics of the 152 studies analyzed in this review. Its purpose is to outline the research evolution and structural landscape of ontology–AI integration in construction by examining temporal publication trends, major sources, and geographical distributions. Specifically, this analysis identifies how academic interest in the field has developed over time, which journals and conferences have served as the key dissemination channels, and how research contributions are distributed across countries.

Figure 2a depicts the annual number of publications on ontology–AI integration in construction from 2014 to April 2025. Between 2014 and 2017, publication activity remained modest—averaging five to six papers per year—marking an initial exploratory phase when ontology–AI integration was largely conceptual rather than empirical. In 2018, the annual output nearly doubled (12 papers), signaling the beginning of active research engagement, followed by steady growth through 2021. A significant surge occurred after 2022, with a peak of 35 publications in 2024. As of April 2025, 23 papers have already appeared, suggesting that the momentum of research is likely to continue. This trend demonstrates that ontology–AI integration has emerged as a sustained research focus within the construction domain, evolving from theoretical discussions to practical, data-driven implementations.

Figure 2b presents the distribution of publication sources. Automation in Construction dominates the field with 53 papers (34.9%), followed by Buildings (27, 17.8%) and various conference proceedings (26, 17.1%). Other notable outlets include the Journal of Construction Engineering and Management (11), Journal of Building Engineering (9), and Advances in Civil Engineering (5). Several interdisciplinary journals—such as Computer-Aided Civil and Infrastructure Engineering, Developments in the Built Environment, and Sustainable Cities and Society—have also featured related works, illustrating the cross-disciplinary nature of ontology–AI research spanning architecture, civil engineering, and applied AI systems.

A cross-analysis of publication sources and years further reveals the field’s developmental trajectory. Automation in Construction has consistently led since 2014, serving as an early adopter and core platform for ontology-based AI research in construction. Its publication rate increased notably from 2022 to 2024, reflecting a shift toward empirical automation and semantic modeling studies. Subsequently, Buildings and the Journal of Building Engineering began publishing more actively after 2021, expanding the topic’s reach into architectural and management-oriented domains. Likewise, traditional engineering journals such as JCEM and Advances in Civil Engineering have integrated ontology–AI topics since the early 2020s, indicating the mainstream adoption of AI-based information modeling in conventional construction engineering. Meanwhile, conference proceedings have maintained consistent presence from 2015 to 2024, showing that the field was initially driven by experimental demonstrations and pilot studies before gaining full journal recognition.

Figure 3a illustrates the global distribution of first-author affiliations using a heat map, and Figure 3b compares publication counts across 24 countries. China leads with 57 papers (37.5%), followed by the United States (10.5%), Germany (9.2%), the United Kingdom (8.6%), and South Korea (6.6%). Collectively, these five nations account for approximately 72% of all publications, highlighting a strong regional concentration of research in East Asia, Western Europe, and North America.

3.2. Topic Modeling-Based Trends in Ontology-Integrated AI Applications

To explore thematic trends in ontology-integrated AI applications in construction, Latent Dirichlet Allocation (LDA) topic modeling was conducted. An initial run revealed that overly generic words (e.g., approach, different, potential) appeared as top-ranked keywords, which limited interpretability. To address this, stopwords and low-information tokens were removed, and an additional TF–IDF–based pruning process was applied before re-running the final model. The number of topics (K) was determined using the coherence score (c_v).

Figure 4a presents a comparison of coherence values across K values ranging from 6 to 14. The results indicate that K = 10 yielded the highest coherence score (0.340) and was therefore selected as the optimal number of topics for this study. Figure 4b illustrates the relative size distribution of the ten topics. Each topic accounted for 5.9% to 18.0% of the entire corpus, corresponding to 9 to 26 papers, indicating that no topic was excessively dominant or marginal. This balanced distribution suggests that ontology–AI applications are being investigated across a wide range of subdomains in construction research.

To assign labels to the ten topics, representative papers and key keywords were examined. The analysis revealed that the topics can be systematically explained along two dimensions: Functional Objectives and Data Types. In other words, ontology–AI applications in construction can be characterized both by their intended purpose and by the types of data on which they are based.

The Functional Objectives identified from the analysis were Quality and Regulatory Checking, Decision Support, Hazard Reasoning and Risk Assessment, Knowledge Accumulation and Reuse, and Sustainability and Structural Assessment. The Data Types were categorized into BIM, Image, Text, and Sensor data. This dual classification indicates that ontology–AI applications tend to exhibit consistent patterns based on the combination of research objectives and data utilization.

Table 1. Topics Defined by Functional Objectives and Data Types.

Topic	Functional Objective	Data Type	Label Summary
0	Decision Support	BIM	Decision support for design and construction processes using BIM
1	Knowledge Accumulation and Reuse	Text + Image	Knowledge integration and retrieval using knowledge graphs and large language models, including vision-based applications
2	Quality and Regulatory Checking	Text	Information extraction and management from regulations, codes, and constraints
3	Sustainability and Structural Assessment	Image + Text	Structural damage detection and sustainability-oriented design assessment
4	Hazard Reasoning and Risk Assessment	Image	Image-based hazard assessment and risk interpretation
5	Quality and Regulatory Checking	BIM + Text	Regulatory compliance checking and safety knowledge mapping using BIM and textual data
6	Hazard Reasoning and Risk Assessment	BIM	Hazard identification and semantic reasoning based on BIM models
7	Knowledge Accumulation and Reuse	BIM + Sensor	Knowledge reuse for robotic inspection, bridge maintenance, and prefabricated construction
8	Hazard Reasoning and Risk Assessment	Sensor	Sensor-based environmental monitoring and cost-overrun risk assessment
9	Knowledge Accumulation and Reuse	BIM + Text	Information integration for BIM-based deconstruction planning and cost estimation

summarizes the classification of the ten topics across these two dimensions.

The analysis revealed five major functional objectives of ontology–AI applications in construction. These objectives reflect how ontology-driven approaches are employed to support both operational and strategic goals.

Table 2. Functional objectives of ontology-based AI applications in construction.

Functional Objective	Description	Example
Quality and Regulatory Checking	Evaluating design and construction quality and ensuring compliance with relevant codes and standards	BIM-based compliance checking; automated quality inspection; text-based regulation analysis
Decision Support	Supporting project planning, resource allocation, and process optimization	Prefabrication constructability analysis; resource identification; energy system evaluation
Hazard Reasoning and Risk Assessment	Identifying hazards, predicting accidents, and supporting proactive risk assessment	Image-based hazard detection; BIM-based safety reasoning; sensor-driven monitoring
Knowledge Accumulation and Reuse	Structuring, managing, and reusing project information and expert knowledge	Knowledge graph–based project information management; LLM-based querying; robotic inspection frameworks
Sustainability and Structural Assessment	Evaluating structural performance and sustainability-related aspects	Structural damage detection; low-carbon design assessment; sustainable building conservation

summarizes the definitions of each functional objective along with representative application examples drawn from the reviewed studies.

The topic modeling results also identified four primary data modalities: BIM, text, image, and sensor data. These data types represent the main sources through which ontologies enable semantic integration and reasoning, highlighting their application across different construction tasks.

In this context, “image” data refers specifically to two-dimensional visual image representations used for scene understanding and image recognition tasks, independent of image acquisition platforms. Three-dimensional sensing modalities, such as LiDAR-based point clouds, are therefore treated separately under sensor data due to their distinct data structures and processing pipelines.

Table 3. Data types used in ontology-based AI applications in construction.

Data Type	Examples
BIM	Structured spatial and locational data used for object property extraction, compliance checking, clash detection, design validation, and hazard recognition in construction projects
Text	Extraction of entities and relationships from accident reports, safety manuals, work logs, and technical documents to support hazard reasoning, regulatory compliance, and knowledge reuse
Image	Two-dimensional image–based object detection and recognition of worker status (e.g., PPE use, unsafe behavior), equipment position and utilization, and construction quality assessment, combined with ontology-based semantic reasoning linked to site activities and safety rules
Sensor	IoT-based spatiotemporal monitoring of worker localization, equipment operation, structural vibration, and environmental conditions to support real-time hazard assessment, often within digital twin platforms

summarizes the major data modalities and representative examples from the reviewed studies.

BIM data provide structured spatial and locational information and are widely used to support compliance checking, clash detection, design validation, and hazard recognition. Text data are mainly applied to extract entities and relationships from unstructured documents such as accident reports, safety manuals, and technical guidelines, thereby supporting regulatory compliance, hazard reasoning, and knowledge reuse. Image data are employed in tasks such as PPE compliance recognition, unsafe behavior detection, equipment monitoring, and construction quality assessment. In these studies, detected visual objects are semantically mapped to ontology concepts, enabling rule-based and ontology-driven reasoning for safety management and risk evaluation. Sensor data are utilized for real-time spatiotemporal monitoring, including worker localization, equipment operation, structural vibration, and environmental conditions, often in conjunction with digital twin platforms.

In addition to single-source applications, several studies demonstrated multimodal integration, for example, combining text–image or image–sensor data streams, to improve reasoning accuracy and enrich semantic context. These findings underscore the potential of ontologies not only as mechanisms for semantic enrichment but also as integrative frameworks for cross-modal information fusion and reasoning in complex construction environments.

Overall, the results suggest that ontology-based AI applications are not confined to a single task domain. Rather, they span immediate operational needs such as regulatory compliance and risk assessment, as well as long-term objectives including knowledge reuse and sustainability. This indicates a gradual expansion of ontology–AI applications in construction from short-term, task-specific optimization toward broader contributions to sustainable and intelligent project delivery.

3.3. Integration Types of AI and Ontology

This subsection explains how ontologies and knowledge graphs are integrated with AI methods in practice, focusing on recurring integration approaches rather than algorithmic details. Integration approaches are not included in topic labeling but are presented as a complementary analytical dimension that illustrates the implementation patterns of ontology–AI applications.

While this review does not aim to reproduce code-level details, it synthesizes how different studies positioned ontologies within AI pipelines and highlights common strategies reported in the literature. Although integration approaches do not directly emerge from topic modeling, an examination of representative papers shows close linkages between application themes and integration methods. For example, Topic 2 (information extraction from regulations, codes, and constraints) involves text analysis and ontology mapping, representing a typical NLP-based integration case. Topics 3 and 4 (image-based structural and hazard assessment) are associated with vision-model integration, where CNN or YOLO outputs are mapped to ontology concepts. In addition, Topic 6 (BIM-based risk reasoning) and Topic 9 (BIM-based deconstruction planning) are directly related to graph neural network (GNN)–based integration, which leverages graph-structured learning.

As summarized in

Table 4. Integration strategies of AI and ontology in construction.

Integration Approach	Definition	Examples	Related Topics
Embedding-based Integration	Transforms ontological concepts and relations into vector embeddings that are incorporated into AI model training or inference.	Poincaré embeddings for hierarchical structures; Word2Vec/BERT embeddings of ontology terms	Topics 0, 5, 6, 7, 9 (BIM attribute embeddings, multimodal compliance checking, risk reasoning, robotic applications, deconstruction planning)
NLP-based Integration	Processes textual data (e.g., regulations, reports, manuals) using AI models that reference, generate, or leverage ontological structures.	LLM–ontology integration for contract analysis; named entity recognition (NER) with ontology mapping from regulatory documents	Topics 1, 2, 5, 9 (information integration and retrieval, regulation and code extraction, compliance checking, cost estimation documents)
Vision Model Integration	Analyzes image or video data using vision models (e.g., CNNs, YOLO) and maps recognition outputs to ontology concepts or incorporates ontological knowledge into vision pipelines.	Image-based hazard detection linked to ontology-driven semantic reasoning	Topics 1, 3, 4, 7 (vision–knowledge graph integration for hazard recognition, structural damage detection, image-based hazard assessment, robotic inspection)
Few-shot and Transfer Learning Integration	Exploits ontological class hierarchies or semantic structures to support transfer learning, few-shot learning, or fine-tuning in data-scarce environments.	Ontology-based class hierarchies used as semantic features for transfer learning	Topic 8 (sensor-based environmental monitoring and risk assessment in data-scarce settings)
Graph Neural Network (GNN) Integration	Employs ontology or knowledge graph structures as graph inputs to GNN architectures (e.g., GCN, GraphSAGE, GAT) for relational learning and inference.	Risk-factor relationship graphs analyzed using GNNs	Topics 6, 9 (BIM-based risk reasoning, IFC- and knowledge graph–based deconstruction planning and risk mapping)

, ontology–AI integration in the construction domain is currently concentrated on embedding-based approaches, which transform ontological concepts and relations into vector spaces and thereby facilitate incorporation into deep learning models. In contrast, NLP-based integration is consistently applied in text-heavy domains such as regulatory compliance, while vision-model integration, though less frequent, demonstrates practical value in hazard detection and scene-level interpretation. GNN-based integration, along with transfer and few-shot learning, remains at an early stage but represents emerging directions in which ontologies are coupled with adaptive AI models for complex relational reasoning and data-scarce environments.

These findings indicate that current ontology–AI applications in construction primarily rely on embedding-based approaches but also reveal a growing body of studies exploring vision-based and GNN-based methods. This pattern suggests the importance of diversifying technical pathways and developing more detailed empirical implementations to broaden applicability and strengthen reproducibility.

3.4. Ontology-Based Image Recognition in Construction AI

While many of the 152 eligible studies incorporated some form of image data, this review specifically focuses on the subset that employed image recognition as the primary analytical method. This focus is justified because image recognition constitutes a foundational component of construction image understanding, is directly linked to construction safety and quality management, and offers one of the clearest demonstrations of ontology’s contribution—by mapping detected objects to ontology-defined semantic concepts and applying rule-based reasoning to interpret recognition results. Although image data appeared across multiple topics (e.g., Topics 1, 3, and 4), in most cases they were combined with other modalities and used as Supplementary Information. By contrast, the 16 studies analyzed in this section explicitly adopted image recognition as their central methodological component, warranting their selection for in-depth analysis. These studies collectively illustrate how ontologies enhance image-based AI pipelines by improving interpretability, reducing false positives, and supporting hazard reasoning. The following subsections (Section 3.4.1, Section 3.4.2, Section 3.4.3, Section 3.4.4 and Section 3.4.5) examine this subset across functional roles, ontology construction approaches, vision models employed, integration strategies within AI pipelines, and reported performance outcomes.

3.4.1. Functional Roles of Ontology in Image Recognition

The functional roles summarized in

Table 5. Functional roles of ontology-based reasoning in image recognition systems.

Functional Roles	Definition
Situation Awareness	Semantic interpretation of worker positions, behaviors, equipment interactions, and procedural compliance to assess normal or abnormal site conditions in real time.
Quality Regulation Compliance	Structuring quality regulations within ontologies and mapping them to recognized objects and relationships to automatically determine conformity or non-conformity.
Hazard Identification	Identifying potential hazards from site objects, equipment, and environmental factors and mapping them onto ontology classes and attributes for structured representation.
Risk Assessment	Classifying identified hazards into risk levels (e.g., low, medium, high) using rule-based reasoning or probabilistic models.
Process Tracking	Mapping recognized objects and task activities to BIM- and process-oriented ontologies to infer progress status and completion levels of construction workflows.
Anomaly Detection	Detecting deviations from ontology-based models of normal workflows—such as sequence errors, bottlenecks, or delays—by comparing actual site observations against expected process patterns.

were derived from the image-related topics identified in Table 1. For example, Topic 1 (Knowledge Accumulation & Reuse, Text + Image) corresponds to situation awareness, Topic 3 (Sustainability & Structural Assessment, Image + Text) corresponds to quality regulation compliance and anomaly detection, and Topic 4 (Hazard Reasoning & Risk Assessment, Image) corresponds to hazard identification and risk assessment. In this way, Table 5 refines these topic-level objectives into specific functional roles that capture how ontologies contribute within image recognition–based AI systems.

In such systems, ontologies extend beyond basic object detection by supporting semantic structuring, reasoning, rule interpretation, and process monitoring, thereby enhancing the interpretability and adaptability of recognition outcomes [18]. By modeling hierarchical and semantic relationships among domain-specific entities and concepts, ontologies enable fine-grained interpretation of interactions and scene contexts, allowing raw recognition results to be transformed into structured knowledge that supports intelligent decision-making [19,20]. In construction environments, where work conditions are complex and dynamic, ontologies facilitate real-time reasoning by linking visual recognition outputs to semantic frameworks. This enables monitoring of work status, process flows, compliance with safety standards, and hazard identification [21]. Such integration has proven useful not only for improving recognition performance but also for strengthening system-level interpretability and robustness, particularly in data-scarce environments [21,22].

Lee and Yu (2023) [21] proposed an ontology-integrated image recognition system that mapped construction site entities and safety regulations to real-time recognition outputs. Detected objects were passed to an ontology-based reasoner, which evaluated conditions such as “worker without PPE combined with an elevated work position,” issuing alerts for potential violations or hazardous situations. This example illustrates that ontologies function not merely as knowledge repositories but as active reasoning engines that interpret recognition results, generate explanatory insights, and support actionable decision-making for construction safety management.

However, critical perspectives also exist. Pileggi et al. (2024) [23] noted that many applications utilize ontologies primarily as post-processing tools for knowledge representation or semantic enrichment. They argue that such approaches fail to fully exploit ontologies as integrated reasoning components or active feature generators within AI pipelines. As a result, most systems remain limited to partial semantic enhancement rather than fully ontology-driven architectures, restricting their potential to support interoperability and explainability in human–machine collaboration.

These observations suggest the need for more empirically grounded strategies that embed ontologies more deeply into AI learning processes, enabling integrated reasoning, dynamic feature generation, and enhanced semantic alignment between AI systems and real-world construction contexts.

3.4.2. Ontology Construction Approaches

In image recognition–based AI systems, ontologies can be constructed using three primary approaches: manual, automated, and hybrid. These approaches differ in the degree of AI involvement in schema design (e.g., classes, relations, and attributes), instance creation, update processes, and their overall suitability for different application domains.

The manual approach relies on domain experts to design both the ontology schema and instances from scratch. This traditional method offers high interpretability and strong domain alignment, making it suitable for applications that demand high-precision reasoning, such as safety management or quality control. For instance, Lee and Yu [21] constructed an ontology schema based on safety regulations and linked image recognition outputs to a rule-based reasoner for real-time risk assessment. Zheng et al. [24] similarly developed a fully hand-crafted process ontology to monitor construction workflows. However, manual development is time-consuming, requires extensive expert knowledge, and lacks scalability.

The automated approach employs AI techniques—such as NLP, deep learning, or large language models (LLMs)—to extract concepts and relationships from unstructured data and to generate or update ontology instances. This method offers rapid development, high scalability, and the ability to process large datasets. For example, Pfitzner et al. [22] transformed recognition results into GNN-based embeddings to automatically expand a knowledge graph. Zhang et al. [25] implemented a pipeline that automatically mapped detected objects to ontologies for real-time hazard analysis. Nevertheless, automated methods may face limitations in precision and domain-specific accuracy.

The hybrid approach combines manual schema design with automated instance generation and updates. It seeks to balance manual reliability with automated scalability, making it particularly effective in real-time applications. Fang et al. [26] manually created a hazard hierarchy but automated object instantiation. Wu et al. [18] built a risk management ontology manually but integrated vision-based recognition outputs into an automated reasoning module.

Among the 16 analyzed studies (see

Table 6. Comparison of ontology construction approaches.

Approach	Representative Tools	Advantages	Limitations
Manual	Protégé, OWL	High interpretability; strong domain alignment	Time- and labor-intensive; limited scalability
Automated	NLP pipelines, LLMs, Graph databases	Rapid development; scalable for large datasets	Reduced precision; limited domain-specific accuracy
Hybrid	Protégé combined with automated mapping modules	Balances manual reliability with real-time adaptability	Higher system complexity; increased integration overhead

), hybrid approaches were the most common (7 studies, 43.8%), followed by automated (5 studies, 31.3%) and manual (4 studies, 25.0%). This distribution suggests a gradual shift away from traditional manual development toward hybrid frameworks—particularly in domains such as process monitoring and hazard detection, where both domain alignment and real-time adaptability are essential.

3.4.3. Image Recognition Models Applied

Among the 16 image-based studies analyzed, object detection models were the predominant choice, with the YOLO family being the most widely adopted (see

Table 7. Image recognition models used in construction AI and their key characteristics.

Model Family	Key Characteristics
YOLO (all versions)	Real-time object detection; lightweight and efficient; widely applied to construction safety monitoring and process-related tasks
Mask R-CNN	Precise object segmentation and fine-grained spatial analysis; suitable for structural damage detection and object state monitoring
VRD (VTransE)	Extraction of inter-object relationships aligned with ontology structures; supports contextual reasoning and compliance assessment
Image Captioning (Attention CNN + RNN)	Generation of descriptive scene narratives; enables ontology-based rule matching and semantic interpretation
Others/Unspecified	Includes BLE-based semantic trajectory analysis for location-based awareness and studies where vision models were not explicitly specified but image data were used for semantic enrichment or training

). This prevalence reflects the specific requirements of construction environments, where real-time performance, lightweight deployment, and robustness across diverse site conditions are critical.

By contrast, Mask R-CNN was selectively used in tasks requiring precise instance segmentation. Visual Relationship Detection (VRD) and image captioning models were applied in tasks focused on contextual or relational interpretation, typically using VTransE. In several studies, outputs from vision models were further combined with knowledge graph embeddings or Graph Neural Networks (GNNs), not as standalone classifiers but as components of ontology-driven reasoning pipelines.

YOLO models were favored for their speed, efficiency, and versatility, and were applied in tasks such as regulatory violation detection, risk level classification, and process monitoring. Studies by Fang et al. [26], Pfitzner et al. [22], Wu et al. [18], and Zhong et al. [8] used YOLO for rapid object detection, mapping recognition outputs to ontology concepts for real-time reasoning. Mask R-CNN, with its capacity for object segmentation and spatial delineation, was used in fine-grained analyses such as structural damage detection [27] and object state monitoring [20]. VRD (VTransE) was used in Li et al. [19] to extract inter-object relationships and align them with ontology structures, enabling context-aware interpretation and compliance assessment. Other approaches included BLE-based semantic trajectory analysis (Arslan et al. [28]) for location-based situation awareness. Two studies (Zeng et al. [29], Pedro et al. [30]) did not explicitly specify the vision models used but incorporated image data for semantic enrichment or training dataset development.

In summary, YOLO and related real-time object detection models were most frequently applied to responsiveness-driven tasks. In contrast, Mask R-CNN, VRD, and image captioning models were selectively used for fine-grained analysis or contextual reasoning. Recent trends indicate a gradual shift from single-frame detection toward context-aware reasoning, with hybrid pipelines combining vision recognition outputs with ontology-based embeddings or GNN-based inference. These trends suggest an ongoing evolution from object-level detection toward semantic scene understanding in ontology–AI integration for construction.

Table 7 summarizes the vision model families identified in the reviewed studies, along with their key characteristics.

3.4.4. Ontology Implementation Strategies in AI Pipelines

Ontology integration approaches in image recognition–based AI pipelines can be broadly categorized into two groups based on the stage and manner of integration: training-stage integration and output-stage enhancement. The former introduces ontological knowledge during model training to improve feature representation and learning structure, while the latter applies ontology-based interpretation and reasoning as a post-processing layer to refine model outputs.

For example, training-stage integration embeds semantic knowledge into the learning process itself, such as applying ontology-defined labels or embeddings during preprocessing or model training (e.g., ontology-guided labeling or knowledge graph embeddings in YOLO). In contrast, output-stage enhancement applies ontology-based reasoning after inference, such as mapping detected objects to ontology concepts and applying SWRL- or SPARQL-based rules to infer hazards or validate compliance.

Figure 5 illustrates this framework, showing not only the distinction between the two approaches but also their positions within the image recognition pipeline: training-stage integration informs feature extraction and model learning (left), whereas output-stage enhancement functions as a reasoning layer applied after object detection (right).

Training-stage integration embeds ontological information directly into input processing and learning stages, enabling semantic knowledge to be internalized within model representations. This category can be further divided into three subtypes: preprocessing-based integration, transfer and embedding-based integration, and co-learning integration.

•

Preprocessing-based integration uses ontology concepts and relationships for data preprocessing, candidate region filtering, or label refinement. For example, Arslan et al. [28] mapped BLE-based semantic trajectories to ontology concepts, enriching training data with semantic labels and features.

•

Transfer and embedding-based integration incorporates ontological hierarchies or embeddings into model initialization or transfer learning stages. Pfitzner et al. [22] combined knowledge graph embeddings with a YOLO detector for semantically informed feature learning, while Pan et al. [31] applied zero-shot label embeddings to reduce the gap between label definitions and recognition outcomes.

•

Co-learning integration jointly trains deep learning models and ontology-based knowledge graph models, enabling parallel learning of data-driven features and semantic relationships. Examples include joint training of CNN and GCN architectures or using ontology reasoning outputs as feedback within the training loop. However, no such cases were identified in the construction domain, suggesting that co-learning strategies remain largely experimental due to the high demands of pipeline design and data preparation.

Output-stage enhancement, by contrast, applies ontology-based interpretation to model outputs without modifying the architecture of existing detection or classification models. This approach can be divided into class mapping and rule-based reasoning.

•

Class mapping integration connects detected object labels with ontology concepts, enriching them with attributes and relationships for subsequent reasoning or knowledge accumulation. Zheng et al. [24] mapped Mask R-CNN results to a process ontology to classify work status, while Zeng et al. [29] and Pedro et al. [30] linked image data to ontologies to improve training scenarios and visualization quality.

•

Rule-based reasoning integration applies ontology-defined rules (e.g., SWRL, SPARQL) to recognition outputs for contextual assessment. Lee and Yu [21] combined YOLO outputs with a safety ontology to infer hazardous conditions in real time, while Fang et al. [26] employed rule-based reasoning with a hazard ontology to automatically detect regulatory violations. Although straightforward to implement and highly compatible with existing models, this approach primarily functions as an interpretive layer rather than directly improving detection performance.

Analysis of the 16 studies revealed that output-stage enhancement was substantially more common, with rule-based reasoning being the most frequently adopted strategy. This preference reflects its low implementation burden and high compatibility with widely used models such as YOLO and Mask R-CNN, enabling relatively rapid deployment. In contrast, training-stage integration was less frequently observed due to greater design complexity and data preparation requirements. Nevertheless, it offers the potential advantage of embedding semantic knowledge directly into model representations, which may be particularly valuable for tasks involving complex relational reasoning or contextual understanding.

Recent developments also indicate a growing interest in hybrid strategies that combine training-stage and output-stage approaches. Methods that integrate ontological embeddings into deep feature learning or align pretrained and transfer learning models with ontology structures may enhance semantic interpretability across the pipeline. Such approaches have the potential to improve not only system performance but also explainability and reproducibility in ontology-enhanced AI systems.

3.4.5. Performance Outcomes

In the 16 studies reviewed, the outcomes of ontology–AI integration in image recognition–based systems varied depending on the research purpose, the models employed, and the integration approaches adopted. Overall, performance improvements were not consistently reflected in large numerical gains; rather, the contributions of ontologies were more evident in terms of interpretability, reliability, and practical applicability.

For instance, in studies focused on hazard identification and situation awareness, models such as YOLO or Mask R-CNN were typically combined with rule-based reasoning integration. These approaches consistently reported reductions in false positives, improved real-time hazard inference, and stronger contextual awareness, although absolute improvements in accuracy were often limited. In contrast, studies on process tracking and anomaly detection frequently employed YOLO outputs mapped to ontologies or embedding-based transfer learning. Reported benefits included workflow monitoring, improved classification consistency, and mitigation of misclassifications; however, many of these works remained at the proof-of-concept stage and were highly dependent on ontology quality. In the domain of regulation and standards compliance, several studies combined detection results with regulatory ontologies to automatically assess conformity. These demonstrated practical value for safety management, yet their outcomes were strongly influenced by dataset variability and ontology design choices.

Some studies reported only marginal numerical improvements, such as 2–3 percentage points over baseline models, or results that were statistically insignificant. More importantly, heterogeneity in datasets, evaluation metrics, and experimental setups limited comparability and generalization across studies. Furthermore, ontology development—including schema design, labeling, and structural definition—was frequently identified as a resource-intensive process, restricting scalability and transferability.

A detailed breakdown of all 16 image recognition studies—including their research purposes, employed models, integration approaches, reported effects, and limitations—is presented in

Table 8. Summary of selected 16 studies on ontology–AI integration in image recognition–based construction applications.

Author(s)	Functional Role	Ontology Construction	Image Recognition Models	Integration Approach	Description	Observed Outcomes	Limitations
Lee & Yu [21]	Regulation and Standards Compliance; Hazard Analysis	Manual	YOLO	Rule-Based Reasoning Integration	Recognition outputs mapped to ontology rules for hazard reasoning	Reduced false positives; support for real-time hazard inference; enhanced situation awareness	Limited generalizability
Zheng et al. [24]	Process Monitoring	Manual	Not specified	Class Mapping Integration	Detection outputs mapped to a process ontology for workflow monitoring	Improved data interoperability; semantic reasoning for situation awareness	No direct accuracy gain; dependent on ontology quality
Fang et al. [26]	Regulation and Standards Compliance; Hazard Analysis	Hybrid	Mask R-CNN	Rule-Based Reasoning Integration	Ontology-based reasoning applied to validate compliance with predefined safety rules	Reduced false positives in hazard detection; strengthened semantic reasoning and situation awareness mIoU: 41.22%; Accuracy: 79.98%	Limited generalizability
Pfitzner et al. [22]	Process Monitoring	Automated	YOLO + GNN (post-reasoning)	Transfer & Embedding-Based Integration	Knowledge graph embeddings used for semantically informed feature learning	Enabled process monitoring; improved state classification accuracy Accuracy: +8–9%	Limited generalizability
Zhang et al. [25]	Hazard Analysis	Automated	YOLO	Rule-Based Reasoning Integration	Ontology-based evaluation of compliance and hazard classification	Reduced incorrect hazard associations; improved real-time relevance; enabled risk evolution tracking	No direct accuracy gain; dependent on ontology quality
Pfitzner et al. [46]	Process Monitoring	Automated	YOLO	Class Mapping Integration	Detection streams linked to ontology/knowledge graph for continuous semantic updates	Supported workflow monitoring; enriched semantic data	Limited generalizability
Pan et al. [31]	Process Monitoring	Automated	YOLO	Transfer & Embedding-Based Integration	Zero-shot label embeddings used to extend recognition beyond training classes	Supported workflow monitoring; improved recognition of construction entities Top-1 Accuracy: +30%; mAP: +19%	Proof-of-concept stage
Pfitzner et al. [76]	Process Monitoring	Automated	YOLO	Class Mapping Integration	Semantic enrichment applied to detection results	Supported workflow monitoring; mitigated misclassifications	No direct accuracy gain; dependent on ontology quality
Li et al. [19]	Regulation and Standards Compliance	Manual	VRD (VTransE)	Rule-Based Reasoning Integration	Relational recognition outputs mapped to ontology rules	Reduced false positives; improved hazard identification; enhanced semantic consistency Top-1 Accuracy: 49.91%; Recall@100: 49.13%	Proof-of-concept stage
Wu et al. [18]	Regulation and Standards Compliance; Hazard Analysis	Hybrid	YOLO	Rule-Based Reasoning Integration	Ontology reasoning applied to construction safety violation detection	Improved classification consistency; complemented semantic gaps	Proof-of-concept stage
Hamdan et al. [27]	Work and Situation Awareness	Hybrid	Mask R-CNN	Rule-Based Reasoning Integration	Ontology-based classification of hazard levels in real-time monitoring	Reduced manual effort; improved expert consistency; enabled damage progression tracking	Limited generalizability
Arslan et al. [28]	Work and Situation Awareness	Hybrid	BLE-based semantic trajectory	Preprocessing-Based Integration	Ontology-based semantic labels applied during preprocessing	Enabled unsafe movement detection; enriched trajectory semantics	Limited generalizability
Zeng et al. [29]	Work and Situation Awareness	Manual	Not specified	Class Mapping Integration	Ontology mapping used for semantic augmentation	Decision support for algorithm selection; improved hazard image retrieval precision	Limited generalizability
Pedro et al. [30]	Regulation and Standards Compliance	Hybrid	Not specified	Class Mapping Integration	Ontology applied for training and scenario evaluation	Automated data integration; improved accessibility of accident cases Post-test performance: +3–13%	No direct accuracy gain; dependent on ontology quality
Zhong et al. (2020) [8]	Hazard Analysis	Hybrid	YOLO	Rule-Based Reasoning Integration	Captioning and detection outputs mapped to ontology rules	Reduced false positives; real-time hazard inference; consistent semantic annotation	No direct accuracy gain; dependent on ontology quality
Pan et al. (2024) [32]	Work and Situation Awareness	Hybrid	Not specified	Rule-Based Reasoning Integration	Ontology reasoning layer supports site safety decision-making	Improved situational interpretation accuracy	Limited generalizability

Note. Some studies did not report quantitative performance indicators (e.g., accuracy, recall, mAP) because their primary objective was to demonstrate ontology-based semantic reasoning, compliance checking, or situation awareness rather than numerical benchmarking. In such cases, performance was evaluated through qualitative analysis, expert validation, or scenario-based assessment.

. This supplementary table highlights the heterogeneity of evaluation setups and clarifies that reported outcomes are highly context dependent.

Taken together, these findings indicate that ontology integration does not guarantee consistent numerical performance gains. Instead, its main contributions lie in enhancing explainability, contextual reasoning, workflow monitoring, and regulatory transparency, thereby increasing trust and applicability in construction safety and quality management. These observations underscore the importance of developing standardized evaluation protocols, conducting repeated validations across domains, and implementing comparative experiments in real-world environments to strengthen both reproducibility and practical relevance.

4. Challenges and Future Research on Ontology-AI Integration

4.1. Research Gaps and Limitations

Among the 152 eligible studies, 16 addressed image-based applications. These works consistently reported qualitative gains—such as reduced false positives, improved situational awareness, and enhanced interpretability—but they also revealed deeper structural and methodological limitations that restrict the advancement of ontology–AI integration in construction.

A first limitation is the field’s strong reliance on output-stage enhancement. Approximately 75% of image-based studies applied ontologies only after object detection, primarily for class mapping or rule-based hazard interpretation [18,21,26]. Although this post hoc integration improves semantic clarity and compliance assessment, it does not influence feature representation or model learning. Training-stage integration was rarely attempted; only a small number of studies explored semantically enriched preprocessing [28] or knowledge and zero-shot embeddings [22,31]. No examples of joint learning frameworks—where neural networks and ontology-based reasoning models are trained in tandem—were identified in the construction domain. This imbalance reflects a structural constraint: construction image datasets lack stable object categories, standardized schemas, and consistent annotations, making ontology-informed training costly and technically challenging. As a result, perception (image recognition) and reasoning processes remain weakly coupled, limiting the potential of ontologies to improve generalization, robustness, or domain-shift performance.

A second limitation concerns the restricted scope of semantic reasoning, which remains largely static and frame-based. Most systems focused on object-level image recognition or simple rule-based inference without modeling temporal progression, multi-object interactions, or causal hazard evolution. For example, Li et al. [19] used Visual Relationship Detection (VTransE) to extract pairwise object relationships, and Zhang et al. [25] and Wu et al. [18] combined YOLO detections with ontology-driven classification. However, these approaches remained confined to single-frame reasoning, lacking mechanisms for identifying unsafe sequences, cumulative risk buildup, or evolving situational patterns. This limitation arises from the absence of temporal constructs in existing construction ontologies and the lack of reasoning modules capable of representing procedural dependencies or scenario-level dynamics.

A third limitation lies in the persistent reliance on manual and static ontology construction. More than 80% of the reviewed studies depended on expert-driven schema design and manual instance creation, as seen in Zheng et al. [24], Zhang et al. [25], and Li et al. [19]. Such methods ensure precision at the initial stage but lack scalability and adaptability as site conditions and hazard categories evolve. Attempts to automate ontology creation through NLP or LLMs were limited in scope [29] and did not address the full ontology lifecycle, leading to schema brittleness, high maintenance costs, and divergence between ontology structures and real-world site dynamics.

A fourth limitation relates to the heterogeneity of evaluation methods, which impedes cross-study comparison and generalization. Studies reported inconsistent metrics—accuracy, recall, precision, IoU, F1-score, and mAP—or relied solely on qualitative assessments [22,28,30]. Explainability benefits were frequently mentioned but rarely operationalized into quantitative or reproducible measures. This variation reflects a deeper gap: construction lacks shared taxonomies, benchmark datasets, and standardized annotation policies. As a result, performance outcomes remain highly context-dependent, and evidence for ontology-integrated models is fragmented.

Finally, ontology–AI integration rarely extended across the full AI lifecycle, revealing a gap in end-to-end architectural integration. Most systems incorporated ontologies only at isolated stages, typically post-detection mapping [24,30]. Even more advanced pipelines, such as Pan et al.’s Video2Entities framework [31], lacked feedback loops connecting semantic reasoning to model training or data curation. Without unified schemas, co-learning mechanisms, or iterative update processes, ontologies function only as supplementary interpretive layers rather than central components capable of guiding perception, reasoning, and continuous adaptation. This architectural fragmentation is further exacerbated in BIM-centered workflows by the absence of standardized AI–BIM data-exchange interfaces, which prevents semantic outputs and reasoning results from being reintegrated into upstream data models or learning processes. As a result, ontology-enabled AI systems remain confined to localized enhancements rather than operating as continuously adaptive, closed-loop architectures. This persistent disconnection between data modeling, perception, and reasoning constitutes a major obstacle to scaling ontology-integrated AI systems for real-world construction applications.

4.2. Future Research Directions

The directions in this section are not speculative wish lists but are directly derived from the recurring limitations identified in Section 4.1. By explicitly linking each proposal to documented gaps, these recommendations highlight evidence-based priorities for advancing ontology–AI integration in construction image recognition (see Figure 6, right panel).

The first direction is to embed ontologies into training pipelines. While most existing studies incorporated ontologies only after detection, a few demonstrated the potential of deeper integration through semantically enriched preprocessing [28], knowledge graph embeddings [22], or zero-shot label expansion [31]. Building on these findings, future research should explore co-learning frameworks that jointly train CNNs and GNNs or introduce feedback mechanisms in which ontology-based reasoning influences iterative model updates. Such approaches can strengthen contextual representation, improve generalization under domain shift, and enhance adaptability in data-scarce environments.

A second direction is to advance temporal and scenario-based semantic reasoning. Current systems largely interpret scenes frame by frame, making it difficult to detect evolving hazards or unsafe action sequences. Building on early efforts in relational modeling [19,25], future research should integrate multimodal data—such as images, text, and sensor streams—with ontology-driven reasoning to support scenario-based hazard prediction and behavioral analysis. Temporal scene graphs, ontology-based activity models, and predictive reasoning frameworks could enable proactive identification of hazard escalation in dynamic environments. Recent advances in vision–language models (VLMs) also offer promising opportunities to enhance ontology-based reasoning by jointly interpreting visual content and linguistic descriptions. When combined with ontological constraints, VLMs may support more flexible semantic grounding, cross-modal consistency, and natural language–based explanation of recognition and reasoning outcomes.

A third direction involves automating ontology construction. Section 4.1 showed that manual ontology development dominates current practice, leading to high labor costs and slow update cycles [24,25,29]. Advances in LLMs and NLP offer opportunities to extract concepts from unstructured text, align them with existing schemas, and generate instances automatically. Hybrid pipelines combining expert oversight with automated instance generation and schema evolution may improve scalability, responsiveness, and alignment with real-world site conditions.

A fourth direction is to establish standardized quantitative evaluation protocols. Given the heterogeneity of metrics observed across studies [22,28,30], future work should develop benchmark datasets with unified taxonomies and define evaluation frameworks that incorporate both accuracy-based and explainability-oriented metrics. In addition to conventional detection metrics (e.g., accuracy, precision, recall, mAP), such frameworks should include ontology-specific indicators, such as rule-consistency rates, semantic coverage, reasoning-path completeness, and false-positive reduction ratios attributable to ontology reasoning. Statistical validation—such as confidence intervals and effect sizes—should accompany performance reporting, and domain-shift tests should be adopted to assess robustness. These measures would enhance reproducibility and enable more consistent comparison of ontology-integrated systems.

Finally, future research should focus on developing end-to-end ontology–AI architectures that span data collection, training, inference, reasoning, and iterative feedback. Ontology-guided sampling and labeling, ontology-informed embeddings, rule-triggered relabeling, and MLOps-based model monitoring can support continuous improvement. Integrating such architectures with digital twins, real-time monitoring systems, and multi-agent platforms may enable adaptive and semantically consistent AI systems capable of operating effectively in complex construction environments.

5. Conclusions

This study conducted a systematic literature review of 152 ontology–AI integration studies, including an in-depth examination of 16 works on image recognition–based construction analysis. By adopting a multi-dimensional analytical perspective—covering application objectives, ontology construction methods, integration strategies, and evaluation practices—the review provides a more holistic understanding of how ontologies function as semantic infrastructures within AI-driven construction systems. Rather than treating numerical performance gains as the primary indicator of value, this study emphasizes ontology’s broader role in providing semantic consistency, contextual reasoning, and knowledge integration across heterogeneous data sources.

The findings reveal that ontology integration consistently yielded qualitative benefits such as reduced false positives in hazard detection, improved interpretability, and enhanced situational understanding. However, quantitative improvements in model accuracy or mAP were typically modest, highly context-dependent, and inconsistent across studies. Many ontology-enabled systems remain at the proof-of-concept stage and have not yet undergone rigorous evaluation in large-scale or real-world construction settings. These results highlight both the promise and immaturity of ontology–AI integration: ontologies enrich semantic reasoning but introduce design, integration, and maintenance complexity that must be justified through feasible and scalable deployment pathways.

This study contributes to the field in three important ways. First, by combining PRISMA methodology with topic modeling, it offers an objective map of 152 studies and reveals cross-domain clusters that cannot be captured by narrative synthesis alone. Second, it provides the first structured and detailed review of ontology-integrated image recognition in construction, synthesizing functional roles of ontologies, construction approaches, vision models, integration strategies, and performance outcomes. Third, it develops an evidence-based research agenda that directly links recurring structural limitations—identified through systematic analysis in Section 4.1—to targeted future directions in Section 4.2. This ensures that recommendations are grounded in observed empirical gaps rather than speculative assumptions. Although a formal quantitative meta-analysis was not conducted due to substantial heterogeneity in datasets, tasks, and evaluation metrics, this limitation reflects the current immaturity of standardized evaluation practices in ontology–AI integration research. As benchmark datasets and harmonized evaluation protocols emerge, future studies will be better positioned to conduct systematic quantitative comparisons and establish effectiveness rankings across different integration approaches.

Building on these observations, future research should therefore prioritize deeper integration of ontological knowledge into model training and representation learning; the development of standardized, quantitative frameworks to evaluate explainability and reasoning clarity; the establishment of benchmark datasets and harmonized evaluation protocols; the advancement of semi-automated ontology construction methods using LLM- and NLP-based techniques; and the design of end-to-end, multi-stage architectures that unify data collection, training, inference, and feedback within a semantically consistent pipeline.

Despite the methodological rigor ensured by adherence to PRISMA 2020 guidelines, several limitations must be acknowledged. First, the analysis relied solely on the Web of Science Core Collection, potentially omitting relevant studies indexed in other databases such as Scopus or Google Scholar. Second, the inclusion criteria focused exclusively on English-language, peer-reviewed publications, excluding grey literature, theses, and technical reports that may contain additional insights. Third, the review period (2014–2025) captures recent developments but may overlook earlier foundational research in ontology engineering or AI integration. Finally, because this study primarily investigates ontology-based image recognition in construction, its findings may not generalize to domains where ontologies function differently or where data structures and workflows diverge from those of construction. These limitations should be addressed in future research by expanding database coverage, incorporating grey literature where appropriate, and exploring cross-domain comparative analyses of ontology–AI integration patterns.

Ontology–AI integration in construction holds substantial potential to enhance transparency, contextual reasoning, and adaptive decision-making. Yet its current state is characterized by methodological fragmentation, heterogeneous evaluation practices, and limited real-world validation. The value of ontologies should therefore not be framed as a universal performance-enhancing mechanism, but as a pragmatic semantic framework that supports the development of explainable and context-aware AI systems. By grounding the synthesis and future directions in systematic evidence, this review underscores both the academic significance and practical relevance of ontology-driven approaches, and highlights pathways toward scalable and semantically robust AI applications in construction environments.