From recognition to reconstruction: AI reshapes cultural heritage conservation

Digital transformation is reaching museums, archaeological landscapes, and historic city centers. According to a new study, artificial intelligence-driven methods are expanding the scale and precision of heritage analysis but also introducing new methodological and governance challenges.

In their review, Machine Learning and Deep Learning for Cultural Heritage Conservation: A Bibliometric and Task-Oriented Review, published in Remote Sensing, the study maps the evolution of AI applications in cultural heritage between 2011 and 2025 to identify dominant trends, technological patterns, geographic concentrations, and structural gaps shaping the future of AI-driven conservation.

Recognition tasks lead AI adoption in cultural heritage

The first major domain identified in the review is recognition. This category encompasses tasks such as object detection, classification, segmentation, damage identification, material recognition, and pattern analysis. Recognition tasks account for a significant share of AI applications in cultural heritage, reflecting both technological maturity and practical need.

Deep learning models, particularly convolutional neural networks, have been widely adopted for visual classification tasks involving paintings, frescoes, architectural elements, and archaeological artifacts. These systems can identify stylistic features, detect cracks or deterioration patterns, and classify materials based on image inputs. In archaeological contexts, machine learning models assist in automated site detection using aerial and satellite imagery.

Segmentation models have become especially important in analyzing fine-grained damage patterns. By isolating areas of discoloration, surface erosion, biological growth, or structural fracture, these systems support conservators in prioritizing intervention strategies. Automated recognition reduces manual workload and increases consistency across large datasets.

The authors note that recognition applications are often the entry point for AI integration due to the relative availability of labeled image datasets. However, even in this mature domain, challenges persist. Cultural heritage objects are often unique, and data scarcity can limit model generalizability. Unlike industrial datasets with thousands of standardized samples, heritage datasets may contain only a small number of examples from a specific monument or artifact.

Transfer learning techniques and data augmentation have partially mitigated this limitation, but the review highlights the need for standardized benchmark datasets tailored to conservation contexts. Without shared evaluation standards, performance comparisons across studies remain difficult.

Reconstruction and virtual restoration expand digital preservation

The second major domain identified in the review centers on reconstruction and virtual restoration. This area includes 3D reconstruction, texture completion, structural restoration modeling, and digital reproduction of damaged artifacts or architectural features.

Deep learning methods are increasingly used to reconstruct missing fragments of sculptures, frescoes, and historical structures. Generative models and advanced image inpainting techniques allow researchers to simulate how damaged works may have originally appeared. In combination with photogrammetry and laser scanning, AI-driven reconstruction enhances digital twin models of heritage sites.

Three-dimensional modeling plays a central role in preserving monuments vulnerable to environmental degradation or conflict. Machine learning algorithms can process large point-cloud datasets generated by LiDAR scans, improving segmentation and classification of architectural components. These digital replicas serve as both documentation archives and analytical tools for structural assessment.

The review emphasizes that reconstruction tasks often require multimodal data integration. Visual imagery, spectral data, thermal imaging, and geometric measurements must be combined to achieve accurate representations. Selecting appropriate model architectures depends on the type and structure of input data, underscoring the importance of the authors' data-driven framework.

Despite technological progress, the study cautions that virtual restoration raises interpretative challenges. AI-generated reconstructions must be validated by conservation experts to avoid introducing historically inaccurate elements. The authors advocate for collaborative workflows in which domain experts and AI developers co-design and evaluate reconstruction systems.

Monitoring and prediction support long-term preservation

The third domain identified in the review involves monitoring and prediction. Structural health monitoring, environmental risk assessment, degradation forecasting, and disaster prediction are increasingly supported by machine learning models trained on sensor data and remote sensing inputs.

Predictive analytics can detect subtle patterns in vibration data, temperature fluctuations, humidity variations, and pollutant exposure. These signals may indicate early-stage structural instability or material degradation. By identifying anomalies before visible damage occurs, AI-driven monitoring enables preventative conservation.

Remote sensing technologies have also become central to large-scale heritage monitoring. Satellite imagery and drone-based surveys allow conservation authorities to track land-use changes, erosion patterns, and urban encroachment around heritage sites. Machine learning algorithms classify these environmental signals to assess risk exposure.

The review highlights that predictive modeling supports climate resilience planning. As extreme weather events increase, heritage sites face heightened vulnerability to flooding, wildfires, and heat stress. AI systems can simulate potential damage scenarios and inform mitigation strategies.

However, long-term deployment remains limited. Many monitoring studies are conducted under controlled experimental conditions without sustained operational validation. The authors stress the need for extended field testing to ensure reliability under changing environmental and usage conditions.

Structural gaps and future directions

The majority of AI-based heritage studies originate from Europe and East Asia, reflecting concentrated research investment in these regions. Other regions remain underrepresented, limiting the global generalizability of findings.

The study also identifies a gap between research innovation and practical implementation. Many contributions remain at the proof-of-concept stage, focusing on algorithmic performance rather than deployment scalability. Standardization across datasets, evaluation metrics, and documentation practices remains inconsistent.

Explainable artificial intelligence (AI) emerges as a critical priority. Conservation decisions often carry irreversible consequences, and black-box models may undermine professional trust. The authors advocate for integrating explainable AI techniques that clarify how predictions are generated, enabling domain experts to verify and interpret outputs.

The review further calls for interdisciplinary collaboration. Cultural heritage conservation sits at the intersection of art history, archaeology, materials science, architecture, and computer science. Effective AI integration requires co-design approaches where conservators participate in dataset creation, validation protocols, and interpretation of results.

Data heterogeneity presents an ongoing challenge. Cultural heritage datasets vary widely in format, resolution, and completeness. Building interoperable platforms that integrate visual, geometric, and environmental data remains a complex task. The authors highlight the importance of establishing standardized repositories and benchmark frameworks to support reproducibility.

Ethical considerations also arise. Digitization and AI reconstruction must respect authenticity principles and avoid misrepresentation of cultural narratives. Transparent documentation of AI-assisted interventions is essential to preserve scholarly integrity.