The preservation of large industrial heritage structures requires efficient and accessible monitoring techniques to detect structural deterioration and surface changes over time. Traditional monitoring approaches rely on manual observation of objects, which is highly specialized and time-consuming. This study presents a low-cost and easily transferable framework for multi-temporal 3D change detection using unmanned aerial vehicles (UAVs) and open-source photogrammetry software, with a focus on machine learning-based difference detection.

The framework utilizes structure-from-motion (SfM) and multi-view stereo (MVS) techniques to generate and precisely co-register temporally separated 3D models of technical-historical structures. In contrast to conventional approaches, the proposed method is entirely self-referential, relying exclusively on optical and geometrical features for alignment, thereby eliminating the necessity for specialized surveying equipment. The primary innovation resides in the integration of machine learning techniques for automatic change detection. The integration of machine learning techniques for automatic change detection represents a key innovation. The proposed approach involves the fusion of both temporal model versions into a composite representation, and a trained model is then employed to identify and segment altered regions, distinguishing structural changes such as corrosion from background noise and minor inconsistencies.

This approach enables a highly automated and scalable monitoring, thereby reducing manual inspection efforts while improving detection accuracy. Moreover, the integration of texture-based comparisons enhances the precision of change localization. The framework is entirely built on open-source software and lightweight UAVs, thereby significantly reducing costs while ensuring regulatory flexibility. Due to its ease of implementation, affordability, and adaptability, this method is particularly well-suited for cultural heritage preservation, where financial and logistical constraints often hinder the adoption of advanced monitoring solutions. The incorporation of machine learning into the analysis enhances the accessibility and efficiency of structural monitoring workflows, thereby fostering broader adoption in heritage conservation.