The Influence of Data Quality in Supervised ML-AI Classification Approaches for Historical Heritage

In recent years, the automatic segmentation and classification of digital survey data has been experimented with in built heritage studies. Despite the encouraging progress in the use of Machine and Deep Learning techniques, the semantic segmentation of point clouds is more complex, especially for the historic environment for which, due to the heterogeneity of shapes, it is more difficult to recognize homogeneous regions with similar properties. Given the need to process a large volume of already annotated data for the training and recognition of new scenes, the type and quality of the initial data play a fundamental role in the classification process, as they influence the subdivision into predefined categories that are not always consistent with a decomposition into architectural elements and sub-elements shared by the scientific community. This is an interpretative problem that already emerges from the traditional manual labelling, which, being highly subjective, reduces the reproducibility of the results. The paper focuses on understanding to what extent the recognition of homogeneous regions is influenced by factors such as:—manual labelling carried out by annotators with different specializations;—density value of the point clouds;—type of data acquired depending on the acquisition sensor. These evaluations were conducted by employing the Random Forest algorithm on specific pre-processed point cloud datasets, with reference to the typology of the Franciscan cloister, in order to make the recognition flows more controlled and less ambiguous, aiming at an advancement towards the more efficient modelling and management of the existing architectural heritage.