Transfer Learning Strategies for the Enhancement of Point Cloud Semantic Segmentation Processes in Landscape Heritage Contexts

Abstract. In recent years, the increasing complexity of spatial data related to landscape heritage and urban legacies has led to a growing focus in research on enhancing the efficiency of management processes while simultaneously increasing their level of automation. In this context, a highly relevant solution is the use of so-called artificial intelligence, and more specifically, predictive models generated through Deep Learning techniques. However, a commonly observed critical issue concerns the limited generalisation capability of these models, which often fail to accurately recognise features in datasets that differ from those used during the training phase. In the present contribution, starting from a predictive model trained on airborne LiDAR point clouds belonging to a regional dataset (Sardinia, Italy), a transfer learning approach is proposed using new data (derived from the ISPRS benchmark dataset for semantic segmentation of Hessigheim – H3D) to improve the generalisation capabilities of the model. In order to assess the suitability of the proposed transfer learning strategy, a comparison between the classification performed by the original predictive model and the fine-tuned predictive model has been performed. Furthermore, the evaluation metrics have been calculated, evaluated, and discussed to quantitatively assess the improvement in terms of results, performance, and absolute gain between the different models tested. The proposed workflow supports scalable landscape and urban heritage monitoring by reducing human intervention in data management workflows while maintaining semantic consistency in available airborne laser scanner (ALS) data processing. While the contribution presents class limitations due to the constraints of the training data used for the first predictive model, the study demonstrates how transfer learning strategies can enhance the performance of semantic segmentation models when handling existing sparse data. This aligns well with scientific community efforts toward automated, efficient, and scalable heritage monitoring and documentation using remote sensing techniques.