Publicação
Knowledge-based machine learning approach to indirect prospecting methodologies for monument identification
| Resumo: | Experts use domain knowledge and experience to identify and analyze archaeological monuments from satellite images. However, traditional object detection methods often rely solely on image data and operate as “black boxes,” which frequently result in false positives, especially when detecting small archaeological sites. For machines to effectively leverage domain knowledge, it must be organized in an interoperable format, addressing the challenge posed by scattered and fragmented data, particularly across multiple disciplines. This study tackles this issue by converting domain knowledge from diverse and multidisciplinary sources into a machine-readable format to reduce false positives in automatic object detection. The study links information about archaeological sites and their landscapes by implementing a Knowledge Graph (KG) based on CIDOC-CRM, its CRMgeo extension, and GeoSPARQL ontologies. This KG integrates textual data from semantic records with spatial data from vector topographic maps, encompassing (i) metadata definitions, (ii) general and specific concepts, and (iii) the geometry of each represented entity. This representation can provide insights into elements within a scene that may not be visible in images. Subsequently, the output from an object detection approach was integrated with the KG to train a Knowledge Graph–Machine Learning (KG-ML) model. This model identifies areas of interest (AOIs) where dolmens in Pavia and Mora (Portugal) are likely to be found, using contextual knowledge to exclude images with a low probability of accurate detections. The KG-ML approach effectively reduced false positives, provided contextual information that clarified recognition decisions, and enhanced the understanding of detected sites. |
|---|---|
| Autores principais: | Câmara, Ariele |
| Assunto: | Monumento arqueológico Machine learning -- Machine learning Imagem de satélite Knowledge Graph (KG) Património arqueológico -- Archaeological heritage |
| Ano: | 2025 |
| País: | Portugal |
| Tipo de documento: | tese de doutoramento |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | ISCTE |
| Idioma: | inglês |
| Origem: | Repositório ISCTE |
| Resumo: | Experts use domain knowledge and experience to identify and analyze archaeological monuments from satellite images. However, traditional object detection methods often rely solely on image data and operate as “black boxes,” which frequently result in false positives, especially when detecting small archaeological sites. For machines to effectively leverage domain knowledge, it must be organized in an interoperable format, addressing the challenge posed by scattered and fragmented data, particularly across multiple disciplines. This study tackles this issue by converting domain knowledge from diverse and multidisciplinary sources into a machine-readable format to reduce false positives in automatic object detection. The study links information about archaeological sites and their landscapes by implementing a Knowledge Graph (KG) based on CIDOC-CRM, its CRMgeo extension, and GeoSPARQL ontologies. This KG integrates textual data from semantic records with spatial data from vector topographic maps, encompassing (i) metadata definitions, (ii) general and specific concepts, and (iii) the geometry of each represented entity. This representation can provide insights into elements within a scene that may not be visible in images. Subsequently, the output from an object detection approach was integrated with the KG to train a Knowledge Graph–Machine Learning (KG-ML) model. This model identifies areas of interest (AOIs) where dolmens in Pavia and Mora (Portugal) are likely to be found, using contextual knowledge to exclude images with a low probability of accurate detections. The KG-ML approach effectively reduced false positives, provided contextual information that clarified recognition decisions, and enhanced the understanding of detected sites. |
|---|