Unveiling the environmental accommodation with cultural identity in vernacular houses through knowledge-fused explainable AI

Abstract

Vernacular houses are a physical representation of humans adapting their dwellings to new natural and cultural environments using traditional architectural techniques. The variation among vernacular houses involves long-term human endeavors to balance environmental accommodation and cultural identity. Due to the lack of large-scale data and quantitative measuring techniques for housing features, evaluating holistic and fine-grained similarity among vernacular houses has remained greatly challenging. Deep learning and satellite images recently enabled the intelligent evaluation of vernacular houses. However, the latent features extracted by deep learning are elusive to humans, hampering their application in vernacular housing fields. To address this issue, this study proposed an innovative framework for quantitatively assessing the feature variations among vernacular houses. We collected a large number of satellite images across China and extracted conceptual features using explainable AI from these images. More importantly, by integrating vernacular housing knowledge into image analysis, conceptual features were successfully interpreted into understandable architectural semantics. Furthermore, through reconstructing the historical migration routes in South China, we discover that Hakka migrants adapted to new places by adjusting building density and housing size while maintaining their ethnic identity and architectural culture via persisting sky-well order and enclosed structure. This study provides a new quantitative method for measuring variations among vernacular houses, thereby enhancing our understanding of the architectural strategies employed to balance environmental accommodation with cultural identity.