Normalized RGB imaging and transfer learning predict surface moisture at exposed earthen sites
Abstract Repeated assessment of moisture content is essential for the preventive conservation of open-air earthen sites, yet conventional contact-based methods cannot readily provide dense spatial coverage. This study proposes a non-destructive method that predicts surface moisture content from chromatic features. A laboratory reference model was first developed using normalized RGB features (R_norm, G_norm, B_norm) that are insensitive to illuminance variation, and the model was then adapted to complex field lighting conditions through transfer learning. The transferred model yielded high prediction accuracy for the silty clay and silt tested in this study under conditions of low surface roughness (PEDI ≤0.3). Within the tested conditions, this workflow enables low-cost, non-contact assessment of surface moisture content and offers a practical tool for the field investigation and preventive conservation of open-air earthen sites.
Authors
- Weijia Mao (ORCID: https://orcid.org/0009-0009-6114-2911)
- Chuxin Liang
- Kai Deng
- Manli Sun
Institutions
- Northwest University (CN)
Publication Details
- Journal
- npj Heritage Science
- Published
- 2026-11-13
- DOI
- https://doi.org/10.1038/s40494-026-03110-5
- Primary Topic
- Cultural Heritage Materials Analysis
- Type
- article
- Field-Weighted Citation Impact
- 0.00