Frost action classification

Automated classification of frost microstructures in archaeological thin sections using transfer learning with CNN architectures (MobileNetV2, EfficientNetV2B0, ResNet50) and Bayesian neural networks. Includes Score-CAM interpretability analysis and expert survey comparison to evaluate human-AI alignment in micromorphological interpretation. This repository contains the dataset, models' .h5 files and the classification, survey, and expert-models comparison. The interactive classification tool is avalabl at: https://huggingface.co/spaces/skkou/frost_classifier. This work is part of the ERC Starting Grant MATRIX project (nº101041245 PI: V. Aldeias). Journal article available at: https://doi.org/10.12688/openreseurope.23697.1