Reframing architectural classification as representation: Perceptual, structural, and hybrid encodings under small-data conditions

This study investigates how architectural identity is encoded in computational systems by reframing architectural classification as a comparison of representational encodings. Here, “architectural identity” refers narrowly to measurable authorship-related traces in elevation drawings rather than to the broader disciplinary concept. A dataset of 30 Thai architectural elevations attributed to two masters is analyzed under small-data conditions. Three encoding strategies are examined: perceptual representations derived from a pre-trained convolutional neural network, structural representations based on geometric descriptors obtained through roof–body–base decomposition, and a hybrid representation combining both. Perceptual features are projected into low-dimensional latent spaces using principal component analysis, and all representations are evaluated using a linear classifier. SHAP and feature-group ablation are used to examine feature contributions. The results indicate that perceptual representations achieve near-perfect classification within the dataset in low-dimensional latent spaces, suggesting that class-discriminative visual information is concentrated within a small number of components, although these cannot be assumed to be architecturally meaningful. In contrast, structural representations distribute information across interpretable features, with base-related and spatial descriptors among the strongest contributors. These findings highlight the influence of representational encoding on classification under small-data conditions.