Explore the vast range of titles available.

Large-scale classification of architectural styles in Chinese traditional settlements is important for heritage conservation and geospatial documentation, but scalable deployment remains constrained by the high cost of expert annotation because villages are widely distributed, the imagery is captured from heterogeneous viewpoints, and each architectural tradition exhibits substantial intra-class variation. To address this bottleneck, we propose CTSMatch, a label-efficient semi-supervised framework that combines an ImageNet-pretrained EfficientNetV2 backbone with SoftMatch-based adaptive pseudo-label weighting so that ambiguous but informative unlabeled samples can still contribute to training, thereby reducing reliance on costly expert annotation. We also construct SemiCTS, an extension of the original CTS dataset that adds 4360 unlabeled images. Using only 545 labeled samples, CTSMatch achieves 96.93% accuracy on SemiCTS, outperforming the strongest fully supervised baseline (Dense-TL-Aug) by 2.73 percentage points and two standard semi-supervised baselines (FixMatch and FreeMatch) by 3.06 percentage points. Beyond classification, we further analyze the feature space to examine stylistic lineage through intra-style heterogeneity, inter-style transitions, and outlier detection. The results reveal two broad regional groupings, a northern cluster (Jing, Jin, Su) and a southern cluster (Chuan, Min, Wan), connected by gradual transitions rather than rigid boundaries. Approximately 15% of the samples are identified as atypical cases, including 8.7% comprising regional variants and 6.3% comprising hybrid forms. These findings show that CTSMatch provides a practical label-efficient framework for architectural heritage classification while supporting the interpretable analysis of stylistic diversification and convergence in Chinese traditional settlements.