Abstract
<title>Abstract</title> <p>Tropical heritage façades in coastal urban environments face accelerated chromatic deterioration driven by persistent high humidity, temperature fluctuations, and salt-laden air. Current conservation practice remains predominantly reactive, while environmental monitoring systems operate largely in isolation from computational diagnostic pipelines. This study develops and empirically validates a hybrid AI-IoT framework that integrates continuous environmental sensing with multi-stage machine learning analysis for early warning of chromatic deterioration. The framework combines unsupervised hierarchical clustering in HSV colour space with supervised Random Forest (RF) classification, incorporating microclimatic variables—relative humidity, temperature, and a derived Humidity Stress Factor (HSF)—as predictive features rather than post-hoc explanatory factors. Validation was conducted on three heritage sites in Semarang, Indonesia: Lawang Sewu, Pasar Johar, and Tawang Railway Station, using 48,830 expert-labelled pixels and 3,750 annotated tiles from radiometrically corrected photogrammetric models. The RF classifier achieved a mean macro-F1 score of 0.87 (range: 0.84–0.90) and overall accuracy exceeding 90%. Hierarchical clustering yielded stable partitions with silhouette scores averaging 0.75 (SD = 0.02). The integrated workflow improved spatial Intersection-over-Union (IoU) by 17–51% compared to unsupervised clustering alone. A rigorous ablation study confirmed that the HSF contributes a 9.2% absolute increase in macro-F1. Leave-one-site-out cross-validation yielded F1 scores of 0.79–0.82, demonstrating acceptable cross-site generalisability. Mixed-effects modelling confirmed a significant positive association between cumulative hours of relative humidity exceeding 85% and material deterioration (β = 0.18, 95% CI: 0.13–0.23, p < 0.001 after Bonferroni correction). On retrospective test data, the framework identified 79% of impending high-risk events (recall = 0.79), providing a mean technical lead time of 4.8 to 16.7 hours conditional upon correct detection. When accounting for missed events, the expected lead time ranged from 3.8 to 13.2 hours. The framework is presented as a methodologically reproducible blueprint requiring local recalibration of thresholds and model parameters for transfer to other tropical contexts. A complete reproducibility package including data, code, and calibration artifacts is provided.</p>