A hybrid AI-IoT framework for predictive conservation of tropical heritage façades: empirical validation from a multi-site study in Indonesia

Abstract

Abstract Predictive conservation for tropical built heritage is hindered by the persistent separation of continuous environmental sensing from computational deterioration diagnostics. We address this gap through a fourteen-month hybrid AI-IoT deployment across three heritage sites in Semarang, Indonesia (Lawang Sewu, Pasar Johar, and Tawang Railway Station), integrating unsupervised HSV-space hierarchical clustering with Random Forest classification and a derived Humidity Stress Factor (HSF). The framework was evaluated using 48,686 expert-labelled pixels, achieving a macro-F1 of 0.87 (95% CI: 0.85–0.89) and pixel accuracy of 91.2% (95% CI: 89.4%–93.0%), with inclusion of the HSF producing a 9.2-percentage-point increase in macro-F1 ( p < 0.001). Cumulative exposure to RH > 85% showed a strong association with deterioration (Spearman’s ρ = 0.82, 95% CI: 0.71–0.89, p < 0.001), while leave-one-site-out cross-validation yielded F1 scores of 0.79–0.82 (95% CI: 0.76–0.84), providing evidence of spatial transferability within the study region. Retrospective testing detected 79% of impending high-risk events and yielded conditional technical lead-time estimates of 4.8, 9.2, and 16.7 h for the 6-, 12-, and 24-hour horizons, respectively (expected: 3.8, 7.3, and 13.2 h after accounting for missed events). These lead-time estimates are technical and require operational validation before integration into management practice. The study advances prior work through three methodological refinements: direct integration of environmental sensing as a predictive input, prevention of data leakage through training-set-only clustering, and quantitative estimation of early-warning lead times from retrospective analysis. The resulting workflow provides a reproducible basis for risk-prioritised intervention planning. However, generalisation beyond the tropical coastal context requires independent cross-site validation and operational piloting. The findings therefore provide an empirical basis for further development of anticipatory conservation approaches for tropical heritage façades.