Spatiotemporal estimation of ambient forest phytoncides: Unveiling patterns through geospatial-based machine learning approach

This study investigated biogenic volatile organic compounds (BVOCs) emitted by tree species, with a specific focus on estimating their ambient air concentrations within the Xitou Nature Education Area, Taiwan. Employing geospatial-based machine learning approaches, which are rarely applied in this context, we aimed to estimate the ambient levels of key forest phytoncides which are representative compounds within the BVOC group. Data on phytoncide, including camphene and α-pinene, were directly collected from the study area. Geospatial data including meteorological factors, topography, land cover, and nearby landmarks were additionally collected and set as predictor variables influencing phytoncides. Random Forest (RF), Gradient Boosting (GB), Extreme Gradient Boosting (XGB), and Light Gradient Boosting Machine (LGBM) were integrated with an explainable artificial intelligence tool to develop the model estimates. To evaluate model performance, we conducted overfitting tests, 10-fold cross-validation, and stratified analysis. The results showed that RF and XGB were the most effective algorithms, explaining approximately 83.3% and 98.4% of the spatiotemporal variability in camphene and α-pinene, respectively. The robustness of these models was confirmed through extensive validation. Spatial pattern analysis revealed that variations in these biogenic compound concentrations were linked to meteorological conditions and vegetation types. Finally, this study presented an innovative approach to accurately estimating and mapping the spatial distribution of forest phytoncides, providing valuable insights to support environmental management, urban planning, and public health.