Comparison of different machine learning systems to identify tree VOCs using MOx-sensor networks for forest health monitoring

Advanced gas sensor networks offer new opportunities for detecting forest health changes and supporting early-warning systems. Trees emit volatile organic compounds (VOCs) that reveal their physiological state. This study investigates the potential of metal-oxide (MOx) gas sensors combined with machine learning to create scent profiles that differentiate tree health conditions. We compare two analytical approaches: eXtreme Gradient Boosting (XGBoost), a gradient-boosting algorithm, and Linear Discriminant Analysis (LDA), a classical statistical classification method. MOx sensor arrays were deployed in controlled laboratory and forest settings to record VOC emissions. The data were preprocessed through cleaning, normalization, and feature extraction. XGBoost models were trained with optimized hyperparameters for accuracy and generalization. In parallel, LDA was used to reduce dimensionality and visualize group separations. Both models successfully classified tree health states based on VOC profiles. XGBoost performed well despite the large feature set, although the small dataset size reduced class separation. LDA produced clear separations between conditions and aided interpretability. Both methods demonstrated robustness in detecting early stress indicators, supporting their potential for real-world monitoring. The results confirm that VOC-based scent profiling with MOx sensors is a promising non-invasive method for assessing forest health. Combining machine learning with gas sensor arrays enhances early detection of stress, enabling proactive forest management. Future research will focus on improving sensor selectivity and sensitivity, integrating real-time analytics, and scaling systems for large-scale deployment.