Speaker
Description
An integrated harvesting site monitoring methodology was developed to support data-driven forest management and root rot risk assessment by combining operational harvester data, field measurements, and spatial analysis. The study aimed to evaluate whether terrain and multispectral predictors can reliably distinguish healthy from root rot–damaged stands at the sample plot level.
Root rot occurrence was identified using harvester production report (HPR) data and validated through field measurements in 905 sample plots. Harvested stumps were classified as healthy or root rot–damaged, and plot-level predictors were derived from a 5 m resolution digital elevation model and multi-season Sentinel-2 imagery (2021, pre-harvest conditions). In total, 42 terrain and spectral variables were extracted and used in a structured machine learning workflow.
The dataset was balanced using upsampling and divided into training (80%) and testing (20%) subsets. An XGBoost classification model was trained using 5-fold cross-validation with grid search optimization based on AUC. Model performance was evaluated using accuracy, balanced accuracy, sensitivity, specificity, F1 score, and Cohen’s Kappa.
The model achieved total and balanced accuracy of 74.4%, with moderate agreement (κ = 0.49). Sensitivity (71.1%) and specificity (77.7%) indicate reliable discrimination between healthy and damaged plots. The most influential predictors were the normalized height model and depth-to-water index. Healthy plots were located on slightly higher microsites, while damaged plots were associated with marginally lower depth-to-water values.
The results demonstrate the potential of integrating harvester data and remote sensing predictors for spatially explicit root rot risk assessment and provide a foundation for developing digital forestry decision-support tools.
| Keywords | harvester, StanfordConvertXML, root rot |
|---|
