Prediktering av skogsdöd : En jämförande studie av datadrivna metoder för tidsserier av klimatdata

Forest die-off events are increasing globally due to climate change, yet understanding and predicting these events remains a challenge. Machine learning offers a promising approach to identifying early warning signals in climate data, which could improve forest management and conservation efforts. This study applies machine learning based time-series classification to predict forest die-offs using climate variables from TerraClimate and labeled die-off events from the Global Tree Mortality Database. Three machine learning models, Detach-ROCKET, GRU, and Random Forest, were evaluated to determine their predictive ability, varying the temporal proximity to the event. Feature importance analysis was conducted to identify key climate variables linked to tree mortality. The study also examined the influence of hyperparameters, data preprocessing, and training-test splitting strategies to ensure robust model evaluation. The results highlight heat- and drought-related variables as the strongest predictors of forest die-offs. Among the models, Detach-ROCKET significantly outperformed the others, achieving the highest accuracy and demonstrating predictive ability up to three years before an event. However, performance declined for longer lead times, limiting its practical usefulness for early detection. Additionally, clustered data in the Global Tree Mortality Database introduced challenges, requiring careful data partitioning to prevent information leakage. While the study demonstrates the potential of machine learning for forest die-off prediction, limitations in dataset size, clustering, and negative sample generation impact model performance and generalizability. Addressing these challenges will be crucial for improving predictive accuracy in future research. Nevertheless, this work represents the first global-scale application of advanced machine learning models to predict forest die-offs, providing valuable methodological insights into performance and reliability, while reinforcing the link between climate stressors and tree mortality.