Predicting insect outbreaks using machine learning: A mountain pine beetle case study.
Pouria Ramazi1,2, Mélodie Kunegel-Lion3, Russell Greiner2,4
1Department of Mathematical and Statistical Sciences University of Alberta Edmonton AB Canada.
Ecology and Evolution
|October 14, 2021
Summary
Machine learning models accurately predict mountain pine beetle outbreaks up to seven years in advance. Proper evaluation methods are crucial to avoid misleading results in forest pest management planning.
Area of Science:
- Ecology and Environmental Science
- Forestry and Forest Management
- Computational Science and Machine Learning
Background:
- Accurate prediction of insect outbreaks, such as mountain pine beetle, is essential for effective forest management.
- Machine learning (ML) offers promising solutions for intermediate-term (e.g., 5-year) outbreak prediction.
- Challenges include selecting optimal ML models, relevant covariates (with time lags), and robust performance evaluation.
Purpose of the Study:
- To systematically identify and evaluate machine learning models for predicting mountain pine beetle outbreaks.
- To assess model performance for 1-, 3-, 5-, and 7-year future infestations in the Cypress Hills area.
- To investigate the impact of data splitting strategies and covariate history length on prediction accuracy.
Main Methods:
- Trained nine machine learning models, including generalized boosted regression trees (GBM) and novel mixed models.
- Compared random data splitting versus year-based splitting for training and testing datasets.
- Analyzed prediction accuracy in relation to covariate history length for various ML models.
Main Results:
- GBM models achieved 92% and 88% AUC for 1- and 3-year predictions; mixed models achieved 86% and 84% AUC for 5- and 7-year predictions.
- Year-based data splitting provided more accurate performance evaluations than random splitting, preventing misleadingly high scores.
- Neural network, naive Bayes, and GBM models showed improved accuracy with increased covariate history length, especially for shorter-term predictions.
Conclusions:
- Optimized machine learning models can reliably predict mountain pine beetle outbreaks across multiple time horizons.
- Appropriate data splitting and covariate selection are critical for accurate model evaluation and practical application.
- The developed predictive approach is adaptable for other invasive species and aids in forest pest management and field study planning.
