Loading...

Performance Evaluation of Machine Learning and Statistical Approaches for Wildfire Modeling and Prediction

Mehrabi, Majid | 2023

143 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 55794 (09)
  4. University: Sharif University of Technology
  5. Department: Civil Engineering
  6. Advisor(s): Moghim, Sanaz
  7. Abstract:
  8. Wildfires are complex phenomena with many indeterminate and highly unpredictable driving factors that have remained unresolved. During the last decade, machine learning methods have successfully excelled in wildfire prediction as an alternative to traditional field research methods by elucidating the relationship between historical wildfire events and various important variables. The main purpose of this research is to evaluate the random forest machine learning approach and the logistic regression statistical approach to prepare a wildfire susceptibility map using data related to historical wildfires and effective variables in the Okanogan region in Washington province of the United States and Jamésie region in Quebec province of Canada. Results show that based on the node impurity and the permutation methods, the most critical variables in modeling in both areas are land cover, temperature, wind, elevation, precipitation, and normalized vegetation difference index. Although temperature and precipitation are essential factors of forest fires, the specific topographical factors and climatic conditions to each region will determine the amount and intensity of future wildfires. The obtained area under the receiver-operating curve indicates that all models are reliable and can be used at the regional level to prepare a forest fire probability map. The performance of the random forest model for predicting the validation data in the Okanogan and Jamésie regions is 98.2 and 99.1%, respectively, and for the logistic regression model is 75.3 and 87.4%, respectively. To predict future wildfires, the random forest model and logistic regression performed 55.5 and 54.6% for the Okanogan region, respectively, and 75.5 and 69.8% for the Jamésie region, respectively. In addition, the models' performance is examined in such an approach that the models are trained with the data of one region and validated in another region at the same period. Results show that the random forest and logistic regression models trained with the data from the Jamésie region have a 52.5 and 54.7% performance, respectively, to predict wildfires in the Okanogan region and contrariwise, 66.3% and 55.2% performance, respectively. Our findings confirm that the performance of the models to simulate and predict future wildfires varies in regions with different physical features
  9. Keywords:
  10. Machine Learning ; Random Forest Algorithm ; Logistic Regression (LR)Analysis ; Fire Probabilistic Safety Assesment ; Fire Probability Map ; Wildfire

 Digital Object List

 Bookmark

No TOC