Tang Jing, Weeramongkolkul Manapat, Suwankesawong Supanida, Saengtabtim Kumpol, Leelawat Natt, Wongwailikhit Kritchart
International School of Engineering, Faculty of Engineering, Chulalongkorn University, Phayathai, Pathumwan, Bangkok, 10330, Thailand.
Disaster and Risk Management Information Systems Research Unit, Chulalongkorn University, Phayathai, Pathumwan, Bangkok, 10330, Thailand.
Heliyon. 2024 Jul 2;10(13):e34021. doi: 10.1016/j.heliyon.2024.e34021. eCollection 2024 Jul 15.
Forest fires in Thailand are a recurring and formidable challenge, inflicting widespread damage and ranking among the nation's most devastating natural disasters. Most detection methods are labor-intensive, lack speed for early detection, or result in high infrastructure costs. An essential approach to mitigating this issue involves establishing an efficient forest fire warning system based on amalgamating diverse available data sources and optimized algorithms. This research endeavors to develop a binary machine-learning classifier based on Thailand's forest fire occurrences from January 2019 to October 2022 using data acquired from satellite resources, including the Google Earth engine. We use four gas variables including carbon monoxide, sulfur dioxide, nitrogen dioxide, and ozone. The study explores a range of classification models, encompassing linear classifiers, gradient-boosting classifiers, and artificial neural networks. The XGBoost model is the top-performing option across various classification evaluation metrics. The model provides the accuracy of 99.6 % and ROC-AUC score of 0.939. These findings underscore the necessity for a comprehensive forest fire warning system that integrates gas measurement sensor devices and geospatial data. A feedback mechanism is also imperative to enable model retraining post-deployment, thereby diminishing reliance on geospatial attributes. Moreover, given that decision-tree-based algorithms consistently yield superior results, future research in machine learning for forest fire prediction should prioritize these approaches.
泰国的森林火灾是一个反复出现且严峻的挑战，造成广泛破坏，位列该国最具毁灭性的自然灾害之中。大多数检测方法劳动强度大，缺乏早期检测的速度，或者导致高昂的基础设施成本。缓解这一问题的关键方法之一是基于整合各种可用数据源和优化算法建立一个高效的森林火灾预警系统。本研究致力于利用从包括谷歌地球引擎在内的卫星资源获取的数据，开发一个基于2019年1月至2022年10月泰国森林火灾发生情况的二分类机器学习分类器。我们使用包括一氧化碳、二氧化硫、二氧化氮和臭氧在内的四个气体变量。该研究探索了一系列分类模型，包括线性分类器、梯度提升分类器和人工神经网络。在各种分类评估指标中，XGBoost模型是表现最佳的选项。该模型的准确率为99.6%，ROC-AUC得分为0.939。这些发现强调了建立一个集成气体测量传感器设备和地理空间数据的综合森林火灾预警系统的必要性。反馈机制对于在部署后进行模型再训练也至关重要，从而减少对地理空间属性的依赖。此外，鉴于基于决策树的算法始终能产生更好的结果，未来森林火灾预测机器学习方面的研究应优先考虑这些方法。
