Predicting Risks of Forest Fires using Federated Machine Learning Methods

在全球范围内，数亿人的生计直接依赖当地的森林生态系统。然而，根据加拿大国家林业数据库的数据，1980年至2020年期间，每年发生8500多起森林火灾，每年烧毁200多万公顷土地。2003-2017年期间，森林火灾造成了近50亿加元的保险损失。因此，开发高效的森林火灾综合管理系统（IFFM）以减少损失至关重要。IFFM系统最重要的组成部分之一是森林火险条件预测（FFDC），即探测火灾并预测其蔓延。总体而言，森林ffdc高度依赖于森林的气象变量（MV）、生物物理变量（BV）和地形变量（TG），准确预测森林ffdc是一项复杂的任务。现有的FFDC预测方法只使用一到两种变量，导致预测的准确性较低。研究人员和行业合作伙伴将设计一个软件框架，用于使用所有三种数据，即MV， BV和TG来预测ffdc，以提高准确性。我们将应用机器/深度学习方法来预测ffdc，因为在点燃和蔓延森林火灾的三种数据之间存在复杂的相互作用。我们的框架考虑了所有三种类型的数据和多个优化模型，因此预测精度更高。此外，联合机器学习方法将加速预测过程，为消防员提供额外的宝贵时间来管理火灾。我们将通过使用安大略省和阿尔伯塔省的公开数据来验证该系统。从拟议的系统中获得的专业知识将扩大两个合作伙伴的投资组合，他们将把研究成果作为新的、扩展的服务产品提供给他们的客户——专注于加拿大森林火灾的公共和私营部门公司。该项目将对加拿大的经济和社会产生巨大的影响，研究也可以用于研究洪水和气候变化。作为该计划一部分的HQP培训将在自然资源管理和气候变化等日益增长的领域发挥作用。

Globally, the livelihoods of hundreds of millions of people directly depend upon their local forest ecosystems. However, according to data from the Canadian National Forestry Database, over 8,500 forest fires occurred each year between 1980-2020, burning more than 2 million hectares every year. Forest fires triggered insured losses of almost CAD 5 billion between 2003-2017. Therefore, it is important to develop efficient, integrated forest fire management (IFFM) systems to reduce the losses. One of the most important components of an IFFM system is the forecasting of forest fire danger conditions (FFDC), namely, detecting fires and predicting their spread. In general, FFDCs are highly dependent on meteorological variables (MV), biophysical variables (BV), and topography (TG) of forests, and accurately predicting FFDCs becomes a complex task. Existing FFDC prediction methodologies use only one or two kinds of variables, leading to less accurate predictions. The researchers and industry partners will design a software framework for predicting FFDCs using all the three kinds of data, namely, MV, BV, and TG, for better accuracy. We will apply machine/deep learning methods to predict FFDCs because of the complex interplay among the three kinds of data in igniting and spreading forest fires. Our framework will result in better prediction accuracy because it considers all the three types of data and multiple optimized models. In addition, federated machine learning methods will accelerate the prediction process, giving firefighters extra valuable time to manage fires. We will validate the system by using publicly available data for Ontario and Alberta. The expertise gained from the proposed system will expand the portfolios of the two partners who will offer the research outcomes as new, expanded service offerings to their clients -- public and private sector companies who focus on fighting forest fires in Canada. The project will have a tremendous impact on both the economy and society of Canada, and the research can also be leveraged for the study of floods and climate change. HQP trained as part of the program will fill roles in the growing sectors of natural resource management and climate change.