Collaborative Research: Planning: FIRE-PLAN: Advancing Wildland Fire Analytics for Actuarial Applications and Beyond

不受控制的野火的影响范围很广，从破坏原生植被到财产损失，再到长期健康影响和人命损失。提高野火活动、火灾行为和野火天气预测的准确性是制定更有效的火灾控制策略和降低野火风险的关键。最近的研究表明，人工智能（AI）工具可以通过提供更高的空间和时间火灾天气预报来帮助规划即将到来的规定烧伤，还可以帮助制定更有效的野火风险缓解策略。然而，目前用于预测火灾活动的建模工具在很大程度上受到许多时间或空间约束。例如，大多数用于野火风险分析的深度学习（DL）方法往往局限于系统地捕获以不同时空分辨率记录的多维信息的能力。此外，这样的DL架构本质上是静态的，并没有明确说明复杂的动态现象，这通常是准确评估野火驱动因素的关键。最后，这些模型主要依赖于监督学习方法，其中大量的任务特定标签（例如，火或没有火）。为了解决野火风险分析中的这些挑战，该项目将利用地球系统科学，DL，计算拓扑学，统计学和精算科学的接口固有的跨学科方法。该项目旨在将拓扑数据分析（TDA）的概念引入野火预测建模，并将其与时间感知图神经网络等新兴人工智能机器相结合。由此产生的新方法，预计将更好地捕捉在荒地火灾过程中的时间和空间的形状模式，并协助更可靠的野火风险的统计评估。新的高保真预测方法将有可能在有限，嘈杂或不存在标记信息的情况下，在多个空间和时间尺度上提供火灾行为，火灾活动和火灾天气的预测。为了提高野火分析研究解决方案的实用性，该项目的研究人员将与利益相关者密切合作，特别是关注保险行业。该项目将为从本科生到执业精算师的所有教育水平提供野火科学、人工智能和数学科学的多个跨学科培训机会。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

The impacts of uncontrolled wildland fires range from the destruction of native vegetation to property damages to long-term health effects and losses of human lives. Increasing accuracy in projections of wildland fire activity, fire behavior, and wildland fire weather is the key toward developing more efficient fire control strategies and reducing the risks of wildfires. Recent studies have demonstrated that the tools of artificial intelligence (AI) can help in planning for upcoming prescribed burns by providing higher spatial and temporal fire weather forecasts and can also assist in developing more efficient strategies for wildfire risk mitigation. However, the modeling tools that are currently used to predict fire activity are largely subject to a number of temporal or spatial constraints. For instance, most deep learning (DL) approaches for wildfire risk analytics tend to be restricted in their capabilities to systematically capture the multidimensional information recorded at disparate spatio-temporal resolutions. Furthermore, such DL architectures are inherently static and do not explicitly account for complex dynamic phenomena, which is often the key behind the accurate assessment of wildfire driving factors. Finally, these models primarily rely on supervised learning approaches where a large number of task-specific labels (e.g., fire or no fire) are needed. To address these challenges in wildfire risk analytics, this project will leverage inherently interdisciplinary approaches at the interface of Earth system sciences, DL, computational topology, statistics, and actuarial sciences. The project aims to introduce the concepts of topological data analysis (TDA) to wildfire predictive modeling, coupling them with such emerging AI machinery as time-aware graph neural networks. The resulting new methods are expected to better capture the shape patterns in the wildland fire processes with respect both to time and space and to assist in a more reliable statistical assessment of wildfire risks. The new high-fidelity predictive approaches will have the potential to deliver forecasts of fire behavior, fire activity, and fire weather at multiple spatial and temporal scales under scenarios of limited, noisy, or nonexistent labeled information. To enhance the utility of the research solutions in wildfire analytics, the researchers in this project will work in close collaboration with stakeholders, particularly, focusing on the insurance sector. The project will provide multiple interdisciplinary training opportunities at the nexus of wildfire sciences, AI, and mathematical sciences at all educational levels, from undergraduate students to practicing actuaries.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.