CDI-Type I: Novel machine learning models for predicting species distributions in response to climate change

地球的生态系统因人类活动而迅速变化，威胁着许多物种，但气候变化、土地覆盖变化和其他因素对物种分布的准确影响是一个复杂的问题，到目前为止还没有进行综合分析。机器学习为解决这一生态问题提供了机会，而这种生态问题可以刺激机器学习的进步。在这项提案中，计算机科学家、生态学家和地球科学家联手开发了一种根本不同的方法，使用计算方法来研究两个主要的生态问题：1)物种组合是否会随着气候/土地利用的变化而变化？2)物种地理范围是否会随着气候/土地利用的变化而变化？土地利用、气候和物种相互关系之间潜在的复杂关系可能会影响物种分布和分布的变化。目前对物种分布和变化的分析受到所用模型简单的限制；大多数模型只分析环境因素对单个物种分布的影响。研究人员建议(1)创建一种基于结构化预测和条件主题模型的物种分布建模的新方法，以同时发现和预测鸟类物种组合(共同出现的鸟类物种群)；(2)开发新的对比挖掘算法来发现地理位置(热点？)与不同时间段物种组合的重大变化(即主题的变化)相对应，以及(3)创建创新的对比挖掘算法，以发现单个物种的空间分布发生重大变化的地点，并识别与这些变化最相关的特征。该项目的方法，使用结构化预测、条件主题模型和对比挖掘，将允许测试迄今无法测试的生态学假设：栖息地连通性促进物种对气候变化的反应，同时测试生态学中关于物种对气候和土地利用变化的个体或相互依赖的反应程度的基本问题。就更广泛的影响而言，这项提议将为关于栖息地恢复以缓解气候变化的环境政策贡献重要的科学知识。PIS还将通过参与美国林业局和NSF资助的H.J.安德鲁斯长期生态研究计划之间的研究管理伙伴关系，向联邦土地管理机构提供研究外展服务。

Earth's ecosystems are changing rapidly in response to human activities, threatening many species, but the precise effects of climate change, land cover change, and other factors on species distributions is a complex problem that has so far defied integrated analysis. Machine learning provides opportunities to address this ecological problem, and this ecological problem can stimulate advances in machine learning. In this proposal, a computer scientist, ecologist, and geoscientist join forces to develop a fundamentally different approach using computational methods to examine two major ecological problems: 1) Are species assemblages changing according to climate/land use change? and 2) Are species geographic ranges changing according to climate/land use change? Potentially complex relationships among land use, climate, and species inter-relationships may influence species distributions and changes in distributions. Analyses of species distribution and change are currently limited by the simplicity of the models used; most models only analyze the effect of environmental factors on the distribution of single species. The investigators propose to (1) create a new approach to species distribution modeling based on structured prediction and conditional topic models that will simultaneously discover and predict bird species assemblages (groups of co-occurring bird species); (2) develop novel contrast mining algorithms to discover geographic locations (?hotspots?) corresponding to significant changes in species assemblages (i.e., changes in topics) between time periods and (3) create innovative contrast mining algorithms to discover sites with significant changes in an individual species' spatial distribution and identify features which are most associated with these changes. This project's approach, employing structured prediction, conditional topic models, and contrast mining will allow testing a hitherto untestable ecological hypothesis: habitat connectivity facilitates species response to climate change, while simultaneously testing fundamental questions in ecology about the degree to which species respond individualistically or interdependently to climate and land use change. In terms of broader impacts, this proposal will contribute significant scientific knowledge to environmental policy about habitat restoration to mitigate climate change. The PIs will also provide outreach for their research to federal land management agencies through their involvement in a researchmanagement partnership between the US Forest Service and the NSF-funded H.J. Andrews Long-term Ecological Research program.