DeepGreen: A Deep Learning Based Tree-Ring Width Data Model for Paleoclimatic Data Assimilation

利用对过去一千年多年来气候变率和变化的探测和归因分析，可以将未来几十年和几个世纪的气候预测纳入背景，从而为气候政策提供信息并建设有复原力的社会。这个项目旨在调查什么影响因素导致气候在十年时间尺度上变化，为什么以及如何变化？这项研究将利用现有的树木年轮数据，这些数据具有高度可复制性、精确的年代，并可在全球范围内获得。因此，它们构成了气候模式同化观测的主要来源。然而，使用树木年轮作为模式数据存在挑战，包括：(1)响应的季节性；(2)生物信号与气候信号的区别；以及森林作为气候传感器的阈值反应。在这个项目中，研究人员建议通过组装足够大的数据集进行机器学习，使用深度机器学习来开发、验证和解释树轮（特别是树轮的宽度）的新数据模型。这种解释和吸收古气候模式中树木年轮记录的新方法框架，有可能改善共同时代气候场的重建，从而加速在数年至数十年时间尺度上对气候变率和变化的探测和归因。该项目将通过为博士后科学家提供受监督的研究、教育推广和指导活动，以及通过支持重要的本科生研究经历，来建设科学能力。该项目与美国国家科学基金会项目（“提供地球科学研究和学习的早期机会：PEARLS”）合作，将支持地球科学多样化的努力。将组织一个开放的虚拟研讨会，培训和指导早期职业研究人员，旨在建立古气候学数据建模的深度学习框架。该项目将使用基于深度学习的方法（DeepGreen）来开发、验证和解释树轮宽度（TRW）的新数据模型。研究人员将通过将具有相似响应特征的TRW系列聚类成集合来收集足够大的数据集，用于机器学习。使用伪代理实验，将确定适合TRW建模的最小数据集大小要求和算法。然后将开发来自现实世界TRW网络的数据模型，并相对于现有的线性统计和非线性和多元过程的TRW模型评估它们的技能。通过从深度学习中推导和验证树轮宽度的数据模型，研究旨在：(1)进一步了解现有TRW数据中包含的环境信息；(2)识别deepGreen和过程或模拟模式中的结构和观测不确定性；(3)补充现有的建模工作，针对现有TRW数据的信息含量和观测不确定性；(4)支持通过机器学习识别其他古气候观测数据模型的工作。在工作范围之外，deepGreen数据模型可以在现有的数据同化框架中使用，以开发和评估新的古气候重建。对这些结果的分析可以加速在年和几十年的时间尺度上对气候变率和变化的探测和归因。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Using detection and attribution analyses of past climate variability and change at multidecadal timescale over the last millennium is a means by which climate projections for coming decades and centuries can be contextualized to inform climate policy and build resilient societies. This project aims to investigate what influential factors cause the climate to vary on decadal timescales, why and how? The research will leverage existing Tree-ring data which are highly replicated, precisely dated, and available at global scale. As such, they constitute a major source of observations for assimilation in climate models. However, there are challenges for using tree rings as model data including (1) the seasonal nature of the response; (2) distinction of biological vs. climatic signals; and threshold responses in forests as climate sensors. In this project, the researchers propose to use deep machine learning to develop, validate and interpret new data models for tree rings (specifically, the width of tree rings) by assembling sufficiently large datasets for machine learning. This new methodological framework to interpret and assimilate tree-ring records in paleoclimate models has the potential for improving the reconstruction of climate fields over the common era which in turn could accelerate the detection and attribution of climate variability and change on timescales of years to decades. The project will build capacity for science by providing supervised research, education outreach and mentoring activities for a postdoctoral research scientist, and by supporting a significant undergraduate research experience. In partnership with NSF project (“Providing Early Access to Research & Learning in geoscienceS: PEARLS), this project will support efforts to diversify the geosciences. One open virtual workshop will be organized to train and mentor early career researchers with the aim to establish deep learning framework for data modeling in paleoclimatology.This project will use deep learning-based approach (DeepGreen) to develop, validate and interpret new data models for tree-ring width (TRW). The researchers will assemble sufficiently large datasets for machine learning by clustering TRW series with similar response characteristics into aggregates. Using pseudoproxy experiments, a minimum dataset size requirements and algorithms suitable for TRW modeling will be identified. Data models from the real-world TRW network will then be developed and their skill evaluated relative to that of existing linear statistical and nonlinear and multivariate process-based TRW models. By deriving and validating data models for tree-ring width from deep learning exercises, the research seeks to: (1) further understand the environmental information contained in extant TRW data; (2) identify structural and observational uncertainties in deepGreen and process or mimic models; (3) complement existing modeling efforts targeting the information content and observational uncertainty in extant TRW data; (4) support efforts to identify data models by machine learning for other paleoclimatic observations. Beyond the scope of the work, deepGreen data models may be used within existing data assimilation frameworks, to develop and evaluate new paleoclimate reconstructions. Analysis of the results may accelerate the detection and attribution of climate variability and change on timescales of years and decades.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.