Agroecosystems in a Climate Crisis: Using Big Data to understand the Out of the Ordinary

英国已经观察到破坏性的气候变化，2020年是气温、降水和阳光排名前十的第一年。2011-2020年期间，英国的平均气温比1981-2010年的基线平均气温高0.5摄氏度，比1961-1990年的平均气温高1.1摄氏度。气温每上升2摄氏度，极端降水的频率预计将增加1.5-3倍，并随着气温进一步升高呈非线性增加。因此，气候变化对英国的农业生态系统构成了直接和直接的风险（例如，通过干旱、洪水、害虫、疾病），但我们对这种风险的理解很差。大数据提供了一个解决这个问题的机会。例如，利用卫星遥感在区域或国家尺度上进行推断，或者利用来自数据丰富的研究平台的数据。北威克农场平台（NWFP; http://resources.rothamsted.ac.uk/farmplatform）就是这样一个平台，它是一个世界领先的实验，通过收集数据来表征四种截然不同的农场系统。收集包括：(i)水化学/流动，土壤湿度；（二）温室气体排放；(3)气象;（iv）调查土壤/作物养分、土壤动物、植物区系；(v)作物/牲畜生产性能；这通常与外部收集相结合，例如通过遥感提供的收集。该项目旨在使用统计和机器学习方法检测平台高维数据集中的极端事件和异常观测。我们提出了一系列检测技术，每一种都能捕捉空间、时间和尺度效应，从而产生新的方法。耦合目标是在比较平台系统的恢复能力时，了解已识别的极端/异常对统计推断的影响。在这里，线性混合模型（lmm）和结构方程模型（sem）将以两种不同的方式进行调整，以便更准确地捕获不确定性：(1)对极端/异常具有鲁棒性（降权重），(2)通过极值理论更好地捕获它们（增权重）。

Disruptive climate change is already being observed in the UK, with 2020 being the first year to feature in the top-10 for temperature, precipitation and sunlight. Between 2011-2020, the UK has been, on average, 0.5C warmer than the 1981-2010 baseline climatic mean and 1.1C warmer than 1961-1990. The frequency of extreme precipitation is expected to increase by a factor of 1.5-3 with a 2C temperature rise, and to increase non-linearly with further warming. Climate change therefore poses an immediate and direct risk to the UK's agroecosystems (e.g., through drought, flooding, pests, disease), but our understanding of this risk is poor.Big Data provides an opportunity to address this. For example, using satellite sensing to make inferences at regional or national scales, or using data from data-rich research platforms. The North Wyke Farm Platform (NWFP; http://resources.rothamsted.ac.uk/farmplatform) is one such platform - a world-leading experiment that captures data to characterize four contrasting farm systems. Collections include: (i) water chemistry/flow, soil moisture; (ii) greenhouse gas emissions; (iii) meteorological; (iv) surveys for soils/crop nutrients, soil fauna, floristics; and (v) crop/livestock performance; that are typically coupled with external collections such as those provided via remote sensing.This project aims to detect extreme events and anomalous observations in the platform's high-dimensional datasets using statistical and machine learning methods. We propose a range of detection techniques, each capturing spatial, temporal and scale effects, where new methodologies will result. A coupled aim is an understanding of the impacts of identified extremes/anomalies for statistical inference when the platform's systems are compared for their resilience. Here, linear mixed models (LMMs) and structural equation models (SEMs) will be adapted in two distinct ways so that uncertainties are more accurately captured: (1) to be robust to extremes/anomalies (down-weighting) and (2) to better capture them (up-weighting) via extreme-value theory.