A next-generation approach for quantifying tropical plant diversity across scales

热带森林拥有地球上许多植物物种。我们知道，这种生命多样性的重要性，不仅在于它本身是伟大的自然奇观之一，而且还在于支撑全球生物地球化学循环（例如碳循环），并决定对气候变化的适应能力。然而，尽管经过了几个世纪的研究，我们仍然不确定热带地区有多少种植物，哪些地区有最多的物种，或者这些不同物种的丰度如何随着时间的推移而变化，例如由于气候变化。热带生物多样性的不确定性主要有两个原因：首先，热带森林中有数千种植物（例如，热带拉丁美洲有12万种植物），其中许多看起来非常相似，因此很难（有时是不可能）确定单个植物属于哪个物种。目前基于形态学的物种鉴定方法具有固有的主观性，难以标准化，这意味着鉴定误差很高，而且大多无法量化。其次，热带森林面积巨大，而且往往地处偏远，这意味着生态学家只能对森林总面积的一小部分进行采样，而且大多数热带森林对科学来说仍然是未知的，而且在未来几十年里很可能仍然是未知的。这两项挑战无法通过过去几十年的相同方式收集更多数据来克服，而是需要从根本上改变方法。这项研究的总体目标是建立一套统一的、定量的、可扩展的方法，利用新技术和现有数据集来测量亚马逊地区——地球上最大、最多样化的热带森林——的植物多样性。我建议使用四种独立而又互补的方法来实现这一目标。这个挑战的规模是巨大的；因此，我计划最初只关注最常见的树种和科。由于这些常见物种占亚马逊所有树木的近20%，因此减少几百种物种的不确定性将对我们对亚马逊植物生物多样性的理解产生深远的影响。首先，我将开发一种新的自动化方法，通过测量叶片样品的反射光光谱来识别植物物种，并使用人工智能（AI）技术基于这些光谱对植物进行分类。我将把这种方法应用于亚马逊地区五个常见的植物科，它们加起来约占亚马逊地区树木总数的19%。这将为亚马逊范围内的标准化定量物种鉴定提供一个框架。其次，我将使用基于无人机的传感器在景观尺度（250公顷）上绘制25种常见物种的地图，该传感器测量树冠的反射光光谱。我将把这些无人机图像与主要树冠树种的实地验证位置结合起来，然后使用人工智能方法根据它们在景观中的树冠光谱来学习和绘制这些常见树种。第三，我将测试我们是否可以用这些常见树种的分布来代替稀有树种的分布。利用明确的物种间共变建模方法，利用常见树种的分布来预测稀有树种的丰度。第四，我将测试我们在多大程度上可以利用卫星图像扩大我们对常见冠层树种分布的理解。卫星图像可以提供整个亚马逊流域与植物生物多样性相关的连续信息。我建议使用大量现有的森林清查图数据集来解开卫星生物多样性信号。通过关注相同的大型常见树冠树种，我将分离出由卫星传感器实际检测到的部分植物群落。

Tropical forests hold much of Earths plant species. We know that this diversity of life is important, both in its own right as one of the great natural wonders, but also for underpinning global biogeochemical cycles (e.g. the carbon cycle), and determining resilience to climate change. Yet, despite centuries of research, we still don't know with any certainty how many of species of plants there are in the tropics, which areas have the most species, or how the abundance of these different species are changing through time, for example due to climate change.There are two main reasons for the uncertainty surrounding tropical biodiversity: First, there are thousands of plant species in tropical forests (e.g. 120,000 plant species in tropical Latin America), many of which look extremely similar, making it difficult (sometimes impossible) to identify which species an individual plant belongs too. Current approaches for species identification based on morphology are inherently subjective and difficult to standardize, meaning that identification errors are high and mostly unquantified. Second, tropical forests are vast, and often remote, meaning that ecologists are only able to sample a tiny fraction of the total forest area and most tropical forests remain unknown to science and are likely to remain so in coming decades. These two challenges cannot be overcome by collecting more data in the same way that we have for the past decades, instead a fundamental change in approach is required.The overarching goal of this fellowship is to establish a suite of unified, quantitative, and scalable approaches that use new technologies and existing datasets to measure plant diversity across Amazonia, Earth's largest and most diverse tropical forest. I propose to realize this goal using four independent yet complementary approaches. The scale of this challenge is huge; therefore, I plan to initially focus only on the most common tree species and families. Because these common species account for nearly 20% of all trees in Amazonia, reducing uncertainty in a few hundred species will have a profound impact on our understanding of Amazonian plant biodiversity.First, I will develop a new automated approach for identifying plant species by measuring reflected light spectra of leaf samples and classifying plants into species based on these spectra using artificial intelligence (AI) techniques. I will apply this approach to five common Amazonian plant families that together account for approximately 19% of individual trees in Amazonia. This will provide a framework for standardized quantitative species identifications at Amazon-wide scales.Second, I will map 25 common species at landscape scales (250ha) using a drone-based sensor that measures reflected light spectra of tree canopies. I will combine this drone imagery with field-verified locations of dominant canopy tree species, and then use AI approaches to learn and map these common tree species based on their canopy spectra across the landscape.Third, I will test if we can use the distribution of these common species as proxies for the distribution of rarer tree species. Using a new modelling approaches that explicitly for the covariation among species, I plan to predict the abundance of rare tree species using the distribution of common species.Fourth, I will test the extent to which we can scale-up our understanding of the distribution of common canopy tree species using satellite imagery. Satellite imagery can provide continuous information across the entire Amazon basin that relates to plant biodiversity. I propose to use massive existing forest inventory plot datasets to untangle the satellite biodiversity signal. By focusing on the same large common canopy tree species, I will be isolating the portion of plant communities that are actually detected by satellite sensors.