PANOPS – Revealing Earth´s plant functional diversity with citizen science

关于功能性植物特性（“植物性状”）和功能多样性的全球格局的知识在地理覆盖、分类群、性状及其生态功能方面是有限的。这限制了我们对生物多样性-环境关系和地球系统动力学的理解，以及对全球变化影响的预测。一系列研究试图利用具有全球和空间连续覆盖的环境预测因子，对植物性状数据库（TRY数据库）中收集的性状观测进行空间外推。然而，这种外推的全球特征分布仍然具有很大的不确定性和很少的一致性。此外，卫星地球观测正在成为植被评估的一项关键技术，但在这方面的潜力有限，因为“鸟类视角”只能提供冠层顶部植物和一些功能特征的信息。为了填补这些关于功能多样性的数据空白，揭示整个陆地植物群的生物多样性-环境关系，PANOPS将使用一个完全不同的“视角”：数百万全球分布的、众包的植物照片。由于形式服从功能，在照片中可见的植物形态特性可以直接或间接地告知植物的一些功能特征。申请人的初步研究将数千张众包植物照片（iNaturalist）和植物性状（TRY数据库）与深度学习和计算机视觉技术（卷积神经网络，CNN）联系起来。这不仅可以从单张照片中预测一些被认为是功能多样性关键的植物性状（例如，每面积叶质量、叶面积、株高），还可以通过使用照片的地理位置对数千个这样的预测进行空间聚合，从而预测它们的全球分布。因此，深度学习与来自公民（照片）和专业科学（特征）的数据相结合，为研究宏观生态学研究的焦点问题提供了一个高潜力和协同的工具集。这一愿景在PANOPS中得到了四个中心目标的证实：1)从植物照片中推断出植物性状，结合辅助环境数据和种内性状变异的先验。2)评估模型在图像设置、植物生长形式和生物群落中的泛化，并评估潜在的机制，即，当从照片中预测性状时，模型真正“看到”了什么？3)从空间聚合的植物性状预测中生成和评估植物功能性状和多样性指标的全球图谱。4)拓展生物多样性与环境关系的知识，包括植物性状的地理趋同，功能性状多样性与非生物环境和功能生态系统特性的关系，以及功能多样性对功能生态系统特性的弹性和抗性的影响。

Knowledge on global patterns of functional plant properties ('plant traits') and functional diversity are limited in terms of geographic coverage, taxa, traits, and their ecological functions. This restricts our understanding of biodiversity-environment relationships and Earth system dynamics, as well as to project impacts of global change. A series of studies attempted to spatially extrapolate trait observations curated in plant trait databases (TRY database) using environmental predictors with global and spatially continuous coverage. However, such extrapolated global trait distributions still feature large uncertainties and little agreement. Also, Earth observation with satellites, which is generally becoming a key technology for vegetation assessments, has limited potential in this regard as the ‘bird perspective’ can only inform on plants in top canopy layers and a few functional traits. Aiming to fill these data gaps on functional diversity and reveal biodiversity-environment relationships for the entire terrestrial flora, PANOPS will use an entirely different ‘perspective’: Millions of globally distributed, crowdsourced plant photographs. As form follows function, plant morphological properties visible in photographs can, directly and indirectly, inform on several functional characteristics of plants. Preliminary studies by the applicant linked thousands of crowdsourced plant photographs (iNaturalist) and plant traits (TRY database) with deep learning and computer vision techniques (Convolutional Neural Networks, CNN). This not only enabled to predict several plant traits considered key for functional diversity from single photographs (e.g. leaf mass per area, leaf area, plant height), but even their global distribution by spatially aggregating thousands of such predictions using the photographs’ geolocations. Accordingly, deep learning in concert with data from citizen (photographs) and professional science (traits) provides a high potential and synergetic toolset to study questions in the spotlight of macroecological research. This vision is substantiated in PANOPS by four central objectives: 1) Identification of plant traits inferable from plant photographs, incorporating ancillary environmental data and priors on intraspecific trait variation. 2) Assess the generalization of models across image settings, plant growth forms, and biomes and assess underlying mechanisms, i.e. what do models really ‘see’ when predicting traits from photographs? 3) Generate and evaluate global maps of plant functional traits and diversity metrics from spatially aggregated plant trait predictions. 4) Expand knowledge on biodiversity-environment relationships, in terms of the geographic convergence of plant traits, relationships of functional trait diversity with abiotic environmental and functional ecosystem properties, as well as imprints of functional diversity on the resilience and resistance of functional ecosystem properties.