Mobilization of trait data from digital image files by deep learning approaches

基于植物的数字图像(照片、标本扫描)以及相关的分类单元和特征数据，我们将开发和训练深度学习算法，以识别(A)其他图像文件中的特征，以及(B)将识别出的特征用作相当于传统确定过程的步骤，并再次使用深度学习来识别图像文件中的分类单元。这一方法将用于大规模动员用于生态学和分类学的特征数据，以及用于数字化收藏和公民科学门户网站中的物种自动识别。作为该项目的一部分，将演示如何(1)利用来自公民科学门户网站和照片社区的照片对花卉和水果特征进行自动检测，以记录气候变化对物候的影响；(2)特征识别可用作公民科学门户网站中分类单元识别的第一步；(3)分类单元识别可用于质量控制和数字化收集数据中的扩展文件；以及(4)可用于通过公民科学门户网站和照片社区跟踪入侵物种的传播。新调集的特征数据将通过提供信息的基础设施和GFBio框架内的长期数据储存库提供给科学界。

Based upon digital images of plants (photos, specimen scans) and related taxon- and trait data we will develop and train deep-learning algorithms to recognize (a) traits in other image files and (b) use the identified traits as a step equivalent to a conventional determination process and using deep learning again to recognize taxa in image files. This approach will be used for massive mobilization of trait data for use in ecology and taxonomy, as well as for the automatic identification of species in digitized collections and citizen science portals. As part of the project will be demonstrated how (1) photos from citizen science portals and photo communities can be used by automated detection of flower and fruit traits to document the impact of climate change on phenology, (2) feature recognition can be used as a first step towards taxon recognition in citizen science portals, (3) the taxon recognition can be used for quality control and extended documentation in digitized collection data and (4) may serve to track the spread of invasive species through civil science portals and photo communities. The newly mobilized trait data will be made available to the science community via the contributing information infrastructures and via long term data repositories in the frame of GFBio.