DIG-IT!

The Fraunhofer Institute’s subproject looks at the implementation of Deep Learning methods including its application for image data derived from ecological research. The strong structural form of artificial neural networks enables machine learning to identify target features and characteristics in images. Machine learning methods will significantly increase the efficiency of image feature identification through automated evaluation, enabling ecologists to process much larger digital image databases. Sample data for testing these methods will be supplied by the project’s research groups from the field of biology that include a broad spectrum of data sources from pollen and roots, to bats, ensuring universal applicability in ecological sciences. The analyses (that shall be supported by a doctoral candidate) will focus on the automated detection of previously defined objects in an image, for instance as required when counting pollen in a microscopic image. Furthermore, the investigations will incorporate the subsequent automated segmentation that separate detected objects from background data. The aim is to refine the object margins from the surrounding rectangles towards exact outlines. This would enable isolated objects to be passed on to human experts or downstream software tools. Research tasks in the subproject are focused on “Deep Convolutional Neural Networks” given their proven suitability for evaluating image data. The method focus will shift gradually from “supervised” to “semi-supervised” until it reaches “unsupervised” classification. At this final stage, the training will require little to no data to be previously classified by humans. Additionally, the inclusion of semantical information (e.g. topology of root systems) will be assessed to enhance image analysis. This approach aims to improve the efficiency of image classification and processing, developing new fields of ecological research that rely on universally practicable methods.

Mareike Fischer is managing the subproject “Biomathematics” with the help of a doctoral candidate and a PostDoc. Biomathematics mediates between mathematic theories and biological application, providing new tools to analyse biological data. M. Fischer will mediate between machine-learning theoreticians and ecologists by incorporating biomathematical modelling and statistical evaluation. The subproject “Roots” will focus on modelling root growth temporally by using graph theory and explore whether the level of root growth is statistically suitable for determining nutrient distribution in the soil. Additionally, multidimensional data collected on bats, wood and pollen are further study targets for promoting deep learning methods in ecological research. Baseline data required to train machine-learning approaches are numerous and include DNA (field of expertise for M. Fischer), sound recordings of bats, or the chemical properties of wood and pollen. DIG-IT!’s Deep Learning methods will identify data characteristics that are neither visible to humans nor can be derived from additional information.

Martin Wilmking is project coordinator for DIG-IT! and manages the subproject “Wood Anatomy”. Some of the measures will be performed by a doctoral candidate in collaboration with partners from Fraunhofer IGD and associated master’s dissertations. The interdisciplinary PostDoc working as “Catalyst and Multiplier” will link all members within the consortium (catalyst) and provide expertise from DIT-IT! to external users (multiplier). Further measures will take place in collaboration with ongoing research projects (e.g. “WETSCAPES” project) and in the dendro-ecological laboratory “DendroGreif”. Two members of technical staff are assigned to the coordinator: a part-time position (50 %) for supporting coordination and a full-time position for assisting the users with data acquisition, digital image processing, and data management.

Jürgen Kreyling manages the subproject “Roots”, in which he coordinates the “automated digitisation of dynamics in root systems of plants”. The subproject will establish the automated acquisition of various root characteristics (e.g. length, diameter etc.) from minirhizotron scans in various settings, ranging from simplified laboratory conditions to complex field conditions. The project will be realised by a doctoral candidate in collaboration with partners from Fraunhofer IGD and associated master’s dissertations. J. Kreyling’s working group is equipped with a greenhouse and established research sites in a local beech forest and the nature reserve “Karrendorfer Wiesen”, enabling high-resolution rhizotron scans over a prolonged period of time. Further research collaborations with the “WETCAPES” project are expected.

Gerald Kerth manages the subproject “automated identification of species/individuals for selected animal species”. The project will be realised by a doctoral candidate in collaboration with partners from Fraunhofer IGD and associated master’s dissertations. The required field data will only be collected from the wild, in particular from 10 different wintering grounds of bats in Mecklenburg-Vorpommern. The established methods for the digital determination of species will be applied at research sites belonging to the two DBU (Deutsche Bundestiftung Umwelt) projects "FUN" and "Living underground".

Hans Joosten is managing the subproject “pollen identification” that will be realised by a doctoral candidate in collaboration with partners from Fraunhofer IGD, the Chair of Biomathematics (UG), and associated master’s dissertations.  Laboratory analyses will be predominantly conducted in the laboratories belonging to the working group Peatland Studies and Palaeoecology where specialised equipment is available for everything from coring, sample collection and preparation to pollen analysis. High-resolution images will be provided by microscopes at the Department of Biology’s Imaging Center. The established transect of the working group J. Kreyling will provide samples from beech trees in blossom. Seasonally stratified sediments will be sampled in collaboration with the German Research Centre for Geosciences (GFZ) in Potsdam.