Virtual 3D models as synthetic training data for automated image analysis of insect biodiversity surveys. 3DTRAB

• Applicants: Professor Dr.-Ing. Bernd Freisleben; Professor Dr. Michael Heethoff; Dr. Lea Heidrich; Dr. Ivaylo Kostadinov
• Subject Area: Methods in Artificial Intelligence and Machine Learning ; Ecology and Biodiversity of Animals and Ecosystems, Organismic Interactions ; Systematics and Morphology (Zoology)
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 557404521

The diversity of insects, both in terms of species and individual numbers, has decreased in recent decades, in some cases significantly - in public this is even communicated as the "insect apocalypse". A close-meshed and large-scale survey of the status quo of insect communities is just as necessary as the scientific investigation of possible anthropogenic influences on them. However, current monitoring programmes have to limit the spatial and temporal coverageand resolution, as the determination of bulk samples, as generated from common trap systems, istoo labour-intensive while the number of available taxonomists continues to decrease. The scientific need for (partial) automation of sample processing, i.e. sorting and determination, is therefore enormous in order to be able to study insect diversity with reasonable effort. Artificial intelligence (AI), especially deep learning, is increasingly used for insect identification, but requires large amounts of annotated training data. Obtaining sufficient amounts of this data is not only a problem for rare species, but also in bulk samples where there is a large variation in body posture and orientation. The aim of this project is to develop a user-friendly hardware and software system for the automatic detection and recognition of insects in order to automate the determination of bulk samples as far as possible. To this end, 1) a tool will be developed to generate synthetic training data from 3D models, 2) a workflow and image acquisition system will be developed to process collection samples as automatically and efficiently as possible so that they can be fed to the AI (Heidrich, University of Marburg) and 3) an AI model for the detection and identification of insects in collection samples will be generated, which can be automatically expanded with synthetic training data. In addition, 4) a web-based software system for user-friendly application and management of AI models for automated image analysis of insect biodiversity surveys will be developed. The overall development is an iterative process in which permanent coordination of test data acquisition, initial semantic modelling and enrichments, training data genesis and AI creates an effective and reliable overall workflow that enables users to evaluate extensive monitoring studies.

• DFG Programme: Research data and software (Scientific Library Services and Information Systems)
• Co-Investigator: Dr. Markus Mühling