AI tools for fungal monitoring
Applying and Improving AI tools for fungal monitoring
Automatic visual recognition of fungal species assists mycologists and citizen scientists in species identification in the field. Here we aim to advance the state-of-the-art in AI-based identification of fungi that would strongly encourage and support collection of valuable biodiversity data by improving the ability of amateurs to submit correct species identifications. Usually, such a recognition system requires an image as an input and provides a ranked list of species sorted by likelihood. In mycological practice, species identification typically does not depend only on the visual observation of the specimen but also on other information available to the observer. Therefore, we will develop tools to include non-visual information (metadata) in the identification process. The recording and analysis of associated metadata and AI training by fungal collections supported by DNA sequence data will contribute to performance beyond the-state-of-art by improving the confidence estimates applied for fungal identification. This WP will extend existing collaborations linked to the Danish Fungal atlas to the European scale, offering highly advanced AI identification tools for citizen scientists across the continent.
Objectives
Development of AI-based identification tools to enhance the fungal mapping/recording in Europe. Offering AI-based image recognition services through recording apps, e.g., PlutofGO, will be a substantial step forward in fungal recording across Europe. By making these tools available in mobile applications, we will support the recording of fungi backed by images, specimens and DNA barcodes that will link the records to consolidated UNITE SHs via the PlutoF data management system.
Continuous re-training of AI-models from the stream of acquired specimens. Starting from methods that require human-in-the-loop in the form of data curation, we will implement methods that reduce the need of human intervention in long-term learning, and handle the changing biases and temporal evolution of data statistics over fungal seasons.
Adding side-information on fungal records, e.g., habitat, substrate, location, time, and smell, to improve species prediction. We will incorporate available metadata to improve models for species identification in Fungi. Even though so-called side-information is sometimes crucial for accurate species identification, the research in this direction is neglected in AI research. Researchers usually just exploit estimated geographical priors; thus, the full potential of metadata has yet to be explored.
Developing an AI-based system for fungal species recognition in an open-species-set scenario. For machine learning based systems it is challenging to address the task of identification in an open-set setting, i.e. when “unknown” or “none” (not a fungal specimen) is a possible correct decision, since, by definition, such data are not represented in the training set. The possibility to detect a specimen unlike any other in the training set is very valuable, and we will aim at developing algorithms with this capability.
Tasks
Dataset for deep learning. Creating and continuously updating a dataset representing the European-wide distribution of fungi species. Communication with institutions, combining datasets from citizen science and experts and publishing the created corpus on one of the public data-sharing platforms for both machine learning (e.g., LILA, LINDAT, or COVE) and conservation (e.g., GBIF).
Development of CV/ML methods for geolocated data. Developing algorithms, techniques, and deep neural network models to provide new insights for applications of computer vision and machine learning to geolocated species data, focussing on images with metadata information.
Integration of recognition service into PlutofGO application. Optimization and deployment of methods developed in Task 2.1 into applications (e.g. PlutofGO) with a focus on citizen science, conservation and biodiversity monitoring.