Fungal Fortune Telling: AI Predicts Macrofungal Fruitbody Occurrences

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Jan 10 2024

In a paper published in the journal Scientific Reports, researchers tackled the scarcity of data on using machine learning for predicting macrofungal fruitbody occurrences based on weather conditions. Focusing on the mycorrhizal species of Russula and Amanita in Western Hungary (2015-2020), they employed an artificial neural network (ANN) to forecast fruitbody presence in undisturbed forests.

Their approach, a feed-forward multilayer perceptron with backpropagation, yielded two key outcomes: a species-specific model outperforming the typical genus-level study and an accuracy range of 60–80% in estimating fruitbody formations. This pioneering work marks the first successful use of ANN to predict fruitbody occurrence based on meteorological parameters across an extended vegetation period, holding promise for field mycological studies relying on sporocarp data to determine species abundance.

Related Work

Prior research extensively explored how various environmental factors influence macrofungal fruitbody formation, particularly highlighting weather components like air temperature and precipitation in European habitats. These studies emphasized the rapid responsiveness of fungal fruitbodies to environmental changes, distinguishing them from plants and advocating for their role in environmental monitoring.

Challenges persist in precisely timing fruitbody formations due to environmental fluctuations, prompting a need for species-specific investigations. Long-term studies revealed the complexities of capturing sporocarps and the diverse distribution of fungi in habitats, challenging standardized sampling.

Structured Fungal Forecasting Methodology

This methodology is structured, starting with identifying weather parameters strongly correlated with fruitbody formation at both genus and species levels. These parameters serve as inputs for the ANN, where 70% of the data forms the training set, teaching the ANN the macrofungal formation patterns. The remaining 30% forms the testing set, validating the ANN's forecasts. Verification is crucial to substantiate the reliability of the predictions.

The research delves into two distinct areas in western Hungary: Vendvidék and Vétyem forest reserves, known for their extensive fungal diversity and well-documented mycological literature. The areas' unique climatic and geographical features and near-complete forest cover render them "mushroom paradises." The study zones, roughly 2 km² each, boast diverse forest stands comprising oak, beech, hornbeam, birch, Scots pine, spruce, and alder.

Forest-associated species with mycorrhizal associations and an extended fruiting season take priority when selecting species for the study. These species should be present in both designated locations under examination and backed by ample occurrence data to ensure comprehensive and reliable study parameters. Accordingly, the selection narrows to six Amanita and seven Russula species—known for their ecological significance and roles as edible or toxic fungi in Hungary.

During the peak growing season, conducting field surveys over 28 days—15 in Vendvidék (2018–2020) and 13 in Vétyem forest reserve (2016–2020) identify fruiting fungal species, meticulously documenting their occurrences. Merging observations from both locations allows the creation of a dataset of 364 occurrences across 13 species.

Homogenized climate data from the Hungarian Meteorological Service, including daily mean temperature, relative air humidity, air pressure, and precipitation, are inputs to calculate 19 meteorological parameters crucial for predicting fruitbody formations. Researchers guide the selection process based on the parameters' correlation with fungal occurrences.

Collinearity diagnostics ensure that the chosen meteorological parameters do not exhibit collinearity, confirming the independence of variables crucial for the ANN predictions. The ANN employs the multilayer perceptrons (MLP) with a backpropagation algorithm—a non-linear learning method. This model, structured with input, hidden, and output layers, learns from the training dataset and estimates fruitbody formation based on meteorological inputs, ensuring an accurate estimation through iterations.

To validate the results, researchers use a contingency table and apply five key verification indices—accuracy (ACC), bias (B), probability of detection (POD), false alarm ratio (FAR), and critical success index (CSI). These indices, calculated based on the contingency table, substantiate the accuracy and reliability of the ANN forecasts. In adherence to guidelines and legislation, the experimental research and field studies on plants align with relevant institutional, national, and international standards.

Fruitbody Forecasting: ANN Study

The study aimed to employ artificial neural networks (ANN) for predicting fruitbody formation in Amanita and Russula species based on meteorological parameters. Researchers built ANN models and tested them using genus-level and species-specific weather variables. In the Amanita species, genus-level variables showed varied reliability, while species-specific parameters significantly improved forecast accuracy. The Russula species exhibited higher reliability with species-specific weather data, showcasing varied prediction successes across different species.

Results showed that genus-level meteorological data yielded mixed forecasting reliability for Amanita and Russula species. However, using species-specific weather parameters notably enhanced the accuracy of fruitbody formation prediction in most cases. The study underscored the importance of species-specific variables over genus-level data, with improved prediction outcomes in the ANN models, particularly evident in the blind testing sessions.

Despite demonstrating effectiveness for species with extended growth periods, the study highlighted the approach's limitations for species with shorter fruiting periods. Additionally, it emphasized the need for future investigations into other fungal groups, such as Cortinarius, Lactarius, and Inocybe species, to determine the broader applicability of ANN for fruitbody occurrence prediction.

Overall, the research suggests that while ANN models show promise for forecasting fruitbody formation in certain fungal species based on meteorological data, their reliability varies based on species-specific variables and the duration of growth periods. Further exploration is needed to understand the method's broader applicability across different fungal groups and ecological niches.

Conclusion

In summary, this pioneering study harnesses ANN to predict fruitbody formations in Amanita and Russula species using weather parameters. Applying ANN with species-specific weather data unveils a cost-effective, reliable method for forecasting fungal occurrences, specifically within mycorrhizal groups. However, its adaptability across various fungal types and ecological niches warrants further exploration to enhance predictive accuracy and broaden ecological applications.