Zoology/Biology session

Surveying biodiversity is essential in our changing world to reveal the trends and to understand the underlying mechanisms of community alterations. This is also true for fish, which are one of the most frequently studied taxa in freshwater ecology. However, the rapidly changing environment makes the monitoring and the extensive community modelling relatively hard. This is where Machine Learning (ML) approaches come in perspective, with their exceptional ability to model complex ecological data, like fish-habitat relationships. Beside their promising features, ML methods have not been utilized in floodplain fish community modelling.

In this study, we tried to evaluate a methodological framework based on Cascaded Neural Networks to sufficiently model the floodplain fish community of one of the largest European remnant floodplain ecosystem. With this approach we also aim to provide a proper example of managing problems of biological origin which may appear in Machine Learning based community modelling.

The irregular manner of available fish faunistic data, the different features of the applied sampling methods (eDNA and electrofishing), the complexity of the environmental explanatory variables (e.g. lateral connectivity) and the unequal frequency distribution of fish makes the preliminary analyses and data transformations even more necessary than in other cases. To tackle these problems, we used the two different sampling methods simultaneously with adjusted spatial scale and merged samples. In addition, fish were classified to robust functional groups by their habitat preference, habitat use and lifestyle traits. Based on the previous hydrological models, we delineated two discrete modelling area which correspond to an average and a flood condition.

After the model training and validation, we will attempt to simulate the fish communities of the delineated floodplain waterbodies and the corresponding main channel on a waterbody scale. We would like to prove that Cascaded Neural Networks are able to deal with complex community modelling, with the proper preliminary assumptions and alterations.