Protecting Biodiversity at Lake Neusiedl, Austria
Austria
Overview
The goal is to answer a multitude of questions regarding the current conservation status and ecology of species in the reed belt of Lake Neusiedl, a European biodiversity hotspot for wetland vertebrates.
Partners
Huawei Australia, University of Vienna
Lake Neusiedl National Park
Objectives
Assessing the occurrence and population density of wetland birds and amphibians in the reed belt during their reproductive period, to evaluate their current conservation status.
Identifying the characteristics of ecological niches of reed birds and amphibians across an
environmental gradient from the lake’s siltation zone to the reed edge exposed to the open lake area to predict effects of future spatial habitat shifts caused by climate change.Quantifying the seasonal importance of the lake’s reed belt for birds.
Set up long-term monitoring with acoustics,
Implementation
Impact
Adopted the AI model methodology outlined in LeBien et al. (2020) to develop a regional species classification model.
We implemented a deep learning pipeline that utilizes a Convolutional Neural Network (CNN) model capable of classifying species using spectrogram images.
Our partners used Arbimon’s Pattern Matching (PM) tool to automate the training data acquisition process which allows users to choose an acoustic signal template (i.e., an example of a species' vocalization) and then search through a userdefined playlist of audio recordings to identify signals that match the selected template.
In addition to data labeling through the PM tool, we also employed another analytical method called "Negative Pattern Matching" to gather more absence clips for the training dataset that introduces random background noise into the training data, which can help reduce the number of false positives classifications made by the model.
We compiled training data for 41 classes (class = species + song type) using pattern matching.
Our final trained model was chosen to be the one that best optimized the evaluation metrics. The final model incorporated these data augmentation techniques: random time shifting, blending, and upsampling positive audio clips.
