2023-08-17 –, HS 120
In this project we generate tools to identify birds within the spatial extent of a meteorological radar. Using the opportunities created by modern dual-polarization radars we build graph neural networks to identify bird flocks. For this, the original point cloud data is converted to multiple undirected graphs following a set of predefined rules, which are then used as an input in graph convolutional neural network (Kipf and Welling, 2017, https://doi.org/10.48550/arXiv.1609.02907). Each node has a set of features such as range, x, y, z coordinates and several radar specific parameters e.g. differential reflectivity and phase shift which are used to build model and conduct graph-level classification. This tool will alleviate problem of manual identification and labelling which is tedious and time intensive. Going forward we also focus on using the temporal information in the radar data. Repeated radar measurements enable us to track these movements across space and time. This makes it possible for regional movement studies to bridge the methodological gap between fine-scale, individual-based tracking studies and continental-scale monitoring of bird migration. In particular, it enables novel studies of the roles of habitat, topography and environmental stressors on movements that are not feasible with current methodology. Ultimately, we want to apply the methodology to data from continental radar networks to study movement across scales.
In this project we generate tools and a python package to identify birds within the spatial extent of a meteorological radar. Using the opportunities created by modern dual-polarization radars we develop graph neural networks (GNN) to identify bird flocks. For this, the original point cloud data is converted to multiple undirected graphs following a set of predefined rules. For example, each point of interest needs to have a label and a minimum number of neighbours within a specified range. Graphs are then used as an input in graph convolutional neural network (Kipf and Welling, 2017, https://doi.org/10.48550/arXiv.1609.02907). Each node has a set of features such as range, x, y, z coordinates and several radar specific parameters e.g. differential reflectivity and phase shift which are used to build model and conduct graph-level classification. Model learns hidden layer representations that encode both local graph structure and features of nodes and is based on an efficient variant of convolutional neural network.
This tool will alleviate problem of manual identification and labelling which is tedious and time intensive. Going forward we also focus on using the temporal information in the radar data. Repeated radar measurements enable us to track these movements across space and time. This makes it possible for regional movement studies to bridge the methodological gap between fine-scale, individual-based tracking studies and continental-scale monitoring of bird migration. In particular, it enables novel studies of the roles of habitat, topography and environmental stressors on movements that are not feasible with current methodology. Ultimately, we want to apply the methodology to data from continental radar networks to study movement across scales. This project is a collaboration between the Netherlands eScience Center and the University of Amsterdam. This is an ongoing project and feedback provided by the community is highly valued.
Git Repo of the package: https://github.com/point-cloud-radar/bird-cloud-gnn
ML Applications (e.g. NLP, CV)
Abstract as a tweet –Where is the flock? Or how graph convolutional neural networks help to identify birds from metheorological radar
Expected audience expertise: Domain –some
Expected audience expertise: Python –some
Project Homepage / Git –
Currently I work as Research Software Engineer at the Netherlands eScience Center in Amsterdam. Pior to that I worked as a Research Fellow at the Atlantic Technological University in Galway, Ireland. I have a passion for scientific programming, open source software and linux. I am also a lecturer and a PhD supervisor.
