AI and Machine Learning for predicting species responses to global change

.

Overview

The project brings together advances in biological modelling, Machine Learning, and Artificial Intelligence (AI) to determine whether cutting-edge AI technologies can transform our capacity to monitor, predict and manage biological responses to global change.

Two of the core challenges in understanding species responses to environmental change are the sparsity of the data available and the complexity of building nuanced models that can link knowledge about fine-grained behavioural responses to population-level effects.

This project aims to address both challenges.

Funding organisations

 
 

Project links

Dr Reid Tingley | Macroecology Research

Dr Reid Tingley on Twitter

Professor Carla Sgrò | Genetics, evolution, biodiversity and climate change

Professor Bernd Meyer website

Associate Professor Alan Dorin Research Profile

Project Description

Data Acquisition

The core reason why species distribution data is generally sparse, both temporally and spatially, is that traditional techniques for gathering such data are labour-intensive and thus costly. Our project opens up new avenues for low-cost data collection by utilizing recent developments in AI and Machine Learning to pave the way for continuous and high-density long-term data collection. We start from a range of readily available data sources, integrating data from existing large-scale research infrastructure (e.g., Australian Acoustic Observatory), social media (e.g., Twitter, Instagram), citizen science projects (e.g., FrogID) and targeted observations (e.g., drone imagery). We develop automatic, AI-driven interpretation methods for these data sources, based on recent advances in sound and image interpretation. This enables us to turn the complex unstructured information from such data sources into numerical data that can readily be aggregated and fed into species occurrence databases and species distribution modelling methods.

Data Modelling

Traditional  models for forecasting species’ responses to environmental change work mostly at the macroscopic level. Typical approaches include establishing correlations between observed distributions and environmental variables at the species-level (e.g., correlative species distribution models), explicitly modelling the spatiotemporal dynamics of populations (e.g., metapopulation models), or some combination of these two methods (i.e., hybrid models). However, population-level effects are ultimately an abstraction. The true cause for these effects is found in fine-grained behavioural responses at the individual level, in response to the environment, to conspecifics, and to other species. Modelling only at the population level must necessarily abstract from this and cannot directly integrate our knowledge on individual-level behaviour. By integrating our rich knowledge on individual-level behaviour directly into the modelling process, we should be able to generate models that better reflect biological responses to environmental change and that have greater predictive power. However, this is not possible with standard statistical species distribution models. Our project will overcome this limitation by building agent-based models that simulate individual-level behaviour explicitly from first-principles and derive population-level responses as the aggregation of thousands or millions of individuals rather than relying on population-level statistics.

As a proof-of-concept, the project currently focuses on range shifts in cane toads, an important invasive species in Australia.

Team members

Chief Investigators

Research Associates

  • Dr Christoph Bergmeir, Department of Data Science and AI
  • Dr Sazzad Hossain, Department of Data Science and AI
  • Ms Hannah Burke, (formerly, Department of Data Science and AI)

PhD Students

  • Md Mohaimenuzzaman, Department of Data Science and AI
  • Sesa Singha Roy, Department of Data Science and AI
  • Arman Pili, School of Biological Sciences

Publications

Md Mohaimenuzzaman, C Bergmeir, and B Meyer. Pruning vs XNOR-net: A comprehensive study of deep learning for audio classification on edge-devices. IEEE ACCESS, vol. 10, pp. 6696-67-7, 2022. Link

Md Mohaimenuzzaman, C Bergmeir, IT West, and B Meyer. Environmental sound classification on the edge: A pipeline for deep acoustic networks on extremely resource-constrained devices. arXiv:2103.03483v3. Link

HM Burke, R Tingley, A Dorin, "Tag Frequency Difference: Rapid estimation of image set relevance for species occurrence data using general-purpose image classifiers”, Ecological Informatics Volume 69, July 2022, 101598. Link

MM ElQadi, A Dorin, A Dyer, M Burd, Z Bukovac, M Shrestha, “Mapping species distributions with social media geo-tagged images: Case studies of bees and flowering plants in Australia, Ecological Informatics 39, 23-31, 2017. Link