Ecological Machine Learning

We develop machine learning (ML) and deep learning (DL) methods to understand how ecological communities assemble in space and time, and what determines the interactions between species within them. Ecological patterns, such as community composition, species co-occurrence, and biotic interactions in pollination networks, and food webs are shaped by nonlinear effects, higher-order interactions, and spatial or temporal dependencies. While ML and DL excel at learning such complexity, they are often perceived as black boxes suited only for prediction. We challenge this view by developing methods that leverage ML and DL for both prediction and ecological inference.

Research areas

• Scalable joint species distribution modeling (jSDMs): We develop methods for predicting biodiversity across scales and for understanding community assembly from large biodiversity datasets, including eDNA metabarcoding and insect monitoring. Our R package sjSDM uses PyTorch to fit jSDMs at scale, allowing variation partitioning across environmental, spatial, and biotic components for hundreds or thousands of species.
• Trait-matching: We use ML and DL to infer the importance of trait-matching in ecological networks (Pichler et al., 2020). A particular challenge is to disentangle the trait-matching signal from confounding factors such as species abundances and unknown encounter rates.
• Interpretable ML and explainable AI (xAI): We develop and apply methods that extract interpretable effects from complex ML models, providing new ecological insights and moving beyond prediction toward mechanistic understanding. A particular focus lies on understanding when and under which conditions these effects are reliable, for example under collinearity, and whether they have reliable statistical properties.
• Accessible deep learning tools for ecologists: We build user-friendly software that makes DL accessible to non-experts. Our R package cito allows ecologists to fit deep neural networks using R's familiar formula syntax, without requiring prior expertise in DL. Moreover, cito provides xAI effects as well as statistical properties, and the option to include images (CNN, including transfer learning) and combine multiple data streams in the form of multimodal neural networks.

Team

The AG Pichler is embedded in the Theoretical Ecology Lab (externer Link, öffnet neues Fenster) led by Prof. Dr. Florian Hartig. We share our team, teaching activities, and research seminars with the broader lab, see the links below.

Joining the Group

If you're interested in joining the AG Pichler, here are some general topics we're looking for students: 

• Biodiversity modeling space and time (e.g. based on images and env time series with CNN/RNN/Transformers)
• Trait-matching with ML/DL, predicting plant-pollinator or foodweb networks, and inferring trait-matching signals with xAI
• Implementation/programming of new features in cito and sjSDM

The general application info for the lab can be found here.

Selected Publications

• Pichler, M., & Hartig, F. (2023). Can predictive models be used for causal inference?. arXiv preprint arXiv:2306.10551. [preprint]
• Pichler, M., & Hartig, F. (2023). Machine learning and deep learning—A review for ecologists. Methods in Ecology and Evolution, 14(4), 994-1016. [journal]
• Pichler, M., & Hartig, F. (2021). A new joint species distribution model for faster and more accurate inference of species associations from big community data. Methods in Ecology and Evolution.[journal]
• Pichler, M., Boreux, V., Klein, A. M., Schleuning, M., & Hartig, F. (2020). Machine learning algorithms to infer trait‐matching and predict species interactions in ecological networks. Methods in Ecology and Evolution, 11(2), 281-293. [journal]