M.Sc. Michael Mechenich defends his PhD thesis "Machine Learning Methods for Ecological Modeling Under Global Change" on Friday the 20th of March 2026 at 13 in the University of Helsinki Exactum building, Auditorium CK112 (Pietari Kalmin katu 5, basement). His opponent is Professor Panagiotis Papapetrou (Stockholm University, Sweden) and custos Professor Indrė Žliobaitė (University of Helsinki). The defence will be held in English.
The thesis of Michael Mechenich is a part of research done in the Department of Computer Science and in the Computational Data Analysis group at the University of Helsinki. His supervisor has been Professor Indrė Žliobaitė (University of Helsinki).
Machine Learning Methods for Ecological Modeling Under Global Change
Ecological modeling discovers patterns of association between the biosphere and the abiotic environment, quantifying notions of biotic-abiotic interconnection first expressed, in the early 1800s, as nature's "painting" by Alexander von Humboldt. In its several forms, namely macroecological modeling and single and joint species distribution modeling (SDM and JSDM), ecological modeling makes quantitative predictions which serve to further our scientific understanding of natural systems, and anticipate biotic response to rapid anthropogenic environmental change.
This dissertation proposes machine learning (ML) methods, and contributes a large ML-ready database, for ecological modeling in a changing Earth system: global change in climatic conditions and the species composition of ecological communities, in the geologic past and a future transformed by human activity, necessitate models which generalize to new settings. The application and development of these ML methods yields new scientific insight and advances the practice of two ecological modeling techniques: SDM relates dependent species occurrence or abundance to an independent environment. Ecometric modeling reverses the direction of conditional dependence, relating dependent environment to independent community-level functional traits.
Contributions are made to four steps in the modeling process: first, spatial sampling and dataset compilation are advanced in developing the Eco-ISEA3H database, the first general Earth observation (EO) resource for ecological modeling built on an equal-area discrete global grid system (DGGS). Second, feature construction is addressed in compiling a novel set of target environmental extremes for use in ecometric modeling, designed to represent evolutionary selective pressures. Third, we build upon feature selection in introducing a new transfer learning methodology comprising two parts, namely (1) removing highly correlated features, and (2) identifying a minimum subset of available features with greatest explanatory power, via forward feature selection (FFS) by spatial cross-validation. Finally, model validation is advanced in developing a spatial blocking tool based upon tetromino packings, for use in the spatial cross-validation process.
The methods developed and evaluated in this dissertation together seek to define and relate ecological communities and environment in ways which transfer to new settings. As ML methods are increasingly used in quantitative ecology, and as the predictions of ecological models are increasingly incorporated in land use and conservation planning, these methods facilitate model development based upon rigorous spatial sampling, relevant and informative features and targets, and evaluation which accounts for common structures of dependence in environmental datasets.
Availability of the dissertation
An electronic version of the doctoral dissertation will be available in the University of Helsinki open repository Helda at
Printed copies will be available on request from Michael Mechenich: