Environmentally grounded pseudo-absence sampling for species distribution models : a language guided framework

Miok, Kristian (author), Laza, Antonio V. (author), Škrlj, Blaž (author), Robnik Šikonja, Marko (author), Pârvulescu, Lucian (author)

Aim: Species Distribution Models (SDMs) are widely used in conservation planning, invasive species management and global change assessments. Their reliability depends on both presence and absence data, yet biodiversity databases are dominated by presence records, while true absences are rarely collected and geographically restricted. We introduce a framework that integrates large language models (LLMs) to sample ecologically realistic pseudo-absences under the constraints of dendritic river networks. Innovation: We present the Language-Grounded Multivariate Pseudo-Absence Sampling (LGMPAS) framework, which converts environmental predictor profiles into natural-language eco-narratives that an LLM uses to score and rank pre-filtered candidate locations drawn exclusively from the river network and restricted to unlabelled sites. Retrieval-Augmented Generation (RAG) further anchors eco-narratives in published ecological knowledge. We tested LGMPAS on two ecologically contrasting crayfish species in the Danube basin, the widespread invasive Faxonius limosus and the narrowly endemic Austropotamobius bihariensis, validating outputs against independent field-collected true-absence data using Random Forest predictive performance, spatial overlap of high-suitability areas and predictor-space distances. LLM-derived pseudo-absences closely reproduced true-absence model outputs and consistently outperformed random sampling across both species. Main Conclusions: LGMPAS demonstrates that LLMs can reliably sample pseudo-absences that reproduce the ecological signal of true-absence data, even under complex freshwater network constraints. By reducing dependence on costly absence surveys in contexts where true-absence data are unavailable or spatially restricted, and by avoiding the biases inherent to random sampling, the framework strengthens the robustness of SDMs for conservation applications. Its reproducibility and adaptability across taxa and ecosystems offer particular value for biodiversity monitoring, invasive species management and conservation planning under global change.

• Language: English
• Keywords: species distribution models, pseudo-absence sampling, large language models, ecological modelling, spatial ecology, freshwater ecosystems, crayfish, retrieval-augmented generation
• Work type: Article
• Typology: 1.01 - Original Scientific Article
• Organization: FRI - Faculty of Computer and Information Science
• Publication status: Published
• Publication version: Version of Record
• Year: 2026
• Number of pages: 13 str.
• Numbering: Vol. 32, iss. 5, art. e70199
• PID: 20.500.12556/RUL-183939
• UDC: 004.8:574
• ISSN on article: 1366-9516
• DOI: 10.1111/ddi.70199
• COBISS.SI-ID: 277720323
• Publication date in RUL: 22.06.2026

Record is a part of a journal

• Title: Diversity and distributions : a journal of conservation biogeography
• Shortened title: Divers. distrib.
• Publisher: Wiley
• ISSN: 1366-9516
• COBISS.SI-ID: 30709

Licences

• License: CC BY 4.0, Creative Commons Attribution 4.0 International
• Link: http://creativecommons.org/licenses/by/4.0/
• Description: This is the standard Creative Commons license that gives others maximum freedom to do what they want with the work as long as they credit the author.

Secondary language

• Language: Slovenian
• Keywords: modeliranje razširjenosti vrst, psevdo-odsotnosti, veliki jezikovni modeli, ekološko modeliranje, prostorska ekologija, sladkovodni ekosistemi, raki, priklicno obogatena generacija