Work

Habitat Inference from Biodiversity Field Notes

Year: 2026

Type: conference-paper

Abstract: Systematic land-cover products derived from remote sensing extend back only a few decades, leaving major gaps for earlier periods. In contrast, biodiversity occurrence repositories provide vast numbers of geo-referenced records spanning much longer time horizons, and many records include free-text habitat notes describing local conditions. This motivates habitat inference from text as a scalable route to extending habitat knowledge before the satellite era. However, this problem has thus far received limited attention and is technically challenging. To support work on this topic, we introduce a benchmark that aggregates habitat-note text into spatial grid-cell documents with proxy supervision from land-cover maps, and we present an extensive analysis of text-based predictors and graph-enriched classification pipelines for inferring standard habitat classes. Among others, we find that LLM encoders struggle with the noisy nature of the input texts and the subtle distinctions between the labels, while taking into account geographic proximity and environmental similarity can significantly boost results.

Source: ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)

DOI: 10.1145/3770855.3818941

Authors: Mahsa Hadikhah Mozhdehi, Steven Schockaert, Sarah E. Perkins, Christopher B. Jones

Institution: Cardiff University

Language: English

Topic: Remote-Sensing Image Classification

Subfield: Media Technology

Field: Engineering

Domain: Physical Sciences

SDG: Life on land

Open Access status: hybrid

Funder: Cardiff University