Skip to main navigation Skip to search Skip to main content

Prevalence- and algorithm-dependent effects of resampling on prediction and interpretation in ecological niche models

  • Dae Seong Lee
  • , Da Yeong Lee
  • , Sagar Adhurya
  • , Jihoon Shin
  • , Young Woo Kim
  • , Dong Won Shim
  • , Yoon Kyung Cha
  • , Ihn Sil Kwak
  • , Young Seuk Park

Research output: Contribution to journal › Article › peer-review

Abstract

The reliability of ecological niche models (ENMs) depends on the quality and representativeness of training data, which are often compromised by class imbalance in ecological datasets. This study evaluated how resampling strategies and training prevalence affect ENMs developed under class imbalance. Using freshwater benthic macroinvertebrate data from Korea, we compared non-resampled Control models with six resampling methods: random oversampling, synthetic minority oversampling technique (SMOTE), borderline SMOTE, adaptive synthetic sampling, density-based SMOTE, and random undersampling. These methods were tested across controlled prevalence scenarios and five ENM variants: maximum entropy (MaxEnt), random forest classifiers and regressors, and support vector machine classifiers and regressors. Resampling effects were strongest under severe-to-moderate class imbalance and diminished as training prevalence increased. Synthetic oversampling methods did not always generate the intended number of presence samples, and resampling-induced distributional shifts were largest at very low prevalence. Predictive gains were algorithm- and metric-dependent: MaxEnt showed little improvement, random forest models improved mainly in F1 and true skill statistic (TSS) with random undersampling, and support vector machine models benefited most consistently from SMOTE-based methods. Resampling also changed selected variables, variable-importance profiles, and accumulated local effect curves, particularly at low prevalence. These findings show that resampling is a context-dependent modeling decision that should be evaluated within the full ENM pipeline and selected according to the imbalance structure of the data and the predictive or ecological objectives of the study.

Original languageEnglish
Article number103920
JournalEcological Informatics
Volume97
DOIs
Publication statusPublished - Aug 2026

Bibliographical note

Publisher Copyright:
Copyright © 2026. Published by Elsevier B.V.

UN SDGs

This output contributes to the following UN Sustainable Development Goals (SDGs)

  1. SDG 15 - Life on Land
    SDG 15 Life on Land

Keywords

  • Data augmentation
  • Data imbalance
  • Ecological data
  • Machine learning
  • Species distribution model

Fingerprint

Dive into the research topics of 'Prevalence- and algorithm-dependent effects of resampling on prediction and interpretation in ecological niche models'. Together they form a unique fingerprint.

Cite this