Trait-environment diagnosis of ecological instability in Korean streams using benthic diatoms and machine learning: a UMAP-CDI framework

Abstract

This study presents a trait-integrated, multistressor-sensitive bioassessment framework that combines the Community Dynamics Index (CDI) and machine learning-based nonlinear ordination to quantify ecological instability in monsoon-impacted river systems. Using data encompassing benthic diatom species composition of 457 stream sites across five major Korean river basins (2013-2015), CDI values were calculated between pairs of sampling periods to capture both monsoon-related and longer-term community shifts. Among four ordination methods tested, uniform manifold approximation and projection (UMAP) most effectively resolved nonlinear transitions, outperforming principal component analysis (PCA) and nonmetric multidimensional scaling (NMDS) in silhouette width. Random forest modeling identified motility as a dominant predictor for UMAP-based CDI with monsoonal and non-monsoonal variations. Partial dependence plots indicated that changes in motility were strongly associated with variation in CDI during both monsoonal and non-monsoonal periods, with pronounced nonlinear responses across environmental gradients and distinct season-specific patterns. Functional analyses further indicated that this short-term instability was accompanied by a long-term shift from sensitive, low-profile taxa (e.g., Achnanthes spp.) to tolerant, motile taxa (e.g., Nitzschia or Navicula spp.). This integrated CDI-trait-UMAP approach enables scalable, ecologically interpretable, and nonlinearity-resolving assessment of river health. The framework can be extended to support international biomonitoring goals and offer early warning capacity for biodiversity loss under intensifying climatic and anthropogenic stressors.