Decoding Ecological Condition in Solomon Island Rivers with eDNA and Machine Learning

Environmental DNA (eDNA) metabarcoding offers a sensitive, non‑invasive alternative to conventional biomonitoring, yet its integration with predictive modelling remains underexplored. This study combines eDNA datasets collected during an ADB‑funded Strategic Environmental and Social Assessment (2024-2025) with supervised machine learning (ML) to evaluate ecological condition across eight rivers on Guadalcanal, Solomon Islands.

eDNA samples from 47 sites identified 861 taxa at species level, spanning native forest species, disturbance‑tolerant fishes, invasive plants and animals, and microbial indicators of pollution. Classifiers (decision tree, logistic regression, random forest, eXtreme gradient boosting) were trained on species profiles contextualised by land‑use intensity and buffer zone scales (100–1000 m). Stratified 5-fold cross-validation and permutation testing ensured robust evaluation. Random forest performed best, achieving AUCs up to 0.94 ± 0.04 (p < 0.001) and accuracy of 0.83 ± 0.05 at the 1000 m buffer, with significant AUCs (p < 0.05) across six additional model configurations. Broader buffers improved classification by integrating cumulative landscape disturbances such as logging, agriculture, and urban runoff.

Feature‑importance and SHAP analyses identified ecologically coherent predictors. Sentinel gobies (Sicyopterus stiphodonoides, Rhyacichthys guilberti) and native forest taxa (Fagraea berteroana, Nephrolepis biserrata) were indicative of reference conditions, while disturbance‑tolerant fishes (Giuris margaritaceus, Crenimugil cf. heterocheilos), invasive species (Rhinella marina, Lissachatina fulica, Megathyrsus maximus), and cultivated plants (Oryza sativa) signalled impacted sites. Some taxa showed context‑dependent behaviour: Intsia bijuga predicted both impact and reference depending on scale, reflecting its dual role as a native forest tree and a species under logging pressure, while Areca catechu shifted from reference at broader scales to impact at smaller ones, consistent with cultural cultivation and naturalisation. Unexpected detections (Piper methysticum, Diplarpea paleacea, Conium maculatum, Phyllanthus talbotii) highlight both the potential for novel biogeographic insights and the need for expanded eDNA reference libraries.

Our findings demonstrate that eDNA‑ML frameworks capture spatially coherent disturbance classifications with greater taxonomic resolution than conventional bioassessment. This approach provides a scalable, data-driven basis for freshwater management, enabling agencies to prioritise restoration, delineate scientifically defensible buffer zones, and monitor ecosystem change where field capacity is limited. The findings underscore the transformative potential of eDNA-ML integration for evidence-based river management and policy development in tropical island nations.