Island area and isolation correlate with amphibian diversity in 36-island Zhejiang study

A field survey and interpretable machine-learning analysis of 36 coastal islands in southern Zhejiang, China, found that island area and distance from the mainland were major correlates of amphibian diversity. The researchers documented nine species, identified two broad community groups and found that larger islands tended to have higher model-predicted diversity, while greater isolation was associated with lower predicted values.

The study surveyed the islands from October 2021 through July 2025, covering both amphibian breeding and non-breeding seasons. Researchers established 54 transects and 20 observation points, with sampling methods adapted to island habitats and the ecology of local amphibians. Land-cover data came from the China Land Cover Dataset, which has a spatial resolution of 30 metres.

Twenty-one islands contained two or more recorded amphibian species. The most abundant species were Duttaphrynus melanostictus, with a relative abundance of 0.371; Fejervarya multistriata, at 0.327; and Hylarana guentheri, at 0.177. The highest Shannon-Wiener diversity values were reported on Dongtou Island at 1.60, Damen Island at 1.48, Lingkun Island at 1.43 and Nanji Island at 1.39. These diversity indices describe the sampled communities, while individual zero values for some islands do not by themselves establish that amphibians were absent.

Spearman correlation analysis identified significant positive associations among the three dominant species. The correlation coefficients were 0.65 for D. melanostictus with F. multistriata, 0.72 for D. melanostictus with H. guentheri, and 0.82 for F. multistriata with H. guentheri, each with P < 0.05. Elbow-method clustering separated the islands into two community groups. One included Lingkun, Damen, Dongtou, Niyu, Beiji and Luxi, whose assemblages contained the dominant species together with additional species.

Redundancy analysis showed that island geographic characteristics were significantly associated with differences in community composition at P < 0.05. Variation partitioning attributed 43.2% of the total explained variance to geographic features, with distance from the nearest mainland and settlement presence significantly associated with community differentiation. In a separate analysis using land-cover proportions, water cover and farmland cover were significantly associated with community composition, and the land-cover variables collectively accounted for 30.8% of the variance reported by that analysis. Impervious-surface cover was excluded because of perfect collinearity.

The researchers compared extreme gradient boosting, support vector machines, K-nearest neighbours and linear regression for the Shannon-Wiener index, Simpson index and Pielou evenness index. Extreme gradient boosting had the strongest goodness of fit, with mean R-squared values of 0.97, 0.95 and 0.95, respectively. Cross-validated performance was more modest: R-squared was 0.42 ± 0.07 for Shannon-Wiener diversity, 0.38 ± 0.07 for Simpson diversity and 0.27 ± 0.07 for evenness. Variance-inflation screening removed coastline length, whose value of 65.39 indicated strong collinearity with other geographic variables.

Shapley Additive Explanations, or SHAP, was then used to interpret the selected models. For the Shannon-Wiener index, the geographic predictors were ranked as land area, absence of settlement, shape index, distance from the nearest mainland, maximum elevation and settlement presence. The analysis associated greater land area and higher shape-index values with higher predicted values of all three diversity measures, and greater distance from the mainland with lower predictions. Maximum elevation and settlement status had comparatively small effects. These are model-based associations rather than experimental estimates of causal effects.

Threshold analysis identified cutoffs of 0.18 square kilometres for land area, 23.88 for distance from the nearest mainland, 2.20 for the shape index and 82.90 for maximum elevation; the supplied text did not specify units for the last three values. Mann-Whitney tests found significant changes in SHAP values around the area, isolation and elevation thresholds at P < 0.05, but not around the shape-index threshold. The area result marks a change in the model’s predictions above 0.18 square kilometres within this 36-island dataset; it does not by itself demonstrate that diversity collapses below that value.

Settlement presence was associated with community differentiation but had little direct association with the alpha-diversity measures. The authors interpreted this as a species-screening effect: human disturbance may filter out specialists dependent on relatively undisturbed habitats while favouring generalists able to use irrigation ditches, rice paddies and artificial water bodies. They proposed that such filtering can contribute to community homogenisation without an immediate sharp reduction in alpha diversity. The observations also did not directly test extinction, competition, resource partitioning or other proposed causal mechanisms.

The study was observational and limited to 36 islands, selected geographic and land-cover variables, and a survey period extending from 2021 to 2025. Its models had limited cross-validated predictive performance, especially for Pielou evenness, and the analysis did not incorporate climate change or stochastic events. These boundaries constrain the findings to associations within the surveyed islands rather than universal habitat requirements or predictions for all coastal-island amphibian communities.

Subject of Research: Amphibian diversity and geographic correlates on coastal islands in southern Zhejiang

Article Title: Analysis of amphibian diversity and its influencing factors in the coastal islands of southern Zhejiang

Article References: Qi, W., Li, J., Qin, C., Xie, L., Li, S., & Zhang, Y. (2026). Analysis of amphibian diversity and its influencing factors in the coastal islands of southern Zhejiang. Nature Conservation, 64, 59-76. https://doi.org/10.3897/natureconservation.64.183777

Image Credits: AI Generated

DOI: 10.3897/natureconservation.64.183777

Keywords: amphibians, island biogeography, biodiversity, machine learning, SHAP, southern Zhejiang, coastal islands, land cover, conservation, Analysis, amphibian, diversity