From October 2021 to July 2025, a comprehensive survey of amphibian fauna was conducted on 36 coastal islands in southern Zhejiang (Fig. 1), including Dongtou Island, Banping Island, and other islands. The survey period covered both breeding and non-breeding seasons of amphibians. Geographic information for the surveyed islands is provided in Suppl. material 1: table SS1. A combination of transect and point-count methods was applied. Based on the habitat features of the islands and the ecology of local amphibians, 54 transects and 20 observation points were established in total. Throughout the study, amphibians representing nine species, five families, and one order were recorded. A complete species list is available in Suppl. material 1: table SS2. Land cover data were obtained from the China Land Cover Dataset (CLCD) (Yang and Huang 2021), which includes nine major land cover types: cropland, forest, shrubland, grassland, water, snow/ice, barren land, impervious surfaces, and wetland. With a spatial resolution of 30 m, the dataset fully covers the coastal islands of southern Zhejiang.

In this study, species relative abundance (R) was used to identify dominant amphibian species, while alpha diversity was assessed using the Shannon-Wiener index (H’), the Simpson index (D), and the Pielou evenness index (J) to characterize species diversity within the amphibian communities (Strong 2016; Thukral 2017; Magurran 2021). The corresponding formulas are as follows:

R=NiN (1)

H′=−∑i=1SPiln⁡Pi (2)

D=1−∑i=1s(Pi)2 (3)

J=H′ln⁡S′ (4)

where N_i represents the number of individuals of species i, and N is the total number of individuals of all species. A species is classified as dominant if R > 0.05, as common if 0.005 < R ≤ 0.05, and as rare or occasional if R ≤ 0.005. Here, P_i denotes the proportion of individuals belonging to the ith species, and S refers to the total number of species.

In this study, multiple machine learning algorithms, including eXtreme gradient boosting (XGB; Grekousis 2025), support vector machines (SVM; Pouteau et al. 2012), K-nearest neighbors (KNN), and linear regression (LR), were employed to model and predict the relationships between island geographic features and biodiversity. The explanatory variables considered were distance to the nearest mainland (MID), land area (ARE), maximum elevation (ALT), coastline length (COAS), shape index (SAI), and settlement presence (SETT). The formula for calculating SAI is as follows:

SAI=COAS2ARE∗π (5)

The biodiversity response variables selected were the Shannon-Wiener index (H’), Simpson index (D), and Pielou evenness index (J). Model performance was evaluated using the mean squared error (MSE) and adjusted R², with the optimal model selected based on these regression metrics:

MSE =∑i=1n(yi−y^i)2n (6)

R2=1−(1−∑i(y^i−yi)2∑i(yi−y¯)2)×n−1n−p−1 (7)

where y_i represents the true value, y^i denotes the predicted value, ȳ is the mean of the observed values, n is the sample size, and p is the number of features. A lower MSE value indicates smaller fitting and prediction errors of the model. The R² value ranges from 0 to 1, with values closer to 1 indicating a greater proportion of variance explained by the model and better stability in performance.

Spearman correlation analysis was employed to assess the co-occurrence and associations among amphibian species across the islands (Gotelli and McCabe 2002). The calculation formula is as follows:

ρ=1−6∑di2n(n2−1) (8)

where d_i represents the difference in rank of the ith species, and n denotes the number of islands.

Cluster analysis in conjunction with the elbow method was applied to examine similarity in amphibian community composition among islands (Rousseeuw 1987; McKenna Jr 2003). Detrended correspondence analysis (DCA) was first conducted on the species distribution matrix. Based on the maximum gradient length of the DCA axes, redundancy analysis (RDA) was used if the value was less than 3, canonical correspondence analysis (CCA) was selected if it was greater than 4, and either RDA or CCA could be appropriately applied if it was between 3 and 4 (Palmer et al. 2008). In this study, as the maximum gradient length of all DCA axes for island geographic characteristics and land cover types was below 3, RDA was chosen to evaluate the effects of geographic features and land cover proportions on amphibian community composition.

To address potential multicollinearity among island geographic variables and land cover proportions, which could distort model outcomes (Dastres et al. 2025), variance inflation factor (VIF) analysis was used to quantify variable autocorrelation (Akinwande et al. 2015). Only variables with VIF values less than 10 were retained in the model. The categorical variable indicating the presence of settlement on islands was processed using one-hot encoding for appropriate feature representation (Enireddy et al. 2022).

To elucidate the specific influence of island geographic characteristics on amphibian species diversity, the optimal predictive model was interpreted using the SHAP method (Mo et al. 2023). This approach quantifies the contribution of each geographic feature to model predictions. The SHAP value is computed as follows:

∅i=∑S⊆N(i)"S"!("N"−"S"−1)!"N"!(v(S∪{i})−v(S)) (9)

where Ø_i denotes the SHAP value for feature i, N is the complete set of features, S represents any subset of features excluding i, |S| indicates the number of features in subset S, v(S) denotes the model output using feature subset S, and v(S U {i}) is the model output when feature i is included. A positive SHAP value indicates a positive effect of the geographic feature on biodiversity prediction, whereas a negative value suggests an inhibitory effect (Wang et al. 2024).

Candidate thresholds for SHAP values were generated using percentiles ranging from the 10th to the 90th percentile, in 5% increments. For each candidate threshold, a Mann-Whitney U test was performed to evaluate statistical significance. The threshold with the smallest P-value, i.e., the most statistically significant, was selected as the final cutoff (Qiao et al. 2025). The SHAP method not only quantifies the overall magnitude of influence exerted by island geographic features but also provides a detailed, island-specific interpretation of how each geographic characteristic affects biodiversity.

A total of 21 islands in the coastal area of southern Zhejiang, including Lingkun Island, Damen Island, and Dongtou Island, were observed to host two or more amphibian species. Among these, Duttaphrynus melanostictus (Schneider, 1799) (S2), Fejervarya multistriata (Hallowell, 1860) (S4) and Hylarana guentheri (Boulenger, 1882) (S6) were identified as the dominant species, with relative abundances of 0.371, 0.327 and 0.177, respectively (Table 1). The highest H’ values were recorded on Dongtou Island (1.60), Damen Island (1.48), Lingkun Island (1.43), and Nanji Island (1.39) (Table 2).

Dominant and common amphibian species on these islands share similar ecological environments (Fig. 2). The distribution of Duttaphrynus melanostictus (S2) showed significant correlations with those of Fejervarya multistriata (S4) and Hylarana guentheri (S6), with correlation coefficients of 0.65 and 0.72, respectively (P < 0.05). A strong correlation of 0.82 (P < 0.05) was also observed between the distributions of Fejervarya multistriata and Hylarana guentheri.

The elbow method indicated that the amphibian compositions of the islands in southern Zhejiang could be clustered into two distinct groups (Suppl. material 1: fig. S1). Islands including Lingkun, Damen, Dongtou, Niyu, Beiji, and Luxi shared similar amphibian community structures (Fig. 3). The VIF value for COAS was 65.39, indicating potential strong collinearity with other island features. After removing this variable, the VIF values of all remaining geographic features were below 10 (Suppl. material 1: table S3), confirming the absence of multicollinearity, and these variables were therefore retained as explanatory factors in the model.

When island geographic features were considered influencing factors, the environmental variables significantly explained the variation in amphibian community structure (P < 0.05). Variation partitioning analysis showed that geographic features accounted for 43.2% of the total explained variance, highlighting their role as major drivers of community variation (Suppl. material 1: fig. S2). Among these, MID and SETT exerted significant effects on community differentiation (P < 0.05; Suppl. material 1: table S4).

For the RDA analysis with land cover type proportions as predictors, impervious surface cover was excluded due to perfect collinearity. The remaining land cover variables showed no multicollinearity (Suppl. material 1: table S5) and significantly explained the differences in amphibian community structure (P < 0.05; Suppl. material 1: fig. S3). Both water coverage and farmland coverage ratio had significant impacts on the amphibian communities (P < 0.05; Suppl. material 1: table S6), with land cover types collectively explaining 30.8% of the variance (Suppl. material 1: fig. S3).

The XGB model achieved the highest goodness of fit for H', D, and J, with mean R² values of 0.97, 0.95, and 0.95, respectively. When the XGB model was selected, it exhibited lower prediction error and greater stability in cross-validation compared to other models. The predictive performance, expressed as R² values, was 0.42 ± 0.07 for H', 0.38 ± 0.07 for D, and 0.27 ± 0.07 for J, respectively (Suppl. material 1: fig. S4).

SHAP analysis revealed that, among the island geographic features influencing the diversity index H’ in the coastal islands of southern Zhejiang, the predictive importance ranking was as follows: ARE, SETT_0, SAI, MID, ALT, and SETT_1 (Fig. 4). The direction of influence varied across features: high ARE and high SAI both exhibited positive effects on the diversity indices H’, D, and J, whereas an increase in MID showed inhibitory effects on all three. ALT and SETT had the smallest impact on diversity predictions (Fig. 4).

Given the consistent model performance and superior predictive accuracy for H', further SHAP analysis was conducted for Type II islands to assess feature impacts in detail, including Lingkun, Damen, Dongtou, Niyu, Luxi, and Beiji. ARE strongly promoted H’ on the largest islands, namely Lingkun, Damen, and Dongtou, with SHAP values of 0.57, 0.52, and 0.52, respectively (Fig. 5). In contrast, MID exerted inhibitory effects on H’ for Beiji Island and Luxi Island, with SHAP values of −0.2 and −0.01, respectively (Fig. 5).

SHAP threshold analysis identified the following thresholds: ARE at 0.18, MID at 23.88, SAI at 2.20 (P > 0.05), and ALT at 82.90 (Fig. 6). Significant differences in SHAP values were observed before and after the thresholds for ARE, MID, and ALT (P < 0.05), while SAI exhibited no significant effect. In addition, the larger the MID, the lower the SHAP value of the islands in the group (Suppl. material 1: fig. S5A, C). The SHAP value on islands inhabited by humans was higher than that on uninhabited islands (Suppl. material 1: fig. S5B, D).

Amphibians are widely distributed across the coastal islands of southern Zhejiang. This study documented a total of nine species from five families and one order. Cluster analysis classified the amphibian communities in this region into two distinct types. Type I communities are dominated by single or multiple combinations of Duttaphrynus melanostictus, Fejervarya multistriata, and Hylarana guentheri, with community structure primarily influenced by geographic features. Type II consists of six larger islands, including Damen Island and Luxi Island. These islands support amphibian communities also centered on the three dominant species, accompanied by secondary species, and their distribution shows strong correlations with human activities and island area.

Dominant species exhibit considerable plasticity in adapting to environmental changes. With broad ecological niches, they efficiently utilize human-altered secondary habitats, including irrigation ditches, rice paddies, and artificial water bodies. Under disturbances such as road construction, settlement expansion, and agricultural development, amphibians exhibit behavioral adjustments, for example, shifting breeding phenology (Carter and Rudolf 2019) and switching habitats (Martínez-Abraín and Galán 2018), which help maintain population stability and mitigate effects on alpha diversity. This adaptability enables them to sustain niche dominance on islands experiencing moderate disturbance. However, the high spatial association among dominant amphibian species in southern Zhejiang contradicts the principle of competitive exclusion. This study proposes that their coexistence can be attributed to (1) habitat heterogeneity resulting from human activities such as farmland reclamation and artificial water body construction (Chen et al. 2024), which facilitates fine-scale spatial and resource niche partitioning among dominant species and reduces direct competition; and (2) temporal differentiation in resource use across different life history stages, indirectly mitigating interspecific competition (Gao et al. 2020).

The low alpha diversity of amphibians on small islands in southern Zhejiang can be mainly explained by the following factors: (1) amphibians have limited dispersal capacity and high environmental dependence, resulting in more restricted distributions compared to other vertebrates; (2) human activities and urbanization pose severe threats to amphibian habitats. Island ecosystems are more susceptible to the combined effects of geographic constraints and anthropogenic disturbances than mainland areas; and (3) small islands consist largely of rocky substrates with low habitat heterogeneity. They lack natural water retention capacity, and bare rocks support low primary productivity, leading to limited food resources for amphibians.

According to niche theory, environmental filtering is a major factor influencing differences in species composition within communities (Kristiansen et al. 2012). This study demonstrates that island geographic features significantly affect both amphibian diversity and community structure (P < 0.05). In particular, human settlement (SETT) has a notable impact on community differentiation (P < 0.05), though its direct effect on alpha diversity is minimal. This seemingly counterintuitive result can be interpreted through the “species screening effect” (Dettki et al. 1998). The amphibian communities in the study area exhibit a dual-mode response to human disturbance: as a strong environmental filter (Zhao et al. 2024), SETT rapidly selects against specialist species dependent on pristine habitats, e.g., certain rhacophorid frogs, leading to community succession. Conversely, generalist species, owing to their high ecological plasticity (Náplavová et al. 2021), efficiently utilize human-altered habitats and maintain stable populations on islands with high and moderate levels of development. Competition among these species has not yet reached a level that results in competitive exclusion, and thus alpha diversity does not show a sharp decline. Nonetheless, at the regional scale, this process contributes to community homogenization (Montesinos et al. 2012).

Islands that are larger, less isolated, and have diverse habitat types, e.g., Niyu Island, Dongtou Island, and Lingkun Island, retain a higher proportion of endemic species such as Polypedates megacephalus. In contrast, highly isolated islands such as Nanji Island support endemic aquatic frogs such as Hylarana latouchii. Consequently, amphibian communities on islands are structured primarily through environmental filtering, which takes precedence over competitive exclusion. Since competitive processes among amphibians have not fully unfolded, changes in diversity lag behind community turnover, driving regional reassembly of amphibian communities on coastal islands.

ARE and MID are the primary geographic features influencing diversity (Suppl. material 1: fig. S6). Larger islands provide diverse survival conditions and opportunities for adaptation, facilitate niche differentiation (Lustenhouwer and Riddell 2025), and possess greater resource availability, thereby supporting higher species richness. SHAP analysis indicates that amphibian diversity increases sharply when island area exceeds 0.18 km², suggesting that this value may represent the minimum area threshold for maintaining diversity. Islands below this threshold fail to sustain essential static freshwater breeding habitats and food web integrity, leading to diversity collapse. This finding offers a reference for defining minimum area requirements in island ecological restoration. Furthermore, large islands often contain natural or artificial open water bodies such as reservoirs and streams, which provide improved conditions for amphibian reproduction and dispersal. Although larger islands are theoretically associated with higher amphibian diversity, high connectivity can conversely reduce diversity. For example, Lingkun Island, which is highly developed and experiences substantial human disturbance, demonstrates lower species diversity compared to other islands of similar size. The negative effects of human activities appear to override the positive effects typically associated with larger areas. Road construction on large islands may further diminish predicted species richness, weakening the expected species-area relationship. Currently, the constructed wetland park on Lingkun Island helps protect endemic species such as Polypedates megacephalus and Pelophylax nigromaculatus.

ALT typically influences temperature and humidity, thereby creating distinct ecological niches. However, the coastal islands of southern Zhejiang are generally low in elevation, resulting in environmental gradients that are too narrow to trigger significant niche differentiation. SHAP analysis confirms that ALT contributes minimally to alpha diversity, as the temperature variations associated with altitude differences fall well within the thermal tolerance range of local amphibians. The weak effect of ALT is overshadowed by the habitat heterogeneity driven primarily by ARE. Larger islands encompass both higher elevation areas and lowland water bodies, suggesting that the distribution of freshwater habitats, rather than altitude itself, is the key driver of amphibian diversity.

Machine learning has become an important tool in biogeographical research and can be applied to predict species richness and diversity in designated areas (Andermann et al. 2022). SHAP can decompose complex “black box” model fitting or prediction results into the contribution values of each feature, and the introduction of this method provides convenience for understanding how input features affect species diversity (Riddley et al. 2025). This study used XGB, SVM, KNN, and LR as modeling methods to select the most suitable model to explain the diversity of amphibians in the coastal islands of southern Zhejiang and its relationship with geographic features. The SHAP method was applied to further quantify the relationship between island geographic features and amphibian diversity. Due to the possible strong correlation between geographic features, VIF analysis was used to analyze the autocorrelation of each geographic feature. The VIF value of COAS was found to be too high and strongly correlated with ARE and SAI. Therefore, to establish a prediction model, COAS needed to be removed.

In the case of a small sample size, due to the strong impact of sampling methods on fitting and prediction results, 20 cross-validations were used to compare model performance, aiming to reduce the interference of randomness in the modeling process on prediction results. Through model fitting and cross-validation, XGB was found to have better fitting and cross-validation results compared to other models. This is because XGB introduces a regularization term, randomly samples smaller subdatasets from coastal islands, and independently trains decision trees to aggregate the prediction results of multiple weak learners to obtain the final output. This type of algorithm can effectively reduce the variance of the overall model and lower the risk of overfitting. In addition, compared to linear models, tree models can more clearly express the interaction between explanatory variables and response variables (Gao and Wang 2023).