Assessing habitat suitability of protected migratory birds in coastal wetlands with multi-source data and a probability-averaging ensemble

Introduction: Habitat-suitability modelling supports conservation planning for protected migratory birds in dynamic coastal wetlands, yet single species distribution models (SDMs) can be unstable when presence records are scarce and class imbalance is strong. Here we present a dual-model probability averaging (DMPA) framework that ensembles two standard SDMs--logistic regression and random forest--by simply averaging their predicted occurrence probabilities to improve robustness.

Methods: We apply the framework to the Yancheng coastal wetlands (eastern China) using a pooled presence-background dataset comprising 18 bird species (56 presence records) and multi-source climatic, topographic, and distance-based predictors, with covariates screened for collinearity (|r| > 0.95) and missing values imputed by variable means. Model performance is assessed using cross-validation with held-out predictions, and binary suitability maps are derived using an F1-based operating threshold selected across folds.

Results: Quantitatively, the DMPA ensemble achieves strong discrimination (ROC-AUC = 0.899; PR-AUC = 0.617) and substantially improves classification performance relative to single models (F1 = 0.643 vs. 0.474 for logistic regression and 0.034 for random forest, which collapses under F1-based thresholding due to extreme class imbalance), while maintaining competitive probabilistic accuracy (Brier = 0.036, compared with 0.057 and 0.034) and moderate calibration (ECE = 0.061, compared with 0.082 for logistic regression). Spatial projections concentrate higher suitability along the coastal wetland corridor, and feature-importance analysis highlights distance to coastline/rivers and key bioclimatic variables as leading predictors.

Discussion: Overall, DMPA provides a simple and practical ensemble strategy that improves PR-AUC and F1 under class imbalance without sacrificing overall discrimination, supporting suitability screening and mapping in fast-changing coastal wetlands.

1 Introduction

Coastal wetlands are key stopover and wintering habitats for migratory birds, and they are also among the most rapidly changing ecosystems on Earth. Key nonbreeding waterbird wetlands are increasingly exposed to sea-level rise (Verniest et al., 2024). Global assessments have shown that more than one-third of the world’s wetlands have been lost since 1970, and the remaining ones are under severe pressure from reclamation and climate change (Ramsar Convention on Wetlands, 2018). Land reclamation, aquaculture, and infrastructure development have converted or degraded large areas of tidal flats and salt marshes. These changes can quickly alter the availability and configuration of suitable habitat for waterbirds. For protected migratory species, managers therefore need spatially explicit information on where suitable habitat currently exists and how it is distributed across the landscape. Recent assessments of waterbird habitats in the Jiangsu coastal wetlands have quantified habitat suitability and identified tidal-flat extent, saltmarsh coverage, and hydrological connectivity as key determinants of spatial patterns (Wang et al., 2025).

Species distribution models (SDMs) are now widely used to map habitat suitability and support conservation planning (Elith and Leathwick, 2009). By relating species occurrence records to environmental variables, SDMs can predict the probability of occurrence across unsampled locations. Many algorithms have been applied in this context, including generalized linear models, random forests, boosted trees, and Maxent. These methods have helped identify key habitats for threatened waterbirds in several coastal regions. However, their predictions can still be unstable when data are limited or when environmental conditions vary strongly in space and time.

A common way to improve robustness is to combine several models into an ensemble. Traditional ensemble SDMs often average or stack predictions from multiple algorithms to reduce individual model bias and variance. In practice, these ensembles are usually evaluated with a small set of discrimination metrics such as the area under the receiver operating characteristic curve (AUC) or the true skill statistic (TSS). While useful, these metrics do not reveal whether predicted probabilities are well calibrated, nor do they directly show how models differ under alternative validation schemes, for example, when tested in new years or new spatial blocks.

For migratory birds in dynamic coastal wetlands, both discrimination and calibration are important. Conservation decisions, such as defining core protection zones or assessing habitat loss, depend not only on where the model predicts high suitability, but also on how reliable these probabilities are. Yet few SDM studies on coastal waterbirds have systematically compared single models and ensembles using calibration-aware metrics, or examined how model performance changes under temporal and spatial validation.

In this study, we develop and test a simple dual-model probability averaging (DMPA) framework for habitat suitability modeling of migratory birds in the Yancheng coastal wetlands, eastern China. The framework combines logistic regression and random forest models, fitted to multi-source environmental predictors including climate, topography, and distance-based variables. We use cross-validated predictions, bootstrap confidence intervals, and a set of discrimination and calibration metrics to compare the ensemble with its component models under different validation strategies. Our aim is to provide a practical and interpretable modeling workflow that produces reliable habitat suitability maps for conservation planning in rapidly changing coastal landscapes. Previous work in the same region has modelled waterbird habitat suitability and potential biological corridors using MaxEnt and circuit theory, revealing fragmented but high-value habitat patches and underlining the need for integrative modelling approaches to guide regional conservation (Sun et al., 2023).

2 Related work

Coastal-wetland habitat mapping has become a central application of remote sensing and species distribution models (SDMs) in conservation planning. Global assessments show that coastal wetlands are experiencing rapid morphological change and habitat loss under the combined influence of human disturbance and sea-level rise (Murray et al., 2019; Schuerch et al., 2018). Advances in satellite remote sensing have greatly expanded the ability of SDMs to represent such dynamic environments, providing fine-scale predictors of vegetation, inundation and salinity (He et al., 2015; Pettorelli et al., 2014). Classical SDMs link species occurrences to environmental covariates and then project predicted suitability over space, supporting tasks such as reserve design, impact assessment and habitat restoration. However, coastal systems are characterised by strong spatio-temporal heterogeneity—tidal cycles, salinity gradients, sharp land-cover transitions and intense human disturbance—which can make single-model predictions unstable and sensitive to data partitioning, class imbalance and threshold choice. Previous studies have highlighted that appropriate threshold selection and cross-validation design are critical for avoiding over-optimistic performance estimates in SDMs (Liu et al., 2005; Roberts et al., 2017; Valavi et al., 2018). These challenges have motivated two complementary lines of research: refining individual SDMs through transparent preprocessing and validation, and developing ensemble approaches that combine multiple models to stabilise predictions and better capture uncertainty.

2.1 Single-model SDMs in habitat-suitability mapping

Logistic-regression SDMs remain a widely used statistical baseline for habitat-suitability modelling. They offer coefficient-level interpretability and stable ranking of sites, which is valuable for communicating the direction and relative importance of environmental drivers to managers and stakeholders. In coastal wetlands, logistic SDMs have been applied to a range of waterbird and shorebird species, demonstrating that simple linear responses to climatic and land-cover variables can already recover much of the broad-scale habitat gradient. More broadly, logistic regression has also been used for habitat suitability mapping in other taxa based on long-term monitoring data; for example, Townsend and Aldstadt (2023) modelled five bat species using a 2009–2021 bioacoustic survey dataset and produced catchment-scale suitability maps. Nonetheless, the linear log-odds formulation limits the ability to represent non-linear responses, interactions and thresholds—for example, abrupt changes in suitability across tidal–vegetation ecotones or tolerance limits to inundation. As a result, logistic models tend to smooth over narrow high-suitability belts and may under-represent fine-scale structure at the land–sea interface.

Tree-based machine-learning methods, particularly random forests, address some of these limitations by flexibly capturing non-linearities and high-order interactions without strong distributional assumptions. Random-forest SDMs have shown strong discrimination performance in heterogeneous landscapes and are now commonly used with multi-source predictors that combine bioclimatic, topographic and proximity-based variables. Its effectiveness has been demonstrated in the same Yancheng coastal wetlands for vegetation classification using multi-temporal Sentinel-2 imagery (Wang et al., 2024). Similarly, random forest has been successfully applied to model habitat suitability for diverse waterbird guilds in the same region, identifying key drivers such as habitat type and vegetation cover (Wang et al., 2020). However, their flexibility can also produce locally over-confident predictions: in sparsely sampled or highly heterogeneous neighbourhoods, random forests may generate speckled high-probability pixels that fragment suitable patches once thresholded. In addition, their probabilistic calibration is often imperfect, and the stochastic training procedure can introduce modest variability in predicted patch configurations, complicating precise spatial planning.

Overall, logistic-regression and random-forest SDMs exhibit complementary strengths and weaknesses. Logistic models provide transparent, stable rankings but can underfit complex coastal gradients, whereas random forests improve discrimination but are prone to spatial speckling and over-fitting in data-poor regions (Chiaverini et al., 2023). This complementarity motivates combining both model families rather than selecting a single “best” SDM. This need is particularly acute in our study area, the Yancheng coastal wetlands, where the overall ecosystem stability has been assessed to be in a warning state under significant anthropogenic and environmental pressures (Tian et al., 2022).

2.2 Ensemble modelling for coastal-wetland habitat evaluation

Ensemble forecasting has become a widely accepted strategy for species-distribution prediction, reducing individual model bias and variance by combining multiple algorithms (Araújo and New, 2007). Platforms such as BIOMOD have further standardised these approaches and facilitated their application in ecological research (Thuiller et al., 2009). Previous work has shown that ensembles often outperform single models in both species-level and community-level predictions, especially under environmental heterogeneity and limited sample sizes. Typical strategies include majority voting, rank-based weighting, probability averaging and more complex meta-learning or stacking frameworks. Recent work on habitat-quality assessment further shows that multi-model fusion (e.g., stacking) can improve predictive accuracy and robustness over single learners, delivering higher precision and recall in applied ecological mapping tasks (Yang et al., 2025). Software platforms such as BIOMOD have made these approaches accessible and have helped establish ensemble SDMs as a standard practice in many ecological applications. The resulting ensemble suitability maps can further support spatial conservation planning, such as identifying priority habitats and designing ecological corridors, as demonstrated in freshwater wetlands for wintering crane conservation (Wei et al., 2023). Empirical comparisons of ensemble strategies further confirm their consistent superiority over single-model SDMs (El Alaoui and Idri, 2024).

At the same time, there is growing recognition that ensembles used for conservation planning should remain interpretable and auditable. Highly complex stacking or deep-learning architectures can blur the link between environmental drivers and predicted suitability, making it difficult to explain or defend management decisions based on their outputs. For coastal wetlands in particular, managers often require not only high discrimination but also spatially coherent maps, calibrated probabilities and transparent decision rules for converting continuous suitability into binary habitat classifications. Similar needs and benefits of ensemble modeling have been demonstrated in other ecosystems, such as in prioritizing conservation for the lesser prairie-chicken using ensemble habitat evaluations (Solomon et al., 2025).

In this context, we adopt a deliberately simple dual-model probability averaging (DMPA) ensemble that combines a logistic-regression SDM and a random-forest SDM at the probability level. By averaging their predicted occurrence probabilities for each location, the ensemble leverages the interpretability and ranking stability of the logistic model together with the non-linear response flexibility of the random forest, while avoiding the additional complexity of a stacked meta-learner. The remainder of this paper evaluates how such a minimal, probability-level ensemble performs under realistic validation schemes and whether it improves discrimination, calibration and spatial coherence for habitat-suitability mapping in dynamic coastal wetlands.

3 Materials and methodology

This section outlines the dual-model probability averaging (DMPA) habitat-suitability modelling framework (Figure 1), which integrates two base SDMs (logistic regression and random forest) via probability averaging to generate robust probabilistic predictions. The workflow proceeds from data assembly and predictor preprocessing (collinearity filtering, mean imputation, and standardisation for logistic regression), to model training and validation under one of three alternative designs (temporal hold-out when year information is available, spatial block cross-validation, or repeated stratified cross-validation), and finally to DMPA ensemble prediction with F1-based thresholding for binary mapping and diagnostic evaluation.

Figure 1

Figure 1. Overall technical roadmap of the DMPA habitat-suitability modelling framework. Species samples and environmental predictors are preprocessed (collinearity filtering, mean imputation, and feature-matrix preparation) and used to train two base SDMs (logistic regression and random forest) to produce probabilistic predictions. Model performance is evaluated using repeated stratified cross-validation with out-of-fold predictions, followed by DMPA probability averaging and F1-based thresholding to generate ROC/PR and calibration (ECE) diagnostics and habitat-suitability maps.

The remainder of this section follows the roadmap in Figure 1: Section 3.2 describes the study area and species-occurrence data; Section 3.3 details environmental predictors and preprocessing; Section 3.4 introduces the two single-model SDMs and their formulations; and Section 3.5 presents the ensemble integration, validation design, threshold selection, and evaluation metrics.

3.1 Study area and species occurrence data

The study focuses on the Yancheng coastal wetlands in Jiangsu Province, eastern China (32°34′–34°28′N, 119°48′–121°05′E). This coastal wetland system comprises tidal flats, salt marshes, and aquaculture ponds, providing important stopover and wintering habitats along the East Asian–Australasian Flyway (Figure 2).

Figure 2

Figure 2. Study area and occurrence locations in the Yancheng coastal wetlands, eastern China. The grey polygon delineates the extent of the coastal wetland study area, while red points show the pooled occurrence records of protected migratory birds used in the modelling. The inset map indicates the location of Yancheng within China.

Species-occurrence records were compiled from public biodiversity databases and local monitoring datasets. After removing duplicate records and invalid coordinates, the pooled dataset contained 18 bird species and 56 presence records (Table 1). Background samples were randomly generated within the study area to represent available environmental conditions, and each sample was labelled as presence $=1$ or background $=0$. The resulting occurrence–background dataset was then paired with environmental predictors for subsequent modelling. We adopted a pooled-species modelling strategy to represent general habitat suitability for protected migratory birds rather than species-specific niches. Given the extremely limited number of presence records per species, species-level SDMs would be statistically unstable. Pooling species allows the model to capture shared habitat constraints common to coastal wetland birds, which is appropriate for screening-level conservation planning.

Table 1

Table 1. Bird species and number of presence records (N) included in the pooled occurrence dataset.

3.2 Environmental predictors and preprocessing

A suite of multi-source predictors was assembled to characterise climatic, topographic, and proximity conditions in the Yancheng wetlands. Climatic variables (e.g., temperature and precipitation) were derived from public climate datasets; topographic variables (elevation, slope, and aspect) from digital elevation models; and proximity variables (distance to coastlines, rivers, and roads) from high-resolution geographic layers. All predictor layers were harmonised to a common spatial resolution and projection prior to extraction. Recent studies have shown that integrating multi-source geospatial datasets—combining high-resolution remote sensing, topography, and climate layers—can greatly enhance the accuracy and reliability of ecological modelling inputs, ensuring consistent spatial representation across variables (Lei and Wang, 2024).

Predictors were then screened and preprocessed before modelling. Highly correlated variables were removed using a pairwise Pearson correlation threshold of |r| >0.95 to reduce multicollinearity. Missing values were imputed using variable means. Predictors for the logistic-regression SDM were further standardised (z-score), whereas the random-forest SDM used unstandardised predictor values. The resulting predictor set was used to construct the final modelling dataset for subsequent analyses (Sections 3.4–3.5).

3.3 Model development

3.3.1 Logistic-regression SDM

Logistic regression models the probability of species presence as a logistic function of a linear combination of environmental predictors (Guisan and Zimmermann, 2000; Hosmer et al., 2000):

$p_i=\frac{\exp \left({\mathrm{x}}_{\mathrm{i}}^{\top }\mathrm{\beta }\right)}{1+\exp \left({\mathrm{x}}_{\mathrm{i}}^{\top }\mathrm{\beta }\right)},$

Where $p_i$ enotes the predicted probability of presence for sample $i$, ${\mathbf{x}}_i$ is the predictor vector, and $\beta$ is the vector of regression coefficients estimated via maximum likelihood. This formulation constrains predictions to the interval [0,1] and corresponds to the canonical logit link for binary responses.

3.3.2 Random-forest SDM

Random forest is an ensemble of T decision trees (Breiman, 2001). The predicted probability of presence is the proportion of trees voting for the positive class:

$P_{\text{RF}}\left(y\left.=1\right|{\mathbf{x}}_i\right)=\frac{1}{T}{\displaystyle \sum _{t=1}^T}I\left(h_t\left({\mathbf{x}}_i\right)=1\right),$

Where $h_t\left({\mathbf{x}}_i\right)$ is the prediction of tree t and $I\left(·\right)$ is the indicator function (1 if presence, 0 otherwise). We used T = 400 trees with class-balanced sampling to handle unequal class sizes.

3.4 Ensemble integration and evaluation metrics

3.4.1 Dual-model probability averaging

The DMPA ensemble combines the two base models through arithmetic averaging of probabilities:

$P_{\text{DMPA},i}=\frac{P_{\text{LR},i}+P_{\text{RF},i}}{2},$

Where $P_{\text{LR},i}$ and $P_{\text{RF},i}$ denote the predicted presence probabilities from the logistic-regression and random-forest SDMs for sample i, respectively. This simple averaging scheme preserves the interpretability of the logistic model while adding the non-linear flexibility of the random forest.

3.4.2 Threshold selection for binary mapping

Binary maps were derived by selecting the threshold ${\mathrm{t}}^{*}$ that maximises the F1-score (Sokolova and Lapalme, 2009):

$\mathrm{F}1=2\times \frac{\text{Precision}\times \text{Recall}}{\text{Precision}+\text{Recall}},\text{Precision}=\frac{\text{TP}}{\text{TP}+\text{FP}},\text{Recall}=\frac{\text{TP}}{\text{TP}+\text{FN}}.$

Where $\text{TP}$ , $\text{FP}$ and $\text{FN}$ denote true positives, false positives, and false negatives, respectively. The optimal ${\mathrm{t}}^{*}$ was computed for each cross-validation fold, and the median threshold across folds was used for final mapping.

3.4.3 Evaluation metrics

Model performance was assessed using discrimination, accuracy, and calibration indices computed from out-of-fold (OOF) predictions (i.e., predictions generated for samples not used in fitting the corresponding fold model). Brier score (Brier, 1950) measures mean-squared probability error:

$\text{Brier}=\frac{1}{N}{\displaystyle \sum _{i=1}^N}{\left({\widehat{p}}_i-y_i\right)}^2.$

Where $N$ is the number of samples, ${\widehat{p}}_i$ is the predicted probability, and $y_i\in \left\{0,1\right\}$ is the observed label (presence/background).

Expected calibration error (ECE) (Naeini et al., 2015) quantifies the average gap between empirical accuracy and predicted confidence:

$\text{ECE}={\displaystyle \sum _{m=1}^M}\frac{n_m}{N}\left|\text{acc}\left(B_m\right)-\text{conf}\left(B_m\right)\right|,$

$\text{where\hspace{0.17em}predicted\hspace{0.17em}probabilities\hspace{0.17em}are\hspace{0.17em}partitioned\hspace{0.17em}into\hspace{0.17em}}M$ bins ${\left\{B_m\right\}}_{m=1}^M$, $n_m$ is the number of samples in bin $B_m$, $\text{acc}\left(B_m\right)$ is the empirical fraction of positive labels in bin $B_m$, and $\text{conf}\left(B_m\right)$ is the mean predicted probability in the same bin.

ROC-AUC and PR-AUC represent threshold-free discrimination abilities computed as areas under the receiver-operating characteristic and precision–recall curves. All metrics were calculated using OOF predictions to ensure unbiased evaluation.

4 Experiments and results

4.1 Overall predictive performance

Overall OOF performance of the three modelling strategies is summarised in Table 2. All models achieve reasonably high discrimination, with ROC-AUC values well above random expectation, indicating that the combined climatic, topographic, and distance-based predictors are informative for separating suitable from unsuitable locations. Among the single-model SDMs, the random-forest SDM shows the highest average ROC-AUC, whereas the logistic SDM attains more balanced precision and recall but slightly lower ROC-AUC.

Table 2

Table 2. OOF performance of the logistic SDM, the random-forest SDM, and the DMPA ensemble for protected migratory birds in the Yancheng coastal wetlands. For ROC-AUC, PR-AUC, F1-score, and Brier score, values are reported as the OOF point estimate with the 95% bootstrap confidence interval in parentheses.

The DMPA ensemble combines the strengths of both base learners. It maintains ROC-AUC values comparable to the random-forest SDM, while substantially improving PR-AUC and F1-score relative to both single models. The ensemble also achieves a Brier score comparable to that of the random-forest SDM, suggesting an improved balance between discrimination, classification performance, and probabilistic accuracy compared with either single SDM.

Notably, despite its high ROC-AUC, the random-forest SDM yields an extremely low F1-score. This reflects a known behaviour of tree-based classifiers under severe class imbalance, where predicted probabilities are concentrated near zero for most samples, resulting in overly conservative decision thresholds when optimising F1. This limitation motivates the probability-averaging ensemble.

4.2 Discrimination under class imbalance

Cross-validated summary statistics indicate that all three models achieve good discrimination despite the strong imbalance between presence and background samples (Table 2). ROC-AUC values are consistently high across models, and the DMPA ensemble attains discrimination performance comparable to the stronger single-model baseline. Relative to the logistic SDM, the ensemble also yields higher PR-AUC and F1-score, suggesting improved identification of rare suitable locations under class imbalance.

The threshold-free behaviour of the models is further illustrated by the ROC and PR curves derived from OOF predictions (Figure 3). The ROC curves (Figure 3A) show that all models maintain high true positive rates at relatively low false positive rates, with substantial gains over the diagonal no-skill line. However, the PR curves (Figure 3B) reveal clearer differences under class imbalance. The logistic SDM exhibits the lowest precision across most recall levels, whereas both the random-forest SDM and the DMPA ensemble sustain higher precision at moderate recall. The ensemble closely tracks the random forest in PR space, indicating that probability averaging retains most of the discriminatory power of the stronger base learner while avoiding the weaker model’s loss of precision.

Figure 3

Figure 3. OOF discrimination of the three habitat-suitability models. (A) ROC curves. (B) PR curves. Curves are based on pooled OOF predictions for the logistic-regression SDM, random-forest SDM, and DMPA ensemble. Legends report the corresponding ROC-AUC and PR-AUC (AP) values for each model.

4.3 Decision threshold and error structure

To select an operating threshold for generating binary habitat maps, we examined the threshold–response behaviour of the DMPA ensemble (Figure 4A). F1-score increases as the threshold rises from 0 to around 0.5 and reaches a clear maximum near the F1-optimising threshold (≈0.52). The median threshold selected across cross-validation folds (≈0.61) lies to the right of this optimum, yielding a more conservative operating point that trades lower recall for higher precision. Beyond this range, recall drops sharply while precision improves only marginally, indicating that increasingly strict thresholds would miss many suitable locations without substantial gains in precision.

Figure 4

Figure 4. Threshold sweep and confusion matrices for the DMPA ensemble based on OOF predictions. (A) Threshold–metric curves (F1, precision, recall) with the F1-optimal and median operating thresholds marked. (B) Confusion matrix (counts). (C) Confusion matrix normalised by row.

The confusion matrices in Figures 4B,C summarise the resulting classification trade-off at the mapping threshold. In absolute terms (Figure 4B), the DMPA ensemble correctly classifies the vast majority of background samples, whereas it detects 28 of the 56 pooled presence locations, with the remaining presences falling below the decision threshold. The row-normalised matrix (Figure 4C) emphasises that commission errors are limited, whereas omission errors dominate (specificity ≈1.0 vs. sensitivity ≈0.5). This error structure is consistent with our goal of producing conservative suitability maps for protected migratory birds, prioritising low false-positive rates over maximising recall.

4.4 Uncertainty and robustness

Beyond discrimination, the reliability of predicted probabilities is important for conservation applications. Calibration curves based on OOF predictions (Figure 5A) show that the random-forest SDM tends to be over-confident at intermediate probabilities, whereas the logistic SDM underestimates observed occurrence frequencies at the upper end of the probability range. The DMPA ensemble exhibits intermediate behaviour, with its reliability curve generally closer to the identity line than that of the logistic SDM and broadly comparable to that of the random forest. These patterns are consistent with the Brier scores and ECE values reported in Table 2, indicating low overall probability error with moderate calibration bias for the ensemble.

Figure 5

Figure 5. Model calibration and probability distributions based on OOF predictions. (A) Reliability curves (decile bins) for the logistic SDM, random-forest SDM, and DMPA ensemble. (B) DMPA predicted probability distributions for presence (y = 1) and background (y = 0).

Class-conditional probability distributions of the DMPA ensemble (Figure 5B) further illustrate the separation between presence and background samples. Presence locations concentrate at higher predicted probabilities, while background samples are concentrated near low probabilities, with limited overlap between the two distributions. This separation indicates that the ensemble provides useful probability rankings that distinguish suitable from unsuitable locations.

4.5 Spatial patterns of predicted habitat suitability

The three habitat-suitability models show broadly consistent spatial patterns at sample locations, with higher predicted suitability probabilities concentrated along the coastal wetland corridor of Yancheng (Figure 6). The logistic SDM produces relatively smooth spatial gradients, with medium-to-high probabilities distributed more continuously across the coastal zone. In contrast, the random-forest SDM yields stronger spatial heterogeneity, with high-probability predictions appearing in a smaller number of localised clusters and many sample locations assigned very low probabilities.

Figure 6

Figure 6. Predicted habitat suitability at sample locations in the Yancheng coastal wetlands from the logistic SDM (A), random-forest SDM (B), and DMPA ensemble (C). Points are coloured by predicted suitability probability (0–1).

The DMPA ensemble exhibits an intermediate pattern. Compared with the logistic SDM, the ensemble reduces the extent of diffuse medium-probability predictions, while compared with the random-forest SDM it produces more spatially coherent clusters of higher suitability probabilities at sample locations along the coastal belt. Overall, probability averaging provides a visually clearer spatial signal in predicted suitability while retaining the main coastal concentration shared across models.

4.6 Environmental drivers

Finally, we explored the environmental drivers associated with the predicted suitability patterns. Random-forest feature importance (Figure 7A) suggests that predictive performance is dominated by a subset of variables. Among the highest-ranked predictors are distance-based covariates (e.g., distance to coastline and distance to rivers) together with several bioclimatic variables, indicating that both landscape configuration and climatic conditions contribute to suitability differentiation. Terrain-related variables (e.g., elevation, slope, and roughness) also contribute, although their importance is generally lower than that of the leading distance and bioclimatic predictors.

Figure 7

Figure 7. Random-forest feature importance (A) and predictor correlation structure (B) after collinearity filtering (|r| >0.95) in the Yancheng coastal wetlands.

The correlation heatmap of the retained predictors (Figure 7B) shows that, after collinearity filtering, some bioclimatic variables remain moderately correlated, whereas distance-based and terrain variables exhibit weaker pairwise correlations. This structure is consistent with the modelling strategy adopted in this study: logistic regression provides stable probability responses to the dominant gradients captured by the predictor set, while random forest offers additional flexibility for representing non-linear relationships. The DMPA ensemble combines these complementary behaviours via probability averaging.

5 Discussion

Our results show that a simple probability-level ensemble of two widely used SDMs—logistic regression and random forest—can substantially improve habitat discrimination for protected migratory birds in a complex coastal wetland system. Across the different validation schemes, the dual-model probability averaging (DMPA) ensemble consistently matched or exceeded the performance of its single-model components in ROC-AUC and PR-AUC, while also delivering higher F1-scores under class imbalance. This indicates that a lightweight ensemble, which only averages predicted probabilities from two complementary models, is sufficient to stabilise predictions and reduce sensitivity to data partitioning, without introducing additional model complexity or sacrificing interpretability.

Beyond discrimination, the DMPA ensemble improved the quality of probability estimates in ways that are directly relevant for conservation decisions. Calibration and reliability analyses showed lower expected calibration error and smaller bin-wise deviations for the ensemble than for at least one of the base models, suggesting that its output probabilities are closer to observed occurrence frequencies. Threshold-sweep curves further indicated that the ensemble achieves a more favourable balance between precision and recall over a wide range of operating thresholds. This property is particularly important in regulatory or management contexts, where decision-makers must weigh the costs of false positives (overestimating suitable habitat) against the risks of false negatives (overlooking truly suitable sites).

The analysis of environmental drivers helps explain why combining a linear and a tree-based SDM is beneficial in the Yancheng coastal wetlands. Variable importance results highlight a small set of dominant predictors, including distance to coastline and rivers and several key bioclimatic variables related to temperature and precipitation. These gradients are consistent with known controls on migratory bird distributions in intertidal marsh–reed mosaics, where moisture availability, thermal conditions and proximity to tidal flats jointly shape habitat suitability. At the same time, the correlation structure of retained covariates shows that terrain and proximity variables are largely independent of the main climatic blocks. In this setting, logistic regression provides stable and interpretable responses along major climatic axes, while the random forest captures non-linear interactions between climate, topography and proximity to coastal features. By averaging their probability outputs, the DMPA ensemble inherits these complementary strengths while dampening the tendency of the logistic SDM to oversmooth and the random-forest SDM to produce overly fragmented high-suitability patches.

From a practical perspective, the resulting habitat maps offer a more coherent and ecologically plausible representation of suitable areas along the Yancheng coastal belt. The ensemble emphasises continuous clusters of medium-to-high suitability in key wetland zones, rather than either spreading suitability too widely or restricting it to isolated, speckled patches. This pattern is advantageous for applications such as reserve expansion, buffer-zone planning and the identification of priority restoration sites, where managers require spatially consistent signals that can be linked back to transparent environmental gradients. Because the DMPA framework relies on standard algorithms, automatically derived covariates and a single, reproducible pipeline, it can be readily updated as new occurrence records become available or as climate and land-cover datasets are refined.

The resulting suitability maps can directly inform spatially explicit restoration planning. For instance, in the same Yancheng coastal wetland system, Zhou et al. (2023) developed a restoration framework based on hydrogeomorphic units, which could be effectively integrated with habitat suitability outputs to prioritize restoration actions in areas of high ecological value and connectivity. Building on site-level suitability mapping, recent work in the Yancheng coastal wetlands further shows how SDM-derived habitats can be translated into corridor-network design and scenario-based optimisation for waterbird connectivity (Huang et al., 2025).

At the same time, several limitations of this study should be acknowledged. First, the analyses rely on presence–background data within a single coastal region, and do not explicitly account for detectability or observation bias, which may affect absolute probability values. Second, the environmental covariates are essentially static snapshots and do not represent intra-annual dynamics in water levels, vegetation phenology or human disturbance that may influence migratory behaviour. Similar large-scale analyses have shown that temporal dynamics of coastal wetlands strongly shape migratory-bird habitats across East and South Asia, underscoring the need to integrate time-varying environmental layers in future modelling efforts (Tian et al., 2025). Recent work further emphasises that hydrological connectivity acts as a key ecological driver of habitat suitability in floodplain wetlands, where dynamic water exchange determines habitat quality and spatial configuration (Teng et al., 2025). Recent research highlights that integrating remote sensing indicators with plant functional traits can effectively capture short-term ecological responses and long-term degradation mechanisms in coastal and estuarine wetlands. For example, Cingano et al. (2025) demonstrated that soil salinity and anoxia jointly drive the decline of estuarine reed beds, and that dynamic spectral and structural indices derived from satellite data can improve detection of vegetation stress patterns. This suggests that incorporating temporally updated remote-sensing variables into SDMs could better represent the dynamic environmental context of waterbird habitats. Third, we focus on one taxonomic group and one coastal wetland system; generalisation of the DMPA framework to other regions, seasons and species assemblages remains to be tested. Future work could therefore extend the approach by incorporating dynamic remote-sensing indicators, explicitly modelling spatial autocorrelation and uncertainty, and exploring ensembles that integrate additional SDMs or cost-sensitive thresholds tailored to specific conservation objectives. Advanced hierarchical Bayesian ensemble frameworks have recently been developed to integrate multiple SDMs while explicitly accounting for climate and management uncertainty, offering a more complex yet statistically rigorous pathway for future extensions beyond simple probability-level averaging (Ogawa et al., 2024). An additional frontier direction is the integration of expert-elicited probability maps with survey data through hierarchical Bayesian calibration, which allows explicit quantification and correction of expert bias while improving predictive performance in data-limited contexts (Kaurila et al., 2024). Recent research also suggests that random forest models can be leveraged not only for predictive mapping but for uncovering nonlinear causal relationships among ecological drivers, providing a promising direction for interpreting complex environmental interactions in future SDM-based analyses (Brown et al., 2025). Future iterations could be enhanced by incorporating high-resolution, temporally dense data from the latest remote sensing platforms, which offer advanced capabilities for monitoring coastal wetland dynamics (Zhang et al., 2023; Song et al., 2025). Additionally, active learning approaches could optimize sampling strategies in data-scarce regions (Liu et al., 2022).

Despite these limitations, the present study suggests that carefully designed, probability-averaging ensembles built from simple SDMs can provide robust and decision-ready habitat suitability products for migratory bird conservation in rapidly changing coastal wetlands.

6 Conclusion

This study developed and evaluated a dual-model probability averaging (DMPA) ensemble to improve habitat suitability assessment for migratory birds in the Yancheng coastal wetlands. By combining a logistic-regression SDM and a random-forest SDM, the DMPA approach integrates the interpretability and stable ranking of a linear model with the flexibility of a non-linear classifier, while keeping the ensemble structure simple and transparent.

Using a unified occurrence–background dataset and hierarchical cross-validation, we generated OOF predictions for all models and compared their discrimination, calibration, and operating-point behaviour. Across all validation splits, the ensemble consistently matched or exceeded the performance of the two single-model SDMs. The ROC and PR curves showed that DMPA achieved stronger threshold-free discrimination, and the threshold-sweep results indicated more favourable precision–recall trade-offs. Calibration analyses also demonstrated that the ensemble produced well-behaved probability estimates with reduced over- or under-confidence.

When projected back into geographic space, the DMPA ensemble produced suitability patterns that were spatially coherent and biologically reasonable. Compared with the two base SDMs, the ensemble avoided excessive fragmentation and better highlighted the main wetland complexes known to support migratory birds in Yancheng. The binary habitat maps derived from the cross-validated thresholds further reinforced these patterns and revealed continuous clusters of potentially important habitat patches.

Overall, the results show that a simple probability-averaging ensemble can provide more reliable and ecologically interpretable suitability estimates than either single model alone. Because the method requires no additional meta-learner and uses only models that are widely available in ecological workflows, it offers a practical and reproducible way to improve species distribution modelling in complex coastal wetlands. Encouragingly, restoration and management actions in the neighbouring Rudong coastal wetlands have been reported to yield measurable improvements in waterbird habitat quality, illustrating that science-based interventions can deliver tangible conservation benefits in this coastal region (Duan et al., 2023). The resulting suitability and binary habitat maps can support conservation planning by helping to identify priority zones for monitoring and habitat protection in the Yancheng coastal region.

Data availability statement

Data are available from the corresponding author upon reasonable request.

Author contributions

XS: Conceptualization, Data curation, Formal Analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review and editing. HX: Writing – original draft, Writing – review and editing. YZ: Conceptualization, Data curation, Formal Analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Software, Supervision, Validation, Visualization, Writing – original draft, Writing – review and editing.

Funding

The author(s) declared that financial support was received for this work and/or its publication. Project supported by the National Natural Science Foundation of China (Grant No.32470525). National Key Research and Development Program of China (2024YFF1307204).

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

Araújo, M. B., and New, M. (2007). Ensemble forecasting of species distributions. Trends Ecol. and Evol. 22 (1), 42–47. doi:10.1016/j.tree.2006.09.010

PubMed Abstract | CrossRef Full Text | Google Scholar

Breiman, L. (2001). Random forests. Mach. Learn. 45 (1), 5–32. doi:10.1023/A:1010933404324

CrossRef Full Text | Google Scholar

Brier, G. W. (1950). Verification of forecasts expressed in terms of probability. Mon. Weather Rev. 78 (1), 1–3. doi:10.1175/1520-0493(1950)078<0001:vofeit>2.0.co;2

CrossRef Full Text | Google Scholar

Brown, M. G. L., Peterson, M. G., Tezaur, I. K., Peterson, K. J., and Bull, D. L. (2025). Random forest regression feature importance for climate impact pathway detection. J. Comput. Appl. Math. 437, 116479. doi:10.1016/j.cam.2024.116479

CrossRef Full Text | Google Scholar

Chiaverini, L., Macdonald, D. W., Hearn, A. J., Kaszta, Ż., Can, Ö. E., Haidir, I. A., et al. (2023). Not seeing the forest for the trees: generalised linear model out-performs random forest in species distribution modelling for Southeast Asian felids. Ecol. Inf. 75, 102026. doi:10.1016/j.ecoinf.2023.102026

CrossRef Full Text | Google Scholar

Cingano, P., Moro, D., Pellegrini, E., Asquini, E., Contin, M., Trotta, G., et al. (2025). Integrating remote sensing and functional traits to elucidate estuarine common reed beds decline driven by soil salinity and anoxia. Ecol. Indic. 171, 114294. doi:10.1016/j.ecolind.2025.114294

CrossRef Full Text | Google Scholar

Duan, H., Xia, S., and Yu, X. (2023). Conservation and restoration efforts have promoted increases in shorebird populations and the area and quality of their habitat in the yellow river Delta, China. Int. J. Digital Earth 16, 4530–4550. doi:10.1080/17538947.2023.2264279

CrossRef Full Text | Google Scholar

El Alaoui, O., and Idri, A. (2024). An empirical evaluation of ensemble strategies in habitat suitability modeling. SN Comput. Sci. 5, 501. doi:10.1007/s42979-024-02828-y

CrossRef Full Text | Google Scholar

Elith, J., and Leathwick, J. R. (2009). Species distribution models: ecological explanation and prediction across space and time, annual review of ecology. Evol. Syst. 40, 677–697. doi:10.1146/annurev.ecolsys.110308.120159

CrossRef Full Text | Google Scholar

Guisan, A., and Zimmermann, N. E. (2000). Predictive habitat distribution models in ecology. Predict. Habitat Distribution Models Ecology, Ecol. Model. 135 (2–3), 147–186. doi:10.1016/S0304-3800(00)00354-9

CrossRef Full Text | Google Scholar

He, K. S., Bradley, B. A., Cord, A. F., Rocchini, D., Tuanmu, M. N., Schmidtlein, S., et al. (2015). Will remote sensing shape the next generation of species distribution models? Remote Sens. Ecol. Conservation 1 (1), 4–18.doi:10.1002/rse2.7

CrossRef Full Text | Google Scholar

Hosmer, D. W., Lemeshow, S., and Cook, E. D. (2000). Applied logistic regression. 2nd ed. New York: John Wiley and Sons.

Google Scholar

Huang, G., Hu, W., Du, J., Jia, Y., Zhou, Z., Lei, G., et al. (2025). Identification and scenario-based optimization of ecological corridor networks for waterbirds in typical coastal wetlands. Ecol. Indic. 171, 113147. doi:10.1016/j.ecolind.2025.113147

CrossRef Full Text | Google Scholar

Kaurila, K., Vanhatalo, J., and Lappalainen, A. (2024). Predicting species distributions with ensembles of mechanistic and empirical models. Ecol. Appl. 34, e2971. doi:10.1002/eap.2971

PubMed Abstract | CrossRef Full Text | Google Scholar

Lei, J., and Wang, J. (2024). Research on the extraction method of coastal wetlands based on Sentinel-2 data. Mar. Environ. Res. 198, 106429. doi:10.1016/j.marenvres.2024.106429

PubMed Abstract | CrossRef Full Text | Google Scholar

Liu, C., Berry, P. M., Dawson, T. P., and Pearson, R. G. (2005). Selecting thresholds of occurrence in the prediction of species distributions. Ecography 28 (3), 385–393. doi:10.1111/j.0906-7590.2005.03957.x

CrossRef Full Text | Google Scholar

Liu, P., Wang, L. Z., Ranjan, R., He, G. J., and Zhao, L. (2022). A survey on active deep learning: from model-driven to data-driven. ACM Comput. Surv. 54, 1–34. doi:10.1145/3510414

CrossRef Full Text | Google Scholar

Murray, N. J., Phinn, S. R., DeWitt, M., Ferrari, R., Johnston, R., Lyons, M. B., et al. (2019). The global distribution and trajectory of tidal flats. Nature 565 (7738), 222–225. doi:10.1038/s41586-018-0805-8

PubMed Abstract | CrossRef Full Text | Google Scholar

Naeini, M. P., Cooper, G., and Hauskrecht, M. (2015). Obtaining well calibrated probabilities using bayesian binning. Proc. AAAI Conf. Artif. Intell. 29 (1). doi:10.1609/aaai.v29i1.9602

CrossRef Full Text | Google Scholar

Ogawa, R., Wang, G., Burger, L. W., Strickland, B. K., Davis, J. B., and Cunningham, F. L. (2024). Bayesian integrated species distribution models for hierarchical resource selection by a soaring bird. Ecol. Inf. 83, 102787. doi:10.1016/j.ecoinf.2024.102787

CrossRef Full Text | Google Scholar

Pettorelli, N., Laurance, W. F., O’Brien, T. G., Wegmann, M., Nagendra, H., and Turner, W. (2014). Satellite remote sensing for applied ecologists: opportunities and challenges. J. Appl. Ecol. 51 (4), 839–848. doi:10.1111/1365-2664.12261

CrossRef Full Text | Google Scholar

Ramsar Convention on Wetlands (2018). Global wetland outlook: state of the world’s wetlands and their services to people. Gland, Switzerland.

Google Scholar

Roberts, D. R., Bahn, V., Ciuti, S., Boyce, M. S., Elith, J., Guillera-Arroita, G., et al. (2017). Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography 40 (8), 913–929. doi:10.1111/ecog.02881

CrossRef Full Text | Google Scholar

Schuerch, M., Spencer, T., Temmerman, S., Kirwan, M. L., Wolff, C., Lincke, D., et al. (2018). Future response of global coastal wetlands to sea-level rise. Nature 561 (7722), 231–234. doi:10.1038/s41586-018-0476-5

PubMed Abstract | CrossRef Full Text | Google Scholar

Sokolova, M., and Lapalme, G. (2009). A systematic analysis of performance measures for classification tasks. Inf. Process. and Manag. 45 (4), 427–437. doi:10.1016/j.ipm.2009.03.002

CrossRef Full Text | Google Scholar

Solomon, M. J., Fricke, K. A., Kruse, C., and McNew, L. B. (2025). Ensembled evaluations of habitat suitability for prioritizing lesser prairie-chicken conservation. Ecosphere 16, e70321. doi:10.1002/ecs2.70321

CrossRef Full Text | Google Scholar

Song, B. Z., Wang, D. B., Liu, P., Wang, D. L., Xie, G. L., and Liu, J. B. (2025). Refining remote sensing image segmentation results based on vision foundation models. IEEE Geoscience Remote Sens. Lett. 22, 1–5. doi:10.1109/LGRS.2025.3606951

CrossRef Full Text | Google Scholar

Sun, X., Shen, J., Xiao, Y., Li, S., and Cao, M. (2023). Habitat suitability and potential biological corridors for waterbirds in yancheng coastal wetland of China. Ecol. Indic. 148, 110090. doi:10.1016/j.ecolind.2023.110090

CrossRef Full Text | Google Scholar

Teng, J., Zhu, Y., Duan, H., Yu, X., Xia, S., Wang, R., et al. (2025). Linking hydrological connectivity and waterbirds habitat suitability in floodplain wetlands: implications for wetland management. Ecol. Indic. 171, 113663. doi:10.1016/j.ecolind.2025.113663

CrossRef Full Text | Google Scholar

Thuiller, W., Lafourcade, B., Engler, R., and Araújo, M. B. (2009). BIOMOD – a platform for ensemble forecasting of species distributions. Ecography 32 (3), 369–373. doi:10.1111/j.1600-0587.2008.05742.x

CrossRef Full Text | Google Scholar

Tian, P., Cao, L. D., Li, J. L., Pu, R. L., Liu, Y. C., Zhang, H. T., et al. (2022). Ecosystem stability assessment of yancheng coastal wetlands, a world natural heritage site. Land 11, 564. doi:10.3390/land11040564

CrossRef Full Text | Google Scholar

Tian, P., Zhang, F., Zhang, H., Wang, L., Zeng, H., Liu, Y., et al. (2025). Dynamics of coastal wetlands and their impacts on migratory bird habitats in china–southeast asia–south Asia. Ocean and Coast. Manag. 270, 107883. doi:10.1016/j.ocecoaman.2025.107883

CrossRef Full Text | Google Scholar

Townsend, J. P., and Aldstadt, J. (2023). Habitat suitability mapping using logistic regression analysis of long-term bioacoustic bat survey dataset in the cassadaga creek watershed (USA). Sci. Total Environ. 889, 165077. doi:10.1016/j.scitotenv.2023.165077

PubMed Abstract | CrossRef Full Text | Google Scholar

Valavi, R., Elith, J., Lahoz-Monfort, J. J., and Guillera-Arroita, G. (2018). blockCV: an R package for generating spatially or environmentally separated folds for k-fold cross-validation of species distribution models, bioRxiv, 357798.

Google Scholar

Verniest, F., Galewski, T., Boutron, O., Dami, L., Defos du Rau, P., Guelmami, A., et al. (2024). Exposure of wetlands important for nonbreeding waterbirds to sea-level rise in the mediterranean. Conserv. Biol. 38 (6), e14288. doi:10.1111/cobi.14288

PubMed Abstract | CrossRef Full Text | Google Scholar

Wang, C., Wang, G., Dai, L. J., Liu, H. Y., Li, Y. F., Zhou, Y., et al. (2020). Diverse usage of waterbird habitats and spatial management in Yancheng coastal wetlands. Ecol. Indic. 117, 106583. doi:10.1016/j.ecolind.2020.106583

CrossRef Full Text | Google Scholar

Wang, Y. J., Jin, S. G., and Dardanelli, G. (2024). Vegetation classification and evaluation of Yancheng coastal wetlands based on random forest algorithm from Sentinel-2 images. Remote Sens. 16, 1124. doi:10.3390/rs16071124

CrossRef Full Text | Google Scholar

Wang, X., Wang, L., Wang, X., Chang, Q., and Ding, J. (2025). Suitability assessment of waterbird habitats in Jiangsu coastal wetlands, China. Wetl. Ecol. Manag. 33, 79. doi:10.1007/s11273-025-10096-7

CrossRef Full Text | Google Scholar

Wei, Z. Z., Xu, Z. L., Dong, B., Xu, H. F., Lu, Z. P., and Liu, X. (2023). Habitat suitability evaluation and ecological corridor construction of wintering cranes in Poyang Lake. Ecol. Eng. 189, 106894. doi:10.1016/j.ecoleng.2023.106894

CrossRef Full Text | Google Scholar

Yang, K., Cui, D., Wang, C., Tang, Q., and Miao, L. (2025). Intelligent assessment of habitat quality based on multiple machine learning fusion methods. Eng. Appl. Artif. Intell. 162, 112395. doi:10.1016/j.engappai.2025.112395

CrossRef Full Text | Google Scholar

Zhang, Y. W., Liu, P., Chen, L. J., Xu, M. Z., Guo, X. Y., and Zhao, L. J. (2023). A new multi-source remote sensing image sample dataset with high resolution for flood area extraction: gf-floodnet. Int. J. Digital Earth 16, 2522–2554. doi:10.1080/17538947.2023.2230978

CrossRef Full Text | Google Scholar

Zhou, Y., Zhou, Y., Liu, H. Y., Li, Y. F., Wang, C., and Wright, A. (2023). Coastal wetland restoration framework based on hydrogeomorphic units of the coastal wetland of yancheng, China. Ocean and Coast. Manag. 243, 106737. doi:10.1016/j.ocecoaman.2023.106737

CrossRef Full Text | Google Scholar