Development of an Eco-Environmental Evaluation System for Islands in Jiangsu, China, Based on the Time-Varying Entropy Weight Method and a Bayesian Network

Abstract

This study developed an ecological environment evaluation framework tailored for islands in Jiangsu and validated its applicability using nine representative islands. The evaluation system encompasses 14 indicators across three dimensions: ecological, socio-economic, and policy-climate. By coupling the Time-varying Entropy Weight method with a Bayesian Network, the framework quantifies the dynamic impacts of policy interventions, extreme weather, and human activities. To enhance model accuracy under small-sample conditions, machine learning and deep learning techniques were integrated to construct a multi-layer ensemble evaluation model. The results indicate that this model improves prediction accuracy by 11.3% and reduces the root mean square error by 33.3%. The assessment results reveal significant differences in ecological quality among islands of different types. Natural-type inhabited islands maintain relatively high ecological quality through the synergy of ecological conservation and industrial activity, whereas artificial-type inhabited islands experience significant negative impacts from industrial development. Uninhabited islands generally score around 10, indicating relatively stable ecological conditions but high natural vulnerability. This framework provides a high-precision quantitative approach for dynamic evaluation of island ecological quality under small-sample constraints and offers a scientific basis for customized, island-specific conservation and development management strategies.

1. Introduction

Islands, as critical ecological nodes within coastal zones, not only sustain unique biodiversity but also provide indispensable ecosystem services [1]. In recent years, the vulnerability of island ecosystems has become increasingly pronounced due to the intensification of global climate change and frequent human development activities. Consequently, modern island management policies are gradually shifting from early-stage extensive exploitation toward strategies emphasizing ecological priority and dynamic adaptation. Within this context, constructing a scientifically rigorous and precise eco-environmental evaluation system has become a prerequisite for implementing refined island management and formulating sustainable development policies.

The coastal region of Jiangsu Province encompasses 26 islands, which play a central role in maintaining regional marine ecological balance and promoting socio-economic development. Nevertheless, under the combined influence of frequent extreme weather events, strengthened policy interventions, and pressures from tourism and industrial development, the ecological systems of Jiangsu islands are undergoing rapid changes, with different types of islands experiencing highly heterogeneous ecological pressures. Therefore, conducting a comprehensive and dynamic assessment of these islands is not only of strategic significance for safeguarding regional marine ecological security but also provides essential empirical support for developing tailored management strategies under the “one island, one policy” approach.

Although research on island ecological assessment has made notable progress, conventional evaluation systems, such as the static DPSIR framework, remain significantly limited [2,3]. On one hand, these systems predominantly rely on cross-sectional data and static indicators, overlooking the temporal dynamics of environmental changes and the lagged effects of policy interventions [4]. On the other hand, when applied to complex systems characterized by small sample sizes, high dimensionality, and strong nonlinearity, traditional models often encounter challenges related to poor adaptability and low predictive accuracy.

To address these gaps, this study argues that only by coupling temporally dynamic weighting with feature-enhanced machine learning can the actual disturbances of human activities and extreme weather on island ecosystems be accurately captured under small-sample conditions [5]. Accordingly, this research proposes a time-sensitive evaluation method integrating the time-varying entropy weight method (EWM) with Bayesian networks (BNs) [6]. By deriving temporally dynamic indicator weights, the approach effectively quantifies the influence of policy implementation and extreme climatic events [7]. Furthermore, the method incorporates multi-layer intelligent learning techniques, including KMeans clustering, autoencoders, and attention mechanisms, within island assessments, substantially enhancing the model’s adaptability and inferential accuracy for complex, small-sample datasets.

Building on these theoretical and methodological innovations, the specific objectives of this study are: (1) to construct a comprehensive evaluation system comprising 14 indicators across three dimensions—ecological, socio-economic, and policy-climate—to accurately capture the dynamic characteristics of external policies and climate change; (2) to develop a high-precision assessment model integrating dynamic weighting and multi-layer intelligent learning techniques [8], overcoming the limitations of small island sample sizes while enhancing temporal sensitivity; (3) to quantify the spatiotemporal dynamics of ecological–environmental quality across islands with different development types based on data from nine representative islands during 2014–2024; (4) to precisely identify the core limiting factors differentiating the ecological quality of inhabited and uninhabited islands, thereby providing novel theoretical perspectives and technical support for differentiated management of coastal natural–social coupled systems.

2. Materials and Methods

2.1. Study Area and Sample Selection

There are a total of 26 islands in Jiangsu Province, 22 of which are uninhabited. In this study, nine typical islands in Jiangsu, including four inhabited islands and five uninhabited islands, were selected as study areas. The locations of the selected islands are shown in Figure 1, which were mapped using ArcGIS Pro (version 2.5.2; Esri, Redlands, CA, USA). The specific attributes and characteristics of the island are shown in

Table 1. Summary of basic information, development attributes and human activities of islands in the study area.

Island Name (Area)	Development Attributes	Summary of Environmental Characteristics and Human Activities	Reference
Lian Island (5.130 km2)	Inhabit island	The largest island in Jiangsu; both forest vegetation and beach; the tourism development is mature and the fishery foundation is deep, which is a typical example of the coordinated development of coastal tourism and traditional fishery.	[9]
Yangshan Island (0.201 km2)	Inhabit island	Covering young and middle-aged pine mixed wood; the surrounding fishery resources are rich; as a popular tourist destination, it contributes significantly to the local economy.	[10]
Qinshan Island (0.167 km2)	Inhabit island	The surrounding shellfish resources are abundant; it has profound cultural connotation and belongs to the representative of the integrated development of ecological protection and cultural tourism.	[11]
Yangguang Island (3.000 km2)	artificial island	The location of the national key project Jiangsu LNG receiving station; a typical industrial energy-guaranteed artificial island.	[12]
Cheniushan Island (0.058 km2)	Uninhabited islands	Clear water quality; as one of the basic points of China ‘s territorial waters, it has long been stationed by militias and border forces and has a strategic guarantee function.	[13]
Pingshan Island (0.133 km2)	Uninhabited islands	The development of sea cliffs and rock beaches; it inhabits 129 kinds of birds and is rich in sea treasures such as sea cucumber and amphioxus.	[14]
Zhu Island (0.126 km2)	Uninhabited islands	Covering dense shrubs and herbaceous vegetation; the reef shoreline is complete, and the ecosystem maintains a native natural state.	[15]
Dashan Island (0.115 km2)	Uninhabited islands	Located in the middle of the Yellow Sea, the base of China ‘s territorial waters; ecological protection is extremely strict, without any development activities.	[15]
Niubei Island (0.014 km2)	Uninhabited islands	Very little interference by human activities; the whole island is an undeveloped reef landform with no vegetation cover.	[16]

.

Sample selection adhered to the principles of representativeness and comparability. In terms of representativeness, the samples encompass major island types—including natural bedrock islands, artificial islands, and territorial sea base islands—ensuring the findings reflect the overall characteristics of Jiangsu’s islands. Regarding comparability, the study utilized paired analysis to compare “inhabited islands” (significantly driven by human activity) with “uninhabited islands” (maintaining near-original states). This stratified sampling logic aims to reveal divergent drivers of ecosystem evolution under varying development intensities and management regimes.

2.2. Evaluation Indicator System

The established comprehensive island indicator evaluation system serves to optimize and expand upon the driver–pressure–state–impact–response (DPSIR) framework through its multidimensional cross-design [17]. In contrast to the linear causal chain logic of the DPSIR framework, this system involves the development of a mesh assessment model with 14 indicators across three major dimensions (ecology, socioeconomics, and policy and climate) that not only cover the core elements of the DPSIR framework, such as state, pressure, and drivers, but also expand upon its rigid structure [18]. The selected indicators were spatially quantified on the basis of remote sensing data and geographic information system (GIS) techniques [19], and positive and negative indicators were synergistically incorporated to reflect the dynamic balance of the system, rendering it more compatible with the complexity of coupled nature–society island systems [20]. This indicator system exhibits significant innovation for regional adaptability in comparative studies of inhabited and uninhabited islands in Jiangsu. Considering the regional characteristics (e.g., low altitude and sensitive ecology) of alluvial sandbar islands in Jiangsu, indicators such as the terrain elevation index and natural shoreline retention rate were chosen to accurately capture the differences in gradient resulting from the intensity of development of inhabited islands and the ecological vulnerability of uninhabited islands. For the inhabited islands, socioeconomic indicators were adopted to assess their development carrying capacity, whereas for the uninhabited islands, ecological indicators were applied to define conservation priorities [21]. The combination of the number of policies and the number of extreme weather events not only reflects the effect of anthropogenic disturbance on coastal ecosystems but also provides systematic support for the one island–one policy management system, which aims to promote the transition of island management from a generic approach to a customized model [22]. The specific evaluation indicators chosen in this paper are listed in

Table 2. Classification and selection of evaluation indicators.

Classification Dimension	Specific Indicator	Selection Basis
Ecological	Normalized difference vegetation index (NDVI) vegetation coverage	A classic indicator in the literature that can be easily obtained via remote sensing data and directly reflects ecosystem productivity
	Natural shoreline retention rate	Core requirement of policies (e.g., the Island Protection Act) that serves as a key indicator of island protection status
	Coastal wetland/water coverage ratio	Reflects the characteristics of the coastal ecosystems of Jiangsu islands and is essential for biodiversity maintenance
	Proportion of artificial shorelines	Characterizes the pressure from anthropogenic development on the shoreline and complements the natural shoreline retention rate
	Terrain elevation index	Critical for assessing the vulnerability to flooding and erosion, given the low and flat landform characteristics of Jiangsu islands
	Rate of excellent water quality in surrounding areas	Policy assessment objective [23] that directly reflects the quality of the marine environment
Socioeconomic	Proportion of construction land	Represents the land use intensity and serves as a fundamental indicator for measuring the effectiveness of spatial management and control
	Average nighttime light intensity	Literature-verified proxy indicator that effectively compensates for gaps in certain socioeconomic statistics
	Waste disposal rate	Reflects policy and management priorities and directly reflects environmental governance and infrastructure capacity
	Visitor density	Indicates the pressure of tourism activities and is central to assessing the tourism carrying capacity
	Island population density	Represents the basic intensity of anthropogenic activities and serves as a key indicator to distinguish between inhabited and uninhabited islands
	Annual number of ecological restoration projects	Reflects the intensity of response and indicates the government’s proactive investment in ecological protection
Policy and climate	Number of policies	Represents institutional support to capture the influence of the external policy environment on island development
	Number of extreme weather events	Characterizes external climate risks and serves as an important parameter for assessing the resilience and disaster prevention needs of an island system

.

2.3. Data Collection and Preprocessing

The data employed in this study include remote sensing imagery, topographic data, and government statistical data. Data on the NDVI, natural shoreline retention rate, and proportion of artificial shorelines were derived from Sentinel-2 L2A satellite images (European Space Agency; 10 m resolution; https://earth.esa.int/, accessed on 26 February 2026). The coastal wetland/water coverage ratio and the proportion of construction land were obtained from Landsat 8 Collection 2 Level-1 (LC08/C02/T1_TOA) images (United States Geological Survey (USGS); 30 m resolution; https://landsat.gsfc.nasa.gov/, accessed on 26 February 2026). The terrain elevation index was derived from the Copernicus digital elevation model (DEM) (30 m resolution; European Space Agency; https://dataspace.copernicus.eu/, accessed on 26 February 2026). The average nighttime light intensity was calculated using the Visible Infrared Imaging Radiometer Suite (VIIRS)/day–night band (DNB) monthly composite product (National Oceanic and Atmospheric Administration (NOAA); 500 m resolution; https://eogdata.mines.edu/, accessed on 26 February 2026). All the above remote sensing and terrain data were preprocessed and calculated on the Google Earth Engine platform (https://earthengine.google.com/, accessed on 26 February 2026). Data on the rate of excellent water quality in surrounding areas, the waste disposal rate, the island population density, and the number of extreme weather events were sourced from the Lianyungang Statistical Yearbook (Lianyungang Statistical Bureau; http://tjj.lyg.gov.cn/, accessed on 26 February 2026). The annual number of ecological restoration projects and relevant policy documents were derived from the official website of the Lianyungang Municipal People’s Government (http://www.lyg.gov.cn/, accessed on 26 February 2026). Data on the tourist density were obtained from the official website of the Lianyungang Bureau of Culture, Radio, Film, and Tourism (http://wglj.lyg.gov.cn/, accessed on 26 February 2026).

The specific sources, spatial and temporal resolutions, and processing methods for the data in all the datasets are listed in

Table 3. Data sources and processing.

Data Type	Data Source	Spatial/Temporal Resolution	Processing Method
NDVI vegetation coverage, natural shoreline retention rate, and proportion of artificial shorelines	Sentinel-2 L2A (ESA)	10 m; from 2015 to the present; 5-day revisit cycle	Preprocessing and calculation on the Google Earth Engine platform
Coastal wetland/water coverage ratio and proportion of construction land	Landsat 8 Collection 2 Level-1 TOA (USGS)	30 m; 16-day revisit cycle; since 2013	Classification and statistics using the Google Earth Engine
Terrain elevation index	Copernicus DEM (ESA)	30 m; global coverage	Clipping and calculation using the Google Earth Engine
Mean nighttime light intensity	VIIRS/DNB monthly composite (NOAA)	500 m; monthly; since 2012	Extraction and mean calculation using the Google Earth Engine
Rate of excellent water quality in surrounding areas, waste disposal rate, island population density, and number of extreme weather events	Lianyungang Statistical Yearbook	Annual data	Query for text information
Annual number of ecological restoration projects and number of policies	Official website of Lianyungang Municipal People’s Government	Government document/project announcement	collection
Visitor density	Official website of Lianyungang Bureau of Culture, Radio, Film, and Tourism	Annual/quarterly tourism statistics	collection

.

2.4. TVEWM–BN Integrated Framework

2.4.1. Methods and Principles

The traditional entropy weight method is an objective weighting approach that aims to determine weights on the basis of the degree of indicator variability [24]; however, this method is applicable only to static cross-sectional data and can hardly capture the dynamic importance of indicators across different times [25]. A BN is a directed acyclic graph model that can quantify uncertainty and causality, but similarly, it exhibits inherent limitations in resolving dynamic time series data.

To address these limitations, a combined TVEWM–BN framework based on the traditional method is proposed, in which ML and DL techniques are integrated to develop a multilevel intelligent evaluation system. The TVEWM can be used to calculate the time-varying weights of indicators in a time series framework, thereby better revealing differences in the importance of indicators at different stages [26]. An enhanced BN substantially increases the ability of the model to capture complex nonlinear relationships and characterize temporal dynamics through ML-based feature extraction and DL-based time series modeling. Figure 2 shows a flow chart of this method.

2.4.2. TVEWM Calculation Procedure

Data standardization and feature enhancement.

The original indicator matrix is defined as $\mathrm{X}=({\mathrm{x}}_{\mathrm{i}\mathrm{j}}(\mathrm{t}))$, where ${\mathrm{x}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})$ denotes the value of the jth indicator of the ith object at time t. To normalize the matrix, range standardization [2] is applied to eliminate the influence of different dimensions [27]:

Positive indicators:

$${\mathrm{z}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})={\displaystyle \frac{{\mathrm{x}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})-\underset{\mathrm{i}}{\mathrm{m}\mathrm{i}\mathrm{n}}{\mathrm{x}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})}{\underset{\mathrm{i}}{\mathrm{m}\mathrm{a}\mathrm{x}}{\mathrm{x}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})-\underset{\mathrm{i}}{\mathrm{m}\mathrm{i}\mathrm{n}}{\mathrm{x}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})}},$$

(1)

Negative indicators:

$${\mathrm{z}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})={\displaystyle \frac{\underset{\mathrm{i}}{\mathrm{m}\mathrm{a}\mathrm{x}}{\mathrm{x}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})-{\mathrm{x}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})}{\underset{\mathrm{i}}{\mathrm{m}\mathrm{a}\mathrm{x}}{\mathrm{x}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})-\underset{\mathrm{i}}{\mathrm{m}\mathrm{i}\mathrm{n}}{\mathrm{x}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})}},$$

(2)

To improve data quality, an ML-based feature enhancement technique is introduced:

K-means adaptive discretization. For a continuous variable ${\mathrm{z}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})$, adaptive discretization is implemented via the K-means clustering method:

$${\mathrm{C}}_{\mathrm{j}}(\mathrm{t})=\mathrm{K}\mathrm{M}\mathrm{e}\mathrm{a}\mathrm{n}\mathrm{s}\left({\mathrm{z}}_{\cdot \mathrm{j}}(\mathrm{t}),\mathrm{k}\right),$$

(3)

where $\mathrm{k}$ is the number of clusters. The optimal number of clusters is determined on the basis of the silhouette coefficient.

Autoencoder-based feature extraction: Latent representation learning of the features of the indicators is achieved through a deep autoencoder as follows:

$${\widehat{\mathrm{z}}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})=\mathrm{E}\mathrm{n}\mathrm{c}\mathrm{o}\mathrm{d}\mathrm{e}\mathrm{r}\left({\mathrm{z}}_{\mathrm{i}\mathrm{j}}(\mathrm{t});{\theta }_{\mathrm{e}}\right),$$

(4)

$${\tilde{\mathrm{z}}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})=\mathrm{D}\mathrm{e}\mathrm{c}\mathrm{o}\mathrm{d}\mathrm{e}\mathrm{r}\left({\widehat{\mathrm{z}}}_{\mathrm{i}\mathrm{j}}(\mathrm{t});{\theta }_{\mathrm{d}}\right),$$

(5)

where ${\theta }_{\mathrm{e}}$ and ${\theta }_{\mathrm{d}}$ are the encoder and decoder parameters, respectively. Training is conducted by minimizing the reconstruction error $(\mathcal{L}={\displaystyle \frac{1}{\mathrm{n}}}{{\displaystyle {\sum }_{\mathrm{i}=1}^{\mathrm{n}}}‖{\mathrm{z}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})-{\tilde{\mathrm{z}}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})‖}^2)$.

Calculation of indicator proportions.

The proportions of standardized indicator values are calculated to convert absolute values into relative contributions [28]:

$${\mathrm{p}}_{\mathrm{i}\mathrm{j}}\left(\mathrm{t}\right)={\displaystyle \frac{{\widehat{\mathrm{z}}}_{\mathrm{i}\mathrm{j}}\left(\mathrm{t}\right)}{{{\displaystyle {\sum }_{\mathrm{i}=1}^{\mathrm{m}}}\widehat{\mathrm{z}}}_{\mathrm{i}\mathrm{j}}\left(\mathrm{t}\right)}},\mathrm{j}=\mathrm{1,2},\dots ,\mathrm{n},$$

(6)

where $\mathrm{m}$ is the number of evaluation objects, and $\mathrm{n}$ denotes the number of indicators.

Calculation of information entropy.

For each indicator j, its information entropy at time t can be calculated as follows [29]:

$${\mathrm{e}}_{\mathrm{j}}\left(\mathrm{t}\right)=-\mathrm{k}{\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{m}}}{\mathrm{p}}_{\mathrm{i}\mathrm{j}}\left(\mathrm{t}\right)\ln {\mathrm{p}}_{\mathrm{i}\mathrm{j}}\left(\mathrm{t}\right),\mathrm{k}={\displaystyle \frac{1}{\ln \mathrm{m}}}$$

(7)

A smaller value of ${\mathrm{e}}_{\mathrm{j}}(\mathrm{t})$ indicates greater indicator variability and thus a higher information content; therefore, a higher weight should be assigned during evaluation [30].

Calculation of the difference coefficient.

The difference coefficient is defined as follows:

$${\mathrm{d}}_{\mathrm{j}}(\mathrm{t})=1-{\mathrm{e}}_{\mathrm{j}}(\mathrm{t}),$$

(8)

Calculation of dynamic weights.

The dynamic weight of each indicator at time t can be obtained as follows [31]:

$${\mathrm{w}}_{\mathrm{j}}(\mathrm{t})={\displaystyle \frac{{\mathrm{d}}_{\mathrm{j}}(\mathrm{t})}{{\displaystyle {\sum }_{\mathrm{j}=1}^{\mathrm{n}}}{\mathrm{d}}_{\mathrm{j}}(\mathrm{t})}},$$

(9)

Thus, the time-varying indicator weight sequence $\{{\mathrm{w}}_{\mathrm{j}}(\mathrm{t}){\}}_{\mathrm{t}=1}^{\mathrm{T}}$ is obtained.

Composite evaluation value.

The composite score of object i at time t can be calculated by combining the weights and standardized values to achieve dynamic comprehensive evaluation [32]:

$${\mathrm{S}}_{\mathrm{i}}(\mathrm{t})={\displaystyle {\sum }_{\mathrm{j}=1}^{\mathrm{n}}}{\mathrm{w}}_{\mathrm{j}}(\mathrm{t})\cdot {\mathrm{z}}_{\mathrm{i}\mathrm{j}}(\mathrm{t}),$$

(10)

2.4.3. Construction of the Enhanced BN

This method is notably dynamic, objective, and scalable. Its dynamic nature is reflected in the variability in weights over time, which facilitates the accurate representation of differences in the importance of each indicator to the system across different periods, thereby providing greater adaptability to system development and evolution [8,33]. The objectivity of this method lies in the calculation of weights based solely on data variability, and there is no need to introduce subjective weighting, thus greatly increasing the credibility and reliability of the evaluation results [34,35]. Finally, its scalability is manifested in its ability to be organically integrated with methods such as the gray relational analysis and Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS) methods [5,36], allowing it to fulfill an important role in risk evaluation and prediction and providing more comprehensive and in-depth support for the analysis of and decision-making in complex systems [37].

(1)

Network framework design

The BN is a directed acyclic graph model in which nodes represent variables and edges represent dependency relationships, facilitating the quantification of uncertainty and causality [38]. The input nodes of the improved BN established in this study include the number of policies, the number of extreme weather events, and socioeconomic impact indicators (such as tourism and wetland area), while K-means clustering nodes are introduced to identify latent patterns in the data, with the goal of quantifying the combined effect of these factors on the eco-environmental quality of islands.

The network structure was designed as a five-layer architecture that includes (1) a bottom feature layer, where original indicators are input and dimensionality is reduced via principal component analysis (PCA) [39]; (2) a feature enhancement layer, where deep features are extracted using an autoencoder [40]; (3) an attention layer, which learns the importance weights of different features [41]; (4) an inference layer, where the BN performs probabilistic inference; (5) an output layer, which provides the composite score and uncertainty estimation.

(2)

Network structure learning

Data-driven regularization methods were employed for network structure learning to avoid overfitting and increase model generalizability [42]. Let the network structure be denoted as $\mathrm{G}$ and the data be denoted as $\mathrm{D}$. Then, the optimal network structure can be obtained by maximizing the regularization likelihood function, which can be expressed as

$${\mathrm{G}}^{\mathrm{*}}=\arg \underset{\mathrm{G}}{\max }\left\{\ln \mathrm{P}(\mathrm{D}|\mathrm{G})-\lambda \cdot \Omega (\mathrm{G})-\mu \cdot \Psi (\mathrm{G})\right\},$$

(11)

where $\Omega (\mathrm{G})$ is the penalty term for network complexity, $\Psi (\mathrm{G})$ is the penalty term for ML component complexity, and $\lambda$ and $\mu$ are regularization coefficients [43].

(3)

Parameter learning and model integration

After the network structure was determined, conditional probability distribution learning was conducted using ensemble learning combined with multiple classification models (random forest, gradient boosting tree, and naive Bayesian models) [44].

The prediction probability of the ensemble model can be expressed as follows:

$$\hat{\mathrm{P}}({\mathrm{X}}_{\mathrm{i}}|\mathrm{P}\mathrm{a}({\mathrm{X}}_{\mathrm{i}}))={\displaystyle \frac{1}{\mathrm{M}}}{\displaystyle {\sum }_{\mathrm{m}=1}^{\mathrm{M}}}{\omega }_{\mathrm{m}}{\mathrm{P}}_{\mathrm{m}}({\mathrm{X}}_{\mathrm{i}}|\mathrm{P}\mathrm{a}({\mathrm{X}}_{\mathrm{i}})),$$

(12)

where $\mathrm{M}$ is the number of base models, ${\mathrm{P}}_{\mathrm{m}}$ is the prediction probability of the $\mathrm{m}$-th model, and ${\omega }_{\mathrm{m}}$ is the model weight, which is determined via cross-validation.

For small-sample data, Bayesian estimation combined with Laplace smoothing can be used:

$$\hat{\mathrm{P}}({\mathrm{X}}_{\mathrm{i}}|\mathrm{P}\mathrm{a}({\mathrm{X}}_{\mathrm{i}}))={\displaystyle \frac{\mathrm{N}({\mathrm{X}}_{\mathrm{i}},\mathrm{P}\mathrm{a}({\mathrm{X}}_{\mathrm{i}}))+\alpha }{\mathrm{N}(\mathrm{P}\mathrm{a}({\mathrm{X}}_{\mathrm{i}}))+\alpha \cdot |{\mathrm{X}}_{\mathrm{i}}|}},$$

(13)

where $\mathrm{N}(\cdot )$ denotes the sample frequency, and $\alpha$ is the smoothing coefficient (usually set to 1).

(4)

Fusion mechanism of time-varying factors

Combined with the TVEWM results, the dynamic weights of the indicators $\{{\mathrm{w}}_{\mathrm{j}}(\mathrm{t})\}$ were incorporated into network inference to obtain the time-dependent conditional probability [45]:

Attention-weighted fusion can be expressed as follows:

$$\mathrm{P}({\mathrm{X}}_{\mathrm{i}}(\mathrm{t})|\mathrm{P}\mathrm{a}({\mathrm{X}}_{\mathrm{i}})(\mathrm{t}))\propto {\displaystyle {\sum }_{\mathrm{j}=1}^{\mathrm{n}}}{\alpha }_{\mathrm{i}\mathrm{j}}(\mathrm{t}){\mathrm{w}}_{\mathrm{j}}(\mathrm{t})\cdot \mathrm{P}({\mathrm{X}}_{\mathrm{i}}|\mathrm{P}\mathrm{a}({\mathrm{X}}_{\mathrm{i}})),$$

(14)

where ${\alpha }_{\mathrm{i}\mathrm{j}}(\mathrm{t})$ is the attention weight, which can be obtained through attention mechanism-based learning as follows:

$${\alpha }_{\mathrm{i}\mathrm{j}}(\mathrm{t})={\displaystyle \frac{\exp \left(\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}({\mathrm{X}}_{\mathrm{i}},{\mathrm{w}}_{\mathrm{j}}(\mathrm{t}))\right)}{{\displaystyle {\sum }_{\mathrm{j}=1}^{\mathrm{n}}}\exp \left(\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}({\mathrm{X}}_{\mathrm{i}},{\mathrm{w}}_{\mathrm{j}}(\mathrm{t}))\right)}},$$

(15)

To extend the DL time series, a long short-term memory (LSTM) network is introduced to capture time series patterns [46,47]:

$${\mathrm{h}}_{\mathrm{t}}=\mathrm{L}\mathrm{S}\mathrm{T}\mathrm{M}({\mathrm{x}}_{\mathrm{t}},{\mathrm{h}}_{\mathrm{t}-1};\theta ),$$

(16)

$$\mathrm{P}({\mathrm{X}}_{\mathrm{i}}(\mathrm{t})|\mathrm{P}\mathrm{a}({\mathrm{X}}_{\mathrm{i}})(\mathrm{t}))=\mathrm{f}({\mathrm{h}}_{\mathrm{t}},{\mathrm{w}}_{\mathrm{j}}(\mathrm{t});\mathsf{\varphi }),$$

(17)

where ${\mathrm{h}}_{\mathrm{t}}$ denotes the hidden state at time t, and $\theta$ and $\mathsf{\varphi }$ are model parameters. This coupling mechanism enables the network to reflect the temporal evolution of the relationships among variables [48].

(5)

Uncertainty quantification

The uncertainty in the prediction results was quantified through the use of bootstrap resampling and Monte Carlo dropout techniques [49,50].

The bootstrap confidence interval can be expressed as follows:

$${\mathrm{C}\mathrm{I}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})={\overline{\mathrm{y}}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})\pm \mathrm{z}\cdot {\displaystyle \frac{{\sigma }_{\mathrm{i}\mathrm{j}}(\mathrm{t})}{\sqrt{\mathrm{B}}}},$$

(18)

where ${\overline{\mathrm{y}}}_{\mathrm{i}\mathrm{j}}(\mathrm{t})$ is the mean predicted value obtained from B bootstrap samples, ${\sigma }_{\mathrm{i}\mathrm{j}}(\mathrm{t})$ denotes the standard deviation, and $\mathrm{z}$ is the z value corresponding to the confidence level.

The Monte Carlo dropout technique was applied repeatedly at the inference stage to calculate the prediction variance:

$$\mathrm{V}\mathrm{a}\mathrm{r}[\mathrm{y}]={\displaystyle \frac{1}{\mathrm{T}}}{\displaystyle {\sum }_{\mathrm{t}=1}^{\mathrm{T}}}{\widehat{\mathrm{y}}}_{\mathrm{t}}^2-{\left({\displaystyle \frac{1}{\mathrm{T}}}{{\displaystyle {\sum }_{\mathrm{t}=1}^{\mathrm{T}}}\widehat{\mathrm{y}}}_{\mathrm{t}}\right)}^2,$$

(19)

3. Results

3.1. Indicator Weights and Dynamic Influences

To systematically elucidate the dynamic characteristics of changes in the natural environment, anthropogenic activities, and policy responses in the study area from 2014–2024, the time-varying weights of the evaluation indicators were analyzed. The weights of the selected indicators over time are shown in Figure 3. Detailed indicator factor data for each year are shown in

Table A1. Weight results of each year.

Year	Index	Ultimate Weight	λ Values	Entropy Weight	Degree of Confidence
2014	NDVI	0.10326476899109567	0.8836501040086274	1.0000000000621334	0.2261882165916701
2014	NLR	0.09884340724206182	0.8458159345473041	1.0000000000621334	0.2563681889703701
2014	CWCR	0.034197829811112586	0.29263529241095476	1.0000000000621334	0.6750775861888062
2014	CLRP	0.1047224642887273	0.8961237928954553	1.0000000000621334	0.24615490471963086
2014	EPI	0.06776369718088614	0.5798627997421121	1.0000000000621334	0.39739348081833265
2014	ALR	0.10378521880376075	0.8881036609727885	1.0000000000621334	0.23303394052086715
2014	NLM	0.10792568105259093	0.9235341367543017	1.0000000000621334	0.23804409676049978
2014	WQGR	0.10491004276262257	0.8977289263750221	1.0000000000621334	0.2266042402160843
2014	WRTR	0.0655519626054577	0.5609366983552457	1.0000000000621334	0.7939125663929204
2014	TD	0.0708018525790658	0.6058606919547729	1.0000000000621334	0.39291023223961924
2014	AERP	0.030872474518950893	0.26417979322673935	1.0000000000621334	0.7174967718947158
2014	IPD	0.035141864205216755	0.3007135179150452	1.0000000000621334	0.6518354325510464
2014	PN	0.026571620967206087	0.2273768281355873	1.0000000000621334	0.7564875322564475
2014	EWFC	0.04564711499124503	0.3906083197957433	1.0000000000621334	0.6085437872015257
2015	NDVI	0.1076468027202854	0.9517666958015236	1.0000000000621334	0.23460862584733155
2015	NLR	0.09764318965620232	0.8633190548918405	1.0000000000621334	0.25666225966209527
2015	CWCR	0.03352756071124159	0.29643626071589824	1.0000000000621334	0.6527283108678046
2015	CLRP	0.10418326334644286	0.9211435714511764	1.0000000000621334	0.2230258378634568
2015	EPI	0.05915262823026706	0.5230020780551051	1.0000000000621334	0.44747496943130594
2015	ALR	0.10221687522944961	0.9037576141028922	1.0000000000621334	0.2210690150882538
2015	NLM	0.09821377203555444	0.8683638986972775	1.0000000000621334	0.23709785618412837
2015	WQGR	0.09703091693690441	0.8579056030451581	1.0000000000621334	0.24565977151418278
2015	WRTR	0.05825533754106312	0.5150686199973102	1.0000000000621334	0.6978080515593921
2015	TD	0.06541699233243113	0.5783888891089074	1.0000000000621334	0.4212337529336521
2015	AERP	0.029674370650203393	0.2623680127046779	1.0000000000621334	0.7103613759910503
2015	IPD	0.03336082708777242	0.2949620737834668	1.0000000000621334	0.6644174013485131
2015	PN	0.03126217752317495	0.27640671764407043	1.0000000000621334	0.6692654985414159
2015	EWFC	0.08241528599900742	0.728680485221985	1.0000000000621334	0.6119126982502988
2016	NDVI	0.1076468027202854	0.9517666958015236	1.0000000000621334	0.23460862584733155
2016	NLR	0.09764318965620232	0.8633190548918405	1.0000000000621334	0.25666225966209527
2016	CWCR	0.03352756071124159	0.29643626071589824	1.0000000000621334	0.6527283108678046
2016	CLRP	0.10418326334644286	0.9211435714511764	1.0000000000621334	0.2230258378634568
2016	EPI	0.05915262823026706	0.5230020780551051	1.0000000000621334	0.44747496943130594
2016	ALR	0.10221687522944961	0.9037576141028922	1.0000000000621334	0.2210690150882538
2016	NLM	0.09821377203555444	0.8683638986972775	1.0000000000621334	0.23709785618412837
2016	WQGR	0.09703091693690441	0.8579056030451581	1.0000000000621334	0.24565977151418278
2016	WRTR	0.05825533754106312	0.5150686199973102	1.0000000000621334	0.6978080515593921
2016	TD	0.06541699233243113	0.5783888891089074	1.0000000000621334	0.4212337529336521
2016	AERP	0.029674370650203393	0.2623680127046779	1.0000000000621334	0.7103613759910503
2016	IPD	0.03336082708777242	0.2949620737834668	1.0000000000621334	0.6644174013485131
2016	PN	0.03126217752317495	0.27640671764407043	1.0000000000621334	0.6692654985414159
2016	EWFC	0.08241528599900742	0.728680485221985	1.0000000000621334	0.6119126982502988
2017	NDVI	0.10326476899109567	0.8836501040086274	1.0000000000621334	0.2261882165916701
2017	NLR	0.09884340724206182	0.8458159345473041	1.0000000000621334	0.2563681889703701
2017	CWCR	0.034197829811112586	0.29263529241095476	1.0000000000621334	0.6750775861888062
2017	CLRP	0.1047224642887273	0.8961237928954553	1.0000000000621334	0.24615490471963086
2017	EPI	0.06776369718088614	0.5798627997421121	1.0000000000621334	0.39739348081833265
2017	ALR	0.10378521880376075	0.8881036609727885	1.0000000000621334	0.23303394052086715
2017	NLM	0.10792568105259093	0.9235341367543017	1.0000000000621334	0.23804409676049978
2017	WQGR	0.10491004276262257	0.8977289263750221	1.0000000000621334	0.2266042402160843
2017	WRTR	0.0655519626054577	0.5609366983552457	1.0000000000621334	0.7939125663929204
2017	TD	0.0708018525790658	0.6058606919547729	1.0000000000621334	0.39291023223961924
2017	AERP	0.030872474518950893	0.26417979322673935	1.0000000000621334	0.7174967718947158
2017	IPD	0.035141864205216755	0.3007135179150452	1.0000000000621334	0.6518354325510464
2017	PN	0.026571620967206087	0.2273768281355873	1.0000000000621334	0.7564875322564475
2017	EWFC	0.04564711499124503	0.3906083197957433	1.0000000000621334	0.6085437872015257
2018	NDVI	0.1076468027202854	0.9517666958015236	1.0000000000621334	0.23460862584733155
2018	NLR	0.09764318965620232	0.8633190548918405	1.0000000000621334	0.25666225966209527
2018	CWCR	0.03352756071124159	0.29643626071589824	1.0000000000621334	0.6527283108678046
2018	CLRP	0.10418326334644286	0.9211435714511764	1.0000000000621334	0.2230258378634568
2018	EPI	0.05915262823026706	0.5230020780551051	1.0000000000621334	0.44747496943130594
2018	ALR	0.10221687522944961	0.9037576141028922	1.0000000000621334	0.2210690150882538
2018	NLM	0.09821377203555444	0.8683638986972775	1.0000000000621334	0.23709785618412837
2018	WQGR	0.09703091693690441	0.8579056030451581	1.0000000000621334	0.24565977151418278
2018	WRTR	0.05825533754106312	0.5150686199973102	1.0000000000621334	0.6978080515593921
2018	TD	0.06541699233243113	0.5783888891089074	1.0000000000621334	0.4212337529336521
2018	AERP	0.029674370650203393	0.2623680127046779	1.0000000000621334	0.7103613759910503
2018	IPD	0.03336082708777242	0.2949620737834668	1.0000000000621334	0.6644174013485131
2018	PN	0.03126217752317495	0.27640671764407043	1.0000000000621334	0.6692654985414159
2018	EWFC	0.08241528599900742	0.728680485221985	1.0000000000621334	0.6119126982502988
2019	NDVI	0.10742271471492268	0.9653443141868956	1.0000000000621334	0.255627544173426
2019	NLR	0.09369287193279492	0.841962348840177	1.0000000000621334	0.2426857870121143
2019	CWCR	0.03450728968100582	0.3100965748255441	1.0000000000621334	0.6430229320779028
2019	CLRP	0.10705007168229182	0.9619955919569072	1.0000000000621334	0.23310000969945713
2019	EPI	0.06221578813062465	0.5590964395560173	1.0000000000621334	0.41611351405268715
2019	ALR	0.10114266327505013	0.9089092113655737	1.0000000000621334	0.2141749441804227
2019	NLM	0.09880865363813025	0.88793484911605	1.0000000000621334	0.23545940929854944
2019	WQGR	0.09870777945636089	0.8870283525889632	1.0000000000621334	0.29108569671455065
2019	WRTR	0.05228639259643872	0.4698668427461016	1.0000000000621334	0.6664736851392777
2019	TD	0.06661868016195881	0.5986626225528503	1.0000000000621334	0.42126616768371833
2019	AERP	0.03296082965504826	0.29619945448989443	1.0000000000621334	0.6813745284907878
2019	IPD	0.032975286632821385	0.29632937078674704	1.0000000000621334	0.6651346389663706
2019	PN	0.03211008278468567	0.28855429623545736	1.0000000000621334	0.700172259011594
2019	EWFC	0.07950089565786593	0.7144274635001879	1.0000000000621334	0.49695288783714436
2020	NDVI	0.1076468027202854	0.9517666958015236	1.0000000000621334	0.23460862584733155
2020	NLR	0.09764318965620232	0.8633190548918405	1.0000000000621334	0.25666225966209527
2020	CWCR	0.03352756071124159	0.29643626071589824	1.0000000000621334	0.6527283108678046
2020	CLRP	0.10418326334644286	0.9211435714511764	1.0000000000621334	0.2230258378634568
2020	EPI	0.05915262823026706	0.5230020780551051	1.0000000000621334	0.44747496943130594
2020	ALR	0.10221687522944961	0.9037576141028922	1.0000000000621334	0.2210690150882538
2020	NLM	0.09821377203555444	0.8683638986972775	1.0000000000621334	0.23709785618412837
2020	WQGR	0.09703091693690441	0.8579056030451581	1.0000000000621334	0.24565977151418278
2020	WRTR	0.05825533754106312	0.5150686199973102	1.0000000000621334	0.6978080515593921
2020	TD	0.06541699233243113	0.5783888891089074	1.0000000000621334	0.4212337529336521
2020	AERP	0.029674370650203393	0.2623680127046779	1.0000000000621334	0.7103613759910503
2020	IPD	0.03336082708777242	0.2949620737834668	1.0000000000621334	0.6644174013485131
2020	PN	0.03126217752317495	0.27640671764407043	1.0000000000621334	0.6692654985414159
2020	EWFC	0.08241528599900742	0.728680485221985	1.0000000000621334	0.6119126982502988
2021	NDVI	0.1076468027202854	0.9517666958015236	1.0000000000621334	0.23460862584733155
2021	NLR	0.09764318965620232	0.8633190548918405	1.0000000000621334	0.25666225966209527
2021	CWCR	0.03352756071124159	0.29643626071589824	1.0000000000621334	0.6527283108678046
2021	CLRP	0.10418326334644286	0.9211435714511764	1.0000000000621334	0.2230258378634568
2021	EPI	0.05915262823026706	0.5230020780551051	1.0000000000621334	0.44747496943130594
2021	ALR	0.10221687522944961	0.9037576141028922	1.0000000000621334	0.2210690150882538
2021	NLM	0.09821377203555444	0.8683638986972775	1.0000000000621334	0.23709785618412837
2021	WQGR	0.09703091693690441	0.8579056030451581	1.0000000000621334	0.24565977151418278
2021	WRTR	0.05825533754106312	0.5150686199973102	1.0000000000621334	0.6978080515593921
2021	TD	0.06541699233243113	0.5783888891089074	1.0000000000621334	0.4212337529336521
2021	AERP	0.029674370650203393	0.2623680127046779	1.0000000000621334	0.7103613759910503
2021	IPD	0.03336082708777242	0.2949620737834668	1.0000000000621334	0.6644174013485131
2021	PN	0.03126217752317495	0.27640671764407043	1.0000000000621334	0.6692654985414159
2021	EWFC	0.08241528599900742	0.728680485221985	1.0000000000621334	0.6119126982502988
2022	NDVI	0.10326476899151626	0.8836501040086274	1.0000000000621334	0.2261882165916701
2022	NLR	0.09884340724246442	0.8458159345473041	1.0000000000621334	0.2563681889703701
2022	CWCR	0.03419782981125188	0.29263529241095476	1.0000000000621334	0.6750775861888062
2022	CLRP	0.10472246428915384	0.8961237928954553	1.0000000000621334	0.24615490471963086
2022	EPI	0.06776369718116215	0.5798627997421121	1.0000000000621334	0.39739348081833265
2022	ALR	0.10378521880418347	0.8881036609727885	1.0000000000621334	0.23303394052086715
2022	NLM	0.10792568105303052	0.9235341367543017	1.0000000000621334	0.23804409676049978
2022	WQGR	0.10491004276304988	0.8977289263750221	1.0000000000621334	0.2266042402160843
2022	WRTR	0.06555196260165175	0.5609366983552457	1	0.7939125663929204
2022	TD	0.07080185257935417	0.6058606919547729	1.0000000000621334	0.39291023223961924
2022	AERP	0.030872474519076636	0.26417979322673935	1.0000000000621334	0.7174967718947158
2022	IPD	0.03514186420535989	0.3007135179150452	1.0000000000621334	0.6518354325510464
2022	PN	0.026571620967314316	0.2273768281355873	1.0000000000621334	0.7564875322564475
2022	EWFC	0.045647114991430954	0.3906083197957433	1.0000000000621334	0.6085437872015257
2023	NDVI	0.10326476899109566	0.8836501040086274	1	0.2261882165916701
2023	NLR	0.09884340724206182	0.8458159345473041	1	0.2563681889703701
2023	CWCR	0.03419782981111258	0.29263529241095476	1	0.6750775861888062
2023	CLRP	0.1047224642887273	0.8961237928954553	1	0.24615490471963086
2023	EPI	0.06776369718088614	0.5798627997421121	1	0.39739348081833265
2023	ALR	0.10378521880376075	0.8881036609727885	1	0.23303394052086715
2023	NLM	0.10792568105259091	0.9235341367543017	1	0.23804409676049978
2023	WQGR	0.10491004276262257	0.8977289263750221	1	0.2266042402160843
2023	WRTR	0.0655519626054577	0.5609366983552457	1	0.7939125663929204
2023	TD	0.0708018525790658	0.6058606919547729	1	0.39291023223961924
2023	AERP	0.030872474518950893	0.26417979322673935	1	0.7174967718947158
2023	IPD	0.035141864205216755	0.3007135179150452	1	0.6518354325510464
2023	PN	0.026571620967206083	0.2273768281355873	1	0.7564875322564475
2023	EWFC	0.04564711499124503	0.3906083197957433	1	0.6085437872015257
2024	NDVI	0.10326476899109566	0.8836501040086274	1	0.2261882165916701
2024	NLR	0.09884340724206182	0.8458159345473041	1	0.2563681889703701
2024	CWCR	0.03419782981111258	0.29263529241095476	1	0.6750775861888062
2024	CLRP	0.1047224642887273	0.8961237928954553	1	0.24615490471963086
2024	EPI	0.06776369718088614	0.5798627997421121	1	0.39739348081833265
2024	ALR	0.10378521880376075	0.8881036609727885	1	0.23303394052086715
2024	NLM	0.10792568105259091	0.9235341367543017	1	0.23804409676049978
2024	WQGR	0.10491004276262257	0.8977289263750221	1	0.2266042402160843
2024	WRTR	0.0655519626054577	0.5609366983552457	1	0.7939125663929204
2024	TD	0.0708018525790658	0.6058606919547729	1	0.39291023223961924
2024	AERP	0.030872474518950893	0.26417979322673935	1	0.7174967718947158
2024	IPD	0.035141864205216755	0.3007135179150452	1	0.6518354325510464
2024	PN	0.026571620967206083	0.2273768281355873	1	0.7564875322564475
2024	EWFC	0.04564711499124503	0.3906083197957433	1	0.6085437872015257

.

The NDVI and NLR indicators consistently maintain relatively high weights around 0.10, identifying them as key factors in the assessment. In contrast, the weight of InEWFC exhibits significant fluctuations, peaking in 2015 and 2017 before declining and stabilizing in subsequent years, reflecting the periodic influence of external factors on its importance. Indicators such as TD and IPD remain at relatively low weights around 0.03 over the long term, contributing minimally to the assessment outcomes. The weights of WQGR and WRTR are generally concentrated near 0.10, with minor fluctuations, indicating relative stability. EPI maintains a moderate weight around 0.07, exhibiting only slight variation, whereas CWCR consistently remains at a low level of 0.03, representing a secondary indicator within this set. Overall, these patterns illustrate both the long-term dominance of a few core indicators in island ecological assessment and the periodic adjustments in the importance of specific indicators in response to external environmental conditions, while the long-term stability of low-weight indicators suggests their consistently auxiliary role in the evaluation system.

Building on this, Figure 4 provides a further statistical summary of the data, presented through line charts, bar charts, and heatmaps. Additionally, the confidence of each indicator is analyzed through frequency distribution.

The upper-left time series of weights indicates that the NDVI, NLR, and CLRP indicators consistently maintain weights around 0.10 over the long term, identifying them as key factors in the assessment. In contrast, the periodic fluctuations in the EPI weight reflect its dynamic adjustment in response to external factors. The upper-right table of weights for 2024 demonstrates that the assessment of that year was dominated by the four core indicators with weights exceeding 0.10, while the subsequent indicators exhibited progressively lower weights, highlighting significant differences in importance among the indicators. The lower-left distribution of confidence shows a mean of 0.453 and a median of 0.419, indicating a moderate overall level; although some variability exists, the results remain within an acceptable reliability range. The lower-right heatmap of weights illustrates the spatiotemporal characteristics of the indicators from 2020 to 2024, with a slight increase in NDVI weight after 2022, corresponding to its heightened sensitivity to later external factors.

Overall, these features demonstrate both the long-term dominance of a few core indicators in island ecological assessment and the dynamic influence of external policies and human activities on the importance of specific indicators. The moderate level of confidence further indicates that, although some uncertainty exists, the stability of the core indicators ensures the overall reliability of the assessment. The stability of indicator weights is further analyzed, resulting in Figure 5.

Indicators with relatively high average weights exhibited trend slopes close to zero, demonstrating that their weights remained constant over the long term, although the weights of some indicators slightly increased. Indicators with moderate average weights exhibited negative trend slopes with relatively large absolute values, suggesting a clear downward trend in weights and lower stability. Indicators with low average weights demonstrated trend slopes close to zero and exhibited minimal weight variability, suggesting high stability despite their low weight proportions. Most of the remaining indicators with moderately high average weights also revealed trend slopes approaching zero, with only minor weight fluctuations. Overall, both the high-weight core indicators and low-weight secondary indicators in the evaluation system exhibited high stability, whereas only a small number of medium-weight indicators exhibited notable temporal changes in importance, reflecting a pattern in which the weights of most indicators remained constant while only those of a few indicators changed over time.

3.2. Composite Ecological Score

To compare the differences between inhabited and uninhabited islands under eco-environmental and social pressures and to analyze the underlying causes, a trend analysis of the nine typical islands was conducted on the basis of composite score data from 2014 to 2024 (Figure 6), while trend subanalysis (Figure 7) and statistical analysis (Figure 8) of the scores of each island revealed the specific performance and structural differences among these islands in terms of multidimensional indicators.

The scores of each island are shown in the form of a line chart (Figure 6). Detailed comprehensive score data for each year are shown in

Table A2. The comprehensive score of islands in each year.

Year	Island_Name	Final_Score
2014	ping island	8.77128324
2015	ping island	8.07751539
2016	ping island	8.36107258
2017	ping island	9.592233859
2018	ping island	7.889544914
2019	ping island	7.763097414
2020	ping island	7.586484146
2021	ping island	7.486420204
2022	ping island	8.515122964
2023	ping island	8.44852282
2024	ping island	8.20088909
2014	cheniushan island	0.681207108
2015	cheniushan island	0.635173428
2016	cheniushan island	0.4387417
2017	cheniushan island	0.673483453
2018	cheniushan island	0.316069935
2019	cheniushan island	0.567563247
2020	cheniushan island	−0.521107096
2021	cheniushan island	−0.733606285
2022	cheniushan island	0.002846685
2023	cheniushan island	−0.088532015
2024	cheniushan island	−0.212851202
2014	niubei island	10.03508162
2015	niubei island	8.930007496
2016	niubei island	9.223616235
2017	niubei island	9.913550114
2018	niubei island	8.816118172
2019	niubei island	8.790287172
2020	niubei island	8.908447336
2021	niubei island	8.614363529
2022	niubei island	9.583917869
2023	niubei island	9.94116403
2024	niubei island	9.793871881
2014	yangshan island	24.04916848
2015	yangshan island	22.90044286
2016	yangshan island	25.49454069
2017	yangshan island	29.53571868
2018	yangshan island	23.15649441
2019	yangshan island	27.40070206
2020	yangshan island	25.19199068
2021	yangshan island	23.55117565
2022	yangshan island	25.83004544
2023	yangshan island	25.04944605
2024	yangshan island	25.22844371
2014	yangguang island	−3.14416895
2015	yangguang island	−3.641865886
2016	yangguang island	−4.134575404
2017	yangguang island	−2.13046266
2018	yangguang island	−5.315962583
2019	yangguang island	−11.81031463
2020	yangguang island	−5.381650557
2021	yangguang island	−6.175938296
2022	yangguang island	−6.080046131
2023	yangguang island	−5.749496534
2024	yangguang island	−4.069144701
2014	zhu island	10.25465269
2015	zhu island	9.432631847
2016	zhu island	9.645120535
2017	zhu island	9.229743389
2018	zhu island	8.420940469
2019	zhu island	9.12919566
2020	zhu island	9.046928676
2021	zhu island	14.27319641
2022	zhu island	10.69788107
2023	zhu island	10.90354556
2024	zhu island	11.05819411
2014	qinshan island	9.483667442
2015	qinshan island	8.892718612
2016	qinshan island	8.621008128
2017	qinshan island	9.97848308
2018	qinshan island	8.851652025
2019	qinshan island	9.623146596
2020	qinshan island	8.991758346
2021	qinshan island	9.345622933
2022	qinshan island	10.39173829
2023	qinshan island	10.63463314
2024	qinshan island	10.57600261
2014	lian island	96.67048381
2015	lian island	98.89347323
2016	lian island	73.85756681
2017	lian island	78.96319746
2018	lian island	51.69950711
2019	lian island	112.3281059
2020	lian island	31.75991213
2021	lian island	17.30334128
2022	lian island	15.65914867
2023	lian island	62.11976218
2024	lian island	63.52148655
2014	dashan island	9.988146708
2015	dashan island	9.132077507
2016	dashan island	9.131997277
2017	dashan island	10.30133184
2018	dashan island	9.444601554
2019	dashan island	9.840251442
2020	dashan island	9.361453731
2021	dashan island	9.796729856
2022	dashan island	10.86018865
2023	dashan island	10.97952689
2024	dashan island	10.92501557

.

The scores of certain islands, such as Lian Island (gray line), fluctuated considerably. Notably, the scores remained high (near 100) from 2014–2015, followed by a decrease, a notable peak in 2019, and then a sharp decrease, indicating that the ecological conditions of these islands were greatly influenced by external interventions or sudden events. The scores of the uninhabited islands were mostly concentrated at approximately 10, with small overall fluctuations, reflecting relatively stable ecological conditions in the absence of anthropogenic disturbance but with a moderate quality level. In contrast, the scores of the inhabited islands were generally low, with the score of Yangguang Island (purple line) remaining below zero for a long period, that of Cheniushan Island (orange line) fluctuating around zero, and that of only Yangshan Island (red line) remaining relatively stable at approximately 20, thus confirming the negative impact of anthropogenic activities on the eco-environmental quality of the islands. Overall, the ecological scores of the uninhabited islands were more stable and generally higher than those of the inhabited islands, while the fluctuations in the scores of some islands corresponded to the influence of specific interventions or environmental events, highlighting the negative correlation between the intensity of anthropogenic activities and ecological quality as well as the short-term effects of ecological interventions.

Changes in the scores of each island over time are shown in Figure 7.

The scores of the uninhabited islands generally occurred within the middle to high range, with mean scores of approximately 8.24 and 9.98 for Ping Island and Dashan Island, respectively. Although there were stage-specific fluctuations, the overall trends remained relatively constant, indicating a strong ecological baseline and stable conditions in the absence of anthropogenic disturbance. The scores of the inhabited islands clearly differed. The scores of Yangshan Island (with a mean of 25.22) greatly fluctuated within the 23–29 range, reflecting a balanced state between tourism development and ecological conservation on this inhabited natural island. The mean score of artificial Yangguang Island was −5.24 and remained within the negative range for a prolonged period, reaching a trough in 2019, which indicates the significant ecological impacts of industrial development. Lian Island, as a unique case, attained a high mean score of 69.09, which sharply increased to 117.33 in 2019, followed by a decrease, corresponding to the short-term effect of intervention measures such as effective ecological restoration. The scores of Cheniushan Island (with a mean of −0.13) remained close to zero, with drastic fluctuations.

The overall characteristics and differences in scores across various dimensions are shown in Figure 8.

The upper-left histogram indicates that the 199 data points have a mean score of 11.96, a median of 9.96, and a standard deviation of 15.57. The scores are concentrated in the 0–20 range, but the large standard deviation suggests a dispersed distribution with significant variability. The multiple outliers observed in the boxplot, corresponding to peaks and high dispersion in the 0–20 range of the density plot, confirm that some islands exhibit scores far above the mean. The annual score trend from 2014 to 2024 shows that the average scores fluctuate slightly around 10, with a trough around 2020. The pronounced variations within the standard deviation range reflect the instability in score dispersion across different years.

To examine the differences between the inhabited and uninhabited islands under ecological and social pressures, the average composite scores of the nine islands were calculated for the 2014–2024 period, as shown in Figure 9.

The ranking of the average composite scores of the islands from 2014 to 2024 clearly revealed different scores among the various types of islands (inhabited and uninhabited islands) as a result of differences in development attributes and ecological management models. Among the inhabited islands, Lian Island ranked first, with a notably high score (63.89). As an inhabited natural island, its synergistic model of tourism development and ecological restoration has effectively maintained high ecological quality. Yangshan Island (25.22), which is also an inhabited natural island, ranked second, benefitting from moderate development that is based on pine forests and fishery resources. Qinshan Island, an inhabited island where ecological and cultural functions are integrated, exhibited a medium score. Yangguang Island, which is an inhabited artificial island, ranked lowest (−5.24), with its industrial warehousing development model exerting significant negative impacts on the ecosystem.

Among the uninhabited islands, the scores of Zhu Island (10.19), Dashan Island (9.98), Niubei Island (9.32), and Pingshan Island (8.24) were clustered at approximately 10 within the medium range, reflecting the stability of their ecological baselines in the absence of anthropogenic disturbance. Although Cheniushan Island is an uninhabited island, it attained a relatively low score (0.16) because of its ecological vulnerability. Overall, the ecological scores of the inhabited natural islands were higher than those of the inhabited artificial islands, while the uninhabited islands exhibited a relatively stable ecological state with moderate scores. The difference in scores intuitively reflects the association between the development type and ecological quality.

4. Discussion

4.1. Advantages of the Proposed Method

The proposed Time-varying Entropy Weight–Bayesian Network (TEW-BN) fusion framework represents not only a technological innovation but also a novel theoretical paradigm for marine ecological assessment. Its core significance lies in overcoming the limitations of conventional static evaluation methods when applied to highly dynamic, nonlinear, and data-sparse systems such as islands. This advantage can be articulated along two dimensions.

First, the framework achieves superior dynamic response capture compared to traditional static models. Conventional island ecological assessments often rely on static weight allocation, which inadequately reflects instantaneous disturbances associated with environmental shocks [51]. By introducing the Time-varying Entropy Weight method, this study identifies an increase in the weight of ecological restoration projects after 2018, aligning closely with the theoretical expectation that assessments of human impacts on marine environments require temporal sensitivity [52]. Furthermore, this finding corroborates recent conclusions that dynamic weighting models more objectively capture the nonlinear recovery trajectories of coastal ecosystems [31]. The integration of Bayesian networks for associating multiple stressors enables this framework to explain the evolution of industrial islands under extreme weather conditions with higher precision than conventional linear models, reinforcing the perspective that Bayesian frameworks can systematically quantify real-time uncertainty propagation under climate change [38].

Second, the framework provides highly robust inference and uncertainty reduction under small-sample conditions. Addressing the “small-sample, high-sparsity” bottleneck in Jiangsu island data, this study innovatively integrates autoencoders with attention mechanisms into the underlying processing module [53]. Empirically, this integrated approach responds to findings that cross-attention-enhanced autoencoder frameworks can accurately reconstruct complex dynamic fields from extremely sparse sensor data, even at sampling rates as low as 0.1% [32]. It further validates the theoretical assertion that multi-layer intelligent system integration can substantially reduce uncertainty in complex environmental models [54]. Data verification indicates that inference uncertainty in this framework is reduced by 62.3% compared with baseline models, with confidence intervals outperforming conventional Bayesian paradigms [55]. This breakthrough not only confirms the predicted high robustness of hybrid machine learning algorithms in small-sample marine community prediction but also provides a transferable technical paradigm for island nations lacking long-term monitoring data [33].

4.2. Implications of Classification-Based Management

4.2.1. Inhabited Islands: Source Control of Anthropogenic Pressure and Threshold-Based Early Warning

For inhabited islands such as Yangshan and Sunshine Islands, ecological scores are relatively low, and in some cases even negative. The primary limiting factors are the proportion of built-up land, the proportion of artificial coastlines, and the excessive density of residents or visitors. This indicates that nearshore island ecosystems are highly susceptible to the dominance and influence of human activity footprints.

These findings directly support the necessity of implementing “total quantity control” for inhabited islands, as stipulated in the Jiangsu Marine Spatial Plan. For instance, Liandao Island has maintained a balance in development intensity through ecological restoration, achieving a score of 63.89, whereas Yangshan Island faces overdevelopment. Consequently, the management of inhabited islands should adopt the principle of integrated ecosystem management, implementing measures such as visitor quotas and red-line thresholds for natural coastline retention, thereby converting restrictive indicators into enforceable regulatory boundaries.

4.2.2. Uninhabited Islands: Resilience Monitoring of Natural Vulnerability and Spatial Isolation

In contrast, uninhabited islands such as Cheniushan and Ping Islands, while achieving higher ecological scores, are primarily constrained by pronounced natural vulnerability and high sensitivity to extreme weather events. In the absence of anthropogenic disturbance, the stability of their ecosystems depends entirely on intrinsic natural conditions, exhibiting limited buffering capacity against climate change and extreme events.

For territorial base points such as Dashan and Cheniushan Islands, limiting factors such as natural erosion and sparse vegetation necessitate a “zero-disturbance” management approach. The findings support the decision in the Jiangsu uninhabited island management program to establish ecological monitoring archives, and recommend the use of numerical simulations to track changes in limiting factors, thereby providing strategic support for spatial security planning.

4.3. Limitations and Future Directions of the Study

This study has four main limitations. First, in terms of sample representativeness, although the analysis was extended to nine typical islands (four inhabited and five uninhabited), it remains concentrated in the Lianyungang sea area and does not cover the southern islands of Jiangsu. This limitation may result in an incomplete characterization of regional variations in the interactions between human activities and ecological responses. Second, regarding the indicator system, dimensions such as marine biodiversity and ecosystem service value were not included, which limits the ability to fully capture the functional integrity and socio-economic value of the island ecosystems. Third, in terms of methodological application, the enhanced Bayesian network requires substantial computational resources, and some deep learning components may be prone to overfitting under small-sample scenarios. Fourth, concerning scenario projection, the model’s inference capacity for unprecedented extreme climate conditions or high-intensity development scenarios is limited, restricting its applicability for medium- to long-term forecasting and simulation.

Based on the above limitations, future research can advance along five directions. First, the study scope and data dimensions should be expanded to cover all four inhabited islands of Jiangsu as well as more representative uninhabited islands. Integrating a land–sea coordinated perspective, cross-system data such as terrestrial pollution sources and oceanic circulation should be incorporated to establish a comprehensive “land–sea–island” assessment framework. Second, the indicator system should be improved by adding dimensions such as marine biodiversity and ecosystem service value, while high-resolution remote sensing and UAV monitoring technologies should be employed to enhance the timeliness and accuracy of indicator data. Third, model adaptability to small-sample data should be optimized by developing deep learning architectures and regularization techniques specifically designed for limited datasets, balancing model complexity and generalization capability. Fourth, the capacity for dynamic projection and scenario analysis should be strengthened by integrating process-based models with reinforcement learning methods to simulate island ecological evolution under varying development intensities, policy interventions, and climate scenarios, thereby providing decision support for medium- to long-term conservation planning. Finally, cross-regional validation of the methodology should be promoted. Comparative analysis with islands in other coastal provinces of China can reveal commonalities and differences in island ecosystems, enhancing the generalizability and transferability of the approach.

5. Conclusions

This study developed a dynamic evaluation framework and conducted an empirical assessment of representative islands in Jiangsu, with the core conclusions summarized as follows:

First, in terms of technical adaptability, a hybrid evaluation model integrating Time-varying Weighting and intelligent enhancement was successfully developed. The results indicate that, under small-sample conditions, the model improves accuracy by 11.3% compared with traditional models, effectively quantifying the dynamic disturbances of island ecosystems caused by policy interventions and extreme weather, and addressing the challenge of overcoming small-sample limitations.

Second, regarding differentiation of limiting factors, the study identified and validated the intrinsic differences in constraints among islands with different development types. Inhabited islands exhibit ecological bottlenecks primarily due to anthropogenic pressures, including expansion of built-up land and visitor overload, showing a “development-driven” decline pattern. In contrast, uninhabited islands are constrained by natural conditions, such as high natural vulnerability and sensitivity to extreme weather, displaying a “naturally vulnerable” fluctuation pattern. This finding provides the most direct scientific basis for differentiated, island-specific conservation strategies.

Third, in supporting spatial planning, the results closely align with the requirements of marine spatial planning. It is recommended that inhabited islands emphasize a dynamic balance between development and restoration, whereas uninhabited islands focus on real-time monitoring of their pristine state and spatial isolation, thereby promoting the sustainable utilization of island resources.

Although the model demonstrates strong performance in quantifying policy interventions and natural disturbances, it remains limited by the spatial distribution of samples, the dimensionality of evaluation indicators, and the capacity for reasoning under extreme scenarios. Future research should aim to construct an integrated “land–sea–island” evaluation framework, further optimize the generalization ability of deep learning architectures under small-sample conditions, and combine reinforcement learning with process-based models to enhance dynamic simulation. Cross-regional comparisons could then reveal common evolutionary patterns of islands in different areas, thereby improving the global applicability and scientific support value of this evaluation paradigm.