Spatiotemporal Patterns of Carbon Storage in Hainan Bamen Bay Mangroves Based on a Decision Tree Classification

Abstract

Mangroves play a vital role in climate change mitigation due to their exceptional carbon sequestration capacity, as a highly productive blue carbon ecosystem. Current research on mangroves in Bamen Bay has been limited to short-term observations, lacking systematic analysis of long-term spatiotemporal dynamics and carbon storage. This study developed a decision tree method integrating SWIR1, NDVI, and NDMI, achieving high-accuracy mangrove mapping. Spatiotemporal dynamics from 2000 to 2020 were analyzed using the dynamic degree model, standard deviation ellipse, centroid model, and landscape pattern indices. Carbon storage was quantified through the InVEST model, grey prediction model, and scenario analysis. The results reveal significant mangrove expansion, with substantial net growth. Spatial aggregation strengthened despite persistent fragmentation, characterized by a shrinking standard deviation ellipse and northeastward centroid migration. Carbon storage increased considerably over the two decades. Under the baseline scenario, carbon storage would continue to grow by mid-century. Among alternative scenarios, the Green Revival scenario achieves the highest carbon storage, outperforming the baseline, while the Hard Preservation scenario achieves slightly above the baseline. The Missed Opportunity and Ecological Collapse scenarios project declines. This study provides a valuable framework for mangrove monitoring, carbon assessment, and ecological restoration, supporting regional conservation and carbon neutrality goals.

1. Introduction

Mangroves are woody plant communities distributed in tropical and subtropical intertidal zones, providing critical ecological services including coastal protection, water purification, climate regulation, and biodiversity conservation [1,2]. As typical coastal blue carbon ecosystems, mangroves play a crucial role in the global carbon cycle despite their limited spatial extent [3]. Despite covering only 0.5% of the world’s coastal area, they account for 10%–15% of coastal sediment carbon storage and contribute significantly to mitigating climate change by sequestering and storing large amounts of atmospheric carbon dioxide over long timescales [4,5]. This exceptional carbon sequestration capacity makes mangrove conservation and restoration a critical nature-based solution for achieving carbon neutrality goals [6,7]. With carbon storage ranging from 79.2 to 242.2 tons per hectare, mangrove ecosystems possess substantial carbon emission reduction value [8]. Consequently, mangrove carbon sinks have emerged as a key focus of blue carbon research, with remote sensing technology providing essential methodological support for blue carbon mapping and quantification [9].

Understanding the spatiotemporal patterns of mangrove carbon storage is essential for effective blue carbon management and climate change mitigation. Carbon storage in mangrove ecosystems varies spatially due to factors such as species composition, forest structure, tidal dynamics, and historical land-use changes [10,11,12]. Temporally, carbon storage fluctuates in response to natural succession, anthropogenic disturbances, and restoration interventions [13]. Quantifying these patterns—how carbon is distributed across space and how it changes over time—provides critical insights for prioritizing conservation areas, setting restoration targets, and assessing the effectiveness of management policies. However, systematic analyses linking long-term spatiotemporal mangrove dynamics to carbon storage outcomes are still scarce, particularly for key regions such as Bamen Bay. Specifically, four critical gaps persist in mangrove carbon storage research. First, most classification methods for long-term mangrove mapping have not systematically evaluated the potential of SWIR1 band integration with moisture-sensitive indices, leading to suboptimal accuracy in distinguishing mangroves from adjacent terrestrial vegetation in complex estuarine environments. Second, while dynamic degree models, standard deviation ellipses, and landscape pattern indices have been applied separately in different regions, an integrated analytical framework that combines these tools to link spatiotemporal dynamics directly to carbon storage outcomes remains undeveloped. Third, carbon storage predictions for mangroves have largely relied on simple area-based extrapolation without incorporating scenario analysis that captures policy and environmental uncertainty. Fourth, for Bamen Bay specifically, despite its status as a high-diversity hotspot, no study has systematically quantified the resulting carbon storage implications under alternative future pathways.

Nevertheless, in recent decades, global mangrove ecosystems have suffered substantial degradation due to the combined impacts of human activities and climate change [14]. Mangroves face a variety of climate threats, including rising sea levels, increased storm intensity, changes in precipitation patterns, rising temperatures, and ocean acidification, which directly damage their viability and carbon sequestration potential [15]. At the same time, continuous human interference, such as coastal infrastructure development and aquaculture expansion, has further exacerbated the risk of ecological degradation [16]. Therefore, systematic long-term monitoring and carbon stock assessment of mangroves have become critically urgent.

In recent years, remote sensing technology has become the primary tool for long-term, large-scale mangrove monitoring due to its advantages in providing synoptic coverage, consistent observations, and repeatable measurements. Researchers worldwide have developed global mangrove databases with high spatial and temporal resolution, achieving remarkable progress in both technological innovation and regional applications [17]. However, most existing methods have not fully exploited the potential of the Short-Wave Infrared 1 (SWIR1) band in feature selection. SWIR1 is sensitive to the water content of vegetation, which can effectively enhance the distinction between mangroves and surrounding vegetation, showing great potential in the accurate identification of mangroves [18]. As a mature spectral index, the Normalized Difference Vegetation Index (NDVI) has shown good performance in mangrove identification and has been successfully applied to the extraction of mangroves in North Halmahera Regency and southern West Lombok [19,20,21]. In addition, the Normalized Difference Moisture Index (NDMI) improves the classification accuracy by enhancing the spectral separability between mangroves and land vegetation. Its effectiveness has been verified in cases such as mangrove extraction in the Marine National Park in Jamnagar, Gulf of Kutch [22,23].

In studies of mangrove spatiotemporal dynamics, multiple models and indices have been widely applied. The dynamic degree model has been used to quantify mangrove change rates in Guangxi’s Zhenzhu Bay over the past three decades [24]. The standard deviation ellipse has effectively revealed spatial dynamics in the mangrove–Spartina alterniflora Loisel. ecotone of Zhangjiang Estuary [25]. The centroid model has been employed to track spatial migration patterns of mangroves in the tidal segment of the Jinggu River from 1987 to 2021 [26]. Additionally, landscape pattern indices have been applied to assess fragmentation trends and ecological service value changes in various regions, including Zhanjiang’s Tongming Bay and the Dongzhai Harbour National Nature Reserve [27,28]. While these individual methods have been applied in various regions, a systematic and integrated approach combining them to understand both the spatiotemporal dynamics and the resulting carbon storage implications, particularly in a key biodiversity hotspot like Bamen Bay, is still lacking.

This study investigates mangrove forests in Bamen Bay, Hainan Province, using Landsat time-series imagery from 2000 to 2020. The specific objectives are as follows:

(1)

To develop a decision tree classification method integrating SWIR1, NDVI, and NDMI for high-precision mangrove mapping;

(2)

To systematically analyze the spatiotemporal evolution of mangrove distribution over the past two decades by integrating the dynamic degree model, standard deviation ellipse, centroid model, and landscape pattern indices;

(3)

To quantify carbon storage changes and predict future trends by combining the InVEST model, grey prediction model, and scenario analysis, thereby revealing potential development pathways under different management scenarios.

This study advances beyond existing work in three ways. The decision tree method achieves 94.5%–96.5% accuracy by leveraging SWIR1’s sensitivity to vegetation water content, surpassing the current study finding of 91% accuracy [29]. This study explicitly links spatiotemporal dynamics and carbon storage to reveal that the 38.35% area increase translated into a 38.6% carbon storage increase, with landscape fragmentation peaking in 2010 before reversing. Finally, unlike studies that assume linear future trends, our scenario analysis quantifies the wide divergence between Green Revival (0.438 million tons by 2050) and Ecological Collapse (0.075 million tons), providing empirically grounded pathways for policy evaluation.

2. Study Area and Datasets

2.1. Study Area

Hainan Qinglan Port Mangrove Provincial Nature Reserve is located in Wenchang City, Hainan Province. Founded in 1981, it is a key nature reserve in China, mainly focusing on the protection of mangrove ecosystems. The region has a tropical monsoon oceanic climate with abundant rainfall, an average annual precipitation of 1749.5 mm, and an average annual temperature of 24.1 °C. The protected area has a rich diversity of mangroves, with 24 species including Bruguiera sexangula (Lour.) Poir., Sonneratia caseolaris (L.) Engl., Sonneratia alba Sm., and Bruguiera gymnorrhiza (L.) Lam. It accounts for 92.31% of the total number of mangrove species (26 species) in Hainan Province, 85.71% of the number of mangrove species (28 species) in the country, and 27.91% of the global number of mangrove species (86 species) [30].

According to the Adjustment Demonstration Report on the Scope and Functional Zoning of the Hainan Qinglan Mangrove Provincial Nature Reserve and the Master Plan for the Hainan Qinglan Mangrove Provincial Nature Reserve (2021–2030), the reserve is divided into four sectors. The primary sector is located along the coast of Bamen Bay (Qinglan Port) in southeastern Wenchang City, referred to as the Bamen Bay sector. The second sector lies in the coastal waters of Puqian Port and Luodou in northwestern Wenchang City, termed the Puqian sector. The third sector is situated along the coast of Huiwen Town in southern Wenchang City, known as the Huiwen sector. The fourth sector is located north of Tonggu Ridge in Longlou Town, southeastern Wenchang City, referred to as the Longlou sector [30].

Hainan Island harbours the highest mangrove species diversity in China, serving as a critical hotspot for mangrove conservation. Bamen Bay is a typical lagoon estuary wetland, characterized by calm waters, soft bottom sediments, and rich mangrove diversity. It is a key area for biodiversity protection in Hainan and an important habitat for migratory birds. Under the background of accelerating global climate and environmental changes, the systematic study of the spatiotemporal dynamics and carbon potential of the mangroves in Bamen Bay is of great scientific significance for the protection and ecological restoration of regional mangroves.

The study area is shown in Figure 1, which focuses on the mangroves within the Bamen Bay sector. Bamen Bay is located in the southeast of Wenchang City, Hainan Province, China, adjacent to Wencheng Town, Dongjiao Town, Wenjiao Town, Longlou Town, and Dongge Town. The bay constitutes a funnel-shaped embayment that extends deep inland, characterized by a constricted mouth and an expansive interior, with its opening oriented toward the southeast. Its coastal sediment mud and wave power are weak, which provides suitable conditions for the growth and development of mangroves [31].

2.2. Data Sources and Preprocessing

The research period is from 2000 to 2020. In order to systematically analyze the spatial and temporal evolution characteristics of mangroves, remote sensing images are obtained every 5 years, and a total of 5 scenes are obtained through the geospatial data cloud platform. The data were acquired from Landsat 5 Thematic Mapper (TM), Landsat 7 Enhanced Thematic Mapper Plus (ETM+), and Landsat 8 Operational Land Imager/Thermal Infrared Sensor (OLI/TIRS) sensors. To minimize the impact of seasonal changes on the classification results, all images were selected from the April to September period in each respective year, during which the mangrove phenology in the study area remained relatively stable and the cloud cover was generally low. In addition, in order to ensure the data quality, images with more than 10% of cloud cover are excluded. The remote sensing data source is shown in

Table 1. Remote sensing data source.

Year	Datasets	Number of Scenes
2000	Landsat 5 TM	1
2005	Landsat 5 TM	1
2010	Landsat 5 TM	1
2015	Landsat 7 ETM+	1
2020	Landsat 8 OLI/TIRS	1

Note: Each scene represents a single Landsat acquisition covering the study area after preprocessing.

.

To accurately extract the spatial distribution information of mangroves, all images were preprocessed. First, radiometric calibration was performed to convert the original digital number (DN) values to surface reflectance, thereby minimizing the influence of different sensors and imaging conditions. For Landsat 7 ETM+ data, the gap fill plug-in is used to repair the stripes caused by SLC-off failures. Then, based on the vector boundary of the research area, the image was cropped to obtain a unified research area centred on the mangrove area of Bamen Bay. ENVI 5.6 was used for the data in 2000–2015, and ENVI 5.6.2 was used for the data in 2020 to complete all preprocessing steps, laying a high-quality data foundation for the subsequent accurate extraction of mangrove information.

3. Methods

3.1. Analytical Framework

The flowchart of this study is shown in Figure 2. This study is analyzed through four steps by using decision tree classification, dynamic degree model, centroid model, standard deviation ellipse, landscape pattern indices, the InVEST model, grey prediction model, and scenario analysis.

3.2. Decision Tree Classification

This study employs the decision tree classification method, using Short-Wave Infrared 1 (SWIR1), Normalized Difference Vegetation Index (NDVI), and Normalized Difference Moisture Index (NDMI) as classification characteristics to improve classification accuracy and reliability. The SWIR1 band effectively distinguishes land cover types, such as bare soil, most vegetation, and construction areas [32]. As an important indicator of vegetation coverage, NDVI has been proven to be effective in mangrove identification [33]. NDMI captures the moisture information of the vegetation canopy layer to further improve the separation accuracy between mangroves and other vegetation types [34].

$$\mathrm{NDVI}={\displaystyle \frac{{\mathrm{L}}_{\mathrm{NIR}}-{\mathrm{L}}_{\mathrm{RED}}}{{\mathrm{L}}_{\mathrm{NIR}}+{\mathrm{L}}_{\mathrm{RED}}}}$$

(1)

$$\mathrm{NDMI}={\displaystyle \frac{{\mathrm{L}}_{\mathrm{NIR}}-{\mathrm{L}}_{\mathrm{SWIR1}}}{{\mathrm{L}}_{\mathrm{NIR}}+{\mathrm{L}}_{\mathrm{SWIR1}}}}$$

(2)

where L_NIR, L_RED and L_SWIR1 represent reflectance in the Near-Infrared, Red, and Short-Wave Infrared 1 bands, respectively.

Threshold values for SWIR1 reflectance, NDVI, and NDMI were determined using training samples. Training pixels representing mangroves and other land cover types were first selected through visual interpretation of high-resolution imagery available for the study area. The reflectance values of these training samples in the SWIR1, NDVI, and NDMI bands were then analyzed to identify the optimal ranges that maximized the separability between mangroves and non-mangroves. The thresholds were selected at the points where the classification accuracy for the training set was highest. To ensure reproducibility while accommodating inter-annual variability, thresholds were not fixed a priori but were determined independently for each year using the training samples. This study selects the threshold that maximizes the separability between mangroves and non-mangroves in the training set. Researchers applying this method to other regions or time periods are encouraged to follow the same sample-based calibration protocol rather than adopting pre-defined thresholds, as optimal values may vary with local conditions and image characteristics.

To rigorously evaluate classification accuracy, an independent validation dataset was created by randomly generating 200 verification sample points for each study year, with land cover types determined through visual interpretation. Based on the confusion matrix comparing the classified maps with these validation points, the overall accuracy and Kappa coefficient were calculated for each year. This assessment confirms the robustness of the classification method and ensures the reliability of subsequent spatiotemporal and carbon storage analysis.

3.3. The Dynamic Degree Model

The dynamic degree model aims to quantify the rate of change in mangrove area as a key indicator for analyzing the time trend of mangrove range [24]. It is expressed as:

$$\mathrm{K}={\displaystyle \frac{{\mathrm{U}}_{\mathrm{b}}-{\mathrm{U}}_{\mathrm{a}}}{{\mathrm{U}}_{\mathrm{a}}}}\times {\displaystyle \frac{1}{\mathrm{T}}}\times 100\%$$

(3)

where K represents the dynamic degree of mangrove area change (%), U_a and U_b denote the mangrove area at the beginning and end of the study period, respectively, and T is the time interval (years).

3.4. Standard Deviation Ellipse and Centroid Analysis

In order to characterize the evolution of the spatial pattern of mangroves, this study adopts standard deviation ellipses and centroid models for quantitative analysis.

The standard deviation ellipse effectively characterizes the spatial concentration or dispersion of mangroves through parameters including area and orientation angle. The elliptical area reflects the degree of dispersion of the distribution of mangroves, while the orientation angle indicates the dominant direction trend, thus revealing the spatial aggregation or expansion pattern [35].

The centroid model analyzes the movement trajectory of mangrove spatial centroids to visually reflect the migration patterns of distribution centres, thereby investigating the primary directions of mangrove expansion or retreat [36]. The centroid coordinates are calculated as:

$${\mathrm{X}}_{\mathrm{t}}={\displaystyle \frac{{\displaystyle {\sum }_{\mathrm{t}=1}^{\mathrm{n}}}\left({\mathrm{C}}_{\mathrm{tj}}{\mathrm{X}}_{\mathrm{tj}}\right)}{{\displaystyle {\sum }_{\mathrm{t}=1}^{\mathrm{n}}}{\mathrm{C}}_{\mathrm{tj}}}}$$

(4)

$${\mathrm{Y}}_{\mathrm{t}}={\displaystyle \frac{{\displaystyle {\sum }_{\mathrm{t}=1}^{\mathrm{n}}}\left({\mathrm{C}}_{\mathrm{tj}}{\mathrm{Y}}_{\mathrm{tj}}\right)}{{\displaystyle {\sum }_{\mathrm{t}=1}^{\mathrm{n}}}{\mathrm{C}}_{\mathrm{tj}}}}$$

(5)

where ${\mathrm{X}}_{\mathrm{t}}$ and ${ \mathrm{Y}}_{\mathrm{t}}$ represent the longitude and latitude of the mangrove centroid in year t, respectively; ${\mathrm{C}}_{\mathrm{tj}}$ denotes the area of the j-th mangrove patch in year t; ${\mathrm{X}}_{\mathrm{j}}$ and ${ \mathrm{Y}}_{\mathrm{j}}$ indicate the longitude and latitude of the j-th patch, respectively; and n is the total number of patches.

3.5. Selected Landscape Pattern Indices

Landscape pattern indices effectively characterize landscape structure and spatial fragmentation, making them suitable for analyzing the dynamic evolution of mangrove landscapes. This study uses ArcGIS 10.8.1 and Fragstats 4.2.1 to calculate the landscape pattern indices. Aligned with the research objectives and mangrove landscape characteristics, five landscape pattern indices were selected: Number of Patches (NP), Patch Density (PD), Aggregation Index (AI), Landscape Shape Index (LSI), and Edge Density (ED).

Table 2. Landscape Pattern indices and implications.

Index	Formula	Description of Symbols	Description and Ecological Interpretation
NP	NP = N	N = total number of patches of mangroves	Total number of patches. Higher NP indicates greater fragmentation.
PD	PD = (N/A) × 100	N = number of patches A = total landscape area	Patches per 100 hm2. Eliminates area effect; higher PD = more fragmented.
AI	$\mathrm{AI}=\left[{\mathrm{g}}_{\mathrm{u}}/(\max \to {\mathrm{g}}_{\mathrm{u}})\right]\left(100\right)$	${\mathrm{g}}_{\mathrm{u}}$ = number of like adjacencies $\left(\max \to {\mathrm{g}}_{\mathrm{u}}\right)$ = maximum possible like adjacencies given the patch size	Reflects the aggregation or dispersion degree among patches. Higher AI = more aggregated (less fragmented).
LSI	$\mathrm{LSI}=0.25{\displaystyle \frac{\mathrm{E}}{\sqrt{\mathrm{CA}}}}$	E = total edge length CA= total class area	Reflects the shape complexity of a specific patch type. Higher LSI = more irregular shape.
ED	$\mathrm{ED}=(1/\mathrm{A}){\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{M}}}{\displaystyle \sum _{\mathrm{j}=1}^{\mathrm{M}}}{\mathrm{p}}_{\mathrm{ij}}$	pij = length of edge between patch i and j M = number of patch types A = total landscape area	Indicates the level of landscape division. Higher ED indicates more habitat boundary, often associated with higher fragmentation or edge effects.

shows landscape pattern indices applied in this article and their implications [37].

3.6. Mangrove Carbon Storage Calculation Method

The InVEST model evaluates the impacts of land use changes on ecosystem service functions [38].

This study uses the “Carbon Reserves and Sequestration” module of InVEST 3.14.2 to calculate the carbon storage of the mangroves in Bamen Bay. This module divides ecosystem carbon storage into four basic carbon pools: above-ground biomass carbon, underground biomass carbon, soil organic carbon and dead organic carbon. The specific calculation formula is as follows:

$$\mathrm{C}={\mathrm{C}}_{\mathrm{above}}+{\mathrm{C}}_{\mathrm{below}}+{\mathrm{C}}_{\mathrm{dead}}+{\mathrm{C}}_{\mathrm{soil}}$$

(6)

$${\mathrm{C}}_{\mathrm{total}}={\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}}\mathrm{C}\times \mathrm{S}$$

(7)

where ${\mathrm{C}}_{\mathrm{total}}$ represents the total carbon storage; C denotes the total carbon density, comprising aboveground carbon density (${\mathrm{C}}_{\mathrm{above}}$), belowground carbon density (${\mathrm{C}}_{\mathrm{below}}$), dead organic matter carbon density (${\mathrm{C}}_{\mathrm{dead}}$), and soil carbon density (${\mathrm{C}}_{\mathrm{soil}}$); and S is the area of the region.

This article adopts the following carbon density values, derived from field sampling and modelling for Hainan Island’s mangroves [39], which were considered the most representative available for the Bamen Bay area: ${\mathrm{C}}_{\mathrm{above}}$ = 140.71 t·hm⁻², ${\mathrm{C}}_{\mathrm{below}}$ = 51.682 t·hm⁻², ${\mathrm{C}}_{\mathrm{dead}}$ = 0 t·hm⁻², and ${\mathrm{C}}_{\mathrm{soil}}$ = 34.133 t·hm⁻².

A critical assumption underlying this calculation is that carbon density per unit area remains constant across all years and all mangrove stands, regardless of age or species composition. Consequently, temporal changes in total carbon storage are driven entirely by changes in mangrove area, not by changes in stand structure, biomass accumulation over time, or species composition shifts. This assumption is made explicit here, and its ecological limitations are addressed in Section 5.4. Therefore, the carbon storage estimates should be interpreted as a real carbon stock change rather than stand-level carbon accumulation.

3.7. Grey Prediction Model

In this study, the GM(1,1) model is applied for two purposes: (1) to forecast future mangrove area and carbon storage under the baseline scenario, and (2) to provide a reference trajectory against which alternative management scenarios (Green Revival, Hard Preservation, Missed Opportunity, Ecological Collapse) are compared. The model is fitted using the five observed carbon storage values from 2000 to 2020, and then extrapolated to 2050. The results are presented in Section 4.4.

The Grey Prediction Model (1,1) [GM(1,1)] conducts predictive analysis through cumulative generation sequence and differential equation modelling, which is suitable for small sample research and short-term prediction [40]. This study applies it to predict the changes in the area of mangroves and carbon storage in Bamen Bay. The specific modelling process is as follows:

Let the original sequence be:

$${\mathrm{X}}^{\left(0\right)}=\{{\mathrm{x}}^{\left(0\right)}\left(1\right), {\mathrm{x}}^{\left(0\right)}\left(2\right), \cdots ,{ \mathrm{x}}^{\left(0\right)}\left(\mathrm{n}\right)\}$$

(8)

where ${\mathrm{x}}^{\left(0\right)}\left(\mathrm{k}\right)$ represents the observed data at time k.

The accumulated generating operation (AGO) is applied to ${\mathrm{X}}^{\left(0\right)}$ to obtain ${\mathrm{X}}^{\left(1\right)}$, where

$${\mathrm{x}}^{\left(1\right)}\left(\mathrm{k}\right)={\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{k}}}{\mathrm{x}}^{\left(0\right)}\left(\mathrm{i}\right),\mathrm{k}=1, 2 , \dots \dots , \mathrm{n}$$

(9)

The whitenization equation is given by:

$${\displaystyle \frac{\mathrm{d}{\mathrm{x}}^{\left(1\right)}\left(\mathrm{t}\right)}{\mathrm{dt}}}+\mathrm{a}{\mathrm{x}}^{\left(1\right)}\left(\mathrm{t}\right)=\mathrm{b}$$

(10)

The solution is obtained as:

$${\mathrm{x}}^{\left(1\right)}\left(\mathrm{k}\right)=\left({\mathrm{x}}^{\left(0\right)}\left(1\right)-{\displaystyle \frac{\mathrm{b}}{\mathrm{a}}}\right){\mathrm{e}}^{-\mathrm{a}(\mathrm{k}-1)}+{\displaystyle \frac{\mathrm{b}}{\mathrm{a}}}$$

(11)

where a is the development coefficient and b is the grey input.

The predicted values are restored through the inverse accumulating generation operation (IAGO):

$${\mathrm{x}}^{\left(0\right)}\left(\mathrm{k}\right)={\mathrm{x}}^{\left(1\right)}\left(\mathrm{k}\right)-{\mathrm{x}}^{\left(1\right)}(\mathrm{k}-1)$$

(12)

Model accuracy was evaluated using relative error tests and grade ratio deviation tests, posterior variance ratio, and small error probability.

4. Results

4.1. Accuracy Assessment of Mangrove Extraction

Accuracy validation was conducted using 200 randomly sampled points through visual interpretation and confusion matrix analysis. These points were compared pixel-by-pixel with the classified mangrove map. From this comparison, a 4 × 4 confusion matrix was derived, from which overall accuracy, Kappa coefficient, producer’s accuracy, and user’s accuracy were calculated. This procedure was repeated independently for each of the five study years.

Table 3. Classification accuracy evaluation results.

Year	2000	2005	2010	2015	2020
Overall Accuracy	96%	96%	94.5%	95%	96.5%
Kappa Coefficient	0.9131	0.9235	0.9074	0.9164	0.9332
User’s accuracy	100%	100%	98.41%	96.7%	99.13%
Producer’s accuracy	87.8%	92.65%	88.57%	95.65%	98.28%

demonstrates that the overall accuracy of mangrove classification for the five periods (2000–2020) ranged between 94.5% and 96.5%, with Kappa coefficients ranging from 0.9074 to 0.9332. Across the five study years (2000–2020), mangrove user’s accuracy ranged from 96.7% to 100%, and producer’s accuracy ranged from 87.8% to 98.28%. For comparison, an existing study employing the Support Vector Machine (SVM) method for the same region reported an overall accuracy of 0.91 and Kappa coefficient of 0.89 [29]. The classification accuracy in this study showed improvement compared to these results, confirming the feasibility of the proposed decision tree classification method based on SWIR1, NDVI, and NDMI. This method effectively extracts the mangroves in Bamen Bay, providing a reliable data basis for subsequent spatiotemporal evolutionary analysis.

4.2. Temporal Changes in Mangroves

The classification model based on SWIR1, NDVI, and NDMI was used to extract the Bamen Bay mangroves. As shown in

Table 4. Areas and changes in mangroves from 2000 to 2020.

Year	Area (hm2)	Area Change (hm2)	Dynamic Degree (%)
2000	756.54	—	—
2005	804.51	+47.97	1.27
2010	870.48	+65.97	1.64
2015	964.26	+93.78	2.15
2020	1046.7	+82.44	1.71

Note: Positive values indicate net increase. Negative values indicate decrease. Dynamic degree is calculated using Equation (3) and represents the annual rate of change relative to the starting area of each period.

, area change refers to the net increase or decrease in mangrove area relative to the previous 5-year period. All observed changes were positive, indicating continuous expansion. Dynamic degree is the annual rate of change (Equation (3)), expressed as a percentage of the starting area per year. Mangrove area in Bamen Bay continued to grow over the study period, with a cumulative net increase of 290.16 hm² (38.35%). The overall dynamic degree over the 20-year period was 1.92%. The growth rate fluctuated, accelerating from 1.27% (2000–2005) to 2.15% (2010–2015), before slowing to 1.71% (2015–2020).

4.3. Spatial Changes in Mangroves

4.3.1. Overall Spatial Evolution of Mangroves

Figure 3 shows the spatial distribution of mangroves from 2000 to 2020. Between 2000 and 2005, mangrove degradation was primarily concentrated in the northeastern part of the bay. Newly established mangroves during this period appeared as relatively scattered patches with small patch sizes and low connectivity. Overall growth at this stage remained relatively slow.

From 2005 to 2010, previously degraded areas in the western and northeastern regions showed signs of stabilization, while new mangrove patches emerged in the northern and eastern coastal areas.

Between 2010 and 2015, despite some local degradation in the central and eastern regions, newly formed mangroves in the east transitioned from a scattered to a more clustered distribution, forming a considerable-scale recovery area. The overall area of mangroves continued to grow steadily.

From 2015 to 2020, historically degraded areas in the western region exhibited significant restoration, forming more continuous communities, while the newly expanded area in the east continued to grow and connect with existing forests, showing a recovery trend of scale expansion and structural optimization.

4.3.2. Standard Deviation Ellipse

Through the analysis of standard deviation ellipses and centroid migration, the evolutionary characteristics of the spatial pattern of mangroves in Bamen Bay from 2000 to 2020 are revealed.

Table 5. Parameters of standard deviation ellipse.

Year	2000	2005	2010	2015	2020
Ellipse Area (hm2)	5611.11	5024.19	4112.92	4966.65	3536.88
Orientation Angle (°)	72.94	65.45	70.39	68.06	65.31

and Figure 4 show the changes in the standard deviation ellipse. From 2000 to 2020, the area of the standard deviation ellipse decreased from 5611.11 to 3536.88 hm², indicating increased spatial concentration. The orientation angle fluctuated during the study period, decreasing from 72.94° in 2000 to 65.31° in 2020, with an increase to 70.39° in 2010.

The centroid of the mangroves migrated during the study period. From 2000 to 2010, the centroid moved northeastward. Between 2010 and 2015, it shifted southwestward. From 2015 to 2020, the centroid moved northward again. Overall, the centroid exhibited a net northeastward migration from 2000 to 2020.

4.3.3. Landscape Pattern Indices

Although 5-year land cover maps were available, landscape pattern indices were analyzed at 10-year intervals because the primary interest was in detecting decadal-scale trends in fragmentation and aggregation, which are less sensitive to short-term fluctuations. Preliminary analysis of 5-year intervals revealed that NP and PD fluctuated within a narrow range between consecutive 5-year periods, while 10-year intervals captured the clear transition from fragmentation (2000–2010) to consolidation (2010–2020). This approach also reduces redundancy in presentation. To effectively identify the evolution trend of the landscape patterns with a small change in 5 years, this study conducted a comparative analysis of landscape pattern changes at 10-year intervals from 2000 to 2020.

Table 6. Landscape pattern indices of mangroves from 2000 to 2020.

Year	NP (Patches)	PD (Patches/100 hm2)	AI (%)	LSI (m/hm2)	ED (m/hm2)
2000	79	10.44	88.51	11.38	38.97
2010	117	13.44	85.23	15.36	59.88
2020	106	10.13	86.87	15.01	58.01

shows the changes in landscape pattern indices.

The Number of Patches (NP) increased from 79 in 2000 to 117 in 2010, then decreased slightly to 106 in 2020. Patch Density (PD) followed a similar pattern, increasing from 10.44 patches/100 hm² in 2000 to 13.44 patches/100 hm² in 2010, then decreasing to 10.13 patches/100 hm² in 2020. The Aggregation Index (AI) decreased from 88.51% in 2000 to 85.23% in 2010, then increased slightly to 86.87% in 2020. The Landscape Shape Index (LSI) increased from 11.38 in 2000 to 15.36 m/hm² in 2010, and remained relatively stable at 15.01 m/hm² in 2020. Edge Density (ED) increased from 38.97 m/hm² in 2000 to 59.88 m/hm² in 2010, then decreased slightly to 58.01 m/hm² in 2020.

4.4. Evolution and Prediction of Carbon Storage in Bamen Bay Mangroves

4.4.1. Carbon Storage Evolution

Following the constant carbon density assumption, total carbon storage in Bamen Bay mangroves was calculated as the product of mangrove area and the sum of carbon densities. Therefore, the temporal trend in carbon storage directly mirrors the trend in mangrove area. From 2000 to 2020, the carbon storage of the Bamen Bay mangroves was 0.171, 0.182, 0.197, 0.218, and 0.237 million tons, respectively. During this period, carbon storage increased by 0.065 million tons, from 0.171 to 0.237 million tons, representing a 38.6% increase over the two decades.

Parameter estimation through the grey prediction model yielded a development coefficient a of −0.089 and a grey variable b of 0.158, resulting in the prediction model:

$${\stackrel{\mathrm{~}}{\mathrm{x}}}^{\left(1\right)}(\mathrm{k}+1)=(0.171-{\displaystyle \frac{0.158}{-0.089}}){\mathrm{e}}^{0.089\mathrm{k}}+{\displaystyle \frac{0.158}{-0.089}}=1.95{\mathrm{e}}^{0.089\mathrm{k}}-1.78$$

(13)

The results show that the average relative error of the model is 0.303% (less than 10%), the average grade ratio deviation is 0.014 (less than 0.1), the posterior variance ratio C is 0.035 (below 0.35), and the small error probability P is 1 (higher than 0.95). These indicators meet the first-level accuracy standard defined by the grey system theory, indicating that the model has high reliability and is suitable for prediction.

Figure 5 and

Table 7. Grey model GM(1,1) predictions and verification.

Year	Observed (Million Metric Tons)	Predicted (Million Metric Tons)	Relative Error (%)	Grade Ratio Deviation
2000	0.171	0.171	—	—
2005	0.182	0.181	0.55	0.027
2010	0.197	0.198	0.51	0.01
2015	0.218	0.217	0.46	0.012
2020	0.237	0.237	0	0.006

show the high fitting accuracy of the GM(1,1) model of carbon storage in the mangroves of Bamen Bay. As presented in Table 7, the relative errors between observed and predicted carbon storage values range from 0% to 0.55%, with an average relative error of only 0.303%. Additionally, the grade ratio deviations are consistently below 0.03, well within the acceptable threshold. This high fitting accuracy not only verifies the ability of the model to capture the potential change patterns of carbon storage in mangroves but also supports its use for stable medium- and long-term prediction. The performance of the model provides practical value for guiding mangrove conservation and climate adaptive management strategies.

4.4.2. Carbon Storage Prediction

The carbon storage capacity of the mangroves in Bamen Bay is jointly affected by natural factors such as temperature and precipitation and human activities such as policy intervention and wetland protection, which has brought great uncertainty to the future development path.

In order to systematically evaluate the evolution of potential carbon storage, this study has developed a baseline scenario based on the grey prediction model, as well as four alternative scenarios: Green Revival, Hard Preservation, Missed Opportunity, and Ecological Collapse. The baseline scenario represents the continuation of the current trend, and its forecast value can be used as a benchmark for assessing the effectiveness of potential carbon storage changes and policy response under different natural-anthropogenic interaction intensities. As shown in Figure 6, these scenarios together provide a structured visualization of possible carbon storage paths, supporting a more detailed understanding of the risks, resilience and potential intervention points of mangrove ecosystem management.

The growth rates for different scenarios were derived from observed mangrove area changes in Bamen Bay across specific historical periods, selected based on their alignment with scenario narratives. The Green Revival scenario adopts the highest 5-year growth rate observed in the study period (10.77%, 2010–2015), which coincided with the implementation of the Hainan Mangrove Conservation Plan (2006–2015) and the South Mangrove–North Willow initiative, conditions of strong policy support and favourable natural conditions that mirror the scenario’s assumptions. The Hard Preservation scenario uses the average growth rate of all periods (9.59% per 5 years), representing sustained conservation efforts without the peak recovery rates. The Missed Opportunity scenario adopts the lowest positive growth rate (6.34%, 2000–2005), a period characterized by nascent conservation efforts, incomplete legal protection, and continued pressure from aquaculture expansion, reflecting delayed action despite suitable natural conditions. The Ecological Collapse scenario adopts a negative growth rate of −17.49% per 5 years, derived from the documented severe degradation period in Bamen Bay between 1998 and 2003 [30], when mangroves experienced intensive shrimp pond conversion and extreme weather impacts. This rate represents the worst-case trajectory under combined anthropogenic and climatic stress.

Figure 7 shows the prediction results in different scenarios. Under the baseline scenario, carbon storage is expected to reach 0.405 million tons by 2050. The Green Revival scenario envisions successful global climate governance, climate change mitigation, enhanced public environmental awareness, and favourable natural conditions. Under this scenario, effective protection and management measures would enable carbon storage to exceed the baseline value, reaching 0.438 million tons by 2050. The Hard Preservation scenario is characterized by severe climate challenges and high-intensity ecological protection investment. It shows the resilience of the ecosystem through continuous protection measures. The fluctuation of carbon storage is close to but slightly above the baseline (0.411 million tons). In the Missed Opportunity scenario, despite relatively suitable climatic conditions, delayed conservation actions and an imperfect governance system mean that natural succession alone cannot offset human-induced losses, resulting in carbon storage far below the baseline (0.343 million tons). The Ecological Collapse scenario, characterized by drastic climate change and serious anthropogenic ecosystem destruction, under the double pressure of harsh natural conditions and weak protection, has caused carbon storage to plummet to 0.075 million tons below the baseline level and continue to decline.

According to these forecasts, the development of carbon sinks in the mangroves of Bamen Bay should actively guide the Green Revival scenario, while strictly preventing the Ecological Collapse scenario.

Achieving this goal requires the systematic reduction in human interference in coastal areas to reduce climate-related stress and create suitable natural conditions for the diffusion of mangroves and carbon storage. At the same time, it is essential to strengthen the effective protection and comprehensive management of the mangrove ecosystem globally, focusing on the restoration of ecological functions. This can be achieved by improving conservation policies, ensuring their strict implementation, and promoting cross-sectoral cooperation between government agencies, research institutions and local communities. In addition, the implementation of a strong monitoring and assessment framework supported by remote sensing and field surveys will help to track mangrove conditions, carbon dynamics and the effectiveness of interventions. This adaptive management strategy ensures continuous improvement of conservation measures, thereby enhancing the long-term resilience of mangroves as important natural climate solutions and biodiversity centres.

5. Discussion

5.1. Comparison with Previous Studies and Implications of Findings

5.1.1. Comparison of Mangrove Area Estimates

The mangrove area estimates derived from this decision tree classification method provide an opportunity to compare with previous studies in Bamen Bay. A previous study documented a declining trend in mangrove area from 1987 to 2003, followed by a gradual recovery from 2003 to 2017, using a combination of supervised classification and visual interpretation [30]. Our estimates of 756.54 hm² in 2000 and 964.26 hm² in 2015 are consistent with this observed trajectory of decline followed by recovery. Discrepancies in specific values likely reflect differences in classification methods, image acquisition timing, and the spatial resolution of source imagery. Similarly, another study documented an expansion trend of mangroves on Hainan Island from 2000 to 2020, which aligns with our observed increase in Bamen Bay [41]. This consistency across independent studies reinforces the conclusion that mangrove forests in this region have experienced substantial recovery over the past two decades.

A previous study employing a SVM approach for Bamen Bay reported an overall accuracy of 91% with a Kappa coefficient of 0.89 [29]. In comparison, our decision tree method achieved higher classification accuracies ranging from 94.5% to 96.5% with Kappa coefficients from 0.9074 to 0.9332. This improvement suggests the potential advantage of integrating SWIR1, NDVI, and NDMI in a hierarchical rule-based framework for this study area, particularly in complex coastal environments where spectral confusion between mangroves and other vegetation types is common.

5.1.2. Implications of Spatiotemporal Dynamics

The observed spatial concentration of mangroves is reflected in the shrinking standard deviation ellipse, which decreased from 5611.11 hm² in 2000 to 3536.88 hm² in 2020. The northeastward centroid migration during 2000–2020 is consistent with the trend reported in a previous study [30], reinforcing the robustness of this directional shift. This consistency across different time periods suggests that the northeastward migration represents a sustained spatial dynamic, likely shaped by the interplay of local geomorphological conditions and anthropogenic interventions.

A previous study reported that NP increased from 65 in 1987 to 164 in 2017, while AI declined from 97.05% to 94.63% over the same period [30]. These trends indicate that mangrove expansion was accompanied by sustained fragmentation and the proliferation of small patches. Similarly, this study observed an increase in NP from 79 to 117 between 2000 and 2010, with a corresponding decline in AI from 88.51% to 85.23%, reflecting an analogous early-stage fragmentation pattern following the cessation of aquaculture activities. However, a notable shift occurred in the later stage of this study. Between 2010 and 2020, NP decreased to 106 and AI increased to 86.87%. This transition from fragmentation to consolidation suggests that restoration efforts initiated after 2010 in the study area have effectively promoted patch coalescence. The emergence of this two-phase pattern—initial fragmentation followed by consolidation—highlights the positive impact of sustained conservation interventions in reversing landscape degradation and enhancing habitat connectivity, providing empirical evidence for the effectiveness of mangrove restoration policies implemented in recent years.

5.1.3. Implications of Carbon Storage Trends

The carbon storage increase from 0.171 million tons in 2000 to 0.237 million tons in 2020 represents a 38.6% increase, closely mirroring the 38.35% increase in mangrove area. This proportional relationship reflects the model’s assumption of uniform carbon density across all mangrove stands regardless of age. In reality, carbon accumulation in younger mangroves may differ from that in mature forests, and the actual carbon storage trajectory of restored areas could be more complex. Therefore, this finding should be interpreted with caution, and future studies incorporating age-specific carbon density would improve estimation accuracy.

The baseline projection of 0.405 million tons by 2050 implies a continued growth trajectory, though the rate of increase is expected to moderate as available suitable habitat becomes limited. The wide divergence among scenarios—from 0.438 million tons under Green Revival to 0.075 million tons under Ecological Collapse—highlights the critical role of management decisions in determining future carbon outcomes. These projections align with broader assessments emphasizing that mangrove conservation outcomes are highly sensitive to policy effectiveness and climate conditions [6,15].

5.2. Drivers of Mangrove Dynamics in Bamen Bay

5.2.1. Drivers of Area Expansion

The observed continuous expansion of mangrove area from 2000 to 2020 can be attributed to a combination of policy interventions and natural processes. In the early 2000s, the growth rate was relatively slow at just 1.27% during 2000–2005, reflecting a period when conservation efforts were nascent and lacked a systematic framework, leading to continued degradation under developmental pressures. Growth accelerated to 1.64% during 2005–2010 and peaked at 2.15% during 2010–2015, coinciding with the implementation of the Hainan Mangrove Conservation Plan (2006–2015) and major ecological projects such as the “South Mangrove–North Willow” initiative. In the most recent period of 2015–2020, although the growth rate moderated slightly to 1.71%, this sustained expansion demonstrates the cumulative benefits of long-term planning and large-scale implementation under national policies like the National Coastal Shelterbelt System Construction Project Plan.

5.2.2. Factors Influencing Spatial Pattern Evolution

The spatial evolution of mangroves in Bamen Bay reflects the interplay between human interventions and natural recovery processes. The initial northeastward centroid migration (2000–2010) was likely influenced by land-oriented interventions such as artificial restoration and compensatory planting in inland areas. The brief southwestward shift (2010–2015) coincided with intensified natural recovery processes under improved protection policies, with mangroves expanding seaward. The subsequent northward movement (2015–2020) resulted from the combined effects of continuous policy implementation and local artificial restoration efforts, demonstrating the spatial reconstruction process jointly driven by human intervention and natural recovery.

5.2.3. Implications of Landscape Pattern Changes

The landscape pattern analysis reveals a complex picture of mangrove ecosystem dynamics. The initial increase in NP and PD (2000–2010) suggests that anthropogenic disturbance fragmented the mangrove landscape. However, the slight decrease in these indices after 2010 indicates that subsequent conservation measures began to reverse this trend by promoting patch coalescence. The initial decline in AI (2000–2010) followed by a rebound (2010–2020) suggests that while habitat fragmentation initially increased, restoration projects have successfully improved the connectivity of some patches. The continuous increase in LSI and ED throughout the study period indicates that landscape shape has become more complex due to the irregular boundaries of both disturbed and restored areas. This suggests that human interference with landscape structure has not been completely eliminated.

Landscape fragmentation and consolidation have distinct ecological implications for Bamen Bay mangroves. During the fragmentation phase, the increase in NP, PD, and ED indicated that mangrove habitat became more isolated. This likely exacerbated edge effects, including higher light penetration, wind disturbance, and invasive species pressure, which can reduce per-area carbon density and lower habitat quality for interior-dependent species. The subsequent consolidation phase, characterized by decreasing NP and PD and recovering AI, suggests improved patch connectivity. Consolidated patches facilitate pollen and seed dispersal, enhance genetic exchange, and provide larger core habitats that support higher biodiversity and greater resistance to external disturbances.

5.3. Methodological Contributions and Model Performance

5.3.1. Decision Tree Classification Method

This study introduces a decision tree classification method that integrates SWIR1, NDVI, and NDMI as key classification parameters. Unlike conventional approaches that often overlook the potential of the SWIR1 band, the proposed method leverages its sensitivity to vegetation water content to enhance discrimination between mangroves and surrounding terrestrial vegetation. The combination of these three indices in a hierarchical rule-based framework achieved overall classification accuracies of 94.5%–96.5% and Kappa coefficients of 0.9074–0.9332 across five study periods, demonstrating competitive performance compared to a previous SVM application in the same region [29], though direct comparability is limited by differences in training data and classification parameters. This methodological innovation provides a robust, transparent, and easily replicable approach for accurate mangrove mapping, addressing a critical need in long-term ecosystem monitoring.

5.3.2. Performance of the Grey Prediction Model

The GM(1,1) grey model demonstrated excellent performance in fitting historical carbon storage data, with an average relative error of 0.303% and a posterior variance ratio of 0.035, meeting the first-level accuracy standard of grey system theory. This confirms the model’s reliability for carbon storage prediction and its value for informing coastal ecosystem management and blue carbon policy formulation.

5.3.3. Methodological Advancements

The integration of multiple analytical methods represents a significant methodological advancement. Specifically, the combination of the grey prediction model with scenario analysis integrates quantitative prediction with qualitative evaluation of carbon storage trends under different policy and environmental assumptions. This framework not only improves the predictability of future ecosystem changes but also provides scientific support for developing conservation and recovery strategies by illustrating potential outcomes under different management pathways. Together with the decision tree classification method and the comprehensive spatiotemporal analysis framework, this study offers an integrated methodological toolkit for mangrove ecosystem assessment that can be adapted and applied to other coastal wetland systems globally.

5.4. Limitations and Future Research Directions

5.4.1. Remote Sensing Data

Several limitations should be acknowledged. While the total area estimates align broadly with existing research [30,41], discrepancies in specific yearly values highlight the sensitivity of results to image acquisition time, preprocessing procedures, and classification parameter selection.

A key limitation concerns spatial resolution. The reliance on Landsat imagery (30 m) is appropriate for regional-scale, long-term analysis of mangrove extent and broad spatiotemporal trends. However, this spatial resolution is less suitable for detecting fine-scale changes, narrow mangrove fringes, small patches, and detailed edge dynamics. Specifically, narrow mangrove fringes narrower than 30 m may be omitted entirely or mixed with adjacent water, reducing NP and PD counts. Small patches are undetectable. Consequently, landscape pattern metrics systematically underestimate true fragmentation levels. The NP and PD values reported in Table 6 should therefore be interpreted as conservative estimates of true fragmentation. The actual number of patches is likely higher, and the decline in NP after 2010 may represent coalescence of detectable patches while new small patches remain undetected. Using higher-resolution imagery for a subset of years would be necessary to quantify the magnitude of this underestimation.

The assumption of uniform carbon density also masks spatial heterogeneity driven by species composition and tidal position. The use of average values for Hainan Island may overestimate carbon in areas dominated by interior species and underestimate in seaward zones. A sensitivity analysis or uncertainty quantification can be further explored in subsequent studies.

Future research should integrate higher-resolution data from sources such as Sentinel-2 or Planet to improve detection of fine-scale spatial patterns. Additionally, incorporating field-based carbon measurements and age-specific carbon density values would enhance the accuracy of carbon storage assessments.

5.4.2. Classification Methods

In terms of classification methods, although the three parameters SWIR1, NDVI and NDMI effectively extract the overall distribution of mangroves, they do not distinguish between different mangrove species. Future research should explore species-level classification using hyperspectral data or advanced machine learning techniques. Additionally, future work could combine more spectral and texture features with methods such as XGBoost and Random Forest to further improve classification accuracy.

5.4.3. Prediction Model

The GM(1,1) model was chosen over fuzzy logic approaches for three reasons. First, grey models require only four to five data points to generate reliable predictions, making them suitable for our five-period (2000–2020) dataset, whereas fuzzy methods typically require more observations for parameter stabilization. Second, grey models provide transparent, closed-form solutions that explicitly show the development coefficient a and grey input b, enabling clear interpretation of growth trends. Third, for short-term (10–30 year) predictions of systems without abrupt regime shifts, grey models have demonstrated accuracy comparable to more complex methods. Fuzzy approaches would be advantageous for longer-term predictions or systems with known non-linear thresholds, which is why this study complements GM(1,1) with scenario analysis to bound uncertainty.

Importantly, the GM(1,1) grey model, despite its strong performance in fitting historical data (average relative error of 0.303%), is a simplified representation of a complex socio-ecological system. The grey prediction model assumes that the underlying system follows an exponential growth or decay pattern, which is mathematically expressed as a first-order linear differential equation. This assumption fails to capture three complexities relevant to mangrove ecosystems. First, mangrove area expansion is ultimately constrained by available suitable habitat. Once all available intertidal zones are occupied, growth must plateau, creating an S-shaped rather than exponential trajectory. Second, the model treats each time step independently, missing autocorrelation. For these reasons, the 2050 baseline projection (0.405 million tons) should be interpreted as an extrapolation of historical trends under the implicit assumption that current conditions persist, not as a prediction that explicitly accounts for climate change or habitat saturation. Third, GM(1,1) cannot incorporate external drivers such as sea-level rise, increased storm frequency, or changes in sediment supply, all of which could fundamentally alter the system’s behaviour. These drivers were excluded because the model’s mathematical structure does not accommodate multiple independent inputs. This exclusion was to maintain model parsimony given the limited 20-year time series, not an assumption that these drivers are unimportant. Future research should integrate biophysical models, such as the Sea Level Affecting Marshes Model, with our carbon accounting framework to explicitly simulate climate driver impacts.

This approach is valuable for highlighting risks, identifying potential intervention points, and guiding adaptive management, even if it cannot capture all future complexities. Follow-up research should explore more advanced models capable of handling nonlinear data and integrating a broader range of biophysical and anthropogenic factors to improve fitting performance and prediction accuracy.

5.4.4. Assumption of Proportional Area-Carbon Relationship

A key assumption underlying our carbon storage estimates is the proportional relationship between mangrove area and total carbon storage, which follows directly from the InVEST model’s use of constant carbon density values (Equation (7)). While this assumption is standard for landscape-scale carbon assessments, it does not account for age-related carbon accumulation dynamics. In Bamen Bay, the 290.16 hm² net increase includes areas restored at different times (2000–2005, 2005–2010, 2010–2015, 2015–2020). Young mangroves in the most recently expanded areas likely have carbon densities lower than our constant values. Consequently, our carbon storage estimates for 2020 may be overestimated, though the direction of temporal trends of increasing storage remains robust. Future research should integrate age-specific carbon density curves derived from chronosequence studies to refine these estimates.

5.4.5. Uncertainty Quantification

This study does not report conventional confidence intervals for carbon storage estimates or dynamic degree calculations because the predictive framework is scenario-based rather than statistically sampled. The baseline and alternative scenarios are defined by assumed growth rates derived from historical periods, not by probability distributions. Consequently, traditional confidence intervals are not directly applicable. We acknowledge that uncertainty exists in carbon density values, classification accuracy, and scenario assumptions. A quantitative uncertainty range could be explored in future research by integrating methods such as Monte Carlo simulations, fuzzy set approaches, or multi-model ensembles.

5.5. Management Implications and Future Perspectives

These findings provide a quantitative basis for prioritizing the Green Revival scenario through integrated, adaptive management approaches that combine robust remote sensing monitoring with proactive policy implementation and community engagement, thereby maximizing the long-term resilience and carbon storage potential of mangrove ecosystems. Achieving this goal requires the systematic reduction in human interference in coastal areas to reduce climate-related stress and create suitable natural conditions for mangrove expansion and carbon storage. At the same time, it is essential to strengthen the effective protection and comprehensive management of mangrove ecosystems, focusing on the restoration of ecological functions. This can be achieved by improving conservation policies, ensuring their strict implementation, and promoting cross-sectoral cooperation between government agencies, research institutions, and local communities. In addition, the implementation of a strong monitoring and assessment framework supported by remote sensing and field surveys will help to track mangrove conditions, carbon dynamics, and the effectiveness of interventions. This adaptive management strategy ensures continuous improvement of conservation measures, thereby enhancing the long-term resilience of mangroves as important natural climate solutions and biodiversity centres.

6. Conclusions

Based on the Landsat series of images from 2000 to 2020, this study has established a decision tree model with SWIR1, NDVI, and NDMI as classification characteristics, and extracted the spatial distribution of the Bamen Bay mangroves in 2000, 2005, 2010, 2015 and 2020. By integrating the dynamic degree model, standard deviation ellipse, centroid model, and landscape pattern indices, the spatial distribution information of mangroves in Bamen Bay was systematically analyzed, revealing the spatiotemporal evolution patterns of mangroves in the study area over the past 20 years. In addition, using the InVEST model, grey prediction model, and scenario analysis, this study calculated carbon storage and predicted future trends. The main findings are as follows:

(1)

The proposed decision tree classification model based on SWIR1, NDVI, and NDMI achieved high accuracy. This performance demonstrates competitive accuracy compared to a previous SVM application in the same region [29], though direct comparability is limited by differences in training data and classification parameters, thereby validating the effectiveness of this combination of spectral indices as a robust alternative for mangrove extraction.

(2)

From 2000 to 2020, the area of mangroves in Bamen Bay showed a continuous growth trend, with an increase of 38.35%. The overall dynamics reached 1.92%, indicating that the mangrove ecosystem in the region is in a stable recovery stage.

(3)

Over the past 20 years, the area of the standard deviation ellipse has decreased, while the orientation angle has generally shown a declining trend. The distribution of mangroves is more concentrated, and the expansion of the western region is obvious. The centroid moves to the northeast as a whole. Although landscape fragmentation remains a problem, it has been partially alleviated through continuous conservation measures.

(4)

Under the baseline scenario, carbon storage is expected to reach 0.405 million tons by 2050, assuming continuation of historical trends and no habitat saturation. Under the Green Revival scenario, carbon storage will exceed the baseline, reaching 0.438 million tons. The Hard Preservation yields 0.411 million tons, which is slightly higher than the baseline. In the Missed Opportunity scenario, carbon storage will drop to 0.343 million tons, below the baseline. The result of the Ecological Collapse scenario is only 0.075 million tons, far below the baseline, and shows a continuous downward trend.

The mangrove extraction method proposed in this study offers an approach for remote sensing classification technology. The constructed carbon storage prediction model and scenario analysis framework can provide scientific support for mangrove conservation and restoration efforts, contributing significantly to global blue carbon research. To ensure the sustainable development of mangrove ecosystems, it is essential to effectively control human disturbances in coastal environments and mitigate climate change to create suitable natural growth conditions for mangroves. Global collaborative governance and systematic protection of mangroves should be strengthened, promoting the development of Bamen Bay mangroves toward a Green Revival scenario. By establishing ecosystem management paradigms with favourable natural conditions and effective protection measures, this study provides scientific support for achieving the “Carbon Peak and Carbon Neutrality” goals.