Estimating Mangrove Aboveground Biomass Using Sentinel-2 and ALOS-2 Imagery: A Case Study of the Matang Mangrove Reserve, Malaysia

Abstract

Mangroves play a critical role in global carbon sequestration, biodiversity conservation, and climate change mitigation. Accurately quantifying mangrove biomass is essential for sustainable forest management and carbon accounting. Yet, the structural complexity and species diversity of mangrove ecosystems pose significant challenges for accurate estimation. In this study, we developed an integrated model that combines multispectral imagery and radar data. Using Sentinel-2 and ALOS-2 satellite imagery combined with field measurements, these data were used to construct linear regression and random forest models for the Matang Mangrove Reserve, Malaysia. We further analyzed the relationships between vegetation indices, radar polarization modes, and biomass. Results indicate that the average biomass is approximately 146 t/ha. The Optimized Soil-Adjusted Vegetation Index (OSAVI) and horizontal–vertical (HV) polarization showed the strongest correlation with field-measured biomass, with an R² of 0.735 and a root mean square error (RMSE) of 46.794 t/ha. This study provides a scientific basis and technical support for mangrove carbon stock assessment, ecosystem management, and climate change mitigation strategies, and highlights the potential of integrating optical and radar remote sensing for large-scale mangrove biomass monitoring.

1. Introduction

Mangroves typically thrive in subtropical and tropical regions and represent an ecologically important forest type. Mangroves provide multiple ecological functions, including enhancing coastal resilience against tsunamis, purifying seawater, sequestering carbon, and providing habitats for diverse marine organisms [1,2]. Nonetheless, mangrove ecosystems have experienced continuous global decline, largely due to natural disturbances and human activities. Over the past five decades, approximately one-third of mangroves have been lost, mainly as a result of agricultural expansion and aquaculture development [3]. Monitoring mangroves is therefore critical for the protection of coastal ecosystems.

Aboveground biomass is one of the important indicators of mangrove ecosystems. The importance of aboveground biomass as a key indicator for mangrove ecological studies has been increasingly emphasized in the literature [4,5,6,7]. However, the conventional method for assessing mangrove biomass has mainly relied on field-based manual techniques. The most precise approach entails quantifying indicators through destructive sampling (tree felling), although this contradicts the conservation objectives of mangrove ecosystems [8]. Other manual methods, such as soil sampling and laboratory analysis, are less destructive but remain time-consuming and costly. While manual measurements exhibit a high degree of accuracy, they frequently encounter challenges, including limited spatial coverage, high costs, and lengthy timelines [9]. In recent years, remote sensing has been increasingly applied in mangrove research due to its advantages of broad spatial coverage and rapid data acquisition [10,11].

There are many ways to monitor mangroves through remote sensing [12]. For example, optical remote sensing satellites such as Landsat and active remote sensing satellites such as the Advanced Land Observing Satellite-2 (ALOS-2) are both widely used [4,5,6,7,13]. Several studies have explored biomass modeling using different satellite datasets and algorithms. For example, Baloloy (2018) [6] constructed predictive models using conventional linear and multiple regression techniques with Sentinel-2, RapidEye, and PlanetScope data to identify the most effective biomass prediction model for each platform. Castillo (2017) [5] used Sentinel-1 and Sentinel-2 data to develop a biomass prediction model for the southern coast of Honda Bay, Palawan (Philippines), applying both traditional regression and machine learning methods to generate a biomass map. Jachowski (2013) [4] utilized GeoEye-1 and Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) satellite data in combination with support vector machine approaches to cartographically represent biomass distribution throughout mainland Southeast Asia. Pham and Brabyn (2017) [7] monitored mangrove biomass changes in Vietnam using SPOT imagery and support vector machines. Ghosh (2021) [13] used Sentinel-1 and Sentinel-2 data with machine learning algorithms such as gradient boosting (GBM) and extreme gradient boosting (XGB) to map aboveground biomass in India.

Despite progress, mangrove aboveground biomass (AGB) estimation remains constrained by three major challenges. First, saturation effects: in dense stands with closed canopies or high biomass, commonly used optical indices (e.g., NDVI, EVI) and radar backscatter lose sensitivity. Previous studies have shown that their prediction accuracy drops sharply in high-biomass conditions, with reported Coefficient of Determination (R²) values often below 0.5 when AGB exceeds 200 t/ha [14,15,16]. Second, species and structural discrimination: mangrove species often display strong spectral similarity, while tidal dynamics and water backgrounds further complicate reflectance signals. Even UAV-based hyperspectral studies have concluded that “spectral information alone is inadequate for species-level separation”, highlighting the necessity of incorporating structural or textural features for improved discrimination [17]. Third, reliance on single-sensor approaches: optical-only methods or those based on a single radar frequency (e.g., C-band) are strongly affected by cloud cover and long revisit cycles, and they fail to capture canopy heterogeneity and vertical structure [18,19]. To overcome these shortcomings, our study develops an integrated framework that combines Sentinel-2 multispectral data and ALOS-2 L-band Synthetic Aperture Radar (SAR) backscatter, leveraging their complementary strengths to reduce saturation effects, improve species discrimination, and enhance robustness.

Sentinel-2 is a multispectral satellite recently launched under the Copernicus program of the European Space Agency (ESA). Sentinel-2 provides 13 multispectral bands, including three vegetation red edge bands and two infrared bands, as well as visible light and near-infrared bands [20]. These red-edge and SWIR bands are particularly sensitive to canopy chlorophyll concentration, leaf water status, and other biochemical traits, which are critical for biomass estimation in vegetation studies. Compared with Landsat, Sentinel-2 provides finer spatial (10–20 m) and spectral resolution, enabling more accurate detection of heterogeneous mangrove stands [6]. ALOS-2 is a radar satellite launched by the Japanese Earth Observation Satellite Program in recent years. Its operating band is the L band (1.2 GHz band). Unlike optical sensors, ALOS-2 is largely unaffected by atmospheric conditions and penetrates cloud cover, providing reliable data under diverse weather conditions [21]. Its L-band wavelength penetrates deeper into dense mangrove canopies than shorter-wavelength C-band radar, capturing trunk density, woody biomass, and vertical forest complexity [22,23]. C-band SAR typically saturates at ~100–150 t/ha, whereas L-band SAR remains sensitive up to 200–250 t/ha, extending the dynamic range of AGB estimation [24]. Together, Sentinel-2 and ALOS-2 provide complementary biochemical and structural information: while Sentinel-2 is sensitive to chlorophyll and moisture, ALOS-2 penetrates vegetation to capture woody structure. This optical–radar synergy effectively reduces saturation, improves representation of canopy heterogeneity, and enhances robustness in tropical regions frequently affected by clouds [25]. In addition, Sentinel-2 (5-day revisit) and ALOS-2 (14-day revisit) together allow complementary temporal monitoring of mangrove biomass dynamics, supporting the detection of seasonal variations and disturbance events. While most previous studies relied predominantly on optical data (Baloloy et al., 2018 [6]; Castillo et al., 2017 [5]), our framework explicitly validates the integration of multispectral and radar imagery, which is particularly critical in dense and cloud-prone ecosystems such as the Matang Mangrove Reserve [5,6].

Therefore, this study aims to develop a novel mangrove biomass prediction model by integrating active and passive remote sensing imagery. Specifically, the combination of Sentinel-2 multispectral data and ALOS-2 L-band SAR backscatter was designed to overcome the limitations of single-sensor approaches by jointly capturing canopy biochemical traits and forest structural attributes. Furthermore, we examined the applicability of various vegetation indices and polarization modes for developing the biomass prediction model. Accordingly, this study is guided by two hypotheses:

H1. 

The integration of Sentinel-2 optical data and ALOS-2 L-band SAR data yields more accurate and robust estimates of mangrove aboveground biomass than either sensor used alone, by reducing saturation effects and capturing complementary biochemical and structural information.

H2. 

The inclusion of vegetation indices and radar polarization features further enhances prediction accuracy and species discrimination in heterogeneous mangrove stands.

Building upon these hypotheses, the broader significance of this study extends beyond methodological contributions. Beyond methodological improvements, this research also supports mangrove conservation and sustainable management. By generating accurate and spatially explicit biomass estimates, our framework provides critical inputs for blue carbon accounting initiatives (e.g., REDD+ programs, national greenhouse gas inventories) and contributes to international climate commitments under the UNFCCC and nationally determined contributions (NDCs). At the regional scale, these biomass maps can inform mangrove restoration planning, guide ecosystem service payment schemes, and support evidence-based coastal management policies, thereby linking scientific advances with practical applications. In the case of the Matang Mangrove Reserve, this integrated framework achieved significantly higher accuracy (R² = 0.735) compared with typical optical-only models (R² often <0.5 in dense stands [5,6]), demonstrating a clear methodological advance over previous studies.

2. Materials and Methods

2.1. Study Area

In Malaysia, the Matang mangrove reserve is situated in the coastal waters between 4°09′–4°54′ N and 100°30′–100°38′ E (Figure 1). A vast expanse of mangroves is intricately connected to shallow yet fertile intertidal mudflat coastal regions through an intricate network of rivers and estuarine waterways. Covering an extensive area of more than 40,000 hectares, the Matang mangrove reserve ranks as the largest mangrove reserve in Peninsular Malaysia.

The climate of the Matang mangrove reserve in Malaysia is equatorial, characterized by warmth and high humidity. Average annual temperatures range from 23.7 to 33.4 °C, accompanied by humidity levels spanning 76.5% to 83.5%. Rainfall patterns typically range between 2000 and 3000 mm [26]. The Malaysian government manages the Matang mangrove forest, with about 80% of the area designated for a sustainable yield system that includes charcoal and pole production under a 30-year rotation cycle.

2.2. Data

2.2.1. Field Data

The field data for this study were obtained from the Matang Mangrove Forest Reserve (MMFR), where a total of 46 circular sample plots were established across different management zones (productive, protective, restrictive productive, and unproductive forests) to ensure representativeness. The sampling design followed the Matang Working Plan (2010–2019) and was intended to capture variability across different stand conditions and management stages (e.g., plantation, thinning, and final harvesting). Plot locations were determined using Trimble GPS to ensure precise georeferencing.

Each plot consisted of a central circular plot with a 10 m radius, within which 5 m and 2 m subplots were nested. In the 2 m subplot, saplings were enumerated; in the 5 m subplot, trees with Diameter at Breast Height (DBH) of 5–9.9 cm were measured; and in the 10 m subplot, all trees with DBH ≥ 10 cm were recorded. For each tree, DBH, height, species, and condition/status were documented. This stratified sampling approach ensured coverage of environmental variability, including proximity to tidal channels and differences in canopy density. A schematic layout of the plots is shown in Figure 2.

Across the 46 circular plots, the total inventoried area was 14,451 m² (plot radius = 10 m). Measured tree diameters ranged from 5.0 to 68.4 cm DBH, and tree heights from 2.1 to 27.6 m. Saplings (DBH < 5 cm) were enumerated within the 2 m nested subplots. Species-specific wood density values were obtained from Komiyama et al. (2005) [27] for Rhizophora and Avicennia species, supplemented by the Global Wood Density Database for less common taxa. A concise summary of plot size, DBH, and height ranges, and wood-density sources is provided in Supplementary Table S1.

2.2.2. Image Data

This study estimated the aboveground biomass in the Matang mangrove Reserve using Sentinel-2 and ALOS-2 satellite data after preprocessing (

Table 1. Imaging data used in the study.

Image Type	Image Name
Sentinel 2	S2A_MSIL1C_20170723T032541_N0205_R018_T47NPF_20170723T034239
ALOS 2	ALOS2180920080-170928_FBDR1.5RUA

).

(1)

Sentinel-2 Data: Sentinel-2, operated by the European Space Agency under the Copernicus program, is an optical imaging satellite. It orbits at an altitude of 786 km and acquires imagery across 13 spectral bands, ranging from visible and near-infrared to shortwave infrared. Among these, bands 2, 3, 4, and 8 provide 10 m resolution over a swath width of 290 km, while three additional red-edge bands are particularly sensitive to vegetation chlorophyll and are therefore valuable for biomass estimation. After acquisition, Sentinel-2 Level-1C products were atmospherically corrected to surface reflectance using the Sen2Cor processor (v2.5), producing Level-2A outputs. Orthorectification and co-registration were applied to minimize geometric distortions, and cloud and shadow contamination were mitigated using the Scene Classification Layer (SCL) [20].

(2)

ALOS-2 Data: The Advanced Land Observing Satellite-2 (ALOS-2), developed by the Japan Aerospace Exploration Agency (JAXA), operates in the L-band (1.2 GHz) and provides data with a spatial resolution of approximately 3 m in range and 1 m in azimuth. Unlike optical sensors, it can acquire data under all-weather and day–night conditions. Its L-band wavelength penetrates dense canopies and is sensitive to forest structural attributes, making it particularly valuable for mangrove biomass estimation [21].

Both Sentinel-2 and ALOS-2 datasets underwent a standardized preprocessing workflow to ensure geometric consistency and radiometric reliability. For Sentinel-2 imagery, Level-1C products were atmospherically corrected to surface reflectance using the Sen2Cor processor (v2.5), producing Level-2A outputs. Orthorectification and co-registration were applied to minimize geometric distortions, and cloud and shadow contamination were mitigated using the Scene Classification Layer (SCL) [28]. For ALOS-2 PALSAR-2 SAR data, preprocessing included radiometric calibration to sigma naught (σ⁰) backscatter coefficients, orthorectification using the Shuttle Radar Topography Mission (SRTM) DEM at 30 m resolution, application of a refined Lee filter to suppress speckle while preserving edges and texture, and multi-looking to enhance radiometric stability and reduce speckle variance [29]. Finally, the processed SAR images were co-registered with Sentinel-2 surface reflectance data to enable integrated optical–SAR feature analysis. All preprocessing operations were performed using the ESA’s Sentinel Application Platform (SNAP) and ArcGIS, ensuring that both datasets were harmonized for subsequent feature extraction, model training, and biomass estimation.

2.3. Method

To ensure methodological transparency and reproducibility, the analytical workflow in this study was systematically structured, as shown in the schematic representation in Figure 3. Each block of the diagram denotes a distinct methodological component, and is elaborated in the following subsections. Specifically, the procedure commenced with the preprocessing of remote sensing imagery. Sentinel-2 data underwent radiometric calibration, atmospheric correction, image sharpening, and spatial subsetting, whereas ALOS-2 was processed with multi-looking, co-registration, spatial subsetting, and noise filtering. Following preprocessing, a suite of vegetation indices was calculated from Sentinel-2 data to generate vegetation index maps, while polarization indices were extracted from ALOS-2 imagery to construct polarization index maps. These spectral and radar-based predictors were subsequently employed in biomass modeling, wherein both linear regression and random forest algorithms were utilized to establish quantitative relationships with field-based biomass observations. Model performance was evaluated using cross-validation and statistical metrics. Thereafter, inverse distance weighting (IDW) interpolation was applied to generate spatially continuous biomass surfaces. The final output was a biomass estimation map characterizing the spatial heterogeneity of aboveground biomass across the study area.

2.3.1. Biomass Calculation

Biomass estimation for individual trees was conducted using the allometric equation established by Komiyama [27]:

Biomass, B(kg) = 0.251pDBH^2.46

(1)

where p is wood density. DBH stands for “Diameter at Breast Height”, a standard measure of tree diameter taken at approximately 1.3 m above ground (or another height defined for a given study). DBH is an important metric for assessing tree biomass, growth rates, and overall health. The total biomass within each plot was calculated as the sum of individual tree biomass values within the plot.

To assess sensitivity to allometric choice, we additionally tested two widely used alternatives: (i) Ketterings (2001) [8] for mixed tropical secondary forests and (ii) Chave (2014) [30] for moist tropical forests. Biomass was recalculated for all 46 plots using these equations, and deviations from Komiyama’s model were quantified in terms of bias, mean absolute error (MAE), and root mean square error (RMSE) [8,30].

2.3.2. Feature Selection

Several studies have demonstrated a correlation between biomass and vegetation indices, as well as radar polarization [31,32,33,34,35,36]. In this study, predictor selection followed a transparent three-step procedure: (i) initial screening of widely used vegetation indices (NDVI, NDRE, SAVI, GNDVI, OSAVI, EVI) and radar polarization modes (HV, HH) based on prior evidence of their sensitivity to canopy greenness, chlorophyll content, soil background, and structural attributes [37,38,39,40]; (ii) collinearity control, where highly correlated variables (|r| > 0.80, VIF ≥ 10) were excluded; and (iii) importance-based pruning, retaining only predictors consistently ranked among the top contributors in regression and random forest models (MDA > 5%).

This process reduced the predictor set to a compact and interpretable combination dominated by OSAVI and HV polarization. These two variables consistently achieved the highest rankings in both correlation and variable importance analyses, and their effectiveness is supported by theory: OSAVI minimizes soil background influence and enhances sensitivity to vegetation cover [37], while HV polarization is particularly responsive to canopy structure and woody biomass [38], Formulas for the vegetation indices are provided in

Table 2. Partial vegetation index formula.

Vegetative Index	Algorithm Formula	Author
NDVI	(NIR − RED)/(NIR + RED)	[39,40]
NDRE	(NIR − RED EDGE)/(NIR + RED EDGE)	[41]
SAVI	(1 + L) ∗ (NIR − RED)/(NIR + RED + L)	[42]
GNDVI	(NIR − GREEN)/(NIR + GREEN)	[43]
OSAVI	(1 + L) ∗ (NIR − RED)/(NIR + RED + L)	[37]
EVI	2.5 ∗ ((NIR − RED)/((NIR) + (C1 ∗ RED) − (C2 ∗ BLUE) + L))	[44]
	Where L = 0.5

. By adopting this structured selection pathway, we ensured that variable choice was both empirically supported and theoretically justified, while avoiding redundancy and overfitting.

Predictor screening and pruning. To ensure a transparent predictor-selection pathway, we applied a three-step procedure. (i) Screening: all candidate vegetation indices and SAR backscatter variables were first checked for data completeness and prior relevance to biomass estimation. (ii) Multicollinearity checks: we computed pairwise Pearson correlation coefficients and removed highly correlated variables (|r| > 0.800); in addition, variables with a variance inflation factor (VIF) ≥ 10 were excluded to avoid redundancy and inflated importance. For example, NDVI and EVI showed strong collinearity, and only one index was retained. (iii) Importance-based pruning: the remaining predictors were ranked by Random Forest mean decrease in accuracy (MDA); variables with consistently low contribution (MDA < 5%) were discarded. This pathway (screening → collinearity control → importance-based pruning) yielded a parsimonious, interpretable, and robust predictor set. The full list of tested predictors and preprocessing steps is provided in Supplementary Table S2, and the final importance ranking in Figure S1.

Rationale for OSAVI & HV. Consistent with prior work and our own ranking, OSAVI (optical) and HV backscatter (L-band SAR) emerged as the most informative features, reflecting complementary sensitivity to canopy biochemical and structural attributes.

2.3.3. Linear Regression

Linear regression was employed to examine relationships between selected predictors and field-based biomass. All predictors were standardized (z-score) prior to modeling. To reduce collinearity, predictors with |r| > 0.80 or VIF ≥ 10 were excluded. A bidirectional stepwise selection procedure (p-entry = 0.05, p-removal = 0.10) was used to retain only statistically significant predictors, with the intercept included in all models. Residual diagnostics were conducted using the Shapiro–Wilk test for normality and the Breusch–Pagan test for homoscedasticity. Model performance was evaluated with leave-one-out cross-validation (LOOCV), consistent with Section 2.3.5.

2.3.4. Random Forest

Random Forest regression was applied to model non-linear relationships between predictors and biomass. The algorithm was implemented in Orange 3.36.1 (scikit-learn backend). Hyperparameters were optimized through grid search, yielding a final configuration of 500 trees, maximum depth = 20, and √p features per split. All other parameters followed scikit-learn defaults unless otherwise specified. Feature importance was quantified using mean decrease in accuracy (MDA). Full tuning ranges, final hyperparameters, and feature-importance outputs are provided in Supplementary Table S4 and Figure S1.

2.3.5. Leave-One-Out Cross-Validation Method

In machine learning, assessing the performance of a model is crucial. In this study, we explicitly adopted an LOOCV strategy instead of k-fold cross-validation or a simple train/test split. The key reason lies in the limited number of field plots (n = 46). A simple train/test division would substantially reduce the effective training set, while even a 5-fold cross-validation would leave only 37 plots per fold for training, and a 10-fold cross-validation would leave 41 plots. Such reductions may yield unstable or biased estimates under small-sample conditions. By contrast, LOOCV ensures that every single sample is used both for training and validation, thereby maximizing the utility of the dataset and providing a nearly unbiased estimate of model performance. Although LOOCV is computationally more intensive and may yield high variance in large datasets, under our small-sample setting, it represents the most appropriate choice [45].

Cross-validation is a commonly used evaluation method aimed at examining a model’s generalization ability, specifically its capability to adapt to previously unseen data. LOOCV represents a form of cross-validation that aims to thoroughly utilize all samples within a dataset for model validation [46].

For a dataset comprising n samples, LOOCV individually reserves each sample as a validation set while employing the remaining n − 1 samples for training. With each sample, the current sample serves as the validation set, while the remaining n − 1 samples from the training set. The model is trained using the training set data, then applied to the reserved validation sample to evaluate the model’s performance. Performance metrics of the model on that sample are recorded. This process is executed for every sample within the dataset.

LOOCV possesses the advantage of fully utilizing all dataset samples for model validation. However, due to each sample being a potential validation set, it incurs high computational costs. This becomes particularly challenging when dealing with large datasets, necessitating the training and validation of numerous models, resulting in increased computational complexity. It may also lead to high variance.

In addition to LOOCV, we conducted supplementary robustness checks to ensure that model performance was not inflated by the limited sample size. Specifically, (i) repeated random hold-out validation (80% training, 20% testing, with 50 repetitions) produced results comparable to LOOCV (R² = 0.720–0.740; RMSE = 47–49 t/ha); (ii) model sensitivity tests showed that the integrated optical–SAR model consistently outperformed both optical-only and SAR-only models; (iii) sensitivity to allometric equations was evaluated by recalculating plot-level AGB using Komiyama et al. (2005) [27], Chave et al. (2014) [30], and Ketterings (2001) [8], which shifted absolute biomass values but only marginally affected accuracy (ΔR² ≈ 0.02–0.03; ΔRMSE ≈ +5 t/ha) [8,27,30]; and (iv) subsampling experiments using 60%, 70%, 80%, and 90% of the plots produced stable outcomes (R² = 0.705–0.736; RMSE = 47–50 t/ha). These additional analyses confirm that the superiority of the integrated optical–SAR Random Forest model is consistent across validation schemes, predictor sets, and allometric assumptions, demonstrating the robustness of our conclusions. The results are summarized in Supplementary Table S6, with details provided in Supplementary Tables S2 and S3.

2.3.6. Spatial Interpolation and Map Production (IDW)

To generate spatially continuous AGB distribution maps across the study area, the field plot–derived AGB estimates were interpolated using the Inverse Distance Weighting (IDW) method. IDW is a deterministic interpolation technique that assumes spatial autocorrelation, whereby the influence of a sampled point diminishes with increasing distance from the prediction location. The interpolated value at each unsampled location is calculated as a weighted average of surrounding observations, with weights assigned inversely proportional to the distance between the prediction point and the sampled plots.

In this study, the power parameter was set to 2, which balances local detail and surface smoothness, while a fixed search radius including a minimum of 12 neighboring points was applied to ensure stable predictions. This approach enabled the production of spatially explicit biomass surfaces that reflect both local variability and broader spatial trends. The resulting interpolated AGB maps were subsequently validated against withheld plot-level observations and cross-checked with model outputs to confirm spatial consistency. Final cartographic products were generated in ArcGIS 10.8, where standardized symbology and legend scales were applied to facilitate interpretation and comparison across the study area.

IDW was selected as a pragmatic solution given the limited number of field plots (n = 46), which constrains the reliability of pixel-level model-based mapping. Unlike model-based prediction, IDW avoids extrapolation beyond the sampled range and provides a transparent, reproducible interpolation under data-scarce conditions. To mitigate the risk of over-interpreting spatial precision, we further incorporated an uncertainty surface. Specifically, cross-validated residuals from the Random Forest model were interpolated using the same IDW procedure to generate a credibility layer, highlighting areas with higher or lower confidence. This dual representation (AGB surface and uncertainty surface) ensures that the resulting maps are interpreted as indicative patterns rather than precise pixel-level predictions.

Final cartographic products were generated in ArcGIS, where standardized symbology and legend scales were applied to facilitate interpretation and comparison across the study area. The uncertainty surface is provided as Supplementary Figure S2.

3. Results

3.1. Generating Biomass Prediction Models Using Multispectral Images

3.1.1. Linear Regression Relationship Between Vegetation Index and Multispectral Imagery

The results of the linear regression analysis between the vegetation indices conducted using Orange software are shown in Figure 4.

The six vegetation indices (NDVI, NDRE, SAVI, GNDVI, OSAVI, and EVI) exhibited varying degrees of correlation with biomass. This is due to the distinct connotations associated with each of these six vegetation indices. The relationships among ground data (biomass data), vegetation indices, and root mean square error (RMSE) are summarized in

Table 3. The relationship between biomass and NDVI, NDRE, SAVI, GNDVI, OSAVI, and EVI using Linear Regression correlation and root mean square error (RMSE).

Vegetation Index	R2	RMSE
NDVI	0.030	98.309
NDRE	0.008	99.402
SAVI	0.036	97.966
GNDVI	0.031	98.227
OSAVI	0.036	97.966
EVI	0.035	98.058

. The R-squared (R²) values of the six vegetation indices were notably low. Given the exceedingly low R² values and substantial RMSE values, the predictive error of the model was unacceptably high. Consequently, it is advisable to discontinue using this model.

3.1.2. Random Forest Regression Relationship Between Vegetation Index and Multispectral Imagery

The results of the random forest analyses of the vegetation indices and biomass are illustrated in Figure 5.

All six vegetation indices (NDRE, EVI, GNDVI, NDVI, OSAVI, and SAVI) exhibited varying degrees of correlation with biomass. Given the distinct nature of each vegetation index, its correlation with biomass varied. The relationships between ground data (biomass data), vegetation indices, and RMSE are summarized in

Table 4. The relationship between biomass and NDVI, NDRE, SAVI, GNDVI, OSAVI, and EVI using Random Forest correlation and RMSE.

Vegetation Index	R2	RMSE
NDVI	0.501	70.530
NDRE	0.474	72.402
SAVI	0.533	68.201
GNDVI	0.478	72.103
OSAVI	0.579	64.738
EVI	0.586	64.235

. Although the R² values for the six vegetation indices reached approximately 0.5, compared with those of the linear regression, the RMSE values remained high. These outcomes did not provide sufficient support for subsequent stages of the study.

3.2. Generating Biomass Prediction Models Using Synthetic Aperture Radar (SAR) Images

3.2.1. Linear Regression Relationship Between Vegetation Index and SAR Image

As the model constructed using multispectral imagery failed to meet the prerequisites for the subsequent phases of the study, a 2017 ALOS-2 image equipped with synthetic aperture radar (SAR) was chosen for further investigation and two polarizations, HH and HV, were selected. HH and HV values were initially sampled for the respective ground data points. Subsequently, linear regression analysis between HH, HV, and biomass was conducted using Orange software. The results are presented in Figure 6.

The HH and HV values correlated with biomass. The strength of the correlation with biomass varied, as HH and HV represent distinct polarization patterns. Furthermore,

Table 5. The relationship between biomass and HH and HV using Linear Regression correlation and RMSE.

Polarization	R2	RMSE
HH	0.146	108.323
HV	0.189	105.580

provides a summary of the relationships between ground data (biomass data), HH, HV, and RMSE. Notably, HH and HV had low R² and high RMSE. These results do not offer adequate support for advancing subsequent phases of the study.

3.2.2. Random Forest Regression Relationship Between Vegetation Index and SAR Image

A follow-up experiment opted for the 2017 ALOS-2 image equipped with SAR because the model constructed using multispectral imagery failed to meet the criteria set for subsequent phases. Two polarizations, HH and HV, were selected from the ALOS-2 image. HH and HV values were initially sampled to correspond to ground data points. Subsequently, a random forest analysis of HH, HV, and biomass was conducted using Orange software. The results are presented in Figure 7.

As depicted in

Table 6. The relationship between biomass and HH and HV using Random Forest correlation and RMSE.

Polarization	R2	RMSE
HH	0.620	72.246
HV	0.627	71.619

, the HH and HV values correlated with biomass. The strength of the correlation with biomass varied according to the distinct polarization patterns indicated by HH and HV. The table also provides a summary of the relationships between ground data (biomass data) HH, HV, and RMSE. Notably, the R² values for HH and HV were significantly higher than those for the linear regression, although they remained relatively large, reaching approximately 0.6. However, the RMSE remained high. These findings do not offer sufficient support for advancing to the subsequent phases of the study.

3.3. Generating Biomass Prediction Models Using Synthetic Aperture Radar (SAR) Images

In light of the unsatisfactory performance observed in biomass prediction models constructed solely using multispectral and SAR images, the current study endeavored to enhance the predictive capacity by amalgamating both image types. Drawing upon the conclusions derived in Section 3.2, this section explores the interplay between HV, vegetation index, and biomass via bivariate analysis. The random forest model was selected as the preferred modeling approach, given its superiority over the linear regression model.

Ground data points were systematically sampled for the vegetation index and HV. Subsequently, a random forest model was constructed using Orange software to establish the interrelationships among vegetation index, HV, and biomass. The results are presented in Figure 8.

As depicted in

Table 7. Relationships between biomass and NDVI, NDRE, SAVI, GNDVI, OSAVI, EVI, and HV using Random Forest correlation and RMSE.

Vegetation Index	R2	RMSE
NDVI	0.530	65.356
NDRE	0.579	61.829
SAVI	0.673	54.488
GNDVI	0.656	55.879
OSAVI	0.796	43.638
EVI	0.725	49.969

, all six vegetation indices (NDRE, EVI, GNDVI, NDVI, OSAVI, and SAVI) were highly correlated with the biomass. However, the strength of the correlation varied given the distinct nature of each vegetation index. The relationships between ground data (biomass data), vegetation indices, and RMSE are summarized in Table 7.

Notably, the R² for the six vegetation indices generally exceeded 0.6, with the OSAVI demonstrating the highest R² of 0.735 and the lowest RMSE of 46.794. Thus, OSAVI was the vegetation index with the most robust correlation to ground data (biomass) and was thus selected for the subsequent phase of biomass prediction.

3.4. Sensitivity of AGB Estimates to Allometric Equations

The choice of allometric equation significantly influenced AGB estimates. Compared with Komiyama (2005) [27], Chave (2014) [30] produced consistently higher biomass values, particularly for trees with DBH > 30 cm, whereas Ketterings (2001) [8] underestimated smaller stems (DBH < 10 cm) [8,27,30]. At the plot level, the average bias reached +15.7% for Chave and −12.3% for Ketterings relative to Komiyama. MAE values were 13.2% and 11.5%, respectively, with RMSE differences of 10–15 t/ha (

Table 8. Comparison of AGB estimates using different allometric equations relative to Komiyama (2005) [27].

Equation	Formula	Bias vs. Komiyama (2005) [27]	MAE (t/ha)	RMSE (t/ha)
Komiyama (2005) [27]	$AGB=0.251\times p\times {DBH}^{2.46}$	—	—	—
Chave (2014) [30] (moist forest model)	$AGB=0.0673\times {(\rho \times {DBH}^2\times H)}^{0.976}$	+15.7%	13.2	61.8
Ketterings (2001) [8]	$AGB=0.11\times \rho \times {DBH}^{2.62}$	−12.3%	11.5	59.7

).

When these alternative AGB estimates were used in the Random Forest models, performance was moderately affected: R² decreased from 0.735 (Komiyama) to 0.701–0.715, while RMSE increased by 5 t/ha. Despite these variations, OSAVI and HV polarization remained the most influential predictors, indicating that the main ecological interpretations were not altered. Detailed numerical comparisons are provided in Supplementary Table S7.

3.5. Model Performance and Robustness Checks

The Random Forest (RF) model integrating Sentinel-2 indices with ALOS-2 L-band SAR backscatter achieved the best performance, with R² = 0.735, RMSE = 46.8 t/ha, and MAE = 35.4 t/ha under LOOCV (Table S5). Linear regression was considerably weaker (R² < 0.4). Supplementary 80/20 hold-out validation (50 repetitions) yielded comparable results (R² = 0.720–0.740; RMSE = 47–49 t/ha), confirming the robustness of the RF model.

Sensitivity analyses further highlighted the advantage of multi-sensor integration. Optical-only and SAR-only models performed poorly (R² = 0.452–0.501; RMSE > 62 t/ha), whereas the combined model consistently outperformed both. Subsampling experiments with reduced training sizes (60%–90%) produced stable outcomes (R² = 0.705–0.736; RMSE = 47–50 t/ha), suggesting that model accuracy is not strongly dependent on the full dataset. Finally, recalculating plot-level AGB with alternative allometric equations shifted biomass estimates by up to ±20%, but RF accuracy changed only marginally (ΔR² ≈ 0.02–0.03; ΔRMSE ≈ +5 t/ha). Detailed LOOCV and hold-out validation outcomes are provided in Supplementary Tables S4 and S6. Given the limited number of field plots (n = 46), these R² values may be somewhat optimistic; therefore, RMSE and MAE are emphasized as complementary indicators of predictive uncertainty.

3.6. Matang Mangrove Biomass Prediction Maps

Extrapolation of mangrove biomass inversion maps using established models. Employing Random Forest for the development of biomass inversion models and biomass inversion maps (Figure 9). The Random Forest algorithm captures complex and highly non-linear relationships, often achieving performance levels that surpass polynomial regression. Biomass has been proven to exhibit both linear and non-linear associations with vegetation indices.

The mangrove biomass within the Matang mangrove forest ranged from 48 to 240 t/ha, with an average biomass of approximately 146 t/ha.

3.7. Biomass Mapping and Uncertainty

The spatial distribution of AGB in Matang, derived through IDW interpolation of plot-level predictions, revealed pronounced heterogeneity. High biomass values were concentrated in long-rotation compartments dominated by mature stands, whereas recently harvested or regenerating zones exhibited notably lower values. Hydrological gradients were also evident, with relatively higher AGB in permanently inundated sites compared to more disturbed or drainage-modified areas.

The corresponding uncertainty surface (Figure S2) indicated that prediction errors were smallest in central compartments with dense sampling coverage and largest along reserve boundaries with sparse plots. These patterns underscore the need to interpret IDW maps as indicative rather than exact, with confidence varying spatially according to sampling density and model residuals.

4. Discussion

4.1. Comparison with Global and Regional Mangrove Biomass

Mangrove biomass in Matang has declined markedly compared with both historical levels and global averages. The mean biomass between 10° S and 10° N is 206.7 t/ha [47,48,49], whereas the current average in Matang (~146 t/ha) is considerably lower. Managed secondary mangroves in Vietnam and Thailand (150–220 t/ha) [49], place Matang within the regional mid-range, while arid-zone and Caribbean mangroves (<150 t/ha) provide a low-end benchmark. This positioning suggests that Matang has shifted from a historically high-biomass system to one that now more closely resembles moderately exploited forests in Southeast Asia.

The observed decline reflects both ecological and management drivers. Rotational harvesting (30-year cycles), selective extraction of large stems, and disturbance–recovery dynamics reduce stand-level aboveground biomass (AGB), while hydrological modification and salinity stress further constrain regrowth. Importantly, this pattern is not unique to Matang: long-term harvesting in Indonesia and Vietnam has also been shown to depress biomass stocks compared with pristine reference sites. However, the persistence of a mid-range biomass in Matang highlights that management can sustain relatively high productivity despite continuous exploitation, albeit below the global optimal level. These insights underline the need to contextualize biomass values not only as ecological signals but also as outcomes of socio-economic forestry practices.

4.2. Model Performance and Predictor Evaluation

Integrating Sentinel-2 multispectral indices with L-band ALOS-2 SAR significantly improved biomass estimation. Random Forest achieved the highest accuracy (R² = 0.735, RMSE = 46.8), outperforming linear regression and aligning with other machine-learning approaches (e.g., SVR, R² ≈ 0.73) [50]. OSAVI and HV polarization consistently emerged as the most informative predictors.

The transparent three-step predictor-selection framework (screening, collinearity control, importance pruning) proved valuable for both robustness and interpretability. For example, NDVI and EVI exhibited high collinearity, and only one was retained in the final model. This process minimized redundancy and prevented artificial inflation of variable importance. Methodologically, this approach contributes to broader applications in remote sensing, where multicollinearity often hampers model stability. By combining statistical diagnostics with machine-learning ranking, we demonstrate a replicable pathway that balances predictive power with ecological interpretability.

Beyond methodological novelty, the model’s performance carries ecological implications. The dominance of OSAVI indicates that leaf area and canopy greenness remain strong proxies for AGB, while HV polarization underscores the importance of structural scattering in mangroves. These findings suggest that optical–SAR synergies can overcome saturation problems more effectively than either data source alone, especially in mid-range biomass conditions. This contributes to a growing body of literature advocating for multi-sensor integration in tropical forest monitoring.

4.3. Systematic Deviations and Model Limitations

Systematic biases emerged at biomass extremes. Dense stands (>250 t/ha) were underestimated due to saturation in optical indices and SAR backscatter, while young regenerating plots (<100 t/ha) were overestimated, likely from soil and background reflectance effects. HV polarization showed strong biomass sensitivity, while HH had limited explanatory power, consistent with previous studies [51,52,53]. These patterns align with general limitations of passive and active sensors in capturing extreme vegetation structures.

A key limitation lies in the modest sample size (n = 46), which constrains the model’s ability to represent the full disturbance gradient in Matang. Although cross-validation demonstrated robustness, the relatively small dataset highlights the need for expanded sampling campaigns that incorporate diverse stand ages, hydrological conditions, and species compositions. Additionally, despite improvements from multi-sensor integration, persistent saturation at very high biomass suggests that LiDAR or GEDI canopy-structural metrics will be indispensable for accurately capturing upper-end variability.

Finally, plot-level uncertainty is compounded by temporal mismatches: satellite acquisitions and field measurements were not perfectly synchronized, which may introduce noise in dynamic mangrove environments. These limitations should be carefully considered when applying the model to operational biomass monitoring and carbon accounting.

4.4. Uncertainty from Allometric Model Selection

Biomass estimates are inherently sensitive to allometric choice. Komiyama (2005) [27], tailored for Southeast Asian mangroves, produced the most regionally relevant results, while Chave (2014) [30] and Ketterings (2001) [8] yielded deviations up to ±20% at plot level. Propagating these alternatives reduced RF accuracy slightly (ΔR² ≈ −0.02 to −0.03; ΔRMSE ≈ +5 t/ha) but did not alter predictor rankings, confirming robustness of ecological interpretations [8,27,30].

This result highlights two key points. First, while absolute biomass values may vary with allometric assumptions, relative patterns and spatial drivers remain stable, suggesting that ecological and management inferences are not overly sensitive to model choice. Second, the quantification of allometric uncertainty is itself an important contribution, as many previous studies have overlooked this factor, potentially inflating confidence in biomass maps. By explicitly propagating multiple allometries, this study provides a more transparent framework for uncertainty assessment in mangrove biomass estimation.

Validation analyses further support reliability: LOOCV, repeated hold-out (R² ≈ 0.72–0.74), and subsampling (60–90% of plots, R² = 0.705–0.736) all produced stable performance. These results indicate that RF outcomes are not artifacts of overfitting despite the small sample size, strengthening confidence in the generalizability of the approach.

4.5. Validation Adequacy and Robustness

IDW-based maps are presented as indicative patterns rather than pixel-level predictions. High biomass zones (>200 t/ha) occur in older protective stands, while low values (<100 t/ha) align with recently harvested compartments. Hydrology and disturbance history further explain spatial heterogeneity. To aid cautious interpretation, a residual-based uncertainty surface was added, highlighting areas of higher and lower confidence.

These spatial patterns align with silvicultural cycles, selective harvesting, and hydrological modifications in Matang, underscoring the ecological and management relevance of remote sensing for sustainable forest management, carbon accounting, and conservation planning. More broadly, the visualization of uncertainty adds practical value: managers can identify zones requiring additional field verification, policymakers can prioritize areas of high carbon density for conservation, and REDD+/blue carbon initiatives can integrate uncertainty layers into credit accounting.

Looking ahead, expanding the framework to incorporate higher-resolution UAV imagery, GEDI structural data, and long-term time series will enhance both accuracy and temporal representativeness. Moreover, comparative application in other mangrove regions (e.g., Indonesia, the Philippines, the Caribbean) would test the generalizability of the workflow and strengthen its contribution to global mangrove biomass monitoring.

5. Conclusions

This study demonstrates that integrating Sentinel-2 multispectral indices with L-band ALOS-2 SAR backscatter provides a reliable and scalable framework for estimating mangrove aboveground biomass (AGB). The Random Forest model (R² ≈ 0.73; RMSE ≈ 47 t/ha) outperformed linear regression and single-sensor models, with OSAVI and HV polarization consistently identified as the most informative predictors. These findings confirm the added value of optical–SAR synergies, particularly the role of L-band in reducing saturation and improving sensitivity across a wide biomass range.

Beyond predictive performance, the study contributes methodologically by establishing a transparent predictor-selection framework, incorporating multiple validation schemes, and explicitly propagating allometric uncertainties. These refinements enhance reproducibility and provide a benchmark for future biomass mapping efforts in mangroves and other forest ecosystems.

Ecologically and managerially, the results reveal that spatial gradients in AGB closely track silvicultural cycles, selective harvesting, and hydrological modifications in the Matang Mangrove Reserve. This underscores the potential of remote sensing not only for biomass estimation but also for guiding sustainable forest management, informing carbon accounting, and supporting conservation planning in intensively managed mangrove systems. The inclusion of uncertainty surfaces further enhances the operational value of the approach, allowing managers to identify high-priority areas for monitoring and intervention.

Nevertheless, certain limitations remain. The restricted number of field plots constrains model generalizability, and saturation effects persist at very high biomass levels. Future research should expand field sampling, integrate canopy-structural metrics from LiDAR or GEDI, and evaluate transferability across contrasting mangrove regions (e.g., arid, deltaic, and island systems). Incorporating UAV-based observations and multi-temporal analyses could further improve sensitivity to fine-scale disturbance–recovery dynamics. Such advances will contribute to more precise and policy-relevant assessments of mangrove carbon stocks, strengthening their role in global climate mitigation and blue carbon initiatives.