The study area (Figure 1) was located in the coastal area of Wenzhou and Taizhou, Zhejiang Province (120°26′E-121° 59′E, 27°6’N-29° 11’N), which belongs to the subtropical monsoon climate area, with the annual average temperature around 18°C and the average annual precipitation of 1750 mm. The total area of the study area was 271,606 ha, of which the land area was 140,504 ha. This area was the northern boundary of mangroves planted in China, among which mangroves were all artificial introduced species, such as Candelaria and Camelia, etc., with a total area of about 393.79 ha, among which the area of mangroves in Wenzhou was about 257.01 ha and that of mangroves in Taizhou was about 129.76 ha (Hao et al., 2024).

Hyperparameters in machine learning are options that are set outside of model training and are not optimized or changed during training. The selection of different hyperparameters has a great influence on the performance and accuracy of the model, so it is necessary to select the optimal hyperparameters for training. In this study, multiple hyperparameters need to be adjusted for different machine-learning algorithms. The detailed information is shown in Table 4. To obtain the optimal hyperparameters, this study used the 5-fold cross-validation method (Hajihosseinlou et al., 2023) to randomly divide the training set into five subsets of the same size. The data of four subsets were used for training the model, and the remaining subsets were used for testing the model. Then the five possible choices were repeated, and the hyperparameters with the highest average accuracy on the five verification sets were obtained by the grid search method.

The max_depth within the hyperparameters serves as a critical parameter for controlling model complexity. Excessively shallow depth can easily lead to underfitting, failing to capture underlying patterns in the data, while excessive depth significantly increases the risk of overfitting and computational overhead. Therefore, this study defined a range of max_depth covering both commonly used shallow and medium-depth trees. In addition, the n_estimators also influences model fitting performance: too few trees may weaken the ensemble effect, whereas too many result in diminishing marginal returns and a sharp increase in computational cost. Accordingly, a range of 20–200 was selected, encompassing typical intervals from basic to large-scale ensembles, with a step size designed to facilitate efficient identification of the optimal value through grid search. For the SVM kernel function selection, multiple commonly used kernel types were employed to better capture both nonlinear and linear relationships. The selection of other parameters was based on relevant domestic and international studies to enhance the rationality and scientific rigor of parameter configuration (Huang et al., 2024; Li et al., 2023; Prakash et al., 2022).

The evaluation indicators of AGB estimation accuracy used in this study were root mean square error (RMSE) and R². RMSE is a standard deviation that measures the difference between the predicted value and the true value and is commonly used to evaluate the predictive performance of regression models. R² is a common regression analysis metric used to evaluate a model’s ability to explain data variation, and its value ranges from 0 to 1, with a closer to 1 indicating the model’s better fit to the data. The calculation formulas are as follow in Equations (2) and (3):

where A i for the first 𝑖 sample AGB per unit area of predicted value, A ri for the first 𝑖 sample AGB per unit area of the measured values, A r ¯ to sample all the average of the measured values, 𝑛 for the total sample area. In addition to accuracy evaluation metrics, this study also assessed the computational efficiency of various machine learning models, focusing on training time, training set prediction time, test set prediction time, and total computational time. This analysis provides a clearer reflection of the computational cost associated with the models’ learning of data patterns.

In this study, sample labeling for classification was conducted using relevant high-resolution imagery, visual interpretation, and field survey samples. A total of 7,451 samples were selected. With reference to the ratio commonly adopted in related classification studies, the samples were divided into training and testing sets at a ratio of 3:1. The specific number of samples for each feature category is presented in the Table 5. This study used the maximum likelihood method (Atkinson and Lewis, 2000) to obtain the classification results of six land types in the region, and the distribution information of mangrove forests. The area of the mangrove forest was 522.31 ha, as shown in Figure 2. This study selected 1,545 ground feature test samples to evaluate the accuracy of classification results and to obtain the recall rate and precision rate of each land use type. This study calculated the overall accuracy and Kappa coefficient, as shown in Table 6. The overall accuracy was 92.39% and the Kappa coefficient was 0.9087, indicating a high accuracy. At the same time, the recall rate of mangroves was 94.82% and the precision rate was 86.43%. As can be seen from the distribution map of mangrove extraction, mangroves were mainly distributed in the estuary area of several rivers and near islands, such as Ou River, Ao River, and Dongtou Islands, and are also distributed in the coastal areas of Yueqing Bay, Sanmen Bay, Yanpu Bay.

This study selected the importance of 304 variables consisting of spectral features, vegetation index, and texture features by RFIR. Figure 3 shows the %IncMSE of the first 30 variables after 100 runs using the RFIR. The texture variable was abbreviated as the base spectral band feature name window size. It can be seen from the figure that the correlation between texture characteristics and AGB was strong, followed by spectral characteristics.

According to the machine learning hyperparameter configurations, the parameters of the machine learning algorithms were optimized by using a 5-fold cross-validation method and grid search hyperparameter setting. This study used the minimum root-mean-square value as the scoring standard. The optimal hyperparameters of the five algorithms are shown in Table 7.

This study used the optimal hyperparameters to predict. From the AGB prediction accuracy of the training set and test set of the five algorithms, it can be seen that XGboost after hyperparameter optimization had the best performance on the training set, R² was 0.998, RMSE was 0.634 Mg ha⁻¹, and RMSE of the RF was 4.736 Mg ha⁻¹. At the same time, the accuracy of LCE was not higher than XGboost, RMSE was 3.928 Mg ha⁻¹. The R² of the five algorithms on the training set was >90%, indicating that the model had a good fitting effect between the remote sensing characteristics and the AGB. XGboost worked best on the test set, with R² of 0.682 and RMSE of 6.851 Mg ha⁻¹. In addition, the accuracy of CATboost was also better, with R² of 0.697 and RMSE of 6.92 Mg ha⁻¹. On the test set, the effect of RF and SVM was poor, while the R² of the other three integrated algorithms was >0.65. Therefore, XGboost, which was effective in the training set and test set, was selected for the inversion of regional mangrove AGB in this study (Figures 4, 5).

Furthermore, this study compared the computation times of different models on both the training and test sets, as shown in the Table 8. The results indicate that the LCE model incurred the highest computational cost, particularly requiring the longest training time—approximately 300 times longer than other models. It was also the slowest in both training set prediction and test set prediction, indicating higher model complexity and lower computational efficiency compared to the other models. Among the five models, SVM was the fastest, taking only 0.01 s, while the other three models also completed within 1 s, demonstrating high computational efficiency. With the exception of the LCE model, the other four models achieved both training and testing times within 1 s, indicating strong real-time performance.

The following figure shows the results of remote sensing estimation of mangrove AGB in coastal Zhejiang Province based on the XGboost. In addition, this study extracted the inversion information of regional mangrove AGB. The average AGB was 11.35 Mg ha⁻¹, and the total AGB was 5930.28 Mg. From the perspective of spatial distribution, the mangrove AGB per unit area the Dongtou Islands is higher, with an average of more than 16.82 Mg ha⁻¹, followed by the mangrove AGB along the coast of the estuary of Ao River and Sanmen Bay, while the AGB of the Ou River estuary in the middle is lower, with an average of <10 Mg ha⁻¹ (Figure 6).

This study employed the RF algorithm to evaluate the importance of 304 input feature variables and selected the top 30 important features. The results indicate that textural features made the most significant contribution (28 in total), while spectral features and vegetation indices had relatively limited influence. Among the selected textural features, Correlation-type features accounted for the largest proportion (12 features), followed by Entropy and Second Moment (5 features each). The prominence of textural features is closely related to the structural complexity of the mangrove ecosystem. This distribution suggests that indicators representing textural uniformity, directionality, and heterogeneity collectively influence the inversion accuracy of mangrove biomass (Shili et al., 2017). Textural information, which is associated with the spatial structure of vegetation, effectively reflects internal canopy shadows and structural characteristics, demonstrating considerable potential for estimating forest AGB (Kraus et al., 2009; Laginha Pinto Correia et al., 2017; Sarker and Nichol, 2011). Furthermore, among the selected textural features, those extracted using a 3 × 3 window predominated (16 features), which may be attributed to the spatial resolution of the remote sensing imagery and the distribution area of the mangroves. Smaller windows help capture detailed characteristics of mangroves, and the extracted textural information is less susceptible to noise interference, better representing random heterogeneity and more sensitively responding to biomass changes, thereby enhancing the correlation with biomass (Cao et al., 2021).

Additionally, due to the optical remote sensing’s limited penetration capability, it struggles to capture vertical vegetation structure information and is susceptible to spectral saturation (Avtar et al., 2012; Mermoz et al., 2014). The sensitivity of vegetation indices decreases in high-biomass regions, and saturated effects may occur in locally high-density areas within the study area, resulting in weaker correlations with AGB. Furthermore, in mangrove habitats, tidal inundation, mudflats, and underlying soil reflectance can interfere with the relationship between vegetation indices and actual AGB. As this study was based on single-temporal imagery, vegetation phenology and seasonal variations cause changes in vegetation color, also contributing to the lower correlation of vegetation indices. Textural features, influenced by surface properties and illumination conditions, are independent of color variations and can reflect spatial patterns that single spectral pixels cannot express, thereby compensating for the limitations of traditional vegetation indices in estimating crop growth parameters (Li et al., 2020; Zhang et al., 2021). Different species and growth stages of vegetation exhibit distinct textural characteristics, with the canopy and foliar structures of large vegetation forming unique textural patterns in imagery, thereby providing higher-order spatial information that enhances biomass estimation accuracy. Although previous studies have suggested that vegetation indices may play a more critical role than original bands in Sentinel-2-based biomass models (Chrysafis et al., 2019), numerous studies have demonstrated that textural features more effectively mitigate the impact of spatial heterogeneity and improve the estimation accuracy of biomass and carbon storage (Lu and Batistella, 2005).

This study used the Sentinel-2 satellite image data combined with ground sample survey data to construct mangrove AGB models by using five machine learning algorithms. Of the five machine learning algorithms, XGboost achieved the best results, with higher accuracy than the latest LCE.

On the test set, both XGboost and CATboost achieved higher accuracy than the other three algorithms, where the RMSE of XGboost was lower than that of CATboost, and the R² of XGboost was lower than that of CATboost. RF was the least effective of the five algorithms, because RF made predictions by integrating multiple decision trees, and the independence between its trees often limits its expressiveness (Pham et al., 2020a). In addition, LCE’s R² on the test set was slightly lower than XGboost and CATboost, and RMSE was also higher than both models. This was since LCE’s local modeling strategy may be adaptable to specific data, but faced challenges when generalizing to different test sets. LCE captured local features of the data through local modeling, but the capabilities of such local models may be limited. Both XGboost and Catboost are gradient-lift tree-based algorithms that optimize prediction results by progressively improving the model while handling complex nonlinear relationships and having greater expressibility in feature Spaces (Zhai et al., 2023). In addition, XGboost included regularization terms (L1 and L2 regularization), which can effectively prevent overfitting, thereby improving the generalization ability of the model. In related studies, the application of the XGboost algorithm combined with differential evolution for mangrove AGB estimation yielded superior performance compared to other machine learning models, achieving an R² of 0.84 (Shen et al., 2025). Another study (Pham et al., 2020b) integrated XGboost with a genetic algorithm for mangrove biomass estimation, reporting an R² of 0.683 and an RMSE of 25.08 Mg ha⁻¹. Additionally, research combining Sentinel-2 imagery and CATboost for mangrove AGB estimation (Pham et al., 2021) achieved an R² of 0.665 and an RMSE of 18.41 Mg ha⁻¹. In comparison, this study utilized Sentinel-2 data and XGboost for biomass estimation, achieving a comparable R² of 0.682, while demonstrating a significantly lower RMSE of 6.851 Mg/ha, indicating improved precision in AGB prediction.

Furthermore, factors such as image quality, model selection, and hyperparameter settings can introduce uncertainties in the estimation of aboveground mangrove AGB (Lu et al., 2016). This study utilized calibrated Sentinel-2 imagery and selected cloud-free time periods to effectively minimize uncertainty related to image quality. The uncertainty analysis in this research primarily focused on model selection and hyperparameter configuration. During model selection, cross-validation and grid search were employed to identify optimal hyperparameters and select the best-performing model on the test set for AGB inversion. However, due to the inherent randomness and systematic errors associated with XGboost, a repeated random prediction approach was adopted for uncertainty analysis to evaluate prediction errors under different configurations. Specifically, biomass inversion was performed 100 times using different random seeds, and the per-pixel standard deviation (STD) of the predictions was calculated (Chen et al., 2015). The final statistical results are shown in the Table 9. As illustrated, the STD is mostly below 6 Mg ha⁻¹, accounting for approximately 90% of cases, with a mean standard deviation of 3.44 Mg ha⁻¹ and a maximum not exceeding 20 Mg ha⁻¹. These results indicate that the error caused by randomness is relatively low, suggesting minimal uncertainty in the model predictions.

However, this study still has certain limitations that need to be addressed and improved in future research. First, regarding sample representativeness and spatial heterogeneity, the 340 sample plots used in this study cover a total area of 522.31 hectares, which may not fully capture the highly complex spatial heterogeneity within the mangrove ecosystem. The model may underestimate or overestimate AGB in these regions. To enhance model accuracy and generalizability, future studies should consider increasing the number of samples, especially by conducting additional surveys in sparsely sampled areas. Second, the spatial resolution of Sentinel-2 data remains a limitation in areas with high mangrove density or strong heterogeneity. In regions with complex canopy structures and high species mixing, intra-pixel spectral mixing is significant, which may lead to biases in biomass estimation—particularly in transition zones between forest and water, where non-vegetation signals in boundary pixels further affect inversion accuracy. Future research could enhance the perception of vertical structural information by integrating multi-source remote sensing data, or use the Google Earth Engine platform to incorporate multi-temporal imagery to reduce noise caused by phenological and seasonal variations.