Mangrove species classification using novel adaptive ensemble learning with multi-spatial-resolution multispectral and full-polarization SAR images

ABSTRACT

Mangroves are one of the important components of Earth's carbon sinks. The current problems of base-model composition strategy of ensemble learning and image features combination are still major challenges in mangrove species classification. This paper constructed two novel adaptive ensemble learning frameworks (AME-EL and AOS-EL) to explored the effect of combing different spatial-resolution optical and SAR images on classification performance, and evaluated the ability in mangrove species classification between dual-polarization and full-polarization SAR images. Finally, we used the SHAP method to explore the effects of different feature interactions on mangrove species classification. The results indicated that: (1) AME-EL and AOS-EL achieve the fine classification of mangrove species with overall accuracies between 77.50% and 94.77%. (2) Combination of Gaofen-7 multispectral and Gaofen-3 SAR improved the classification accuracy for Kandelia candel, with the F1 score increasing from 26.4% to 40.2%. (3) The VV/VH polarization performed better in the classification, with the F1 scores for Aegiceras corniculatum and Kandelia candel were higher than those of HH/HV and AHV polarization by 7%−16.1% and 5.9%−16.1%, respectively. (4) SAR features interacted well with other spectral features, which made a strong contribution to the classification accuracy of mangrove species, and effectively affect the prediction results.

KEYWORDS:

• Mangrove species classification
• adaptive ensemble learning
• base-model composition strategy
• optical and SAR combination
• feature interactions

1. Introduction

Mangroves are one of the most carbon-rich ecosystems on Earth and have received much attention for their crucial role in regulating the global carbon cycle (Zhu and Yan 2022). Mangrove areas cover about 145,068 km² (Jia et al. 2023) with 84 species globally. Mangrove areas in China cover about 28,010 ha and account for 1/3 of global mangrove species. The dominant species include Avicennia marina, Aegiceras corniculatum, Kandelia candel, etc. (Zeng et al. 2023). Mangroves have important social, economic, ecological values, and provide a wide range of ecosystem services (Wang et al. 2023a; Liu et al. 2022). Mangroves play multiple and significant roles towards the UN’s Sustainable Development Goals (SDGs), such as regulating carbon storage (SDG13), providing habitats for fishes and marine organisms (SDG14), and providing fish products as well as important fishing grounds for coastal communities (SDG2) (Sasmito et al. 2023). Enormous global environmental changes and excessive interference from human activities have seriously affected mangrove habitats and environments (Xu et al. 2023). From the 1950s to the present, the global area covered by mangroves has declined by approximately 30% (Jia et al. 2022). The continued degradation of mangroves has led to a steady decline in their carbon sequestration and storage capacity, overloading the coastal ecological carrying capacity. Therefore, the fine classification of mangrove species and the timely as well as accurate evaluation of the health of mangrove ecosystems are essential for their management and conservation.

Multispectral satellite images have been widely used for mapping and monitoring mangrove species (Halder et al. 2021; Kamal et al. 2021; Navarro et al. 2020). For example, Wang et al. (2018) used Random Forest (RF) algorithms to classify mangrove species in the Dongzhaigang National Nature Reserve by combining Sentinel-2, Landsat 8, and Pléiades-1 multispectral satellite images, with the overall accuracy (OA) ranging from 68.57% to 78.58%; Peng et al. (2020) used Gaofen-2 and RapidEye-4 images as data sources to map the mangrove species on Qi'ao Island with ensemble learning, with the OA ranging from 81.15% to 92.01%. However, the usability of the optical images was seriously affected by the cloudy and foggy days in the coastal areas. Concurrently, the similarities in the spectral characteristics of vegetation canopies limited the identification of species of populations (Fu et al. 2023; Huang et al. 2022). Optical images and SAR images can complement each other and provide information on aspects such as the spectra, texture, spatial structure, and other features of mangroves to improve the accuracy of mangrove interspecific classification (Hu et al. 2020; Lu and Wang 2021). For example, Sharifi, Felegari, and Tariq (2022) used RF combined with Sentinel-1A and Sentinel-2A images to classify and map mangroves, and the F1 score of combined optical and SAR images was 0.95, which was a 5% improvement over only using an optical image. Zhen, Liao, and Shen (2018) combined Landsat 8 OLI and Radarsat-2 SAR images to map mangrove species in the wetland reserve of Hainan Island with the Support Vector Machines (SVM) method. The results showed that the OA improved from 83.5% when only using the optical image to 95.0% when using the combination of optical and SAR images; however, due to the complexity of mangrove composition and configuration, the individual pixels of low- and medium-spatial-resolution data may contain mixed information from different species (Pham et al. 2019), thus affecting the accuracy of mangrove classification. Concurrently, some studies have demonstrated that combined high-spatial-resolution optical and SAR images can achieve the fine classification of mangroves (Ghorbanian et al. 2021); however, the impact of the increased spatial resolution of optical and SAR images on improving the classification accuracy of mangrove species is still not fully explored. Different polarization modes and polarimetric decomposition parameters (PP) of SAR images can help improve the model performance for wetland vegetation classification significantly (Chen et al. 2020). For example, Tu et al. (2021) used full-polarization SAR image with Gaofen-3 and Zhuhai-1 OHS hyperspectral images, and achieved an OA of 97%, which was significantly higher than the classification accuracy of a single SAR or hyperspectral image. Fu et al. (2021) used RF to explore the synergy between multispectral and SAR images in marsh vegetation mapping. The results showed that the RF model integrating multispectral data, backscatter coefficients, and polarization decomposition parameters obtained the highest OA of 91.16%; however, due to the polarization modes of SAR images having different effects on the interaction with mangrove biophysical properties (size, geometry, leaves, trunks, etc.) (Aslan et al. 2016), the ability of polarization methods to the identify mangrove species needs to be specifically verified.

Machine learning algorithms, such as RF, SVM, Extreme Gradient Boosting (XGBoost), and Convolutional Neural Network (CNN) (Li, Kit Wong, and Fung 2019; Toosi et al. 2019; Wei et al. 2023), are widely used for vegetation classification due to their high applicability and ability to generate accurate as well as consistent results (Liu et al. 2021). For example, Ou et al. (2023) used XGBoost to classify mangrove species in the Pinglu Canal estuary, with an OA of 71.28%−96.78%. Gilani et al. (2021) used RF to map Pakistani mangroves, with an OA of 90%–97%. The above studies suggested that a single machine learning algorithm has high classification results in vegetation classification. In comparison, however, ensemble learning algorithms can effectively reduce the risk of model overfitting and improve classification accuracy by integrating multiple base models (Cai et al. 2020; Cao et al. 2021). For example, Yao et al. (2022) combined RF, XGBoost, and categorical boosting (CatBoost) algorithms to construct an ensemble learning model for wetland vegetation classification, and the results showed that the OA of the ensemble learning model reached the 92.27%, which was better than that of a single machine learning model. Long et al. (2021) used minimum distance, Bayesian, SVM, RF and classification and regression tree (CART) algorithms as the base model for stacking ensemble learning models to map the spatial distribution and species composition of wetland vegetation in the Dongting Lake. The results showed that the OA and Kappa of the ensemble learning model were 94.59% and 0.92, respectively, in which the OA was higher than single models of RF and SVM by 13.36% and 9.33%, respectively. Integrating multiple machine learning algorithms also comes with problems such as increased computational complexity and memory (Krawczyk et al. 2017). Therefore, the composition strategy of the base model in the mangrove ensemble learning model needs to be further explored. While the ensemble learning model improves the accuracy and robustness of the classification model, it was difficult to account for the effects of different feature variables on the model classification ability (Ma et al. 2023). The SHapley Additive exPlanations (SHAP) method calculates the marginal contribution of each feature to the model output and assigns a specific predictive importance value to each feature to ensure good global and local interpretability (Lin et al. 2023). This method has gradually been applied to the evaluation of the importance of wetland vegetation classification features (Zhang et al. 2022a); for example, Fu et al. (2022) used the SHAP method to calculate the importance of different features in mangrove classification models, and the results showed that multispectral bands, vegetation indices and DSM data contributed to the model. Although the above studies demonstrated the feasibility of the SHAP method in evaluating the classification of mangrove species with different characteristics, using it to evaluate the effect of the interaction between different characteristics on the classification accuracy of mangrove species requires further examination.

In order to cope with the above-mentioned challenges, we took the Maowei Sea Mangrove Nature Reserve wetland in Qinzhou City, Guangxi Zhuang Autonomous Region, as the study area. We constructed two new adaptive ensemble learning models with different base-model strategies for mangrove species classification using Sentinel-2A, ZY-3 GF-7 multispectral images, and GF-3 full-polarization SAR images. The main contributions of this paper include: (1) developing adaptive base-model multicollinearity elimination ensemble learning (AME-EL) and adaptive optimal base-model selection ensemble learning (AOS-EL) models, to explore the effect of different base-model strategies on mangrove species classification; (2) integrating multi-source multispectral and SAR images to quantitatively evaluate their ability to improve the classification accuracy of mangrove species; (3) exploring the effects of different polarization modes on mangrove species mapping based on dual-polarization (HH/HV, VV/VH) and full-polarization (AHV) SAR images; and (4) quantifying the interaction effect and contribution rates of optical as well as SAR features, further evaluating the influence of the interaction between multi-source features on model performance in mangrove species classification.

2. Study area and data source

2.1. Study area

The Maowei Sea Mangrove Nature Reserve is located in Qinzhou City, Guangxi Zhuang Autonomous Region, in southern China (Figure 1), which is characterized by a subtropical oceanic monsoonal climate. The average annual temperature is 22.4°C. The average annual precipitation is 2100 mm. The tide type is an irregular diurnal tide with an average tidal height of 2.5 m. Mangroves are usually distributed in the intertidal zone of estuaries with a mixture of sea and fresh water; however, the area is the confluence of three runoff streams, namely the Maoling River, the Qin River and the Dalan River, and forms a typical sea fork topography with rich resources and a shallow water depth. The tidal flats cover about 80% of the total area of the reserve, which is conducive to the deposition of sediment into the sea. The suitable climate as well as terrestrial sediment and nutrient provide good conditions for mangrove growth (Zhang et al. 2023). The total area of the reserve is 2784 ha, including 1892.7 ha of mangroves, which is the largest and most typical mangrove area in the Chinese island group (Jiang et al. 2020). Additionally, the dominant mangrove species in the study area are Aegiceras corniculatum, Cyperus malaccensis, Kandelia candel, Sonneratia apetala, etc.

Figure 1. Geographic location of study area (108°28′42″E-108°36′45″E,21°49′43″N-21°54′24″N). (a) ZY-3 true-color images and the distribution of sample points, (b) GF-3 false-color images with the VV/VH polarization decomposition parameters (red: Shannon entropy; green: entropy; blue: alpha), and (c) field survey photos.

2.2. Data acquisition and preprocessing

2.2.1. Multispectral images

The Sentinel-2A multispectral image (Level-2A product) used in this study was obtained from the United States Geological Survey (https://earthexplorer.usgs.gov/). The Gaofen-7 (GF-7) and Ziyuan 3-02 (ZY-3) images were obtained from the China Centre for Resources Satellite Data and Application (http://www.cresda.com/). The images were acquired on 23 January 2021 (ZY-3), 24 February 2022 (Sentinel-2A), and 14 April 2022 (GF-7). The preprocessing included: (1) the radiometric calibration, atmospheric correction, and ortho-correction of the GF-7 and ZY-3 images used the ENVI 5.6 software to obtain surface reflectance images; (2) the fusing of the 0.65 m panchromatic and 2.6 m multispectral images of GF-7 to obtain a new multispectral image with a spatial resolution of 0.65 m and resampling to 3 m using the spectral sharpening tool of ENVI 5.6 software; (3) the resampling of the Sentinel-2A image to 10 m using the S2 resampling processor tool of SNAP 9.0 software; and (4) the georeferencing of all multispectral images and cropping those using ArcGIS 10.6 software to unify the projected coordinate system as WGS 1984 UTM Zone 49 (Table 1).

Table 1. Summary of the main parameters of the three multispectral sensors.

Satellite	Spectral band	Spectral range (nm)	Spatial resolution	Satellite	Spectral band	Spectral range (nm)	Spatial resolution
GF-7	Blue	450–520	2.6 m	Sentinel-2A	Blue	458–523	10 m
	Green	520–590			Green	543–578
	Red	630–690			Red	650–680
	NIR	770–890			NIR	785–900
	Pan	450–900	0.65 m		Narrow NIR	855–875	20 m
ZY3-02	Blue	450–520	5.8 m		Vegetation Red Edge 1	698–713
	Green	520–590			Vegetation Red Edge 2	733–748
	Red	630–690			Vegetation Red Edge 3	773–793
	NIR	770–890			SWIR 1	1565–1655
	Pan	500–800	2.1 m		SWIR 2	2100–2280
-	-	-	-		SWIR-Cirrus	1360–1390	60 m
-	-	-	-		Coastal aerosol	433–453
-	-	-	-		Water vapor	935–955

2.2.2. Polarimetric SAR images

In this study, we selected C-band Gaofen-3 (GF-3) SAR images of with three polarization modes, including: (1) a quad-polarimetric SAR (AHV) image from 20 December 2021, with a spatial resolution of 5 m; (2) a dual-polarized (HH/HV) image with a spatial resolution of 3 m from 1 March 2020; (3) a dual-polarized (VV/VH) image with a spatial resolution of 3 m from 15 February 2021. The specific preprocessing of GF-3 SAR images included: (1) calculating the backscattering coefficients of different polarization modes (HH, HV, VH, and VV) using ENVI 5.6 software; (2) extracting the covariance matrix (C11, C22) and coherence matrix (T11, T22, T33) based on PolSARpro 6.0 software through multi-looking processing and Refined_Lee filtering; (3) collecting the polarimetric parameters from dual- and quad-polarimetric SAR images using four decomposition methods (Table 2); and (4) georeferencing as well as subsetting all SAR images and their derivative features, and unifying their projected coordinate system as WGS 1984 UTM Zone 49.

Table 2. Description of polarimetric decomposition methods of GF-3 SAR images.

Decomposition types	Specific decomposition methods	Number of features
Coherent decomposition	Touzi Decomposition (TSVM)	35
	Aghababaee Decomposition (Aghababaee)
Two-component decomposition	Huynen Decomposition (Huynen)	12
	Barnes 1 Decomposition (Barnes 1)
	Barnes 2 Decomposition (Barnes 2)
	Unified Huynen Decomposition (UHDx)
Scattering model decomposition	Freeman 2 Component Decomposition (Freeman 2)	48
	Freeman 3 Component Decomposition (Freeman 3)
	Yamaguchi 3 Component Decomposition (Yamaguchi 3)
	Yamaguchi 4 Component Decomposition (Yamaguchi 4)
	Bhattacharya & Frery 4 Component Decomposition (BF4)
	Neumann 2 Component Decomposition (Neumann)
	Arii 3 Component Decomposition (Arii3_NNED)
	Singh 4 Component Decomposition (Singh_G4U2)
	L.Zhang 5 Component Decomposition (MCSM)
	Singh-Yamaguchi 6 Component Decomposition (Singh_i6SD)
Vector feature decomposition	HAAlpha Decomposition (HAAlpha)	44
	Cloude Decomposition (Cloude)
	Holm1 Decomposition (Holm1)
	Holm2 Decomposition (Holm2)
	An & Yang 3 Component Decomposition (An_Yang3)
	An & Yang 4 Decomposition (An_Yang4)
	Van Zyl 3 Decomposition (VanZyl3)

2.2.3. Field measurements

In this study, we acquired the sample data through field surveying and the visual interpretation of ultra-high-spatial resolution images of belt transects from an unmanned aerial vehicle (UAV). We used a five-point sampling method to conduct field measurements. We firstly determined several field surveying areas with 10 m × 10, 5.8 m × 5.8, and 3 m × 3 m three different scales to correspond to the spatial resolution of Sentinel-2A, ZY3, and GF-7 remote sensing images, respectively, and then uniformly constructed multiple 1 m × 1 m sample plots in each field surveying area. We ensured that the vegetation type of 1 m × 1 m sample plots was the same. The center geographic position of the three scales of each sample plot was recorded by theV90 real-time kinematic global positioning system (Hi-Target, GZ, China). The UAV images were collected from a DJI P4M multispectral drone flying at 40 meters on cloud-free days with an 80% heading overlap and 80% side overlap between 24 and 25 November 2021 (10:00 am–15:00 pm, Beijing time). The original UAV images were processed via Pix4DMapper software, and produced a 0.05 m DOM image with a projected coordinate system of WGS 1984 UTM Zone 49N. We collected a total of 487 samples by visually interpreting the combined UAV image and field surveying in seven land cover types. The training and validation data was divided with a ratio of 7:3 (Table 3).

Table 3. Summary of training and testing samples in the study area.

Dataset	Aegiceras corniculatum	Cyperus malaccensis	Kandelia candel	Mudflat	Other	Sonneratia apetala	Water	Total
Training data	80	38	44	32	18	103	30	345
Validation data	38	14	18	16	6	40	10	142

3. Method

This study combined different-spatial-resolution multispectral images and full-polarization SAR images, and constructed two adaptive ensemble learning models to classify mangrove species in the Maowei Sea. The main steps (Figure 2) of the experiment were as follows: (1) the preprocessing of multispectral and SAR images to obtain their derivative features, and segmenting the multi-source image features to construct 12 multidimensional feature datasets; (2) conducting data dimensionality reduction of high-dimensional datasets based on a recursive elimination algorithm (RFE) to remove high-correlation features, and extracting the optimal feature combination in each scenario; (3) constructing two adaptive ensemble learning models for mangrove species classification, and examining the effects of different base-model composition strategies as well as combinations of optical and SAR images on classification accuracy; and (4) interpreting the contribution of active and passive image features to the classification accuracy of mangrove species, and exploring the effect of interactions between features on model classification results.

Figure 2. The workflow of this study.

3.1. Image segmentation and creation of multidimensional feature datasets

In this study, we used eCognition Developer 9.4 software to segment GF-7, ZY-3, and Sentinel-2A multispectral images and GF-3 polarimetric SAR images with a segmentation scale of 20, shape and color factors of 0.3/0.7, and smoothness/compactness factors of 0.5/0.5. All segmentation parameters were determined by the 20 instances of trial and error of image segmentation. We then calculated the vegetation indices and textural characteristics of the segmented images of the GF-7, ZY-3, and Sentinel-2A sensors. Combination of multispectral bands, vegetation indices, textural features, polarimetric decomposition parameters, and backscattering coefficients (Table 4) were used to construct eight multidimensional feature datasets and twelve classification scenarios (Table 5).

Table 4. Multidimensional feature datasets and specific parameters.

Feature types	Descriptions	Number of features
Multispectral bands	Blue, Green, Red, NIR, Max_diff	5
Vegetation index	ARVI, BGRI, BNDVI, CMRI, DVI, ENDVI, EVI, EXG	17
	GNDVI, NDVI, NDWI, NGBDI, NGRDI, OSAVI, RGRI, RVI,
	VARI
Textural features	Mean, Variance (Std), Homogeneity, Contrast, Dissimilarity, Entropy, Second moment (Ang), Correlation	32
Coherence matrix	T11, T22, T33	3
Covariance matrix	C11, C22	2
Polarimetric decomposition parameters (PP)	TSVM, Agh, Huynen, Barnes1, Barnes2, UHDx, Freeman2, Freeman3, Yamaguchi3, Yamaguchi4, BF4, Neumann, Arii3_NNED, Singh_G4U2, MCSM, Singh_i6SD, HAAlpha, Cloude, Holm1, Holm2, An_Yang4, VanZyl3, etc.	139
Backscattering coefficients	${\sigma }_{\mathrm{HH}}^0$, ${\sigma }_{\mathrm{HV}}^0$, ${\sigma }_{\mathrm{VV}}^0$, ${\sigma }_{\mathrm{VH}}^0$	4

Table 5. Twelve classification scenarios with different data combinations.

Scenarios and Resolution	Feature combinations	Scenarios and Resolution	Feature combinations	Scenarios and Resolution	Feature combinations
Scenario 1 (10 m)	Sentinel-2A	Scenario 5 (5.8 m)	ZY-3	Scenario 9 (3 m)	GF-7
Scenario 2 (10 m)	Sentinel-2A + PP ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0$	Scenario 6 (5.8 m)	ZY-3 + PP ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0$	Scenario 10 (3 m)	GF-7 + PP ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0$
Scenario 3 (10 m)	Sentinel-2A + PP ${\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$	Scenario 7 (5.8 m)	ZY-3 + PP ${\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$	Scenario 11 (3 m)	GF-7 + PP ${\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$
Scenario 4 (10 m)	Sentinel-2A + PP ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0+{\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$	Scenario 8 (5.8 m)	ZY-3 + PP ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0+{\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$	Scenario 12 (3 m)	GF-7 + PP ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0+{\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$

3.2. Data reduction and feature selection

In order to obtain the optimal feature variables for each classification scenario, we chose the RFE method for performing data reduction and feature selection on the multidimensional feature datasets of twelve classification scenarios. Before feature selection, the input training dataset was subjected to high correlation deletion with the correlation coefficient set to 0.95. Three methods, including RFE-RF, RFE-TreeBag, and RFE-NB, were selected to train the dataset for 10 iterations after removing high correlations (Figure 3). Counting the precision of each model training and the number of output variables, we found that the RFE-RF algorithm produced the highest training accuracy. We then selected RFE-RF for subsequent feature optimization. The optimal feature variables for each classification scenario were summarized in Table 6.

Figure 3. Number of input variables and training accuracy of three RFE algorithms.

Table 6. Summary of the number of optimal features for twelve scenarios based on the RFE-RF algorithm.

Scenarios	All feature variables	Number of variables after removing high correlation	Number of variables after RFE-RF selection	Training accuracy (%) of RFE-RF model
Scenario 1	54	34	9	80.42
Scenario 2	81	46	19	84.54
Scenario 3	81	48	22	79.20
Scenario 4	200	102	10	84.55
Scenario 5	54	35	9	88.39
Scenario 6	81	45	22	88.16
Scenario 7	81	47	15	89.23
Scenario 8	200	103	11	88.17
Scenario 9	54	29	23	88.51
Scenario 10	81	40	10	90.60
Scenario 11	81	42	10	91.20
Scenario 12	200	92	49	90.34

3.3. Construction of adaptive ensemble learning models

Ensemble learning has been utilized for the wetland vegetation classification task, with good results (Berhane et al. 2018; Deng et al. 2022; Martínez et al. 2021). Currently, the mainstream ensemble learning methods are Bagging, Boosting and Stacking (Zhang, Liu, and Shen 2022). In particular, stacking can further improve the accuracy of ensemble learning models by combining multiple base models of different types to strengthen their classification performance, compensate for cross-model errors, and enable broader model integration (Jiang et al. 2023; Zhang et al. 2021); however, combining more base models in ensemble learning did not imply an improvement in classification ability (Zhang et al. 2023). Therefore, in order to investigate the effects of different base-model composition strategies on the classification accuracy of mangrove species, we constructed two adaptive ensemble learning models with different base-model composition strategies (Equations (1) and (2)). In this study, we chose seven machine learning algorithms as the base models, including RF, SVM, XGBoost, Adaptive Boosting (Adaboost), CatBoost, Light Gradient Boosting Machine (LightGBM), and Multilayer Perceptron (MLP). The grid search method was used to determine the optimal parameter values of each base model to realize the high-precision classification of mangrove species. (1) $\mathrm{h}0=\{\mathrm{BM}1,\mathrm{BM}2,\dots ,\mathrm{BMj}\}$(1) (2) $\mathrm{h}1=\{\begin{array}{c}\mathrm{rank}(\mathrm{BMC}\mathrm{orr}\mathrm{elation})\le n\\ \mathrm{Max}(\mathrm{OABM}(1,2,\dots ,n))\end{array}$(2) where h₀ denotes the initial set of input base models and j denotes the number of base models; h₁ denotes the final set of base models after conditional selection and n denotes the number of final selected base models.

3.3.1. Adaptive base-model multicollinearity elimination ensemble learning model

It has been demonstrated that a high correlation between data affects the efficiency and classification accuracy of ensemble learning models (Mahdavi̇fard et al. 2023). In order to improve the classification efficiency of stacking ensemble learning models and eliminate data redundancy, we designed an adaptive base-model multicollinearity elimination ensemble learning model (AME-EL) to eliminate high correlation between the input data (Figure 4). The training process was divided into two steps: (1) The segmented dataset was trained using five-fold cross-validation. Each base model prediction {A₁, A₂, A₃, A₄, A₅, A₆, A₇} was computed for relevance calculation, and i base model classification results with low relevance were retained (i = 3 or 5). (2) The base model with the highest classification accuracy was selected as the metamodel, and the retained base model predictions were composed into a new dataset {A₁, A₂, … , A_i}, which was input into the metamodel for training.

Figure 4. Structure of the AME-EL model.

3.3.2. Adaptive optimal base-model selection ensemble learning model

The classification performance of the base model can always influence the final meta model classification results (Mienye and Sun 2022). In order to improve the classification results of the metamodel, we designed an adaptive optimal base-model selection ensemble learning mode. The core idea was to compute the overall accuracy to get the optimal base model of each input data to construct an adaptive ensemble learning model (Figure 5). The training process was divided into two steps: (1) The partitioned training data were randomly divided into i equal parts and input into the base model for cross-validation training, and the optimal base model for each subset was calculated. (2) The base model with the highest classification accuracy was selected as the meta model, and the results predicted by the optimal model for each subset were composed into a new dataset {A₁, A₂, … , A_i}, which was input into the meta model for training.

Figure 5. Structure of the AOS-EL model.

3.4. Accuracy assessment

In this study, four indicators were used to quantitatively evaluate the accuracy of mangrove species classification models (Purwanto et al. 2021; Xiao et al. 2021), overall accuracy (OA), average accuracy (the average of the user's accuracy and the producer's accuracy, AA), and F1 score: The overall effectiveness of different ensemble learning models for mangrove species classification was evaluated via OA. The effects of different spatial resolution combinations of active and passive remote sensing images on the recognition accuracy of mangrove species were evaluated via AA and F1 score. The differences in the classification accuracy of different mangrove species with different combinations of active and passive remote sensing features were explored.

In order to explore the contributions of different multidimensional feature variables to the identification of mangrove species, we used the SHAP method to explain the importance of different feature variables to the classification of mangrove species. The SHAP method used an additive feature attribution method combined with a machine learning algorithm to calculate SHAP values to quantify the extent to which each feature contributes to a model's predictive power (Zhou et al. 2022). The function used was defined in Equation (3). In addition, in order to explore the potential impact of interactions between multidimensional feature variables on the model's predictive results. We further quantify the interactions between the multidimensional features through Equations (4) and (5) to estimate the joint contribution of the multidimensional features to the model (Meng et al. 2021). (3) ${\varphi }_{\mathrm{i}}=\sum {{}_{S\in }}_N{\displaystyle \frac{|S|!(N\text{-}|\mathrm{S}|\text{-}1)!}{N!}[f(S\cup \{\mathrm{i}\})\text{-}\mathrm{f}(\mathrm{S})]}$(3) In Equation (3), ϕ_i denotes the contribution of the i feature, N denotes the full feature dataset, S denotes the full subset of N, and f(S) denotes the model results with or without the i feature, respectively. (4) ${\varphi }_{\mathrm{i},\mathrm{j}}=\sum _{S\subseteq N\{i,j\}}{\displaystyle \frac{|S|!(M\text{-}|\mathrm{S}|\text{-}2)!}{2(M-1)!}}{\mathrm{\nabla }}_{\mathrm{ij}}(S)$(4) when i ≠ j (5) ${\mathrm{\nabla }}_{\mathrm{ij}}(\mathrm{S})=\mathrm{f}(\mathrm{S}\cup \{\mathrm{i},\mathrm{j}\})-\mathrm{f}(\mathrm{S}\cup \{\mathrm{i}\})\text{-}\mathrm{f}(\mathrm{S}\cup \{\mathrm{j}\})+\mathrm{f}(\mathrm{S})$(5) In Equations (4) and (5), ϕ_i denotes the contribution of the i feature, M denotes the number of input features, and S denotes the full subset of N.

4. Results

4.1. Ensemble learning-based mangrove specie classifications

4.1.1. Classification performance of AME-EL and AOS-EL models

In this study, we constructed twelve classification scenarios to evaluate the effects of two adaptive ensemble learning models on the classification accuracy of mangrove species. We counted the overall classification accuracy (OA) of mangroves based on the AME-EL and AOS-EL models in Figure 6. The highest OA of AME-EL was 0.948, the highest OA of AOS-EL was 0.941, and their OAs were over 0.7. We found that the AME-EL and AOS-EL models have similar classification performance in same scenarios, the difference in average OA in scenarios 1–4 and scenarios 5–8 and scenarios 9–12 were ranged from 0.5% to 0.6% and 1.5% to 3.0% and 0 to 0.2%, respectively, and their average OAs were over 0.8. These above results demonstrated that the AME-EL and AOS-EL models both have good classification performance in mangrove species.

Figure 6. Overall classification accuracy of mangroves based on the AME-EL and AOS-EL models. (a) Three-base models, (b) Five-base models.

We compared the mangrove classification performance of AME-EL and AOS-EL models and other traditional machine learning algorithms in Tabel 7 and Figure 7. The OA of the XGBoost and LGBoost algorithms ranged from 0.756 to 0.941 and 0.775 to 0.928, respectively; the OA of the Stacking algorithm ranged from 0.781 to 0.928. However, the OA of the AME-EL and AOS-EL models was higher than that of the XGBoost, LGBoost, and Stacking algorithms in general, and the OA were between 0.789 and 0.948. Compared to the machine learning algorithms and ensemble learning algorithms, the AME-EL model had a higher OA than other algorithms in eleven scenarios, and had a max increase in OA of 6.3%; the AOS-EL model had a higher OA than other algorithms in eight scenarios, and had a max increase in OA of 5.0%. The above result shows that, compared with other traditional machine learning algorithms, the AME-EL and AOS-EL models both have different degrees of improvement in mangrove classification performance, which demonstrated the advantages of the AME-EL and AOS-EL models in mangrove classification (Tables 7 and 8).

Figure 7. Comparison of the overall accuracy (OA) between our classification model and traditional machine learning algorithms.

Table 7. Summary of the OA between our model and traditional machine learning algorithms.

Scenarios	XGB	LGB	Stacking	AME	AOS	Scenarios	XGB	LGB	Stacking	AME	AOS
Scenario 1	0.756	0.794	0.794	0.794	0.806	Scenario 7	0.789	0.775	0.789	0.838	0.789
Scenario 2	0.788	0.781	0.781	0.794	0.794	Scenario 8	0.824	0.830	0.863	0.882	0.863
Scenario 3	0.808	0.821	0.808	0.834	0.834	Scenario 9	0.922	0.909	0.922	0.922	0.915
Scenario 4	0.794	0.781	0.819	0.825	0.831	Scenario 10	0.889	0.869	0.902	0.895	0.902
Scenario 5	0.830	0.837	0.850	0.850	0.830	Scenario 11	0.941	0.928	0.928	0.948	0.941
Scenario 6	0.856	0.869	0.856	0.869	0.850	Scenario 12	0.922	0.922	0.882	0.928	0.915

Table 8. Comparative analysis of the classification accuracy of mangrove species under two base-model composition strategies.

	F1 score	AC	CM	KC	MF	OT	SA	WB
Base model (7–3)	Min	0.675	0.625	0.333	0.919	0.500	0.872	0.897
	Max	0.912	0.944	0.857	1.000	0.923	0.989	1.000
	Mean	0.796	0.833	0.587	0.946	0.740	0.938	0.964
	Median	0.800	0.842	0.552	0.938	0.769	0.935	0.966
Base model (7–5)	Min	0.683	0.714	0.375	0.923	0.545	0.857	0.900
	Max	0.943	0.944	0.880	1.000	0.923	0.977	1.000
	Mean	0.800	0.837	0.588	0.952	0.732	0.935	0.962
	Median	0.806	0.857	0.552	0.944	0.727	0.938	0.947

4.1.2. Comparison of classification accuracy of two base-model composition strategies

In order to explore the classification performance of ensemble learning models with different base-model composition strategies for mangrove species classification, we counted the min, max, mean, and median F1 scores of all land cover types in the twelve scenarios in Tabel 8. The five-base-mode composition strategy obtained the max F1 score of Aegiceras corniculatum and Kandelia candel of 0.943 and 0.880, respectively. The three-base-mode composition strategy obtained a max F1 score of Sonneratia apetala of 0.943. The median F1 score of the five-base-model composition strategy for Aegiceras corniculatum, Cyperus malaccensis, and Sonneratia apetala was higher than that of the three-base-mode composition strategy; however, compared to the mean F1 score, it can be seen that the mangrove classification accuracies of the two base-model composition strategies were close to each other for the same mangrove species type, the difference in mean F1 score ranged from 0.1 to 0.4%.

In order to further explore the effects of different base-model composition strategies on mangrove species classification, we plotted the radar plots of classification accuracy for the twelve scenarios in Figure 8. The classification performance of both base-model composition strategies for mangrove species in scenarios 9–12 was relatively stable; the F1 scores of Aegiceras corniculatum, Cyperus malaccensis and Sonneratia apetala were over 0.8. However, in scenarios 1–8, there was a large difference in the classification accuracy of the ensemble learning models of the two base-model composition strategies for some mangrove species. By observing the changes in the shape of the radargrams (Figure 8), Kandelia candel had a lower classification accuracy in the four mangrove species. In the AME-EL model, the classification accuracies of Cyperus malaccensis using the five base-models were slightly better than those of the three base-models in scenario 2 and scenario 6; and the classification accuracies of Kandelia candel using the three base-models were better than five base-models in scenario 5 and scenario 7; In the AOS-EL model, the classification accuracies of Kandelia candel and Cyperus malaccensis using the three base-models were better than five base-models in scenario 2 and scenario12; and the classification accuracies of Aegiceras corniculatum using the five base-models were better than those of the three base-models in scenario 1 and scenario 3. However, they have almost equal ability to recognize Sonneratia apetala. The above results show that, when using the same ensemble learning model, the mangrove species classification accuracies of different base-model composition strategies have some differences, and both of these have good performance in mangrove species classification.

Figure 8. F1 scores of mangrove species for twelve classification scenarios based on two base-model composition strategies.

4.2. Mangrove specie classifications of multi-source feature combinations

To explore the mangrove specie classification accuracy of multi-source feature combinations, we counted the classification results of the AOS-EL model based on the three base-model composition strategy in Table 9. Compared to a scenario using a single multispectral image, the scenario of incorporating optical and SAR images can improve the classification accuracy of an ensemble learning model for mangrove species, and the overall accuracy (OA) improvement ranged from 0.6% to 5.6%. Scenarios 2–4 consisted of the Sentinel-2A multispectral image and GF-3 SAR image, and there OA ranged from 0.794 to 0.831; scenarios 6–8 consisted of ZY-3 multispectral image and GF-3 SAR image, and the OA was between 0.798 and 0.856; scenarios 10–12 consisted of GF-7 image and GF-3 SAR image, and the OA was over 0.9, which was higher than that for scenarios 2–4 and 6–8, and the OA improvement ranged from 5.9% to 10.3%. The above results show that the used of optical and SAR images with a higher spatial resolution can effectively improve mangrove classification performance.

Table 9. Classification accuracy of mangrove species based on different image combinations.

Scenario	Feature combinations	Metrics	AC	CM	KC	MF	OT	SA	WB
Scenario 1	Sentinel-2A	F1 score	0.684	0.688	0.429	0.971	0.759	0.911	0.950
		OA	0.775
Scenario 2	Sentinel-2A + PP ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0$	F1 score	0.723	0.800	0.421	0.919	0.545	0.951	0.923
		OA	0.794
Scenario 3	Sentinel-2A + PP ${\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$	F1 score	0.811	0.759	0.444	0.929	0.667	0.920	0.897
		OA	0.821
Scenario 4	Sentinel-2A + PP ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0+{\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$	F1 score	0.762	0.788	0.514	0.971	0.769	0.938	0.974
		OA	0.831
Scenario 5	ZY-3	F1 score	0.806	0.800	0.593	0.941	0.615	0.872	0.936
		OA	0.830
Scenario 6	ZY-3 + PP ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0$	F1 score	0.765	0.857	0.552	0.944	0.615	0.935	0.917
		OA	0.837
Scenario 7	ZY-3 + PP ${\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$	F1 score	0.693	0.788	0.375	1.000	0.923	0.886	1.000
		OA	0.789
Scenario 8	ZY-3 + PP ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0+{\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$	F1 score	0.794	0.889	0.552	0.971	0.714	0.935	0.936
		OA	0.856
Scenario 9	GF-7	F1 score	0.862	0.889	0.714	0.929	0.727	0.977	1.000
		OA	0.909
Scenario 10	GF-7 + PP ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0$	F1 score	0.841	0.889	0.692	0.929	0.800	0.966	1.000
		OA	0.902
Scenario 11	GF-7 + PP ${\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$	F1 score	0.912	0.882	0.857	0.963	0.800	0.966	1.000
		OA	0.919
Scenario 12	GF-7 + PP ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0+{\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$	F1 score	0.853	0.944	0.692	0.929	0.800	0.977	1.000
		OA	0.915

We compared the overall classification accuracy of multi-source feature combination scenarios in Figure 9(d). The integration of GF-7 optical image and GF-3 polarimetric SAR image (scenarios 10–12) obtained an OA ranged from 0.902 to 0.919; the integration of ZY-3 optical image and GF-3 polarimetric SAR images (scenarios 6–8) obtained an OA ranged from 0.789 to 0.837; the integration of Sentinel-2A optical image and GF-3 polarimetric SAR image (scenarios 2-4) obtained an OA ranged from 0.794 to 0.831. Combined GF-7 optical image and GF-3 polarimetric SAR images also achieved a high F1 score for all four mangrove species. Furthermore, Kandelia candel have the most significant classification accuracy enhancement, the F1 score increased by 0.482 (Figure 9(a)); the F1 score of Aegiceras corniculatum increased by 0.218 (Figure 9(b)); the F1 score of Cyperus malaccensis increased by 0.157 (Figure 9(c)); and the F1 score difference of Sonneratia apetala was relatively stable, ranging from 0.03 to 0.08. The above results show that different multi-source feature combination scenarios have difference effects on improving the classification accuracy of mangrove species, the enhancement of mangrove classification accuracy was ranked as follows: Kandelia candel > Aegiceras corniculatum > Cyperus malaccensis > Sonneratia apetala.

Figure 9. The difference in the F1 score and OA between three multispectral images. (a) The difference in the F1 score between scenarios 2, 6 and 10, (b) The difference in the F1 score between scenario 3,7 and 11, (c) The difference in the F1 score between scenario 4,8 and 12, and (d) Overall accuracy of three multispectral images.

We used the AOS-EL model with multi-source feature combination scenarios (4, 8, and 12) to map the spatial distribution of mangrove species. As can be seen from Figure 10, all three spatial scales of images accurately delineated the spatial distribution of mangrove species, and were able to distinguish the species composition of the study area well. Among the three multi-source feature combination scenarios, combined GF-7 optical image and GF-3 polarimetric SAR image (scenario 12) was the most effective in classifying and clearly distinguishing the boundaries between mangrove species (Figure 10(c)). Combined Sentinel-2A optical image and GF-3 polarimetric SAR image (scenario 8) has suffered from some Aegiceras corniculatum misclassification into Kandelia candel and Sonneratia apetala (Figure 10(a)), and combined ZY-3 optical image and GF-3 polarimetric SAR image has also suffered from a small amount of Aegiceras corniculatum misclassification into Cyperus malaccensis (Figure 10(b)). The above results show that the recognition effect of mangrove species varies with the improvement of the spatial resolution of remote sensing images, and high-spatial-resolution images were necessary for the classification of mangrove species.

Figure 10. Mangrove species mapping based on multi-source combination images. (a) Scenario 4 (Sentinel-2A + GF-3 polarimetric images), (b) Scenario 8 (ZY-3 + GF-3 polarimetric images), and (c) Scenario 12 (GF-7 + GF-3 polarimetric images).

4.3. Comparison of classification accuracy of different polarimetric SAR modes

In order to evaluate the effect of SAR images with different polarimetric modes on the classification accuracy of mangrove species, we summarized the classification results based on the AOS-EL model using a combination of GF-7 optical image and GF-3 quad-polarization SAR image. As can be seen from Table 10, dual-polarization and full-polarization SAR images can achieve the high-precision classification of mangrove species, with the average AA and F1 score above 0.8. The VV/VH polarization mode has the highest average classification accuracy, with an average AA and average F1 score of 0.917 and 0.912, respectively, which were 2.5%−3.6% and 2.7%−3.9% higher than those of the AHV and HH/HV polarizations. The VV/VH polarization was adept at identifying Aegiceras corniculatum and Kandelia candel with F1 score of 0.912 and 0.857; The AHV polarization was good at identifying Cyperus malaccensis and Sonneratia apetala with F1 score of 0.944 and 0.977. Furthermore, combined with Figure 11 and Table 8, it can be seen that the VV/VH polarization has the largest improvement in the classification accuracy of Aegiceras corniculatum and Kandelia candel, and the F1 score was higher than that of the HH/HV and AHV polarizations by 7.1%−16.1% and 5.9%−16.1%, respectively (Figure 11(b and c)). Additionally, the AHV polarization obtained the highest F1 scores in the identification of Cyperus malaccensis and Sonneratia apetala, which were 5.6%−6.5% and 1.1% higher than those of the HH/HV and VV/VH polarizations (Figure 11(a and b)).

Figure 11. Differences in the classification accuracy of mangrove species based on HH/HV, VV/VH, and AHV polarization. (a) The difference in average accuracy between scenario10 and 12, (b) The difference in average accuracy between scenario11 and 12, and (c) The difference in average accuracy between scenario10 and 11.

Table 10. Classification accuracy of combining GF-7 multispectral with dual – and quad-polarization SAR images.

Scenario	Metrics	AC	CM	KC	MF	OT	SA	WB	Mean Accuracies
Scenario 10	AA	0.842	0.892	0.696	0.933	0.833	0.967	1.000	0.881
	F1 score	0.841	0.889	0.692	0.929	0.800	0.966	1.000	0.873
Scenario 11	AA	0.913	0.882	0.857	0.964	0.833	0.966	1.000	0.917
	F1 score	0.912	0.882	0.857	0.963	0.800	0.966	1.000	0.912
Scenario 12	AA	0.854	0.947	0.696	0.933	0.833	0.978	1.000	0.892
	F1 score	0.853	0.944	0.692	0.929	0.800	0.977	1.000	0.885

In order to further investigate the differences in the classification accuracy of mangrove species based on three polarimetric modes of GF-3 polarimetric SAR images, we displayed the classification results by combining GF-7 multispectral and three polarimetric SAR images (Figure 12). We found that the classification results based on three polarimetric SAR images can clearly distinguish the boundaries of four mangrove species. In particular, three polarization modes were able to accurately delineate the Sonneratia apetala distribution. The AHV polarization displayed good performance for Sonneratia apetala mapping, but there were still some misclassifications of fine patches. HH/VH and AHV polarization still produced Aegiceras corniculatum and Kandelia candel misclassification. While the VV/VH polarization was more accurate in mapping Aegiceras corniculatum and Kandelia candel, it also appears to misclassify some Aegiceras corniculatum patches as Kandelia candel. Meanwhile, the AHV polarization was more accurate in mapping the distribution of Cyperus malaccensis, and can better separate Cyperus malaccensis from the surrounding land cover types. The HH/VH polarization performed less well in Cyperus malaccensis recognition, and misclassified some Sonneratia apetala patches as Cyperus malaccensis. These results indicated that all three polarimetric modes of GF-3 polarimetric SAR images can accurately classify mangrove species, but there was still some difference in the recognition ability of different mangrove species.

Figure 12. Mangrove species mapping by combining GF-7 multispectral and three polarimetric SAR images based on the AOS-EL model.

4.4. Interaction effect of optical and SAR image on mangrove specie classifications

In order to assess the influence of active and passive image features on the classification accuracy of mangrove species, we used the SHAP method to evaluate the importance of the optimal features from the combination scenarios of optical and GF-3 polarimetric SAR images, and to quantify the influence of the interactions between optical and SAR image features on the classification results based on ensemble learning model, as shown in Figure 13(a–c). In scenario 2, 6, and 10, the top features contributing to Aegiceras corniculatum classification accuracy were NIR, BGRI, and EXG. For Cyperus malaccensis classification accuracy, EVI, VARI, and NGRDI contributed more than other features did. The top features contributing to Kandelia candel classification accuracy were DVI, NIR, and BGRI. Furthermore, RVI, Entropy_shannon_I, and L1 significantly contributed to Sonneratia apetala classification accuracy. Additionally, we mapped interaction diagrams (Figure 13(d)) based on the top nine important features in scenario 10 (GF-7 + GF-3 HH/HV SAR images), which have the highest classification accuracy for mangrove species. As shown in Figure 13(d), we found that most of the features in the blue box displayed good interaction effects, such as EXG and NIR, which have a good negative interaction (SHAP values increase negatively with increasing feature values), which indicated that EXG and NIR features can present a positive effect on the prediction results of an ensemble learning model. Furthermore, the EXG and NIR features have significant contribution to the Aegiceras corniculatum classification accuracy. This means that, compared with other features, the interaction of EXG and NIR bands effectively promotes the ability of the ensemble learning model to recognize Aegiceras corniculatum. The orange box showed the features with almost no interaction effect, such as the ARVI and NGRDI features, whose interaction value was almost close to 0, and have not effectively influence the classification accuracy of mangrove species.

Figure 13. Characteristic SHAP maps based on the combined scenarios of multispectral images with GF-7 and GF-3 HH/HV polarimetric images. (a) Scenario 2 (Sentinel-2A + PP + ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0$), (b) Scenario 10 (GF-7 + PP + ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0$), (c) Scenario 6 (ZY-3 + PP + ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0$), and (d) Scenario 10 features interaction. Horizontal coordinates were SHAP values, dots indicate feature sample values, red color indicate larger feature values, and blue color presents the opposite.

As illustrated in Figure 14(a–c), in scenario 3, 7 and 11, the top features contributing to Aegiceras corniculatum classification accuracy were DVI, NIR, and BGRI. For Cyperus malaccensis classification accuracy, RVI, EVI, and EXG have more contribution than other features. The top features contributing to Kandelia candel classification accuracy were BNDVI, NIR, and EVI. And the RVI, BGRI, and VARI significantly contributed to Sonneratia apetala. In addition, as shown in Figure 14(d), the L2 feature showed a good positive interaction with NIR, BNDVI, and RVI (i.e. as the feature value increased, the SHAP value also increased). Although the L2 feature has low contributions for four mangrove species (Figure 14(b)), it has a good interaction with the NIR, BNDVI, and RVI features, which made the above three features have a good contribution to the identification of Aegiceras corniculatum, Cyperus malaccensis, and Kandelia candel. In contrast, the Blue, Green and VH features were unsatisfied in their interactions with other features, and themselves also showed lower contributions to the classification accuracy of mangrove species.

Figure 14. Characteristic SHAP maps based on the combined scenarios of multispectral images with GF-7 and GF-3 VV/VH polarimetric images. (a) Scenario 3 (Sentinel-2A + PP +${\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$), (b) Scenario 11 (GF-7 + PP + ${\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$), (c) Scenario 7 (ZY-3 + PP + ${\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$), and (d) Scenario 11 features interaction. Horizontal coordinates were SHAP values, dots indicate feature sample values, red color indicate larger feature values, and blue color presents the opposite.

As shown in Figure 15(a–c), in scenarios 4, 8 and 12, the top features contributing to Aegiceras corniculatum classification accuracy were An_Yang3_Dbl, NGBDI, and TSVM_alpha_s1. For Cyperus malaccensis classification accuracy, EVI, HV, RGRI have more contribution than other features. The top features contributing to Kandelia candel classification accuracy were An_Yang3_Dbl, GLCM_Con_1, and GLCM_Mean_2. Additionally, NGRDI, TSVM_alpha_s1, GLCM_Cor_3 significantly contributed to Sonneratia apetala classification accuracy. Furthermore, as shown in Figure 15(d), the An_Yang3_Dbl feature has a better positive interaction with the VARI feature among all feature interactions. Compared with other features, the An_Yang3_Dbl feature obtained the highest contribution to Aegiceras corniculatum and Kandelia candel, and the VARI feature obtained a high contribution to the Cyperus malaccensis classification accuracy. In addition, the CLGM_Cor_3 feature also achieved good interactions with some of other features, and the CLGM_Cor_3 feature obtained the highest contribution to Sonneratia apetala classification accuracy. Some of the other features such as Holm2_T33, GLCM_Mean, MCSM_Hlx have low interactions with each other, as well as unsatisfactory contributions to the classification accuracy of the four mangrove species. The above results show that the An_Yang3_Dbl, and VARI features have a good interaction with other features, which can effectively affect their contribution to the classification accuracy of mangrove species.

Figure 15. Characteristic SHAP maps based on the combined scenarios of multispectral images with GF-7 and GF-3 AHV polarized images. (a) Scenario 4 (Sentinel-2A + PP + ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0+{\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$), (b) Scenario 12 (GF-7 + PP + ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0+{\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$), (c) Scenario 8 (ZY-3 + PP + ${\sigma }_{\mathrm{HH}}^0+{\sigma }_{\mathrm{HV}}^0+{\sigma }_{\mathrm{VV}}^0+{\sigma }_{\mathrm{VH}}^0$), and (e) Scenario 12 features interaction. Horizontal coordinates were SHAP values, dots indicate feature sample values, red color indicate larger feature values, and blue color presents the opposite.

5. Discussion

This study constructed novel AME-EL and AOS-EL models, explored the effect of different base-model composition strategies on their identification of mangrove species, and quantitatively evaluated their differences in classification accuracy. Two base-model composition strategies achieved fine mangrove species classification (Figure 7); however, we found that different base-model compositions had different abilities to distinguish mangrove species. In the AME-EL model, the classification accuracies of Kandelia candel based on the three base models were higher than those of the five base models. In the AOS-EL model, the classification accuracies of Cyperus malaccensis based on the three base models were higher than those of the five base models. These results indicated that the selection of optimal base models is crucial to constructing an ensemble learning framework, and fewer base models can reduce the complexity of ensemble learning, improve training efficiency, and acquire the satisfactory classification results. This conclusion was consistent with that of Uhl, Græsdal Rasmussen, and Oppelt (2022). Uhl, Græsdal Rasmussen, and Oppelt (2022) delineated the spatial distribution of coastal vegetation by constructing an ensemble learning model with six machine learning algorithms, and used the five base-models composition strategy to achieve the highest classification accuracy (overall accuracy reaches 0.98). Furthermore, previous studies demonstrated that the classification performance of an ensemble learning model was also limited by the inherent diversity of its base models (You et al. 2022). Therefore, future research can attempt to stack multiple different machine learning algorithms into the ensemble learning framework, compare it with the current mainstream deep learning models, and verify their classification robustness in different mangrove species and other wetland scenarios.

Due to the spectral similarity of vegetation species, weather conditions and tides influence, it is using difficult to conduct the fine classification of mangrove species using only a single multispectral image. Previous studies have confirmed that the combination of optical and SAR images can effectively resolve the aforementioned challenges (Upakankaew et al. 2022). This study combined different-spatial-resolution multispectral images (Sentinel-2A, ZY-3, GF-7) and full-polarization GF-3 SAR images to construct 12 classification scenarios. Our results found that the integration of GF-3 polarimetric SAR and three multispectral images obtained outstanding classification results, especially in combining GF-7 and GF-3 SAR images (scenarios 9–12), which obtained the highest F1 score of mangrove species (85.70%−97.70%). This conclusion was consistent with that of Aja, Miyittah, and Angnuureng (2022). In addition, our study further found that there are differences in the range of improving classification accuracy of mangrove species based on different image combinations. Kandelia candel obtained the highest accuracy improvement with the F1 score increasing by 16.1%, and the F1 score of Aegiceras corniculatum, Cyperus malaccensis and Sonneratia apetala improved by 7%, 6.5%, and 1.1%, respectively. These results indicated that the integration of high-spatial-resolution multispectral and polarimetric SAR images is an effective approach to perform mangrove species classification. Future studies can attempt to combine ultra-high resolution UAV images with LiDAR point clouds and their derivatives to classify mangrove species and other wetland vegetation species.

We found that different polarization modes had a significant impact on mangrove species mapping. The VV/VH polarization had the highest classification accuracy for mangrove species (Table 10), compared to the AHV and HH/HV polarization methods, the F1 score of Aegiceras corniculatum and Kandelia candel were higher by 5.9%−16.5% and 7.1%−16.5%, respectively. Compared to the HH/HV and VV/VH polarization methods, the F1 scores of Cyperus malaccensis and Sonneratia apetala of AHV polarization was higher by 5.6%−6.5% and 1.1%, respectively. The HH/HV polarization also had excellent classification accuracy of mangrove species, the F1 score ranged from 0.692 to 0.966 (Figure 11). These results indicated that the polarimetric decomposition parameters of HH/HV, VV/VH, and AHV polarimetric SAR images are the significant features for mangrove species classification. This conclusion was consistent with that of Wang, Tan, and Fan (2023), Wang, Tan, and Fan (2023) combined optical and GF-3 polarimetric SAR images to map mangrove species distribution, and obtained an excellent overall accuracy of 90.13%. Our study demonstrated that the VV/VH and AHV polarization methods are more effective in mangrove species classification.

In this study, we used the SHAP method to investigate the effect of optical and SAR features interaction on mangrove classification performance. We found that the SAR features, optical band features, and vegetation index features had good interaction, which can effectively influence the mangrove classification results (Figures 13–15). The An_Yang3_Dbl feature had good interaction with the VARI feature, and had a high contribution to the classification results of Aegiceras corniculatum, Kandelia candel, and Cyperus malaccensis. The above results indicated that the SHAP method can be used to evaluate the effect of the interaction between features on the prediction results, and that it is a significant method in mangrove species classification; however, some scholars neglected the effect of interactions between features on model prediction results in previous studies (Bi et al. 2020; Feng et al. 2021). Therefore, future studies can attempt to use the features which have good interaction with other features, and input these into a model to validate their effects on model classification performance.

We delineated the spatial distribution of the dominant mangrove species in the study area (Figure 16). We found that Aegiceras corniculatum and Sonneratia apetala are the dominated mangrove species of the study area. Aegiceras corniculatum is widely distributed in the coastal margin and low-tide mudflat with an area of 6.56 km², and has strong salt resistance to adapt to the impact of seawater (Su, Wang, and Sun 2022). Aegiceras corniculatum has greater ability to reproduce and spread seeds (Deng et al. 2009), and is also a dominant species in the mangrove succession of the Maowei Sea Mangrove Nature Reserve; Sonneratia apetala has an area of 5.17 km², is mainly distributed in the northwest of the study area, and is highly adaptable to seawater flooding inundation. However, when the salinity of seawater is too high, its seed germination and growth will be inhibited (Song et al. 2023). Therefore, it is less distributed in the coastal margin. Cyperus malaccensis with the area of 3.32 km², is mainly distributed in the intertidal zone and mostly grows interspersed with mangrove forests; Kandelia candel has an area of 1.20 km², and undergoes typical viviparous reproduction that can adapt well to a mudflat environment (Geng et al. 2021). However, the Kandelia candel is possibly still in an early stage of growth, being less distributed in the study area, and most of the Kandelia candel is sporadically distributed in the Aegiceras corniculatum.

Figure 16. Displaying the spatial distribution of the dominant mangrove species in the study area.

6. Conclusions

We developed two novel adaptive ensemble learning models to classify mangrove species using multi-source optical and polarimetric SAR images. This study explored the effect of two base-model composition strategies on classifying mangrove species and evaluated their classification performance under twelve feature combinations. The study quantified the impact of interactions between optical and SAR images on the identification ability of mangrove species. We concluded that the two ensemble learning models (AME-EL and AOS-EL) proposed in this study under different base-model composition strategies achieved over 90% overall classification accuracy. This study demonstrated that the base-model composition strategy had a significant effect on the classification ability of an ensemble learning model, and further found that the three-base-model composition strategy could achieve equivalent classification results with the five-base-model composition strategy. Among mangrove species of Aegiceras corniculatum, Cyperus malaccensis and Sonneratia apetala, the integration of optical and SAR images significantly boosted the classification accuracy for Kandelia candel with a 48.2% improvement in the F1 score. The VV/VH polarization mode can effectively identify Aegiceras corniculatum and Kandelia candel. The full polarization mode can better identify Cyperus malaccensis and Sonneratia apetala species. The interactions between polarimetric decomposition parameters (An_Yang3_Dbl), the NIR band and other features contributed highly to mangrove species classification, which had an important impact on improving model prediction accuracy. This study contributed to scientific understanding and management practices in the conservation of mangrove communities and ecosystems.

Acknowledgements

We appreciate the anonymous reviewers for their comments and suggestions, which helped to improve the quality of this manuscript. And we are grateful to Prof.Wang of the University of Rhode Island to checked and modified the language expression and scientific contents of this paper.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The data that support the findings of this study are available from the corresponding author, upon reasonable request.

Additional information

Funding

This study was supported by the National Natural Science Foundation of China (grant number 42371341, 42004006), the Natural Science Foundation of Guangxi Zhuang Autonomous Region (Grant number 2024GXNSFAA010351), the Innovation Project of Guangxi Graduate Education (Grant Number YCSW2023353), and the ‘BaGui Scholars’ program of the provincial government of Guangxi (grant Number 2019A30), and in part by Zhejiang Province ‘Pioneering Soldier’ and ‘Leading Goose’ R&D Project (grant Number 2023C01027), the Guilin University of Technology Foundation (grant number GUTQDJJ2017096).

Appendix A1.

Vegetation indices calculation equation

Table

Indices name	Formula
ARVI	$(\mathrm{NIR}-2\times \mathrm{Red}+\mathrm{Blue})/(\mathrm{NIR}+2\times \mathrm{Red}-\mathrm{Blue})$
BGRI	$\mathrm{Blue}/\mathrm{Green}$
BNDVI	$(\mathrm{NIR}-\mathrm{Blue})/(\mathrm{NIR}+\mathrm{Blue})$
CMRI	$(\mathrm{NIR}-\mathrm{Red})/(\mathrm{NIR}+\mathrm{Red})-(G\mathrm{reen}-\mathrm{NIR})/(\mathrm{Green}+\mathrm{NIR})$
DVI	$\mathrm{NIR}-\mathrm{Red}$
ENDVI	$((\mathrm{NIR}-\mathrm{Green})-(2\times \mathrm{Blue})/((\mathrm{NIR}+\mathrm{Green})+2\times \mathrm{Blue}))$
EVI	$2.5\times (\mathrm{NIR}-\mathrm{Red})/[\mathrm{NIR}+(6\times \mathrm{Red}-7.5\times \mathrm{Blue})+1]$
EXG	$2.5\times G\mathrm{reen}-\mathrm{Re}d-\mathrm{Blue}$
GNDVI	$(\mathrm{NIR}-\mathrm{Green})/(\mathrm{NIR}+\mathrm{Green})$
NDVI	$(\mathrm{NIR}-\mathrm{Red})/(\mathrm{NIR}+\mathrm{Red})$
NDWI	$(\mathrm{Green}-\mathrm{NIR})/(\mathrm{Green}+\mathrm{NIR})$
NGBDI	$(\mathrm{Green}-\mathrm{Blue})/(\mathrm{Green}+\mathrm{Blue})$
NGRDI	$(\mathrm{Green}-\mathrm{Red})/(\mathrm{Green}+\mathrm{Red})$
OSAVI	$(\mathrm{NIR}-\mathrm{Red})/(\mathrm{NIR}+\mathrm{Red}+0.16)$
RGRI	$\mathrm{Red}/\mathrm{Green}$
RVI	$\mathrm{NIR}/\mathrm{Red}$
VARI	$(\mathrm{Green}-\mathrm{Red})/(\mathrm{Green}+\mathrm{Red}-\mathrm{Blue})$

Appendix A2. HH-HV, VV-VH and GF-3 SAR image of the study area in the low tide.

Appendix A3. The tide heights corresponding to three optical images in study area (https://www.chaoxibiao.net/tides/151.html)