Comparison of the Applicability of J-M Distance Feature Selection Methods for Coastal Wetland Classification

Abstract

Accurate determination of the spatial distribution of coastal wetlands is crucial for the management and conservation of ecosystems. Feature selection methods based on the Jeffries-Matusita (J-M) method include J-M distance with simple average ranking (JM_ave), J-M distance based on weights and correlations (JM_improved), and heuristic J-M distance (JM_mc). However, as the impacts of these methods on wetland classification are different, their applicability has rarely been investigated. Based on the Google Earth Engine (GEE) and random forest (RF) classifier, this is a comparative analysis of the applicability of the JM_ave, JM_improved, and JM_mc methods. The results show that the three methods compress feature dimensions and retain all feature types as much as possible. JM_mc exhibits the most significant compression from a value of 35 to 15 (57.14%), which is 37.14% and 40% more compressed than JM_ave and JM_improved, respectively. Moreover, they produce comparable classification results, with an overall classification accuracy of 90.20 ± 0.19% and a Kappa coefficient of 88.80 ± 0.22%. However, different methods had their own advantages for the classification of different land classes. Specifically, JM_ave has a better classification only in cropland, while JM_mc is advantageous for recognizing water bodies, tidal flats, and aquaculture. While JM_improved failed to retain vegetation and mangrove features, it enables a better depiction of the mangroves, salt pans, and vegetation classes. Both JM_ave and JM_improved rearrange features based on J-M distance, while JM_mc places more emphasis on feature selection. As a result, there can be significant differences in feature subsets among these three methods. Therefore, the comparative analysis of these three methods further elucidates the importance of J-M distance in feature selection, demonstrating the significant potential of J-M distance-based feature selection methods in wetland classification.

1. Introduction

Coastal wetlands represent zones of interaction between terrestrial and aquatic ecosystems; therefore, the associated wetland ecosystems are dynamic and fragile [1]. They are important for water conservation, regional climate regulation, flood control, and biodiversity protection, etc. [2,3,4]. Owing to factors such as anthropogenic disturbance, pollution, climate change, and invasive species, coastal wetlands are increasingly deteriorating [4,5]. Compared with other ecosystems, wetlands are considered one of the three major ecosystems of the earth, with a wide range of ecological functions and social values. Consequently, accurate information and high-resolution classification of coastal wetlands are important for their monitoring and protection.

Remote sensing (RS) is the most commonly used technique for wetland monitoring and protection. High-resolution multispectral data obtained using RS technology have facilitated wetland characterization, because they can provide more classification information, especially with the emergence of the Google Earth Engine (GEE) platform and machine learning algorithms. Combining GEE and machine learning has enabled rapid and efficient processing of massive amounts of remote sensing data [6,7,8]. However, such data typically contain strong correlations between feature variables, and classification involves irrelevant information [9].

Owing to the abundance of data, selecting an optimal feature subset is crucial to produce efficient and accurate classifications of wetlands. Moreover, feature selection is one of the most important factors affecting the scalability and performance of machine learning classification algorithms [10]. The Jeffries-Matusita (J-M) distance is an important measure for selecting high-quality feature subsets and has a lower computational cost and higher operational efficiency [10,11,12]. The J-M feature subset selections methods involve the following three common types: the J-M distance with an average ranking (JM_ave), J-M distance based on weights and correlations (JM_improved), and heuristic-based J-M distance (JM_mc). In the JM_ave method, spectral, textural, and topographic features are selected and optimized using the separability and thresholds (SEaTH) algorithm. The top two features with J-M values > 1 are then selected to produce optimal results. This method effectively reduce the feature number from 28 to 11 and achieved the 93% overall accuracy [13]. Both the SEaTH algorithm and J-M distance methods consider the separability of feature variables. However, correlations between feature variables are neglected. Therefore, a J-M distance feature subset selection method which preserves the separability of classes while avoiding data redundancy was proposed. This improved method selected the first 4–7 feature variables based on the improved J-M distance’s ability to achieve great classification accuracy, while the traditional J-M distance ranking requires 20 feature variables [14]. Similar to JM_ave, JM_mc also does not consider information redundancy. JM_mc is capable of finding the optimal feature subset from a J-M distance-based ranking feature list for multiclass problems without explicitly using a search technique. In the 37 multiclass datasets, each dataset had different degrees of dimensional compression and produced stable average classification accuracy more than other methods [15].

Coastal wetlands, as important ecosystems, are characterized by diverse land classes and fuzzy boundaries, which pose many classification challenges. The influence and applicability of different feature subset selection methods involving the J-M distance in classification have differences, a comprehensive comparison of these methods was conducted to highlight the advantages of each in wetland classification. The main research objectives are to: (1) evaluate the impact of different J-M distance feature selection methods on feature dimensions, feature types, classification accuracy, and classification results and (2) assess the applicability of different J-M methods and find a combination of feature types and feature selection methods for different coastal wetland classes.

2. Materials and Methods

2.1. Study Area

Fujian Province lies on the southeastern coast of China, with a topography characterized by mountains and hills. The province covers approximately 124,000 and 136,000 km² of land and sea areas, respectively, and its 3752 km coastline is the second longest in China. In this study, the Fujian Province coastline was considered the boundary, and the threefold division method [16] was used to extract the boundary along the coastal zone. The 6 m isobath contour was extracted from the ETOPO1 topographic elevation data, which was then used to extract boundary on the seaward side, while the boundary on the land side was extracted using county-level administrative division data (Figure 1).

2.2. Data

2.2.1. Data and Processing

Sentinel, topographic, and visual interpretation of sample data were utilized. Sentinel and DEM data were acquired and processed using the GEE platform. SAR (Sentinel-1) and optical (Sentinel-2) imagery data with a 10 m spatial resolution acquired in 2019 were used. The Sentinel-1 data were pre-processed using the GEE platform to remove thermal noise, calibrate radiometric, and correct terrain. The Sentinel-1 GRD comprised data collected in IW and VV + VH as the polarization mode. Preprocessing of the level-1C Sentinel-2 data involved cloud masking, image fusion, and mosaic cropping, followed by the calculation of relevant feature indexes.

The 2019 ASTER GDEM v3 data were resampled using the nearest-neighbor method to obtain a spatial resolution of 10 m. Topographic indexes (elevation, slope, aspect, and hillshade) were extracted as feature variables and used for the classification. The 6 m isobath contour was extracted from the ETOPO1 topographic elevation data, which was used to extract boundary of the study area.

Based on field surveys and previous studies [17], the coastal zone in Fujian Province yielded the following land classes: water bodies, tidal flats, mangroves, aquaculture areas, croplands, salt pans, and vegetation. The classification accuracy was high if the pixels of the training samples were 24–30 times the number of bands [18]. Therefore, based on the relationship between the number of training samples and the classification accuracy, 6632 sample points were obtained. These data associated with 2019 Google images were selected using the principle of random uniformity. The data included 4666 training and 1966 validation samples, and these are presented according to class in

Table 1. Distribution of sample points for wetlands in Fujian Province according to class.

Class Code	Wetland Classes	Number of Samples
		Training	Validation	Total
C1	Water body	565	260	825
C2	Tidal flat	587	238	825
C3	Mangrove	599	260	859
C4	Aquaculture	575	256	831
C5	Cropland	588	225	813
C6	Salt pan	573	241	814
C7	Vegetation	576	225	801
C8	Others	603	261	864
Total		4666	1966	6632

.

2.2.2. Feature Selection

Construction of feature variables is crucial for classification based on remote-sensing data because a combination of optimal features can improve the accuracy. In the present study, 35 feature variables were selected (

Table 2. Features used for classification.

Feature Variables	Feature Name	Equation
spectral feature	B2, B3, B4, B5, B6, B7, B8, B8A, B11, B12	-
radar feature	VV, VH, POL	POL = (VV − VH)/(VV + VH)
topographic feature	Elevation, Slope, Aspect, Hillshade	-
texture feature	B3_asm, B3_idm, B3_var, B3_ent, B3_corr, B3_contrast	-
vegetation feature	NDVI	NDVI = (B8 − B4)/(B8 + B4)
water feature	NDWI, MNDWI, AWEI, LSWI	NDWI = (B3 − B8)/(B3 + B8) MNDWI = (B3 − B11)/(B3 + B11) AWEI = 4 × (B3−B11) − (0.25 × B8 + 2.75 × B12) LSWI = (B8 − B11)/(B8 + B11)
mangrove feature	MVI, MFI	MVI = (B8 − B3)/(B11−B3) MFI = [(B5 − ρBλ1) + (B6 − ρBλ2) +(B7 − ρBλ3) + (B8A − ρBλ4)]/4 ρBλi = B12 + (B4 − B12) × (2190 − λi)/(2190 − 665) (λ1 = 705 λ2 = 740 λ3 = 783 λ4 = 865)
red edge feature	NDVIeg1, NDVIeg2 NDVIeg3, NDre1, NDre2	NDVIeg1 = (B8A − B5)/(B8A + B5) NDVIeg2 = (B8A − B6)/(B8A + B6) NDVIeg3 = (B8A − B7)/(B8A + B7) NDre1 = (B6 − B5)/(B6 + B5) NDre2 = (B7 − B5)/(B7 + B5)

,

Table 3. Waveband parameters of Sentinel-2 data.

Band	Central Wavelength (μm)	Resolution(m)
Band 2—Blue	0.490	10
Band 3—Green	0.560	10
Band 4—Red	0.665	10
Band 5—Red-edge 1	0.705	20
Band 6—Red-edge 2	0.740	20
Band 7—Red-edge 3	0.783	20
Band 8—NIR	0.842	10
Band 8a—NIR narrow	0.865	20
Band 11—SWIR1	1.610	20
Band 12—SWIR2	2.190	20

and

Table 4. Description of features and indices.

Feature Type	Feature Name	Description
radar feature	VV	The C-band vertical–vertical polarization backscatter data
	VH	The C-band vertical–horizontal polarization backscatter data
	POL	POL was calculated based on the VV and VH images
topographic feature	Elevation	Elevation extracted by ASTER GDEM
	Slope	Slope extracted by ASTER GDEM
	Aspect	Aspect extracted by ASTER GDEM
	Hillshade	Hillshade extracted by ASTER GDEM
texture feature	B3_asm	Angular Second Moment Matrix is used to measure the degree of uniformity or regularity of the image texture
	B3_idm	Inverse Difference Moment Matrix is used to measure the degree of gray level difference of neighboring pixels in an image
	B3_var	Variance is used to measure the degree of variation in the gray level of pixels in an image or texture
	B3_ent	Entropy is used to measure the complexity and randomness of the texture
	B3_corr	Correlation measures the linear correlation or dependency between pixel gray levels in a texture
	B3_contrast	Contrast is used to measure the degree of contrast between light and dark variations in texture or gray levels
vegetation feature	NDVI	Index of vegetation condition and growth vigor
water feature	NDWI	Index of water distribution and humidity
	MNDWI	Improved normalized difference water body index
	AWEI	Index that widens the difference between water body and non-water body information
	LSWI	Soil moisture change monitoring index
mangrove feature	MVI	Capturing the unique green color and water content of mangrove pixel index
	MFI	Mapping mangrove distribution based on single date images
Red-edge feature	NDVIeg1	Normalized Difference Vegetation Index red-edge 1 narrow
	NDVIeg2	Normalized Difference Vegetation Index red-edge 2 narrow
	NDVIeg3	Normalized Difference Vegetation Index red-edge 3 narrow
	NDre1	Normalized Difference red-edge 1
	NDre2	Normalized Difference red-edge 2

), including spectral, radar, topographic, texture and related vegetation, water [19,20,21], mangrove [22,23], and red edge features [24]. Specifically, the Normalized Difference Vegetation Index (NDVI) is a commonly used index to assess vegetation health and density. NDVIeg1, NDVIeg2, and NDVIeg3 are similar to NDVI, which are calculated in the NIR narrow band. The Normalized Difference Water Index (NDWI) is an index used to detect the presence of water bodies. The modified NDWI (MNDWI) aims to enhance water detection. The Automated Water Extraction Index (AWEI) is designed specifically for water extraction in coastal and inland water bodies. The Land Surface Water Index (LSWI) is used for mapping and monitoring surface water. The Mangrove Vegetation Index (MVI) can assess vegetation vigor and health, while the Mangrove Forest Index (MFI) is designed for mapping and monitoring forests. Differences in spectral features of seawater involving algae and shellfish (seawater associated with aquaculture) and regular seawater were primarily obvious in blue and green bands [25]. Therefore, to adequately distinguish water linked to aquaculture from regular seawater, the green band (B3), which produced large differences in spectral features, was selected for the extraction of textural features.

2.3. J-M Distance-Based Feature Selection Methods

At present, J-M distance-based feature selection characteristics can be categorized as follows: (1) J-M distance based on a simple average ranking [12,13], (2) J-M distance based on a weights and correlations [14,26,27], and (3) J-M distance based on heuristic methods [15]. Based on the usage and popularity of selection rules, in the current study on classification problems, the average-ranked J-M (JM_ave), improved J-M distance (JM_improved), and heuristic J-M distance (JM_mc) methods were utilized.

2.3.1. Average-Ranked J-M Distance

The J-M distance is used to measure the separability between two classes, and thus, it is a better method for distinguishing land classes. The J-M distance can be expressed using Equations (1) and (2).

$${\mathrm{J}=2(1-\mathrm{e}}^{-\mathrm{B}})$$

(1)

$$\mathrm{B}=\frac{1}{8}{{(\mathrm{m}}_1{-\mathrm{m}}_2)}^2\frac{2}{{\mathsf{\sigma }}_1^2{+\mathsf{\sigma }}_2^2}+\frac{1}{2}\ln [\frac{{\mathsf{\sigma }}_1^2{+\mathsf{\sigma }}_2^2}{{2\mathsf{\sigma }}_1{\mathsf{\sigma }}_2}]$$

(2)

where B is the Bhattacharya distance; m_i and ${\mathsf{\sigma }}_{\mathrm{i}}^2$ (i = 1 and 2,) represent the mean and variance of classes C1 and C2, respectively. The J-M distance involves values between 0 and 2, and the closer the value is to 2, the better the separation between samples. The average-ranked J-M distance (JM_ave) is a simple and effective method for feature selection. In this method, portions of each feature with a J-M distance greater than 1 in all class pairs are selected for the averaging process [13], and feature variables are then ranked in descending order.

2.3.2. Improved J-M Distance

Considering the redundancy of information caused by correlations between feature variables, an improved J-M distance based on separability and redundancy (JM_improved) was introduced. Weights of different land classes [28] were then used to calculate the weighted average J-M distance between classes of each feature variable, whereas the Pearson correlation coefficient represented a redundancy measure. The Pearson’s correlation coefficient was obtained using Equation (3).

$${\mathsf{\rho }}_{\mathrm{XY}}=\frac{\mathrm{Cov}\left(\mathrm{X},\mathrm{Y}\right)}{{\mathsf{\sigma }}_{\mathrm{X}}{\mathsf{\sigma }}_{\mathrm{Y}}}$$

(3)

Here, Cov represents the covariance between X and Y, whereas ${\sigma }_X$and ${\sigma }_Y$ denote the corresponding variance. The feature variable with the largest separability is ranked first, and the improved J-M distance is obtained as the ratio of the separability to redundancy. The feature variable with the highest value is then selected as the second. Analogously, the improved J-M distance is calculated between the remaining feature variables, with the highest value is selected, until all variables are ordered [14].

2.3.3. Heuristic J-M Distance

The JM_ave and JM_improved methods are used to produce a ranking of feature variables. Based on the ranked feature lists, an approach (JM_mc) for multiclass problems was then proposed [15]. This approach comprises the following steps: (1) First, J-M distances are calculated (Equations (1) and (2)), and values corresponding to each feature are ranked in descending order to produce an initial feature ranking table. (2) Second, a heuristic approach is employed to compile the final optimal feature subset. The proportion (α%) of features that are extracted in different J-M distance ranges is associated with features for each class pair in the feature ranking list varies. Therefore, maximum J-M distance values had varied proportions in different ranges. If the maximum J-M distance value is >1, >0.5 and <1, <0.5, the top α%, 2α%, and 3α% features of the class pair are correspondingly selected. The effect of the extraction ratios between 1 and 20 on the classification accuracy was tested experimentally. The overall accuracy was highest at an extraction ratio of 3%, and this was associated with the lowest number of features; therefore, this value was considered the feature extraction ratio. Features extracted in the preceding step were then used as the final feature combination according to the relationship between the number of class pairs and features (the total number of features was greater than the number of class pairs).

2.4. Random Forest Algorithm

A random forest (RF) classifier is generated by combining decision trees, and each tree represents a unit vote on the most popular category in the input data [29]. The RF is an ensemble classification method which comprises many classifiers instead of one. This classifier utilizes random subsets of input features or predictor variables to partition the search node to reduce generalization errors. To enhance tree diversity, bagging or bootstrap aggregation is used to grow trees from different subsets of training data in the RF method. The Gini index, a measure of the best split selection for RF, measures the impurity of a given element relative to other classes [30]. The RF algorithm has been widely used for classification applications involving multispectral, hyperspectral, SAR, and multisource data [31]. Compared with other machine learning algorithms, such as the decision tree (DT) and support vector machine (SVM), the RF algorithm is more robust and easier to use [30]. In the pre-study stage, we compared the classification accuracy of the Bayesian classifier and RF based on different feature combinations (Table S1 and Figure S1), and this result is similar to numerous studies [13,32,33]. For the Bayesian classifier, we used the default parameters (Lambda = 0.000001), and for the RF, we used the cross-validation method to record the parameters after each adjustment and set 76 and 0.7 as parameters for numberOfTrees and bagFraction.

2.5. Evaluation Method

A confusion matrix [34] was used to evaluate the classification, including the overall, producer, and user accuracies, as well as the Kappa coefficients. The overall accuracy (OA) highlights the probability of the classification result coinciding with the actual land type on the ground, the producer accuracy (PA) reflects the quality of the method used to generate a classification map, and the user accuracy (UA) represents the credibility of each category in the classification map, that is, the reliability of the map [35,36]. The Kappa was originally proposed [37] as a coefficient of the interjudge agreement for nominal scales. Landis indicated it could measure the observed agreement for categorical data [36]. Further, Congalton [38] proposed that as each measure incorporates various levels of information from the error matrix into its computations, they will reflect different information contained within the error matrix. Moreover, Kappa is often used when the class distribution is imbalanced or when chance agreement needs to be accounted for, while OA is a straightforward measure of overall accuracy and is widely used as a performance indicator.

To make a comprehensive comparative analysis of the three methods and to determine the advantages and disadvantages of each method, we comprehensively compared the feature dimension, feature type and feature importance distribution, classification accuracy, and classification results for the optimal feature subset. We specifically evaluated the three methods in terms of precision and dimension by iteratively adding the optimal subset of features. This evaluation was based on observing the change in precision associated with each feature addition and the overall feature dimension. Then, we counted the different types of feature variables separately. It was used to evaluate the compression dimension of different types of features and to analyze the impact of different feature combinations on the classification results of different land classes. Meanwhile, feature importance scores were used to analyze the impact of different features on classification accuracy. Finally, we visually assessed the classification performance by evaluating the accuracy of different methods and comparing the classification results visually. By considering these aspects, we formed a comprehensive judgement of advantages (such as feature selection, classification accuracy and results, and generalizability) and disadvantages (such as limited scope or sensitivity to certain parameters).

3. Results

3.1. Feature Dimensions in an Optimal Subset

Figure 2 shows the OA and Kappa of each iteration. The plots can be divided into three stages [39]. In the early stage (the first six features), the classification accuracies of the three methods rapidly increase, with values ranging from 37.76 ± 3.35% to 82.95 ± 2.04%. In the second stage, the classification accuracies are essentially stable, and the highest classification accuracy (90.20 ± 0.19%) for the B3 (green band), B4 (red band), and B6 (red edge band) bands is attained. In the third stage, the OA decreases slightly, and subsequently remains relatively stable. The number of optimal feature subsets based on the JM_mc differs significantly from the JM_ave and JM_improved methods. Specifically, JM_ave, JM_improved, and JM_mc yield 28, 29, and 15 optimal feature subsets, respectively. JM_improved produces the largest number of dimensional features, while JM_mc achieves the highest compression of features, reducing dimensions by 57.14%. This compression rate is 37.14% and 40% higher than that of JM_ave and JM_improved, respectively. Therefore, JM_mc outperformed the JM_ave and JM_improved methods in the compression of feature dimensions.

3.2. Comparison of Optimized Feature Types and Importance

Figure 3 shows the differences between the optimal feature subsets in feature types. Both the JM_ave and JM_mc methods encompass all feature types, whereas JM_improved exhibits missing data for certain features. Although JM_improved and JM_ave yield a similar number of features, JM_improved lacks data for the vegetation and mangrove features method. These two methods display the highest number of spectral features. The JM_ave method involves higher water features, whereas the JM_improved method involves more topographic and textural features. The JM_ave method involves approximately half the number of features compared with the JM_mc and JM_ave methods. This method excels in employing spectral features and exhibits minimal changes in other feature types. Altogether, spectral features have the largest number in each method and provide abundant classification information. These results suggest that the JM_mc method is better for both the preservation of classification information and reduction of feature dimensions.

Importance scores of features that surpass the mean values of 0.246, 0.244, and 0.505 for the JM_ave, JM_improved, and JM_mc methods are displayed in Figure 4. A higher importance score indicates a more important feature. The JM_mc method produced the lowest feature dimensions and the highest importance score. The JM_improved method had the highest number of features (16) with importance scores greater than the mean value for each method, whereas the JM_ave and JM_mc methods produced 12 and 10 features, respectively. Regarding all methods, spectral features are the most abundant, aligning with the predominance of spectral features in the feature type. Other features that display high importance scores across all three methods include VH, B3_contrast, B4, AWEI, and NDWI. Important features are also relatively high in the ranking of optimal feature subsets. VH, AWEI, and NDWI are all involved in the early stage of feature addition for the JM_ave and JM_mc methods, while B4 is incorporated in the initial stage for the JM_improved method (Figure 2). Thus, the importance of features varies with the J-M distance feature subset selection method. The radar, texture, spectral, and water body features are vital for wetland classification.

3.3. Classification Results

To assess the accuracy of the optimal feature subsets of various methods, the OA, Kappa, and PA were calculated from the confusion matrix. As depicted in Figure 5a, all three methods achieved OA values above 90%. The features of the JM_ave and JM_improved are comparable; however, the latter outperforms the former (JM_ave: OA = 90.07%, Kappa = 88.64%; JM_improved: OA = 90.12%, Kappa = 88.70%). The JM_mc method produces the highest classification accuracy, with an OA value of 90.42% and a Kappa of 89.05%.

Figure 5b illustrates that PA values for various land classes differ according to the feature selection method. Overall, the PA is greater than 80% for all land classes except aquaculture. The lowest PA values for aquaculture are observed in the JM_ave, JM_improved, and JM_mc methods. The highest PAs for the three methods are obtained from the mangrove dataset. The accuracy values obtained from various methods are essentially comparable. The biggest PA differences occur in aquaculture, where the JM_mc method shows an improvement of 1.96% and 2.35% compared to the JM_ave and JM_improved methods, respectively. The smallest PA differences occur in croplands and salt pans. The JM_improved and JM_mc methods had the same values in cropland, and the JM_ave and JM_mc methods also had the same values in vegetation.

In the confusion matrix (

Table 5. Confusion matrix for the JM_ave, JM_improved, and JM_mc methods.

JMave	C1	C2	C3	C4	C5	C6	C7	C8
C1	230	10	0	13	3	0	2	1
C2	16	210	5	5	1	1	0	0
C3	0	2	253	1	3	0	0	1
C4	23	18	4	192	2	12	0	4
C5	0	0	1	6	204	1	8	5
C6	0	2	0	7	0	230	0	2
C7	6	0	0	0	6	0	205	7
C8	1	2	0	0	12	1	1	244
JMimproved
C1	228	11	0	14	3	0	2	1
C2	14	211	4	5	2	2	0	0
C3	0	2	254	1	1	0	1	1
C4	24	18	4	191	3	13	0	2
C5	0	0	1	6	203	1	9	5
C6	0	2	0	6	0	231	0	2
C7	6	0	0	0	6	0	206	6
C8	1	3	0	1	10	0	1	245
JMmc
C1	232	9	0	12	3	0	2	1
C2	12	212	5	7	1	1	0	0
C3	0	2	252	1	3	0	1	1
C4	20	17	3	197	3	13	0	2
C5	0	0	1	6	203	1	8	6
C6	0	2	0	7	0	230	0	2
C7	6	0	0	0	8	0	203	7
C8	1	3	0	1	8	1	1	246

), the mangrove class exhibits the lowest misclassification rates for all three methods, whereas the highest is the aquaculture class. Overall, the confusion of the methods produced the following order JM_ave > JM_improved > JM_mc, and JM_improved and JM_mc performed well in the wetland classification. The JM_improved method provides a better depiction of the mangrove, salt pan, and vegetation classes and decreases misclassifications between the mangrove and vegetation classes. Relatedly, the JM_mc method is advantageous for recognizing water bodies, tidal flats, and aquaculture classes. Despite the similar spectral characteristics of these areas, different methods could be used to their advantage to distinguish different classes.

To facilitate a visual assessment, the study area was partitioned into three sub-regions, as shown in Figure 6. The circles in black represent differences of the classification maps. In region a1–a3, the circle on the right shows that all three methods are suitable for the classification of marine aquaculture areas. The JM_mc method produced the best classification for the aquaculture class, whereas the JM_ave and JM_improved methods produced a mix of aquaculture and tidal flat classes. On the left circle, JM_improved is dominantly covered by tidal flats, whereas the JM_ave method primarily includes water bodies with a small portion of aquaculture. Compared to the other two methods, the JM_mc method decreases misclassification and enables a better depiction of the aquaculture class. Part of the Zhangjiang Estuary mangrove reserve in region c1–c3 was then selected for analysis. As shown in Figure 6c2, the JM_improved method adequately distinguished mangroves and vegetation. Conversely, the JM_ave method exhibits misclassification of mangrove and vegetation in the upper circle (Figure 6c1), while the JM_mc method shows misclassification of both mangroves and vegetation in both regions (Figure 6c3). These results align with the findings of the JM_improved method for mangroves and vegetation, which also reflect higher PA values compared to the other two methods (Figure 5b). The salt pan region (Figure 6b1–b3) highlights comparable classification results for the three methods. Salt pans were mixed with tidal flats and aquaculture areas, but the highest misclassification is the aquaculture class. Both the JM_improved and JM_mc methods are suitable for distinguishing wetland classes and correctly classifying land classes.

4. Discussion

4.1. Optimal Feature Subsets from Different J-M Methods

The optimal feature subsets were selected using three J-M methods that involve different feature dimensions and types. The three feature selection methods utilized in the present study have effectively chosen optimal feature subsets overall, making them suitable for wetland classification. However, optimal feature subsets differ in many ways. Compared with the JM_ave and JM_improved methods, JM_mc efficiently enables the filtering of several features. According to the range of J-M values, the extracted proportions differed [15], JM_mc can effectively compress the feature dimensions using the extraction proportions. Although JM_improved considers weight and information redundancy issues [14], it mainly affects the ranking of feature variables and is inefficient in reducing feature dimensions. Moreover, the improved classification accuracy is related to the utilization of multiple feature types. These feature types provide more information that minimizes misclassifications [40]. Multiple feature types have been widely utilized in various studies to achieve good classification results [5,21,41,42]. The JM_improved method can yield the highest feature dimensions, while vegetation and mangrove features exhibit low rankings. This is because the correlation coefficients are considered, the ranking of many irrelevant features is high, and mangrove and vegetation features occur in an optimal feature subset at a later stage. JM_ave and JM_mc are based on different methodological strategies for optimal feature subset selection, which have a complete set of feature types. For JM_ave, there are enough features based on the descending order of J-M values so that it can satisfy the selection of all feature types. Relatively, JM_mc iteratively selects all features of each class pair when forming the initial feature table, and this method enables the optimal feature subset to contain all feature types. The importance of features can also reflect their contributions to classification accuracy [43]. Compared with the average importance scores of JM_ave and JM_improve of about 0.2, the average importance score of JM_mc exceeds 0.5, which shows that the optimal feature subset composition of JM_mc can better contribute to the accuracy in wetland classification. Moreover, high-ranking radar, texture, spectral, and water body features were more important than other feature types, which aligns with findings from previous studies. For example, radar data significantly influenced the wetland area assigned to decayed vegetation [44]. Hu et al. [45] combined spectral and texture features data to develop a classification and obtained good results. In general, the optimal feature subsets yielded by all three methods produced better classification accuracy in wetland classification. Despite comparable classification accuracy, the JM_mc method, with its lower feature dimensionality, ensures the inclusion of all relevant feature types. Moreover, its optimal feature subset has a higher importance score than the other two methods and can play an important role in wetland classification.

4.2. Influence of the Feature Extraction Method on Wetland Classification Results

The results show that the overall accuracies of the three J-M methods are comparable, among which the JM_mc method produced the highest accuracy. The JM_improved and JM_mc methods performed well in the classification of wetlands. The JM_improved method is advantageous for the separation of mangroves from vegetation; however, the associated optimal feature subset does not contain these features. Similar result were reported by Chen et al. [42]. The inclusion of vegetation and mangrove features, which are derived from spectral bands, may introduce redundant or irrelevant information, leading to a decrease in accuracy. Therefore, considering the correlation among feature variables can reduce the confusion of information and improve the classification performance, especially when dealing with a large number of spectral features. In contrast, the complete feature types were well preserved in JM_ave and JM_mc. The JM_mc method with fewer feature dimensions yielded better classification accuracy and performed well in water body, tidal flat, and aquaculture classes. The classification accuracy of the proposed method with a different number of relevant features selection was compared by Chormunge et al. [46] and indicated that good classification accuracy may vary based on the datasets. Compared with our present study, this result is suitable for wetland classification. In addition, salt pans are easily depicted in aquaculture areas because of their similar textural characteristics, spectral information, and close spatial distribution. All three methods present comparable classification results. Hou et al. [47] also highlighted this problem and solved it by adding water quality parameters. Therefore, utilizing additional variables with distinctive characteristics can enhance the separability of salt pans from other land classes.

4.3. Limitations and Future Work

A combination of multiple features types is most appropriate for the classification of wetland types [4]. Eight feature types, including spectral, radar, topographic, texture, related vegetation, water, mangrove, and red edge features, were used in this study, which contains all common types of wetland classification. However, it should be noted that this study employed a limited number of feature variables selected from each feature type. It may be one of the reasons for misclassification in various wetland classes. Therefore, data from multi-sensors are preferred [48], and the combined use of multi-source satellite data can provide more wetland information [49]. In addition, in our present work, we evaluated the performances of the three feature selection methods based on the J-M distance using a single dataset, with regard to compression of feature number, selection of feature types, and classifier accuracy. However, the study did not consider the inclusion of additional benchmark datasets for a more comprehensive analysis. Therefore, there is still much to learn, and further validations with more individual datasets are still needed to address the specific effects and mechanism of different J-M distance approaches on wetland classification at different spatial scales, as well as other ecosystems.

5. Conclusions

In this study, the performances of three feature subset selection methods, JM_ave, JM_improved, and JM_mc, for wetland classification were compared based on dimensions, feature types, and importance, as well as the associated classification results. The three methods all compressed feature dimensions while preserving classification accuracy. The JM_mc method exhibited the best performance, achieving a feature dimension compression of approximately 50% compared to the original number of features, even though optimal feature subsets obtained using the three methods involved similar feature types, and JM_improved lacked mangrove and vegetation features. Regarding the classification results, the overall accuracies slightly improved in the JM_ave, JM_improved, and JM_mc methods, with values of 90.07%, 90.12%, and 90.42%, respectively. The accuracy difference between the different methods was reflected in the various wetland classes. JM_improved and JM_mc performed well in wetland classes relative to JM_ave. In addition, JM_improved enabled a better depiction of the mangrove, salt pan, and vegetation classes, while JM_mc was advantageous for recognizing water bodies, tidal flats, and aquaculture classes. JM_mc is a promising choice when dealing with a large number of features in wetland classification. Additionally, it is worth noting that different feature selection methods can be utilized for different wetland classes, depending on the specific requirements and characteristics of the classification task. Overall, J-M methods exhibit a good generalization capability and robustness on wetland land cover classification, with great application potential.