Fine-Scale Mangrove Species Classification Based on UAV Multispectral and Hyperspectral Remote Sensing Using Machine Learning

Abstract

Mangrove ecosystems play an irreplaceable role in coastal environments by providing essential ecosystem services. Diverse mangrove species have different functions due to their morphological and physiological characteristics. A precise spatial distribution map of mangrove species is therefore crucial for biodiversity maintenance and environmental conservation of coastal ecosystems. Traditional satellite data are limited in fine-scale mangrove species classification due to low spatial resolution and less spectral information. This study employed unmanned aerial vehicle (UAV) technology to acquire high-resolution multispectral and hyperspectral mangrove forest imagery in Guangxi, China. We leveraged advanced algorithms, including RFE-RF for feature selection and machine learning models (Adaptive Boosting (AdaBoost), eXtreme Gradient Boosting (XGBoost), Random Forest (RF), and Light Gradient Boosting Machine (LightGBM)), to achieve mangrove species mapping with high classification accuracy. The study assessed the classification performance of these four machine learning models for two types of image data (UAV multispectral and hyperspectral imagery), respectively. The results demonstrated that hyperspectral imagery had superiority over multispectral data by offering enhanced noise reduction and classification performance. Hyperspectral imagery produced mangrove species classification with overall accuracy (OA) higher than 91% across the four machine learning models. LightGBM achieved the highest OA of 97.15% and kappa coefficient (Kappa) of 0.97 based on hyperspectral imagery. Dimensionality reduction and feature extraction techniques were effectively applied to the UAV data, with vegetation indices proving to be particularly valuable for species classification. The present research underscored the effectiveness of UAV hyperspectral images using machine learning models for fine-scale mangrove species classification. This approach has the potential to significantly improve ecological management and conservation strategies, providing a robust framework for monitoring and safeguarding these essential coastal habitats.

1. Introduction

Mangrove ecosystems are vital salt-tolerant communities, thriving in tropical and subtropical intertidal regions along coastlines and estuaries [1,2]. They provide plentiful critical ecosystem services such as shoreline protection, biodiversity maintenance, and noteworthy carbon sequestration [3,4]. However, these ecosystems are severely threatened, with a loss of 35% of the area worldwide during the last two decades of the 20th century [5,6]. Mangrove forests have also experienced a rapid decline in China [7,8]. Considering the reduction and degradation of mangrove forests worldwide, there is an urgent need for effective conservation and restoration of mangroves. Accurate mapping of mangrove forests can assist such efforts [2,9]. However, since diverse mangrove species have different functions due to their morphological and physiological characteristics [10,11], there is a growing need for a precise spatial distribution map of mangrove forests at the species level.

Remote sensing offers a powerful tool for mangrove species mapping [12,13,14]. The initial efforts relied on medium-resolution multispectral satellite imagery, such as Landsat and SPOT [15,16,17]. However, they encountered challenges in distinguishing species with small patches, irregular spatial distribution, or similar spectral signatures, due to the limitations of low spatial and spectral resolution [9,18]. The enhanced spatial detail provided by high-resolution satellites is counterbalanced by the limited spectral resolution, which continues to impede nuanced species classification [19,20]. The utilization of hyperspectral imagery satellites, which provide abundant spectral information, hold great potential for precise species classification [21,22,23], albeit with limited spatial resolutions. Data fusion of space-borne hyperspectral imagery with high-resolution data from other platforms is a viable approach to achieve superior classification accuracy [24,25]. However, there are still practical challenges with the data fusion approach that relies on satellite imagery. Mangroves predominantly thrive in cloudy and rainy areas where tidal conditions vary over time. Given that satellite transit times are typically fixed, obtaining low-tide images in specific areas becomes challenging. Consequently, the classification of mangrove communities is constrained by the availability of high-quality optical remote sensing image data.

UAV remote sensing technology emerges as a promising alternative due to its flexibility, affordability, and data quality [26,27]. UAV imagery delivers high spatial resolution, enabling fine-scale classification of mangrove species at the landscape scale [28,29]. This bridges the gap between large-scale satellite monitoring and field surveys, providing a crucial link for advancing species classification and mapping research [30,31,32]. Most UAV remote sensing applications for mangrove monitoring have utilized multispectral data, achieving promising overall classification accuracies between 85% and 90% [14,33,34]. Conversely, studies employing UAV hyperspectral data for mangrove species classification are scarce, and the reported accuracies often range between 80% and 90%, similar to multispectral data [35,36]. One reason for suboptimal classification performance with the UAV hyperspectral data is that the high-dimensional hyperspectral data are often susceptible to the “Hughes phenomenon”, where classification accuracy suffers due to the inclusion of redundant features [37,38]. Dimensionality reduction is often deemed necessary to overcome the “Hughes phenomenon” and fully untap the potential of UAV-based hyperspectral data.

Various remote sensing-based classification methods have been developed, each with its advantages and limitations [1,21,30,39]. Traditional methods like visual interpretation, pixel-based, and object-oriented classification have drawbacks such as subjectivity, inefficiency, or neglecting spatial information [40,41]. Machine learning algorithms have emerged as powerful tools due to their ability to handle complex data and extract nonlinear relationships [42,43]. Ensemble learning methods like Random Forest (RF) and Gradient Boosting have shown promise by combining multiple classifiers for higher accuracy and robustness [5,24,44]. Deep learning, another branch of machine learning, holds great potential for mangrove species classification but has been limited by data availability and computational complexity [1,13]. Although machine learning algorithms such as RF and Light Gradient Boosting Machine (LightGBM) have demonstrated satisfactory classification accuracy with UAV multispectral data, their generalizability to UAV hyperspectral remote sensing has received scant attention in the literature [44]. Further research is necessary to address existing challenges and fully harness the potential of machine learning for effective mangrove ecosystem management and conservation.

This study investigates the potential of UAV-based hyperspectral and multispectral imagery for mangrove species classification. Yingluo Bay and Pearl Bay were chosen as our study areas due to their contrasting community structures in Guangxi, China. This variation allows us to evaluate the robustness and generalizability of the model for mangrove species identification across diverse ecological settings. The specific objectives of this study encompass (1) comparing the performance of four machine learning algorithms (Adaptive Boosting (AdaBoost), eXtreme Gradient Boosting (XGBoost), RF, and LightGBM) in classifying mangrove species; (2) identifying the key features of multi- and hyper-spectral data derived from UAV imagery for mangrove species classification; and (3) providing recommendations on the suitability of UAV imagery data and classification methods for mapping mangroves accurately and efficiently. The findings will provide valuable insights into the effectiveness of different data sources and classification methods for accurate mangrove species identification. The results will also inform the development of robust and transferable UAV remote sensing protocols for mangrove species classification in diverse geographical and environmental settings.

2. Materials and Methods

2.1. Study Area

Yingluo Bay is a part of Shankou National Mangrove Ecological Nature Reserve (21°28′~21°37′N, 109°37′~109°47′E) in Guangxi (Figure 1A). The bay has a tropical marine monsoon climate with an average annual temperature of 22.9 °C and rainfall of 1573 mm (http://data.cma.cn/data/ (accessed on 1 June 2020)). Both irregular diurnal and semidiurnal tides coexist, with an average tidal range of 2.5 m and a maximum of 6.3 m [45]. This area exhibits denser vegetation with a wider variety of species, including Bruguiera gymnorrhiza, Rhizophora stylosa, Kandelia candel, Aegiceras corniculatum, Avicennia marina, Excoecaria agallocha, Hibiscus tiliaceus, and Spartina alterniflora [46]. Designated as a core conservation area, Yingluo Bay boasts well-preserved and representative mangrove vegetation, making it ideal for studying and classifying mangrove species.

Pearl Bay is within the Guangxi Beilun Estuary National Nature Reserve, which is situated at the mouth of the Beilun River in Guangxi, Southwest China (21°28′~21°37′N, 109°37′~109°47′E) (Figure 1B). Pearl Bay is approximately 3.5 km wide at its mouth and extends 46 km along the coastline. The bay experiences a subtropical marine monsoon climate with an average temperature of 22.5 °C and annual rainfall of 2220 mm (http://data.cma.cn/data/ (accessed on 1 July 2020)). Regular diurnal tides dominate the area, with an average tidal range of 2.2 m and a maximum of 5.1 m [47]. Mangroves cover 939.97 hectares, with dominant species including Bruguiera gymnorrhiza, Kandelia candel, Aegiceras corniculatum, and Excoecaria agallocha [47]. Pearl Bay is also a core conservation area, where there are preserved and representative mangrove vegetation areas to implement research.

2.2. Method

2.2.1. Overview

Our mangrove species identification process follows a five-step workflow (Figure 2): (1) Data acquisition: This stage involves UAV image data acquisition and sample plots inventory. (2) Data preparation: we assemble UAV hyperspectral and multispectral data, pre-process it, generate vegetation index datasets, and incorporate field survey data. (3) Feature extraction and selection: The most informative features were identified from vegetation indices, texture features, and original spectral data employing the Recursive Feature Elimination (RFE) method. (4) Classifier optimization: We evaluate the performance of different combinations of data sources and machine learning methods in mangrove species classification. (5) Modeling assessment: The performance of the chosen models is assessed to evaluate their robustness and generalizability across different regions.

2.2.2. Data Acquisition

We employed a DJI M300 RTK (M300, DJI, Shenzhen, China) quadcopter equipped with a ULTRIS X20 Plus sensor for hyperspectral image capture. A DJI Phantom 4 Multi RTK (P4M RTK, DJI, Shenzhen, China) drone with an integrated multispectral imaging system (visible, red, green, blue, red edge, near-infrared) was used to acquire multispectral data. High-resolution visible light data were obtained using a DJI Phantom 4 RTK (P4 RTK, DJI, Shenzhen, China) drone equipped with a 1-inch, 20-megapixel CMOS sensor. This provides a Ground Sampling Distance (GSD) of 2.74 cm at a flight altitude of 100 m, ideal for acquiring reference data.

Hyperspectral, multispectral, and visible light data were collected between 2022 and 2023 under clear skies, ample sunlight, and low tidal levels (below 3 m) to ensure exposed mangrove areas. Data acquisition during summer was restricted to morning and afternoon to avoid high solar noon conditions that caused water surface glare. Flight missions maintained an 80% forward overlap and 70% side overlap for comprehensive data capture. The flight path adopted a strip pattern extending from inland to coastal areas to ensure complete coverage of mangrove distribution.

Field surveys were performed at Yingluo Bay and Pearl Bay on 22–24 June 2023, and 25–27 November 2023, respectively. A Qianxun SR6 Plus RTK (SP6 Plus, Qianxun Spatial Intelligence Inc., Shanghai, China) instrument was used to collect ground truth data based on the China Geodetic Coordinate System 2000 coordinate system, including geographic location, species names, images, and other relevant vegetation information for sample selection. However, marshy terrain and dense canopy cover within the study area hampered RTK signal reception, impacting accuracy. To address this challenge, tree species surveys primarily followed tidal creek directions for improving accessibility and locational accuracy.

2.2.3. Data Preparation

The acquired hyperspectral images were first fused using Cubert-Pilot 3.0 software (Cubert, Ulmu, Germany), then processed using Agisoft PhotoScan Professional 1.7.3 software (Agisoft LLC, St. Petersburg, Russia) for photo alignment, generating dense point clouds, creating meshes, texturing, and producing orthoimages. Subsequently, geometric correction and image cropping were conducted in ENVI 5.5 software, maintaining a Root Mean Square Error (RMSE) below 0.5 pixels. Visible and multispectral images were stitched and processed into RGB images and Digital Orthophoto Maps (DOMs) for each band using DJI Terra 4.2.3 software (DJI, Shenzhen, China). These five single-band DOMs were composited into a multispectral image using ArcGIS 10.7 software (ESRI, Redlands, CA, USA). All imagery was resampled to 0.06 m using the nearest neighbor method to maintain consistent spatial resolution. Detailed flight parameters are provided in

Table 1. UAV equipment parameters and image data acquisition information.

Parameters	Visible Image	Multispectral Image	Hyperspectral Image
UAV model	DJI P4 RTK	DJI P4M RTK	DJI M300 RTK
Sensor	RGB camera	Multispectral camera	ULTRIS X20 PLUS
Acquisition time	7 November 2022/24 April 2023	20 June 2023/24 April 2023	9 November 2022/25 July 2023
Flight altitude (m)	50	80/100	120/150
FOV/Maximum filed angle (°)	84	62.7	35
Band range (nm)	400~700	450~840	350~1000
Spectral resolution (nm)	—	—	4
Spatial resolution (m)	0.02	0.03/0.05	0.03/0.06
Number of bands	3	5	164

.

Given the slow successional cycles of mature mangrove trees, the spatial distribution patterns were assumed to be stable within a year. A total of 434 and 692 samples were collected in the Yingluo Bay and Pearl Bay areas, respectively (Figure 1A,B). In areas with limited ground access, a DJI P4 RTK drone captured high-resolution RGB images at 50 m flight altitude (parameters in Table 1). These drone images, combined with the collected ground survey data, were used to establish interpretation signs (

Table 2. Field survey photos of mangrove plant species and other cover types in the study area and establishment of interpretation signs from UAV visible image.

Species and Cover Type	File Survey Photo	UAV Image	Interpretation Signs
RS 1			Dark green in color, with a regular texture resembling dense points, occurring in continuous patches.
BG 2			Light green in color, irregular in texture, with blurred boundaries, clustered in patches, and dispersed distribution.
AM 3			Composed of light and tender green colors, with a rough texture, predominantly distributed around river channels.
AC 4			Yellow-green in color, featuring a relatively smooth texture with frequent small gaps, primarily distributed around river channels.
KC 5			Green in color, appearing clustered with a uniform hue.
EA 6			Bright green in color, characterized by rough texture, mostly growing along the coastline in scattered distribution.
HT 7			Yellow-green in color, displaying a mixed hue, with visibly rough texture, mostly growing along the coastline in a patchy distribution.
SA 8			Light green in color, uniform in hue, with smooth texture, clustered in patches, predominantly located near the seaside.
WB 9			Mainly composed of gray and light green colors, with a smooth and delicate texture.
MF 10			Predominantly gray-brown, with a smooth texture and uniform tone.
RD 11			Composed of bright white and light gray colors, displaying a striped distribution.

¹ Rhizophora stylosa, ² Bruguiera gymnorrhiza, ³ Avicennia marina, ⁴ Aegiceras cor-niculatum, ⁵ Kandelia candel, ⁶ Excoecaria agallocha, ⁷ Hibiscus tiliaceus, ⁸ Spartina alterniflora, ⁹ water bodies, ¹⁰ mudflats, and ¹¹ roads.

) for identifying 11 land cover types. These categories included mangrove species Rhizophora stylosa (RS), Bruguiera gymnorrhiza (BG), Avicennia marina (AM), Aegiceras corniculatum (AC), Kandelia candel (KC), Excoecaria agallocha (EA), and Hibiscus tiliaceus (HT), other vegetation Spartina alterniflora (SA), and non-vegetation water bodies (WB), mudflats (MF), and roads (RD). The sample library was enhanced through the incorporation of visual interpretation of visible images for feature selection and model optimization.

2.2.4. Feature Extraction and Selection

Mangrove species exhibit unique spectral characteristics, with variations in their biochemical and physiological properties reflected in their spectral signatures. Preprocessing of the acquired drone-based hyperspectral data revealed significant noise in the first 20 bands (ranging from 400 to 432 nm). Consequently, these bands were removed, resulting in a refined dataset of 144 bands spanning 433 to 1000 nm. Conversely, multispectral data exhibited less noise and retained the original red, green, blue, red edge, and near-infrared bands. Sole reliance on spectral reflectance for mangrove species classification has limitations. To improve species identification accuracy, additional features are crucial. This study incorporated two key feature types: vegetation indices (VIs) and texture features.

Vegetation indices are mathematical transformations of spectral data used to highlight vegetation features and distinguish between plant types [48,49,50,51]. In this study, 49 candidate VIs were selected for analysis. Texture features capture spatial variations in the arrangement of pixels within an image, providing valuable information for species classification. This study utilized high-resolution UAV imagery and the Gray-Level Co-occurrence Matrix (GLCM) method to extract texture information. GLCM statistically describes the spatial distribution of grayscale values between pixels at specified distances and directions [50,52,53].

To reduce redundancy in spectral information, principal component analysis (PCA) was employed [54]. Research suggests using the first principal component image from PCA for GLCM calculations yields better texture feature descriptions compared to single-band approaches. Therefore, eight texture operators (Entropy, Correlation, Dissimilarity, Homogeneity, Contrast, Mean, Angular Second Moment, and Variance) were computed on the first principal component image derived from the UAV multispectral and hyperspectral data. Calculations were performed in eight directions (0°, 45°, 90°, 135°, 180°, 225°, 270°), with a step size of 1 and window sizes of 3 × 3, 5 × 5, and 7 × 7. Ultimately, a total of 192 texture features (1 PCA band × 3 window types × 8 texture operators × 8 directions) were computed for both multispectral and hyperspectral datasets.

When using a large number of features for modeling, the Hughes phenomenon can occur, reducing model efficiency and classification accuracy. To address this issue, the Recursive Feature Elimination (RFE) algorithm was employed. RFE iteratively trains a model and eliminates the least important features, ultimately selecting an optimal feature subset. The Random Forest model was chosen within the RFE algorithm due to its robustness to overfitting, reduced sensitivity to parameters, and fast training speed [33].

2.2.5. Selection of Machine Learning Classification Algorithms

This study evaluated the performance of four machine learning algorithms (AdaBoost, XGBoost, RF, and LightGBM) for mangrove species classification. AdaBoost, an ensemble method, enhances weak classifiers by focusing on misclassified instances, proving particularly effective for task with subtle class distinctions like mangrove species identification [55]. XGBoost, a gradient boosting algorithm with regularization, excels at handling complex, high-dimensional remote sensing datasets [56,57]. RF, a collection of decision trees, is robust for large datasets and offers valuable feature importance estimates [58]. LightGBM, optimized for speed and efficiency, employs a leaf-wise tree growth strategy, often outperforming level-wise algorithms like XGBoost [59,60].

These four machine learning algorithms exhibit certain distinctions [55,61,62,63]. AdaBoost and XGBoost are both boosting techniques, with XGBoost incorporating regularization for improved performance [62]. RF, based on bagging methodology, offers stability but may be less powerful in capturing complex relationships compared to boosting methods. XGBoost and LightGBM, both gradient boosting algorithms, differ in tree growth strategies, with LightGBM potentially achieving higher accuracy but at the risk of overfitting if not carefully tuned [63]. RF excels in feature importance assessment, while XGBoost and LightGBM prioritize error minimization. Both XGBoost and LightGBM employ regularization to prevent overfitting. Ultimately, these algorithms were selected due to their complementary strengths in addressing the challenges of mangrove species classification, including handling high-dimensional data, achieving robust performance, and providing interpretable results.

To optimize model performance, a comprehensive hyperparameter tuning process was conducted using grid search with cross-validation from the scikit-learn library. Hyperparameter tuning is critical in machine learning, aiming to identify the best combination of settings to enhance model performance, generalization ability, and training efficiency. Grid search is a popular method for this task due to its simplicity, ease of use, and ability to leverage parallel processing. The grid search automatically selects the hyperparameter combination that produces the best performance, defining the final model’s configuration.

Open-source Python libraries were used for model implementation: scikit-learn for AdaBoost (https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.AdaBoostClassifier.html (accessed on 1 June 2023)), XGBoost (https://readthedocs.org/projects/xgboost/ (accessed on 1 June 2023)), RF (https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html (accessed on 1 June 2023)), and LightGBM (https://lightgbm.readthedocs.io/en/latest/Python-Intro.html (accessed on 1 June 2023)).

2.2.6. Classification Accuracy Evaluation

This study employed the confusion matrix to evaluate the classification accuracy of mangrove species. The confusion matrix compares and analyzes the agreement between actual ground truth categories and the model’s classified results. This comparison allows for the calculation of several key metrics to assess classification performance: Producer’s Accuracy (PA), User’s Accuracy (UA), Kappa Coefficient (Kappa), and Overall Accuracy (OA), aiming to assess mapping precision and classification effectiveness comprehensively. These metrics provide a comprehensive evaluation of mapping precision and classification effectiveness for mangrove species identification.

3. Results

3.1. Feature Selection Based on REF-RF

An initial set of 237 features was extracted for multispectral imagery in Yingluo Bay. Correlation analysis reduced this to 45 features, and RFE-RF further refined the selection to 17 optimal features. These included nine vegetation indices, six textural features, and two spectral features (

Table 3. The optimal feature set is derived from multi- and hyper-spectral images in two study sites, respectively.

Study Sites	Data Sources	Dimension	Feature Types	Optimal Features
Yingluo Bay	Multi	17	Vegetation Indices	ARI1, GEMI, LAI, MCARI, NDMI, SRI, SGI, TDVI, TGI
			Texture features 1	m_5_0_Con, m_7_180_Con, m_7_270_Mea, m_7_270_Con, m_7_45_Con, m_7_45_ASM
			Spectral features	Red, RedEdge
	Hyper	25	Vegetation Indices	CRI2, SRI, TGI, MRENDVI, VARI, MRESR, SIPI, NDMI, RGRI, PSRI, ARI1, ARI2, REPI, PRI, WBI
			Texture features 2	h_3_90_Mea
			Spectral features 3	h_band52, h_band89, h_band92, h_band98, h_band147, h_band150, h_band153, h_band155, h_band157
Pearl Bay	Multi	12	Vegetation Indices	ARI1, GARI, MCARI, NDMI, RGRI, TGI
			Texture features 4	m_5_45_Mea, m_7_0_Con, m_7_180_Con, m_7_225_Con, m_7_270_Con
			Spectral features	Blue
	Hyper	16	Vegetation Indices	ARI1, ARI2, CRI2, GLI, MRESR, NDMI, SRI, SIPI, TCARI, VREI1, VREI2
			Texture features 5	h_5_180_Var
			Spectral features 6	h_band52, h_band91, h_band101, h_band162

^1,4 The abbreviation “m” represents multispectral data. The numbers 3, 5, and 7 represent the dimensions of the sliding window, the angles 0, 45, 90, 135, 180, 225, and 270 represent the directional movement of the sliding window. Contrast (Con), Mean (Mea), Angular Second Moment (ASM), and Variance (Var). ^2,5 The abbreviation “h” represents hyperspectral data, and the interpretations of additional symbols align with the multispectral data. ^3,6 The “band + numbers” indicates the spectral reflectance values for different bands in hyperspectral data. The abbreviations of vegetation index can be found in Table A1 of Appendix A.

, Appendix A 

Table A1. The equations of vegetation indices were used in this study.

Spectral Indices	Formula	References
Anthocyanin Reflectance Index1 (ARI1)	$ARI1={\displaystyle \frac{1}{{\rho }_{550}}}-{\displaystyle \frac{1}{{\rho }_{700}}}$	[67]
Anthocyanin Reflectance Index1 (ARI2)	$ARI2={\rho }_{800}\times \left({\displaystyle \frac{1}{{\rho }_{550}}}-{\displaystyle \frac{1}{{\rho }_{700}}}\right)$	[68]
Carotenoid Reflectance Index1 (CRI2)	$CRI2={\displaystyle \frac{1}{{\rho }_{510}}}-{\displaystyle \frac{1}{{\rho }_{700}}}$	[69]
Enhanced Vegetation Index (EVI)	$EVI=2.5\times {\displaystyle \frac{NIR-Red}{NIR+6\times Red-7.5\times Blue+1}}$	[70]
Global Environmental Monitoring Index (GEMI)	$GEMI=eta\left(1-0.25\times eta\right)-{\displaystyle \frac{Red-0.125}{1-Red }}$	[71]
Green Atmospherically Resistant index (GARI)	$GARI={\displaystyle \frac{NIR-\left[Green-\gamma \left(Blue-Red \right)\right]}{NIR+\left[Green-\gamma \left(Blue-Red \right)\right]}}, \gamma =1.7$	[72]
Green Leaf Index (GLI)	$GLI={\displaystyle \frac{\left(Green-Red\right)+\left(Green-Blue\right)}{2\times Green+Red+Blue}}$	[73]
Leaf Area Index (LAI)	$LAI=3.618\times EVI-0.118$	[74]
Modified Chlorophyll Absorption Ratio Index (MCARI)	$MCARI=\left[\left({\rho }_{700}-{\rho }_{670}\right)-0.2\times \left({\rho }_{700}-{\rho }_{550}\right)\right]\times \left({\displaystyle \frac{{\rho }_{700}}{{\rho }_{670}}}\right)$	[75]
Modified Red Edge Normalized Difference Vegetation Index (MRENDVI)	$MRENDVI={\displaystyle \frac{{\rho }_{750}-{\rho }_{705}}{{\rho }_{750}+{\rho }_{705}-2\times {\rho }_{445}}}$	[76]
Modified Red Edge Simple Ration (MRESR)	$MRESR={\displaystyle \frac{{\rho }_{750}-{\rho }_{445}}{{\rho }_{750}+{\rho }_{445}}}$	[77]
Normalized Difference Mud Index (NDMI)	$NDMI={\displaystyle \frac{{\rho }_{795}-{\rho }_{990}}{{\rho }_{795}+{\rho }_{990}}}$	[78]
Photochemical Reflectance Index (PRI)	$PRI={\displaystyle \frac{{\rho }_{531}-{\rho }_{570}}{{\rho }_{531}+{\rho }_{570}}}$	[79,80]
Plant Senescence Reflenctance Index (PSRI)	$PSRI={\displaystyle \frac{{\rho }_{680}-{\rho }_{500}}{{\rho }_{750}}}$	[81]
Red Edge Position Index (REPI)	The wavelength of the maximum derivative of reflectance in the vegetation red edge region of the spectrum in microns from 690 to 740 nm.	[82]
Red Green Ratio Index (RGRI)	$RGRI={\displaystyle \frac{{\sum }_{i=600}^{699}{R}_i}{{\sum }_{j=500}^{599}{R}_j}}$	[83]
Simple Ratio Index (SRI)	$SRI={\displaystyle \frac{NIR}{Red}}$	[84]
Structure insensitive Pigment Index (SIPI)	$SIPI={\displaystyle \frac{{\rho }_{800}-{\rho }_{445}}{{\rho }_{800}-{\rho }_{680}}}$	[85]
Sum Green Index (SGI)	${\rho }_{500}~{\rho }_{600}$	[86]
Transformed Chlorophyll Absorption Reflectance Index (TCARI)	$TCARI=3\left[\left({\rho }_{700}-{\rho }_{670}\right)-0.2\left({\rho }_{700}-{\rho }_{550}\right)\left({\displaystyle \frac{{\rho }_{700}}{{\rho }_{670}}}\right)\right]$	[87]
Transformed Difference Vegetation Index (TDVI)	$TDVI=\sqrt{0.5+{\displaystyle \frac{NIR-Red}{NIR+Red }}}$	[88]
Triangular Greenness Index (TGI)	$TGI={\displaystyle \frac{\left(\mathrm{R}\mathrm{e}d-Blue\right)\left(\mathrm{R}\mathrm{e}d-Green\right)-\left(Red-Green\right)\left(Red-Blue\right)}{2}}$	[89]
Visible Atmospherically Resistant Index (VARI)	$VARI={\displaystyle \frac{Green-Red}{Green+Red-Blue}}$	[90]
Vogelmann Red Edge Index1 (VREI)	$VREI1={\displaystyle \frac{{\rho }_{740}}{{\rho }_{720}}}$	[91]
Vogelmann Red Edge Index2 (VREI)	$VREI2={\displaystyle \frac{{\rho }_{734}-{\rho }_{747}}{{\rho }_{715}+{\rho }_{726}}}$	[91]
Water Band Index (WBI)	$WBI={\displaystyle \frac{{\rho }_{970}}{{\rho }_{900}}}$	[92]

The selected bands of Red, Green, Blue, and NIR should be within their respective wavelength ranges. For example, the wavelength of the selected red band must fall between 600–700 nm. In case multiple bands fall within this range simultaneously, choose the band with the closest central wavelength. Here, the ${\rho }_{num}$ represent the reflectivity of each corresponding band where num denotes its central wavelength. The selection principle for bands remains consistent as mentioned above.

). The cross-validation score was 0.81 for multispectral imagery. An initial set of 385 features was extracted for hyperspectral imagery in Yingluo Bay. Correlation analysis reduced this to 49 features, and RFE-RF further refined the selection to 25 optimal features. These included 15 vegetation indices, 9 spectral features, and 1 textural feature (Table 3, Appendix A Table A1). The cross-validation score was 0.94 for hyperspectral imagery.

Similar to Yingluo Bay, correlation analysis and REF-RF were also applied to the Pearl Bay image. In Pearl Bay, the optimal number of features selected for multi- and hyper-spectral were 12 and 16, respectively (Table 3, Appendix A Table A1). The cross-validation scores were 0.75 and 0.91, with the hyperspectral imagery demonstrating a higher score attributed to its enhanced spectral information.

3.2. Mangroves Species Classification and Accuracy Evaluation

This study evaluated the performance of four machine learning algorithms (AdaBoost, XGBoost, RF, and LightGBM) for mangrove species classification using multispectral and hyperspectral data sources. In Yingluo Bay, across both data sources, a consistent ranking of algorithm performance emerged. The ranking from highest to lowest accuracy was as follows: LightGBM > RF > XGBoost > AdaBoost (

Table 4. The mangrove species classification accuracy obtained from four machine learning models based on UAV multi- and hyper-spectral images in Yingluo Bay.

Models	Multispectral Image		Hyperspectral Image
	OA (%)	Kappa	OA (%)	Kappa
AdaBoost	63.05	0.56	82.96	0.79
XGBoost	80.37	0.77	94.26	0.93
RF	80.50	0.77	95.73	0.95
LightGBM	80.96	0.78	97.15	0.97

). LightGBM achieved the highest classification accuracy (OA: 80.96%, 97.15%, Kappa: 0.78, 0.97) in multi- and hyper-spectral images, respectively. Conversely, AdaBoost exhibited the lowest accuracy (OA: 63.05%, 82.96%, Kappa: 0.56, 0.79) in both scenarios. Although LightGBM consistently outperformed other algorithms, the margin of improvement varied based on the data source. Using multispectral data, the difference between LightGBM, RF, and XGBoost was minimal, with all achieving OA values around 80% and Kappa around 0.77. In contrast, LightGBM displayed a significant advantage in hyperspectral data. Compared to RF and XGBoost, LightGBM yielded improvements of 1.42% and 2.89% in overall accuracy, respectively. This highlights the ability of LightGBM to leverage the richer spectral information in hyperspectral data for more accurate classification.

This study also evaluated the performance of each classification algorithm by calculating the PA and UA for each mangrove species class based on the overall confusion matrix (Figure 3). LightGBM achieved the highest or near-highest PA and UA values across most classes, particularly for species with limited distribution areas like HT, EA, and SA. These species posed a challenge in mapping due to their smaller sample sizes. In contrast, the AdaBoost algorithm yielded significantly lower PA and UA for all classes. The AdaBoost algorithm showed significant discrepancies between the two data sources. It achieved zero accuracy for BG and HT when used with multispectral images while demonstrating poor performance for AC and AM with hyperspectral images.

To assess the generalizability of the LightGBM model, we applied it to separate UAV datasets acquired in Pearl Bay. LightGBM based on multispectral data failed to identify EA entirely, highlighting its limitations for certain species (Figure 4). Hyperspectral data displayed significant variation in PA and UA across different species. This suggests that while hyperspectral data generally outperform multispectral data, their effectiveness can vary depending on the specific species. Unlike Yingluo Bay, Pearl Bay exhibited high PA and UA accuracy for HT (>80%), indicating successful identification.

3.3. Mapping the Distribution of Mangrove Species

After performing feature selection and model building, four algorithms were applied to the UAV-acquired multispectral and hyperspectral images of Yingluo Bay (Figure 5). The classification results revealed the advantages of hyperspectral data in effectively mitigating the “salt-and-pepper” noise, a common issue encountered in multispectral data that often leads to misclassification (Figure 5). The utilization of hyperspectral data further facilitated the more precise delineation of vegetation boundaries, thereby enhancing the visual quality of classification maps (Figure 6).

Although visual differences between algorithms using hyperspectral data were subtle (Figure 5), LightGBM outperformed others, with a classification map closely matching the actual spatial distribution of species (Figure 6). The geographical distribution range of BG scattered in a large area of RS is visible, particularly. However, in the coastal area, it was observed that a small portion of AC had been erroneously classified as BG, and some MF had been mistakenly categorized as SA. Non-vegetated features such as RD and MF can still be accurately distinguished. The AdaBoost algorithm exhibits the poorest classification performance, with HT being indistinguishable and only a small portion of EA correctly identified, while most of SA is erroneously classified as the bare beach (Figure 5). These factors primarily contribute to the low accuracy (OA: 82.96%, Kappa: 0.79) observed in the AdaBoost model (Table 4).

The classification results provided insights into the spatial distribution of mangrove species in Yingluo Bay (Figure 5 and Figure 6). HT and EA were often found intermixed near inland roads. RS exhibited a more widespread distribution across the area. BG was primarily scattered along inner beaches. AM was concentrated in outer beaches, while AC was found interspersed with AM in low-tide zones and outer beaches.

The classification maps derived from UAV multispectral and hyperspectral imagery using the LightGBM model in Pearl Bay highlight the advantages of hyperspectral data (Figure 7 and Figure 8). Hyperspectral imagery significantly mitigates the “salt-and-pepper” effect observed in multispectral data. The presence of this noise can result in misclassifications, such as confusion between AC and AM (Figure 7). Hyperspectral data facilitate clearer delineation of species boundaries compared to multispectral data (Figure 7 and Figure 8). This was evident in the improved distinction between bare beaches and vegetation. Notably, LightGBM using multispectral data failed to detect EA entirely. Hyperspectral data, however, offered improved species identification (Figure 8).

The classification maps provided valuable insights into the spatial distribution of mangrove species in Pearl Bay. EA and HT were distributed along the landward edge in strip-like patterns. BG exhibits sparse distribution in the central region. AM was primarily concentrated in the central area and along the coast. KC and AC were intermixed throughout the study area.

4. Discussion

4.1. Feature Selection

The RFE-RF and correlation analysis techniques were employed to investigate the impact of feature selection on mangrove species classification using UAV imagery in Yingluo Bay and Pearl Bay. Feature selection significantly reduced the initial feature sets, confirming stable model performance even with less than 20 features. Across both multispectral and hyperspectral data, vegetation indices consistently ranked as the most influential features for distinguishing mangrove species. This highlights their sensitivity to species identification. Vegetation indices are mathematical combinations of reflectance values from two or more spectral bands. Compared to single spectral bands, vegetation indices exhibit a closer relationship with key plant characteristics, such as growth stage, biomass, and leaf properties, which could enable a more accurate representation of the health and structure of mangrove species [64]. Textural features played a less prominent role in hyperspectral data. Richer spectral information in hyperspectral data allows individual spectral bands to provide valuable discriminatory power, reducing the reliance on textural features. Conversely, multispectral data, with lower spectral resolution, require texture analysis to extract internal canopy details (shadows, leaf attributes, and branch thickness). Textural features thus contributed more to classification than original spectral features for this data type.

The selected features also differed between the two study areas due to variations in mangrove species composition. Similar features can vary in importance depending on the specific species present. Vegetation indices reflect various attributes like structure (biomass, cover, and LAI), biochemistry (water and pigments), and physiology (fluorescence). Different indices target different attributes. Similar to other studies, this research found that widely used traditional indices like NDVI and DVI were rarely selected features [65]. This may be because these indices are not sensitive enough to capture the subtle spectral differences between mangrove species. Regardless of the chosen features, most selected spectral features concentrated in the red edge and near-infrared regions, suggesting these ranges hold key information for species differentiation [66].

4.2. Advantage of Hyperspectral Data in Mangrove Species Identification

This study investigated the influence of data sources (multispectral vs. hyperspectral) and machine learning algorithms on mangrove species classification accuracy. UAV-derived imagery from both sensors was subjected to variable selection and modeling, yielding distinct classification results. Multispectral data exhibited a significant amount of salt-and-pepper noise. In contrast, hyperspectral imagery demonstrated superior performance. Its richer spectral information provided greater clarity of vegetation boundaries, effectively mitigating salt-and-pepper noise and enhancing overall visual quality. This highlights the substantial advantage of hyperspectral data for accurate species identification. For instance, AdaBoost consistently underperformed, particularly in distinguishing species like BG and AC that are sparsely distributed. Notably, hyperspectral data enabled AdaBoost to achieve a high accuracy of 0.93 and 0.67 based on the confusion matrix (Appendix B Figure A1), whereas multispectral data faced challenges for these sparsely distributed species.

Hyperspectral data allowed for more precise species differentiation, which contains abundant spectral information. The LightGBM model achieved high classification accuracies for key species such as BG (0.98) and RS (0.98), as evidenced by the confusion matrix (Appendix B Figure A1). However, species with spectral similarities and mixed distributions, such as EA and HT, yielded less optimal results. Overall, hyperspectral data exhibited a lower probability of misclassification compared to multispectral data, solidifying its superiority for mangrove species classification.

4.3. Model Robustness Evaluation

This study investigated the influence of the study area on classification accuracy for four machine learning algorithms. The four classification algorithms consistently demonstrated high levels of OA and Kappa rankings across both data sources. The order of accuracy rankings, from highest to lowest, was as follows: LightGBM > RF > XGBoost > AdaBoost. The LightGBM algorithm demonstrates the highest classification accuracy across various data sources, with an optimal OA of 97.15% and Kappa of 0.97. On the other hand, the AdaBoost algorithm exhibits the lowest accuracy, with an OA of merely 63.05% and Kappa as low as 0.56, indicating its significant deviation from the performance levels achieved by the other three algorithms. Although the LightGBM algorithm exhibits numerous advantages in classifying mangrove species, there is no statistically significant difference when applied to multispectral data.

Beyond its superior accuracy, LightGBM addresses limitations present in RF and XGBoost. These algorithms can be computationally expensive, requiring high memory consumption and lengthy parameter optimization processes. LightGBM effectively overcomes these drawbacks, offering faster training efficiency, lower memory usage, and the ability to handle large datasets. This combination of speed, efficiency, and superior accuracy makes LightGBM a compelling choice for analyzing high dimensional hyperspectral data.

4.4. Limitations and Implications for Mangrove Species Mapping

This study confirms the feasibility of UAV multispectral and hyperspectral data for mangrove species classification using machine learning approaches. However, several challenges are remaining. Firstly, weather variations, particularly in large study areas, can significantly impact image quality due to changes in light intensity and solar elevation during data acquisition. Although UAVs offer flexible operation times, rapid weather fluctuations can degrade data quality compared to the fixed-rate collection process. This limits UAV remote sensing to scenarios where weather conditions are more unpredictable.

Secondly, dense mangrove vegetation presents challenges for image mosaicking and orthorectification. Establishing precise tie points between flight strips is difficult, hindering accurate image stitching. Additionally, storing and processing high-resolution hyperspectral data, characterized by large data volumes, is computationally demanding, limiting its application in large-scale studies.

Furthermore, mangrove classification is inherently more complex than plantation forests or urban trees. Dense stands with significant species intermixing make canopy boundary delineation difficult. This results in the phenomena of “same object, different spectra” and “different objects, same spectra” within imagery. Current classifications typically encompass fewer than five species, and the number of classified species often inversely affects accuracy. Laser scanning technology offers promise by providing detailed three-dimensional structural information that can complement optical imagery and potentially enhance classification accuracy.

Studies indicate that spatial resolution also influences classification effectiveness, where higher resolution does not necessarily lead to better classification results [22,24,32]. This study focused on a single spatial resolution. Future research should investigate the impact of using different resolutions on feature selection and classification performance. Spatial resolution may not always correlate directly with classification accuracy, highlighting the importance of selecting an optimal resolution for each study.

The present study showcases the efficacy of machine learning ensemble algorithms in handling hyperspectral data, while deep learning holds promise for future advancements. However, challenges such as long training times, significant computational requirements, limited interpretability, and sensitivity to data quality hinder its widespread adoption. Future research can explore the differential effectiveness of deep learning and machine learning ensemble approaches in mangrove species classification.

5. Conclusions

This study investigated the feasibility and limitations of UAV multispectral and hyperspectral imagery, coupled with machine learning algorithms, for fine-scale mangrove species identification. The key findings are summarized below:

• Dimensionality reduction: UAV remote sensing data can contain redundant features, impacting classification accuracy. To address this, we employed correlation analysis and utilized an RF algorithm with recursive feature elimination for feature selection. This approach effectively reduced overfitting and identified vegetation indices as the most informative features for inter-species classification.
• Data source and algorithm performance: Hyperspectral data consistently yielded superior classification results compared to multispectral data. Among the evaluated algorithms, the LightGBM algorithm achieved the highest classification accuracy. RF and XGBoost also demonstrated promising performance for hyperspectral mangrove species classification.
• Spatial variability and classification performance: Comparative analysis between study areas revealed higher overall classification accuracy in Yingluo Bay compared to Pearl Bay, despite using the same data source. This difference likely stems from variations in species characteristics between the two locations. This finding underscores the importance of considering spatial variability in species distribution when optimizing classification strategies.