Provincial-Scale Monitoring of Mangrove Area and Spartina alterniflora Invasion in Subtropical China Using UAV Imagery and Machine Learning Methods

Highlights

What are the main findings?

• The canopy coverage areas for mangrove trees and smooth cordgrass (Spartina alterniflora) were 115.73 ha and 52.96 ha in Zhejiang Province.
• Mangrove tree canopy occupancy was 36.41%, while the invasion rate of smooth cordgrass was 13.70%.

What are the implications of the main findings?

• Smooth cordgrass invasion rates in some districts were higher than 67.3%, suggesting that active control and replanting of mangrove trees are needed.
• Smooth cordgrass invasion can be suppressed when mangrove canopy coverage is greater than 40%.

Abstract

The survival and growth of mangroves along coastal China is threatened by invasive smooth cordgrass (Spartina alterniflora). Due to the high mortality and frequent replanting of mangrove trees and the impacts of invasive smooth cordgrass, the exact mangrove forest area in Zhejiang Province, China, is still unclear. Based on provincial-scale fine-resolution Unmanned Aerial Vehicle (UAV) imagery and a large number of field survey plots, this study mapped the distribution of mangroves and smooth cordgrass in 2023 using three machine learning classifiers, including Classification and Regression Tree (CART), Convolutional Neural Networks (CNNs), and Support Vector Machine (SVM). The accuracy assessment indicated that the CNN algorithm was superior to the other two algorithms and yielded an overall accuracy and Kappa coefficient of 97% and 0.96, respectively. The total areas of mangrove forest and smooth cordgrass were 140.83 ha and 52.95 ha, respectively, in 2023 in Zhejiang Province. The mangrove forest area was mostly concentrated in Yuhuan, Dongtou, Yueqing, and Longgang districts. The mean canopy coverage of mangrove trees was only 36.41%, with lower than 20% coverage in all northern and some central districts. At the spatial scale, the mangrove trees showed a scattered distribution pattern, and over 70.04% of the planting area had canopy coverage lower than 20%. Smooth cordgrass has widely invaded all 11 districts, accounting for about 13.7% of the total planting area of mangrove trees. Over 67.3% and 85.4% of the planting areas have been occupied by smooth cordgrass in Wenling and Jiaoxiang districts, respectively, which necessitates an intensive anthropogenic intervention to control its spread in these districts. Our study provides more accurate monitoring of the mangrove and smooth cordgrass distribution areas at a provincial scale. The findings will help guide the replanting and management activities of mangrove trees, control planning for smooth cordgrass, and provide a data basis for the accurate estimation of carbon stock for mangrove forests in Zhejiang Province.

1. Introduction

Mangrove ecosystems are tidal, flat, wetland, woody plant communities comprising mainly tree and shrub plants in tropical and subtropical regions [1]. Mangrove forests play a pivotal role in the coastal social, ecological, economic, and other ecosystem services [2,3]. There were 147,359 km² of mangrove forests distributed worldwide in 2020, with an area loss of 3.4% during 1996–2020 [3]. During 2000–2016, over 62% of global mangrove forests were lost, primarily due to direct and indirect human impacts [4]. Similarly, approximately 60% of China’s mangrove forests were lost between 1973 and 2000 [5,6], while a 12% area loss was found during 1996–2020. The continued decline of mangrove forests is generally caused by conversion to agriculture, aquaculture, tourism, and urban development and overexploitation [1]. However, invasive species can also threaten the persistence of mangrove forests [7]. Smooth cordgrass (Spartina alterniflora) is one of the most successful invasive species in mangrove wetlands and has dispersed coast-wide to all mangrove sites in China [8], underscoring the need for the integration of an invasive plant management strategy into mangrove forest management [9]. Smooth cordgrass was first introduced to China in 1979. Since then, this species has proliferated along the coast of China, spreading from the southern tropical (i.e., Hainan Province) to the northeastern temperate (i.e., Liaoning Province) coastlines [10]. In 2003, smooth cordgrass was recognized as an invasive species in China by the Ministry of Ecology and Environment [11]. Characterized by high adaptability and rapid reproduction, smooth cordgrass can proliferate over large areas after invasion, thereby encroaching on the habitats of mangroves and affecting ecosystem stability [12,13]. However, the role of the invasion of smooth cordgrass in the degradation of mangrove forests has received little attention [9,14].

Due to the advantages of more extensive spatial coverage, lower cost, and various scales and resolutions, remote sensing tools have been widely applied to classify the smooth cordgrass and mangrove forests along the coastlines in the past few decades [7,14,15,16]. Various remote sensing tools, including spaceborne, airborne, and UAV (drone) platforms, and sensors, including multispectral, hyperspectral, and radar sensors, have been individually applied or combined to derive spectral information for mangroves and smooth cordgrass [15,16,17,18]. Spaceborne platforms such as Landsat, SPOT, Sentinel, ALOS, WorldView, IKONOS, and Gaofen have all been reported to effectively capture coastal land cover information. For example, based on Landsat imagery, Wu et al. [19] monitored the dynamic changes in smooth cordgrass and mangrove area and landscape patterns from 2005 to 2021 in the Zhangjiang Estuary, Zhejiang Province. Based on Landsat imagery and machine learning methods, Jia et al. [5] and Mao et al. [11] detected mangrove distribution and smooth cordgrass invasion, respectively, in China during 1990–2015. Based on Gaofen-1 multispectral imagery, Li et al. [20] monitored and assessed the spread of smooth cordgrass in the mangrove forests of the Shankou Mangrove Reserve, Guangxi Province. Liu et al. [21] applied multiple sources of high-resolution satellite images to monitor the invasion of smooth cordgrass into mangrove forests. Combining the ALOS Phased-Array L-band Synthetic Aperture Radar (PALSAR) with Landsat imagery, Bunting et al. [3] monitored the global mangrove extent and temporal variations during 1996–2020. These satellite platforms and sensors were successfully applied and achieved high accuracy metrics in monitoring the spatiotemporal change patterns of smooth cordgrass and mangrove forests at regional, national, and global scales; however, due to the relatively coarse resolution, these satellite platforms were not accurate enough to detect dynamic changes in scattered and mixed distribution patterns and shrub-like closed canopy mangrove trees. For example, the detected area of mangrove forests in Zhejiang varied from 6.12 ha to 386.77 ha based on classifications from different satellite platforms, such as Landsat, SPOT, Sentinel-2, ALOS, WorldView, and Gaofen-1 [3,5,22,23,24], implying significant uncertainty in the estimated mangrove forest area of Zhejiang Province, mainly due to the compounding effects of mixed pixel issues [22,24]. In recent years, a few studies have begun to apply finer-scale Unmanned Aerial Vehicle (UAV) imagery to detect the invasion of smooth cordgrass into mangrove forests [13,14,17,25]. For example, Kan et al. [14] applied UAV images and machine learning methods to detect the coverage of both smooth cordgrass and mangrove trees in an estuary of Fujian Province and achieved excellent performance, suggesting that the UAV method can be more accurately applied, especially at the plot and ecosystem scales. Based on UAV multispectral and hyperspectral imagery, Yang et al. [25] classified mangrove species and other vegetation types in Yingluo Bay of Guangxi Province. Combining WorldView multispectral and UAV hyperspectral imagery, Jiang et al. [26] classified mangrove and smooth cordgrass distribution in Qinglan Harbor of Hainan Province. By synthesizing over two thousand papers, Xu et al. [17] reviewed drone technology applications in mangrove mapping. However, most of the UAV-based studies have focused on a small area, and few studies have been conducted to detect the coverage of both mangrove trees and smooth cordgrass at the provincial scale.

The classification methods for mangrove and smooth cordgrass have also evolved from regular supervised and unsupervised classifications to machine and deep learning methods [16,17,18]. Many machine learning methods have been applied to map mangrove forests [16,17], including Maximum Likelihood Classifier (MLC), Support Vector Machine (SVM), Random Forest (RF), XGBoost, AdaBoost, Convolutional Neural Network (CNN), Classification and Regression Tree Algorithm (CART), of which SVM was the earliest and most commonly used method [16], and ensemble learning methods such as CNN have been proven to outperform RF and SVM [16,27,28,29]. Except in remote sensing platforms and classification methods, input features, including bands, vegetation indices (VIs), and texture patterns, are also vital for classification precision [30]. Based on Sentinel-1 and -2 and ALOS SAR imagery, Shen et al. [30] applied feature selection methods and four machine learning methods to classify mangrove distribution in Zhanjiang City, Guangdong Province. They concluded that combining satellite imagery with optimally selected features and the XGBoost classifier can more accurately map mangrove distribution. Bands were found to generally be more important than VIs for classification. Similarly, based on UAV multispectral and hyperspectral imagery, Yang et al. [25] combined feature selection and four machine learning methods to classify mangroves in Yingluo Bay, Guangxi Province. They concluded that this combination can significantly improve classification accuracy, especially for hyperspectral images. VIs were found to be more important than bands, and texture information was the least important. Finally, to remove the salt-and-pepper effect, object-oriented classification methods offer improvements in both accuracy and stability, and image segmentation has gradually shifted from pixel-based to object-oriented classification methods [16,18,31]. Multiscale segmentation, feature selection, and machine learning methods could be combined to increase classification accuracy for mangroves based on fine-resolution (less than 1 m) UAV imagery [17].

Zhejiang Province is the northernmost boundary suitable for mangrove forests in China [23]. Due to the limitation of air temperature, the main mangrove species, Kandelia obovata, was artificially cultivated and planted in this province [32,33]. This tree species can grow up to 6 m in the tropical region of China, but in Zhejiang Province, the tree height of mature K. obovata forests has declined from about 3 m on the southern coast to lower than 1 m on the northern coast [34]. The suitable habitats of mangrove trees overlap with smooth cordgrass, and most mangrove trees are planted after smooth cordgrass is mechanically removed [19,20]. Due to the lower height, slow growth rates, and lower competitiveness of K. obovata, smooth cordgrass can easily reinvade these mangrove forests. The invasion of smooth cordgrass can either cause the death of mangrove trees or poor growth due to resource competition or overshadowing. In addition, the high density and extensive root system of smooth cordgrass create a “green desert” in its distribution areas, which poses a significant threat to intertidal biodiversity [8,35]. The local government is striving to recover these mangrove forests and remove the invasive smooth cordgrass. The Special Action Plan for Mangrove Conservation and Restoration (2020–2025) was issued in 2020, and under this plan, Zhejiang Province plans to afforest 690 ha of mangrove forests, accounting for 182% of the current area (385 ha) [33]. To help identify a suitable planting area and mitigate the impacts of smooth cordgrass, accurate monitoring of the present mangrove and smooth cordgrass distribution areas and landscape patterns is urgently needed.

Based on single-date provincial-level UAV imagery, our study aimed to detect the coverage of both mangrove trees and smooth cordgrass and assess the invasion status of smooth cordgrass in mangrove forests. More specifically, our study sought to address three research questions: (1) Which classification method is the most suitable for the detection of both mangrove trees and smooth cordgrass based on UAV imagery? (2) How many mangrove trees are preserved, and where are they located? (3) What is the invasive status of smooth cordgrass and the implications for mangrove management in Zhejiang Province? Our findings will provide accurate and finer-scale mapping of mangroves and smooth cordgrass for Zhejiang Province and further help guide the conservation of and planting planning for mangrove trees and the scientific control of smooth cordgrass invasion in this province.

2. Materials and Methods

2.1. Study Area

The study area is located on the coastline of Zhejiang Province (27°03′–31°04′N, 119°37′–123°25′E) (Figure 1). The study region is characterized by a middle-to-northern subtropical maritime monsoon climate. The annual average temperature is 17.7 °C, the annual average rainfall is 1507 mm, and the frost-free period is 258 days. All mangrove forests in this province were artificially planted [32,36]. Based on the previous inventory and governmental statistical documents, only three cities have planted mangrove trees, namely, Zhoushan, Taizhou, and Wenzhou City. The majority (>95%) of the existing mangrove species is K. obovata in Zhejiang Province [34]. Other mangrove species, including Myoporum bontioides, Aegiceras corniculatum, and Ceriops tagal, were also cultivated and planted but proved to be unsuccessful. At present, only about 6 ha of M. bontioides has survived on Maoyan Island, Yuhuan district [32,36]. K. obovata is an evergreen mangrove tree species belonging to the Rhizophoraceae family and the Kandelia genus. This species was first introduced from Fujian Province, China, in 1957 and was cultivated using the tissue culture method to make it suitable for survival in Zhejiang Province [37]. These mangrove forests are often threatened by extreme low temperatures, high-level sea salt concentrations, and long-term submergence caused by water in this region [36]. Field experiments indicated that the lower limit temperature for mangrove forests is −5 °C, so the most northern distribution area is Zhoushan City [33,37]. Except for the mudflat area at the leeward bay, all distribution areas of smooth cordgrass are also suitable for K. obovata [21].

Smooth cordgrass was first introduced to Zhejiang in 1983 and expanded thereafter to 1560 ha [12]. This grass species was planted to protect seashores from damage from strong winds and waves. However, this species can also cause serious ecological consequences, including altered traditional landscapes, threatened biodiversity, and degraded ecosystem functioning [11]. This invasive species poses a serious threat to mangroves through stronger competition due to its ecological niche overlap with mangrove species, especially in the northern limits of mangrove distribution in Zhejiang Province, where the mangrove trees grow more slowly than smooth cordgrass.

Because the mangrove area is relatively small and widely scattered, we divided the distribution areas into 11 districts for analysis and mapping convenience, namely, Putuo, Sanmen, Jiaojiang, Wenling, Yueqing, Yuhuan, Dongtou, Rui’an, Pingyang, Longgang, and Cangnan districts (Figure 1).

2.2. The Workflow

The workflow of our study method consisted of various phases, as shown in Figure 2. More detailed descriptions of the workflow are shown below.

2.3. UAV Image Collection

The multispectral airborne images were collected during the low tide period from 8 to 15 August 2023, using a DJI MAVIC 3 PRO UAV. The regions of interest for flights were first determined based on the inventory planting area boundary (Figure 1). The flight route was designed and planned by DJI Fly, with an 80% flight overlap and 75% lateral overlap. The exposure time was 1/640 s, the ISO speed was 100, and the flight altitude was 100 m. The UAV is equipped with a multispectral camera, a global navigation satellite system (GNSS), and a real-time kinematic (RTK) instrument. The multispectral sensor has five bands: green (G), red (R), blue (B), red edge (RE), and near-infrared (NIR). The spatial resolution for the RGB bands (NIR and RE) is 4 cm, and for the thermal bands, it is 6 cm. The UAV is equipped with a light intensity sensor that is capable of automatically adapting to ambient light conditions, thereby facilitating the acquisition of enhanced NDVI data during reconstruction. The weather conditions during the data collection were very suitable for flight, with low wind velocity. The UAV images for different levels of invasion are shown in Figure 3.

2.4. Ground-Truth Survey and Sampling Data Collection

A full field inventory was conducted from 23 to 29 August 2023 along the coastal line of Zhejiang Province from the south (Wenzhou City) to the north (Zhousan City). The plot coordinates and sampling point locations were recorded using the survey-grade GPS receivers of Huace RTK (Everest Edition), with 8 mm horizontal accuracy and 15 mm elevation accuracy. Due to the difficulty of separating individual mangrove tree canopies and lower heights, we set up each 5 × 5 m plot to represent the average plot-scale conditions. For mangrove trees, the investigated variables include average and maximum tree height, mean canopy coverage (percent of land area), mean basal diameter, and tree numbers within each plot. For smooth cordgrass, the mean height and coverage within a plot were investigated. The fraction of canopy coverage occupied by mangrove trees and smooth cordgrass in each plot was then manually calculated. In addition, to increase sampling size, we also conducted a visual interpretation to calculate the canopy coverage for all land cover types. In total, 1534 sampling plots in 11 survey sites and randomly selected visual interpretation sites were collected, including 251 plots for mudflats, 149 plots for water bodies, 139 plots for built-up land, 479 plots for mangrove trees, 362 plots for smooth cordgrass, and 154 plots for others (Figure 1). The “others” category represents a sheltered mudflat with grass residue. Smooth cordgrass is removed annually for newly planted mangrove trees to reduce its impacts on mangrove trees in some planting areas, and the residue is discarded locally on the mudflat. The spectral signal of the grass residue is significantly different from that of the mudflat, so we separated it as a single category. This category was combined into the mudflat category after classification. Among all sample plots, 530 samples were randomly selected to evaluate the classification results, and 1004 samples were selected to train the classifiers.

2.5. Image Processing

Image preprocessing was conducted using the DJI Terra software (DJI, Shenzhen, China), which provides comprehensive image preprocessing steps for DJI UAVs, facilitating the acquisition of data. The flight route was processed and stitched to generate an orthophoto map. The atmospheric correction, radiometric calibration, and reflectance correction were performed to remove the distortion. Finally, the UAV images were cropped based on the reported boundary of mangrove planting area.

Previous research has proven that texture features, vegetation indices, and spectral bands can be applied to differentiate mangrove trees and other vegetation types [5,15,16,38]. Therefore, we selected the reflectance values of four bands (R, G, NIR, and red edge), four vegetation indices (Normalized Difference Vegetation Index, NDVI; Normalized Difference Red Edge Vegetation Index, NDRE; Green NDVI, GNDVI; and Leaf Area Index, LAI), and texture variables (the mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment, and correlation) (

Table 1. The extracted features for classification.

Object Features	Description
Spectral Bands	Red (R); Green (G); Red Edge (RE); Near-Infrared (NIR)
Vegetation Indices	Normalized Difference Red Edge Vegetation Index (NDRE), Normalized Difference Green Vegetation Index (GNDVI), Leaf Area Vegetation Index (LAI), Normalized Difference Vegetation Index (NDVI)
Textural Variables	Mean, Variance, Homogeneity, Contrast, Dissimilarity, Entropy, Second Moment, Correlation

). The eight texture variables were calculated for each band using the gray-level co-occurrence matrix (GLCM). The chosen moving window was 7 × 7 pixels.

The Boruta algorithm was employed in the screening of the key and most important classification features. The Boruta algorithm is capable of identifying all quantities that exert a significant influence on the classification outcome [39]. The fundamental premise is to statistically ascertain the significance of actual variables and those that have been randomly introduced. Furthermore, it is capable of considering the relationship between multiple variables. The importance of all 33 input variables for the machine learning classification of land cover types was ranked. The 10 variables with the highest classification capabilities were identified based on the feature importance index. As shown in Figure 4, the five most important features were G_correlation, G_mean, R_correlation, B_mean, and GNDVI, indicating that the texture features are more important than vegetation indices and spectral bands, especially in differentiating mangrove trees and smooth cordgrass.

The land cover types were subdivided into five primary categories, namely built-up land (BU), mangroves (MA), mudflats (MU), smooth cordgrass (SC), and waterbodies (WA).

2.6. Multiscale Segmentation

Due to the high spatial resolution and different sizes of varied objects, multiscale segmentation rather than the pixel-based classification method was applied to remove “salt-and-pepper noise” and increase the feature extraction accuracy [40]. Multiscale segmentation is a step in object-oriented classification that divides pixels into discrete units, thereby facilitating the formation of objects comprising pixels with similar characteristics at a given scale. The FbSP (Fuzzy-based Segmentation Parameter) optimizer in the eCognition 10.3 software was selected to determine the optimal segmentation parameters, including scale, shape/color, and smoothness/compactness. The scale parameter represents the size of the segmentation (the maximum heterogeneity degree of objects), while the shape parameter represents the weight assigned to the segmentation criterion. An elevated compactness index indicates a more compact image [41].

The shape index, compactness, and scale parameters were adjusted to perform a series of tests. The ranges for shape index and compactness parameters were set from 0.1 to 0.9, while the scale parameters were set from 10 to 100. The results showed that when the scale is set to 70, neighboring objects with similar features are often misclassified. When the scale was set to 50 or 30, the segmentation results were too fragile, which affected the efficiency of subsequent image processing (Figure 5a). When the shape index was set as 0.1, 0.3, and 0.7, adjacent objects with some shape similarities were grouped into one category. This included cases such as mangroves and smooth cordgrass (Figure 5b). Visual inspection of the segmentation results showed that when the compactness was 0.5, the objects of each land cover type were more compact, and the overall segmentation result was visually most satisfactory (Figure 5c). Accordingly, when evaluating the clustering characteristics of the distribution of mangrove communities in the study area, the optimal scale, shape, and compactness parameters were determined as 70, 0.3, and 0.5, respectively, based on multiple interactive segmentation trials.

2.7. Classification Methods

Based on the review of the previous literature for mangrove and smooth cordgrass mapping, we selected three machine learning classification methods to detect land cover types in the study area, which is based on optimal segmentation parameters and the selected features. The most commonly used algorithms—Support Vector Machine (SVM), CART decision tree (DT), and Convolutional Neural Network (CNN)—were chosen to evaluate their performance. These methods have all been previously applied to detect mangroves and smooth cordgrass [14,31,42].

SVM is an advanced supervised learning algorithm that can be used as a non-parametric classifier for land cover classification based on multispectral imagery, including the classification of mangrove species. SVM is based on statistical learning theory and aims to identify the optimal decision hyperplane in high-dimensional space, thereby achieving the best possible category separation. In the context of uncertain classification problems with high-dimensional features, SVM has demonstrated consistent performance, even with a limited number of training samples [31].

The decision tree algorithm partitions data into numerous subsets and represents decision rules and classification outcomes in a tree-like data structure, frequently a multi-node tree [43]. CART is a decision tree regression model that can be employed for both classification and regression analysis. CART is mainly applied to land cover mapping, disaster monitoring, and crop identification. This algorithm offers several advantages: the rules are transparent and easy to interpret, the tree can be adjusted effortlessly afterward, and it supports refined multi-class discrimination. Decision trees are split into classification trees and regression trees on the basis of the predictor variables: when the response is discrete, the tree performs classification; when the response is continuous, it performs regression. Starting from the root, samples are routed through successive nodes until a final class label or value is returned.

Convolutional Neural Network (CNN) represents a class of artificial neural networks that are designed to process information in a manner similar to that of the human brain [44]. A distinctive feature of CNNs is their network structure, comprising multiple hidden layers, each of which contains both a convolutional layer and a pooling layer. In this study, the pre-trained convolutional neural network takes 64 × 64-pixel image patches as the input. The architecture consists of 8 image-processing stages and 4 hidden layers whose neuron configurations are 9, 7, 5, and 3 units. For feature extraction, the model applies 12 feature maps to learn multi-level representations of the data. This hierarchical design captures spatially local patterns and, through successive abstraction, extracts high-level semantic features.

To test the best classification accuracy, we further designed two classification experiments: Experiment A: Only multispectral bands are used as input features; Experiment B: Both multispectral bands and the 10 identified most important features are used. Due to the large UAV image size and huge workload for the entire study region, we only selected four mangrove and smooth cordgrass distribution sites with intensive training and validation plots to test the effectiveness of these classification methods. The test regions include Ximen Island (Within Wenling), Wugen Town (Wenling), Taipingtang (Yuhuan), and Nanpu (Yuhuan) (Figure 1).

2.8. Accuracy Assessment Metrics

To assess the precision of classification methods, confusion matrices were constructed based on training sample plots. The overall accuracy (OA), kappa coefficient, user accuracy (UA), and producer accuracy (PA) were calculated based on the confusion matrices and applied to evaluate the performance of the classified land cover types. The calculation equations for OA, UA, PA, and kappa are shown below:

$$U{A}_i={\displaystyle \frac{C_{ii}}{{\sum }_{j=1}^k{C}_{ji}}}$$

(1)

$$P{A}_i={\displaystyle \frac{C_{ii}}{{\sum }_{j=1}^k{C}_{ij}}}$$

(2)

$$OA={\displaystyle \frac{{\sum }_{i=1}^{k}T{P}_i}{N}}$$

(3)

$$\mathit{Kappa}={\displaystyle \frac{N{\sum }_{i=1}^{k}T{P}_i-\sum _{i=1}^k\left(R_i\times C_i\right)}{N^2-{\sum }_{i=1}^k\left(R_i\times C_i\right)}}$$

(4)

where $T{P}_i$ is the number of samples correctly classified as class i, k is the total number of classes, N is the total number of samples, and P_i is the number of true samples of class i; C_i is the number of predicted samples of class i; C_ii denotes the number of samples correctly assigned to class i in the confusion matrix (the diagonal element for class i); C_ji is the number of samples for predicted class i; and C_ij is the number of samples for true class i.

2.9. Analysis Methods

The classified mangroves and smooth cordgrass were isolated from the land cover maps. Due to the scattered distribution patterns and large size of the UAV images and classified results, we divided 11 districts and three cities (Taizhou, Wenzhou, and Zhoushan; Figure 1) for mapping and statistical analysis. The direct classified results were for canopy coverage of all land cover types (canopy area), and then, they were aggregated to 30 m spatial resolution to calculate the area of each land cover type. Due to the lower height of mangrove trees in Zhejiang Province, the mangrove forest area was calculated as the area with canopy coverage equal to or greater than 20% [34]. The occupancy rate of mangrove trees was calculated as the ratio of the existing mangrove area to the total planting area. The planting area and boundary of the mangroves were obtained from [34]. The invasion rate of smooth cordgrass was calculated as the ratio of the existing smooth cordgrass area to the planting area of the mangrove trees. The area of mangroves from previous studies was calculated after clipping Zhejiang Province out of the national or global datasets of mangrove distribution.

3. Results

3.1. Accuracy Assessment Results

At the test regions, five land cover categories were classified, namely, waterbodies (WA), mangroves (MA), mudflats (MU), smooth cordgrass (SC), and built-up land (BU). Confusion matrices were constructed to calculate and compare the accuracy assessment metrics. The results indicated that the CNN method had the highest kappa coefficients of 0.96 and 0.97 and the highest overall accuracy (0.97 and 0.98) based on the all-feature (Experiment A) and selected-feature (B) experiments, respectively, which are significantly higher than that of the CART and SVM methods for both metrics under both experiments (

Table 2. Classification accuracy assessment results in test regions.

Methods	Accuracy Metrics	Experiment A (All Features)	Experiment B (Selected Features)
SVM	Overall Accuracy	75%	83%
	Kappa	0.63	0.76
CART	Overall Accuracy	73%	91%
	Kappa	0.62	0.88
CNN	Overall Accuracy	97%	98%
	Kappa	0.96	0.97

). The selected feature experiment only slightly improved the classification accuracy using the CNN method; however, it can significantly increase the overall accuracy and kappa coefficient under the CART and SVM methods. Finally, the CNN classification method was selected for the land cover classification in the study region. The classified distributions of the five main land cover types are shown in Figure 6. Compared with the CART and SVM methods, the CNN method can more accurately detect waterbodies and reduce misinterpretations between mangrove trees and smooth cordgrass. Based on the CNN method, the land cover types in the entire Zhejiang area were classified, and the classification results were further evaluated using 530 sampling plots. The results indicated that the overall accuracy and kappa coefficient were still very high, reaching 97% and 0.96, respectively (

Table 3. Classification accuracy assessment results for the entire study region using CNN classification methods. Note: UA: user accuracy; PA: producer accuracy; OA: overall accuracy; BU: built-up land; MA: mangroves; MU: mudflats; SC: smooth cordgrass; WA: waterbodies.

Types	SC	MA	BU	WA	MU	Samples	UA
SC	119	4	0	0	0	123	97.86%
MA	2	170	0	0	1	173	97.92%
BU	0	0	47	1	1	49	98.34%
WA	1	0	1	48	3	54	94.11%
MU	1	0	1	2	127	131	96.96%
Samples	123	174	49	51	132	530
PA	97.15%	98.55%	96.73%	93.74%	97.55%
OA	97.34%
Kappa	0.96

).

3.2. Distribution of Mangroves and Smooth Cordgrass

Due to the higher resolution of the UAV imagery, our classified results for mangroves and smooth cordgrass are actually the canopy cover area. The canopy cover area of mangroves in Zhejiang Province was 115.73 ha in 2023, while the area of smooth cordgrass that invaded mangrove plantations was 52.95 ha (Figure 7). The total area of mangrove forests was calculated as 140.83 ha in terms of the shrub-like forest area definition (canopy coverage >= 20%) in China. At the 30 m pixel scale, over 70.04% of the pixels with mangrove tree distributions had canopy coverage lower than 20%, indicating that the mangrove trees were mostly distributed in a scattered manner in the planting area (Figure 8). The lower canopy density mostly appeared in northern Zhejiang, such as Putuo, Sanmen, Jiaojiang, Wenling, and Yueqing districts. About 11.91% of the pixels had coverage greater than 50%. The highest coverage (80–100%) only accounted for about 3.35% of the mangrove distribution area and mostly spreads into southern Zhejiang, such as Yuhuan, Dongtou, Pingyang, Cangnan, and Longgang districts. In contrast, the highest canopy coverage (80–100%) of smooth cordgrass mostly appeared in central and northern Zhejiang, such as Jiaojiang and Wenling districts, with continuous and smaller canopy coverage in southern Zhejiang, such as Dongtou, Rui’an, and Pingyang (Figure 9). Before the planting of mangrove trees, smooth cordgrass is first mechanically removed. This resulted in its sporadic spreading pattern across the areas with mangrove tree distribution. Only 7.03% of pixels with smooth cordgrass had canopy coverage greater than 50%, and about 81.80% of the smooth cordgrass had canopy coverage lower than 20% at the 30 m spatial resolution. By comparison, the smooth cordgrass distribution was more dispersed than that of the mangrove trees.

At the regional scale, the largest mangrove forest area was located in Dongtou district (40.92 ha), followed by Yuhuan (30.50 ha) and Yueqing (23.93 ha) districts (Figure 7). They accounted for 46%, 18.5%, and 17.5% of the total mangrove area, respectively (Figure 10). Within the mangrove planting areas, the largest smooth cordgrass areas are located in Yuhuan (24.35 ha), Dongtou (9.82 ha), and Yueqing (9.26 ha) districts (Figure 7). They accounted for about 29.1, 21.7%, and 17.0% of the total smooth cordgrass area (Figure 10). These three regions had both the highest mangrove and smooth cordgrass areas due to the habitat overlaps between both vegetation types.

3.3. The Occupancy of Mangrove and Invasion Rates of Smooth Cordgrass

Due to the limitation of low air temperature, strong tides, seawater submergence, the invasion of smooth cordgrass, and human activities, most of the planted mangrove trees died in the study region. The existing mangrove trees were actually replanted several times after the first planting date. Even after several instances of replanting, the occupied mangrove canopy within the planting area was still very low (Figure 11). The mangrove canopy only occupied 36.41% of the total planting area (386.77 ha). In other words, 63.59% of the planting area was occupied by other land cover types. Among the districts, the highest occupancy rates were found in Rui’an (72.46%) and Pingyang (72.79%), while the lowest occupancy rates were found in Jiaojiang (0.63%) and Sanmen (3.91%) districts. The occupancy rates in other districts ranged from 7.15% to 54.52% and were mostly lower than 50%. The occupancy rates generally showed a declining pattern from the south to the north of Zhejiang Province. Replanting and, thereafter, intensive human management are definitely needed for most districts to reduce the limiting factors on mangrove tree survival and growth.

Most mangrove trees were planted in the area after smooth cordgrass was removed. Due to its faster growth rate, taller height, and stronger competitiveness than mangrove trees, smooth cordgrass soon reinvaded and became one of the most significant factors influencing mangrove tree growth and canopy occupancy. Our study found that about 13.7% of the mangrove planting area was occupied by smooth cordgrass. The invasion rates were extremely high in Jiaojiang (85.44%) and Wenling (67.26%) districts, indicating that most of the planting area of mangrove trees was occupied by smooth cordgrass (Figure 11). The invasion rates of smooth cordgrass in Yueqing (19.08%) and Yuhuan (20.90%) districts were also relatively high, while the invasion rates in other districts ranged from 0.88% (Cangnan) to 9.26% (Dongtou). The area of smooth cordgrass has surpassed the mangrove area in Jiaojiang and Wenling and was close to the mangrove area in Yuhuan. The ratio of smooth cordgrass area accounted for 99.26% and 90.39% of the total smooth cordgrass and mangrove areas in Jiaojiang and Wenling, respectively. In these three districts, the invasion of smooth cordgrass could be the main cause of mangrove tree mortality or low growth rate.

The different levels of mangrove occupancy rates and smooth cordgrass invasion rates are also illustrated in Figure 12. Generally, two types of invasion patterns of smooth cordgrass were observed. The first type is spreading in the open gaps after the mangrove trees die, and the second type is spreading from the periphery to the central distribution area of mangrove trees. The first type mainly occurs in southern Zhejiang, while the second type is more common in the farther parts of northern Zhejiang. This resulted in the highest coverage of smooth cordgrass in some northern districts.

4. Discussion

4.1. The Mangrove Area in Zhejiang Province

Our study estimated that the mangrove canopy cover area was 115.73 ha, and the mangrove forest area was 140.83 ha in 2023 in Zhejiang Province. Many previous studies have also estimated the mangrove area in Zhejiang Province, showing a range from 6.12 ha to 386.77 ha during 2015–2021 (

Table 4. Comparison of our estimated mangrove forest area with previous studies of Zhejiang Province.

Platforms	Time	Area (ha)	Resolution	References
ALOS SAR imagery	2020	47	23.5 m	[3]
Landsat and Sentinel-1	2015	8.0	30 m	[23]
Landsat	2015	56	30 m	[5]
Landsat	2015	6.12	30 m	[45]
Landsat	2015	55	30 m	[46]
Gaofen-1 and Ziyuan-3	2018	48.68	1 m & 2 m	[22]
Sentinel-1 and -2	2019	39	10 m & 20 m	[24]
Gaofen-2	2020	19.8	1 m & 4 m	[38]
Statistical and inventory	2020	386.77	/	[34]
Sentinel-2	2021	115.5	10 m & 20 m	[47]
UAV multispectral imagery	2023	140.83	4 cm & 6 cm	This study

). For example, based on ALOS SAR imagery, Global Mangrove Watch (GMW) estimated a mangrove forest area of 47 ha in 2020 in Zhejiang Province [3]; based on Landsat imagery, Hu et al. [45] estimated a mangrove forest area of 6.12 ha in 2015; based on finer-scale Gaofen-1 and Ziyuan-3 imagery, Zhang et al. [22] estimated a mangrove area of 48.68 ha in 2018; based on the statistical documents, Wu et al. [34] reported the largest area of 386.77 ha, mainly because they identified the planting area of mangrove trees rather than the canopy cover area; based on Sentinel-2 imagery and a hybrid neural network classification method, Ye and Weng [38] estimated a mangrove forest area of 115.5 ha in Zhejiang Province, which is close to our estimate. This large discrepancy is mainly due to the difference in reporting times, remote sensing tools, and classification methods [16]. Zhao and Qin [24] and Zhang et al. [22] have also explained the large gaps among different studies for mangrove area and attributed them to the compounding effects and mixed pixel effects of other coastal vegetation types, especially smooth cordgrass. Mangrove trees are relatively short and often mixed with smooth cordgrass [43]. Many remote sensing platforms have been applied and achieved high precision in classifying mangrove distribution; however, it should be noted that the mixed pixel problem could be more severe for coarser-resolution images [14,16,17,22]. The classification method is another important factor influencing classification accuracy. Our study found that the CNN algorithm was superior to the SVM and CART algorithms in detecting mangrove and smooth cordgrass. Previous studies with algorithm comparisons also suggested that ensemble learning methods, including CNN, outperformed the RF and SVM algorithms, especially when applied to classifications of UAV-based images [16,27,28]. In addition, the observation time period also significantly affects the estimated mangrove area because the planting area in Zhejiang Province has increased rapidly, especially in the most recent decade [33,34]. Based on a provincial-level inventory in 2020, Wu et al. [34] concluded that 52.06% of mangrove forests were younger than 3 years old in Zhejiang Province. In a government report (http://www.cenews.com.cn/news.html?aid=1185132, accessed on 20 December 2025), the present planting area of mangrove forests was 477.3 ha, increasing by 90.53 ha from 2020 to 2025. By synthesizing all existing study results along the timeline, our estimation of mangrove area (140.83 ha) should be reasonable and represent the real mangrove forest area in this province.

Based on high-resolution UAV multispectral imagery, the identified best classification algorithm, and extensive field inventory data, our approach can more effectively differentiate mangrove trees from smooth cordgrass and provide more robust and updated mangrove area data for Zhejiang Province.

4.2. The Invasion Patterns of Smooth Cordgrass

The invasion of smooth cordgrass into mangrove forests has been reported worldwide [9]. Smooth cordgrass expansion can directly affect mariculture activities and shorebirds that forage on mudflats [33,48]. In addition, its high density, strong root system, and tall height can significantly affect the growth rate of mangrove trees and result in high mortality [35]. Most mangrove trees in Zhejiang Province were planted on smooth cordgrass habitats after removing all smooth cordgrass because these areas have also proven to be suitable for the mangrove tree species K. obovata [33,37]. In addition, the government was encouraged to control its spread. Our study indicated that 52.95 ha of smooth cordgrass exists in the mangrove planting area at present, accounting for about 13.7% of the total planting area. This was under conditions of frequent mechanical and chemical controls within the first several years after mangrove trees were planted [34]. However, control measures cannot guarantee the prevention of the expansion of smooth cordgrass [33]. Based on Landsat imagery and SVM classification methods, Mao et al. [11] indicated that the smooth cordgrass area in Zhejiang Province was continuously increasing during 1990–2015, converted partially from mangroves. We found that smooth cordgrass invasion was more severe in central Zhejiang Province than in the northern region, where the mangrove trees are shorter and smaller. The densest smooth cordgrass coverage (80–100%) was mainly found in the central area (Figure 9), and Jiaojiang, Wenling, and Yuhuan districts have the highest coverage (greater than 20%) (Figure 11). This can be explained by the lower temperature in the north, which is not suitable for smooth cordgrass invasion, whereas the higher temperature and small size of mangrove trees in the central region cannot prevent its sprouting and invasion [21,46]. Specifically, the most suitable habitats are in southern Zhejiang, such as Longgang, Cangnan, and Pingyang districts, but the mangrove trees in this area have become dense forests with near-closed canopies. The denser coverage and height of mangrove trees prevent the invasion of smooth cordgrass [14]. Based on Landsat imagery and the CART algorithm, Lu [12] classified the distribution of smooth cordgrass and its biomass and concluded that the biomass of smooth cordgrass declined from central to south Zhejiang Province, which indirectly confirmed our study results.

4.3. Implications for Mangrove Management and Smooth Cordgrass Control

The mangrove forests along the coastlines have important functions in beach fixation, wave reduction, soil nutrient cycling, hydrological regulation, air purification, carbon sequestration, and biodiversity conservation [3,8,9]. To protect and restore mangrove forests, many countries and organizations have implemented protection plans, policies, or regulations. Such actions could support coastal communities, jobs, and food security and provide global climate mitigation benefits [8]. For example, the Global Mangrove Alliance (GMA) was established in 2021 to convene worldwide governments and organizations to protect mangrove resources. The Chinese government also released the “Special Action Plan for Mangrove Conservation and Restoration (2020–2025)” in 2020 [34]. Under this plan, Zhejiang Province needs to expand the mangrove forest area to 690 ha [33]; however, the existing mangrove forest area is only 140.83 ha, which is far from the targeted area. Therefore, to scientifically guide the protection and restoration efforts in Zhejiang Province, it is necessary to identify the exact distribution information and causes of area losses of mangrove forests. Our study provided a more accurate monitoring of the mangrove distribution area, which will help guide replanting and management planning activities and an accurate estimation of carbon stock for mangrove forests.

The occupancy rates of mangrove trees were extremely low in Zhejiang Province, especially in Jiaojiang, Wenling, and Yuhuan districts, with mangrove tree coverage mostly lower than 20%. Replanting of mangrove trees is definitely needed in these regions. In addition, the major causes of mortality should be identified to avoid damage to planted mangrove trees. Smooth cordgrass invasion is one of the most significant causes [32,33]. Our study clarified the locations and coverage of smooth cordgrass, which could help local decision makers and managers control the spread of smooth cordgrass in mangrove forests in Zhejiang Province. The canopy density of mangrove trees has been reported to significantly affect the invasion rates of smooth cordgrass [14,26]. Kan et al. [14] indicated that 41–64% canopy coverage of mangrove trees could significantly suppress the invasion of smooth cordgrass, and >65% coverage could completely suppress it in Fujian Province. We also observed a large decline in smooth cordgrass coverage when mangrove canopy coverage was greater than 40%. Therefore, in the districts with mangrove coverage over 40%, anthropogenic intervention is not necessary to conduct, such as Longgang (45.45%), Rui’an (62.80%), and Pingyang (68.36%) districts. In contrast, removal of smooth cordgrass is needed in the districts with extremely low mangrove canopy coverage and high invasion rates. For example, Jiaojiang, Wenling, and Yuhuan districts have mangrove coverage rates of 0.38%, 12.31%, and 20.83%, respectively, but invasion rates of 85.44%, 67.26%, and 20.90%, respectively. Anthropogenic control measures are urgently needed in these three districts to protect mangrove trees.

4.4. Limitations and Future Direction

Although our classifications using the UAV image and machine learning methods achieved a high precision, there are several limitations that may confound our results and conclusions. Due to the similar reflectance characteristics between mangrove trees and smooth cordgrass, the tall height of smooth cordgrass, and the small size and shorter height of mangrove trees, some mangrove trees are sheltered by smooth cordgrass plants and cannot be recognized by our methods. UAV images in non-growing seasons are needed to assist in classifying these under-covered mangrove trees because smooth cordgrass is a deciduous species, and the mangrove tree is an evergreen species. In addition, the training and validation plots for the survey and visual interpretation were obtained at a 5 m × 5 m grid scale, which may not match well with the segmented polygons in terms of both area and shape, which will also bring classification errors. Furthermore, only one-period UAV imagery was used to analyze the distribution of mangrove trees and the invasion status of smooth cordgrass. Temporal variations cannot be addressed through our UAV approach; thus, we cannot analyze the effects of smooth cordgrass invasion on the mortality of mangrove trees. Classification errors still existed in discriminating smooth cordgrass from other grass species and mangrove trees due to similar spectral information. UAV-based hyperspectral imagery could be applied to solve this issue [16,17,26]. In the future, we will apply multi-temporal, UAV-based, multispectral, hyperspectral, and LiDAR images to monitor the dynamics of mangrove trees and smooth cordgrass and to estimate mangrove biomass and carbon sequestration in Zhejiang Province.

5. Conclusions

Based on provincial-scale multispectral UAV imagery, this study screened the most important features to differentiate mangrove trees from smooth cordgrass and identified the optimal segmentation-scale parameters based on object-based classification methods. Three machine learning methods were applied to classify land cover types in the test regions, and the CNN algorithm was proven to outperform the other two methods. The land cover types within the planting area of mangrove trees were then detected using the CNN method. Our study indicated that only 140.83 ha of mangrove forests was preserved in 2023 in Zhejiang Province. Through comparison with previous estimates, our estimate was proven to be sufficient, and the data can be adequately used to guide mangrove forest management. The canopy occupancy of mangrove trees was only 36.41% of the planting area. The occupancy in some planting districts was lower than 10%, implying an urgent need to replant mangrove trees to restore mangrove forest area. Smooth cordgrass has expanded to 52.96 ha, and its area has surpassed the mangroves in some districts in central and northern Zhejiang Province, suggesting that anthropogenic intervention is needed to remove or reduce this smooth cordgrass. Smooth cordgrass invasion can be suppressed when mangrove canopy coverage is greater than 40%, which means that anthropogenic intervention is urgently needed for those mangrove planting areas with low canopy coverage. Our study provides an accurate inventory of the mangrove forest area and the distribution of smooth cordgrass, which will help guide local agencies in protecting mangrove forests and removing or controlling the spread of smooth cordgrass. Our results will also provide a data basis for the future planning of mangrove planting and the estimation of carbon stock capacity and potential in mangrove forests.