A Fine-Scale Mangrove Map of China Derived from 2-Meter Resolution Satellite Observations and Field Data

: Mangrove forests are important ecosystems in the coastal intertidal zone, but China’s mangroves have experienced a large reduction in area from the 1950s, and the remaining mangrove forests are exhibiting increased fragmentation. A detailed mangrove dataset of China is crucial for mangrove ecosystem management and protection, but the fragmented mangrove patches are hardly mapped by medium resolution satellite imagery. To overcome these difﬁculties, we presented a ﬁne-scale mangrove map for 2018 using the 2-meter resolution Gaofen-1 and Ziyuan-3 satellite imagery together with ﬁeld data. We employed a hybrid method of object-based image analysis (OBIA), interpreter editing, and ﬁeld surveying for mangrove mapping. The ﬁeld survey route reached 9500 km, and 2650 patches were veriﬁed during the ﬁeld work. Accuracy assessment by confusion matrix showed that the kappa coefﬁcient reached 0.98, indicating a highly thematic accuracy of the mangrove dataset. Results showed the total area of mangrove forest in China for 2018 was 25,683.88 hectares, and approximately 91% of mangroves were found in the three provinces of Guangdong, Guangxi, and Hainan. About 64% of mangroves were distributed in or near the nature reserves established by national or local governments, which indicated that China’s mangroves were well protected in recent years. The new ﬁne-scale mangrove dataset was freely shared together with this paper, and it can be used by local authorities and research groups for mangrove management and ecological planning.


Introduction
Mangrove forests are distributed in the tropical and subtropical areas of the world between approximately 30 • N and 30 • S latitude [1,2]. These forests consist of trees, shrubs, and palms that grow in harsh environmental settings including high salinity, high temperatures, extreme tides, high sedimentation, and muddy anaerobic soils [1]. Mangrove forests provide a wide range of ecosystem services such as shoreline stabilization, water purification, the provision of biological diversity and nursery habitats for marine fisheries, and they are important to the recreation and tourism industry [3][4][5][6].
The mangrove ecosystem is one of the most vulnerable ecosystems on Earth because of human disturbance and climate change. According to reports, 20-35% of global mangrove forest land has been lost due to deforestation since the 1980s [7,8]. The rapid loss of mangroves has forced forest managers and associated scientists to inventory and monitor their spatial extent. A number of efforts have been made for the extensive monitoring of mangrove forests from local to global scales [1,[9][10][11][12][13][14][15][16][17][18][19][20] using imagery including optical, hyperspectral, radar, etc. The most influential datasets at a global scale were produced for were (1) to generate a fine-scale benchmark inventory of mangroves in China (MC2018) [46] using Ziyuan-3 and Gaofen-1 satellite imagery; (2) to present a new hybrid method of object-based image analysis and field verification for mangrove mapping; (3) to compare the MC2018 dataset with existing mangrove datasets both quantitatively and visually; and (4) to discuss the advantages of 2 m images for mangrove mapping, the confusions between mangroves and other vegetation, and the limitations and caveats of the MC2018 dataset.

Study Area
The study area included the coastal mangrove ecosystems in southern China ( Figure  1), comprising the coasts of the Guangxi Zhuang Autonomous Region (referred to as Guangxi), Hong Kong, Macao, Taiwan, and the provinces of Guangdong, Hainan, Fujian, and Zhejiang. The study area was mainly located within the 30 km buffer zone of the coastline. Natural mangrove forests in China are mainly distributed within latitudes between 18° N and 27° N, while planted mangrove forests have extended to Leqing Bay in Zhejiang Province with a latitude of 28°25′ N [47]. Historically, the total area of mangrove forests in China once reached 250,000 ha and, in the 1950s, was about 42,000 ha, which continued to decrease because of extensive aquaculture farming and urbanization [48]. Dominant mangroves species in China include Bruguiera gymnorrhiza, Kandelia obovta, Avicennia marina, Aegiceras corniculatum, Rhizophora stylosa, Acrostichum aureum, and Acanthus ilicifolius [48].

Satellite Remote Sensing Images and Processing
The remote-sensed imagery used in this research was obtained from six satellites, including Ziyuan-3-01, Ziyuan-3-02, Gaofen-1A, Gaofen-1B, Gaofen-1C, and Gaofen-1D. The Ziyuan-3-01 (ZY3-01) satellite was launched on 9 January 2012 and the Ziyuan3-02 (ZY3-02) satellite was launched on 30 May 2016 in China [41]. These two Ziyuan-3 series satellites form a constellation at an altitude of about 505 km, providing pan and multispectral imagery (PMS) with resolutions of 2.1 m and 5.8 m, respectively. The Gaofen-1A satellite was launched on 26 April 26 2013 and the Gaofen-1B, -1C, and -1D were launched on 31 March 2018 [42]. The four Gaofen-1 (GF-1) series satellites form a constellation at an altitude of about 645 km, providing pan and multispectral imagery with resolutions of 2 m and 8 m, respectively. The main parameters for ZY-3 and GF-1 imagery are shown in Table 1. The swath width for the Gaofen-1 satellites was 60 km, and the revisit period was 41 days. The four-star constellation of the Gaofen-1 satellites has shortened the revisit period to 10 days. The Ziyuan-3 satellite had a swath width of 52 km and a revisit period of 59 days, and the double stars constellation shortened the revisit period to 30 days. The radiation resolution of the PMS sensors of the Gaofen-1 and the Ziyuan-3 satellites were both 10 bit. In this study, 2018 imagery obtained from the ZY-3 and GF-1 sensors were screened for the use of mangrove mapping. The rules for image screening were: (1) cloud cover should be less than 5% and (2) imagery obtained at low tide should be preferred. A total of 509 images were screened in the first step. Next, all images were orthorectified to the World Geodetic System 1984 (WGS84) reference system, and the root mean square geometric errors were less than 1 pixel. Each image was normalized for variation in solar angle and earth-sun distance by converting the digital number values to top-of-atmosphere reflectance. Then, the ortho-corrected panchromatic and multi-spectral imagery were fused using the pan-sharpen [49] algorithm and resampled to 2-meter imagery. The images covering the study area were mosaiced into a 2-meter image base, which would be used for mangrove mapping in the following step.

Mangrove Habitat Preparation
Prior to mangrove delineation, mangroves exist within a specific ecological niche, which can be used to eliminate much of the area where they will not be found. Since mangroves need to be tidal inundated periodically, the mangrove growing area should not be too far from the shoreline and the terrain elevation should not be too high. In this study, the mangrove habitat region was obtained by following two steps. First, a 30 km buffer zone was generated based on the coastline of China. Second, a mangrove habitat region was obtained by extracting the areas with a terrain elevation less than 20 m in the buffer zone, as the height of mangrove trees along the coast of China seldom exceeded 20 m [50]. The 30-m Shuttle Radar Topography Mission (SRTM) digital elevation model (DEM) data was used for terrain elevation estimation. Through the mangrove habitat preparation step, it was possible to screen out areas where mangroves were unlikely to occur, thereby improving interpretation efficiency.

Methods for Mangrove Delineation
The overview of the method workflow is shown in Figure 2. First, the satellite imagery was processed and mosaiced into a 2-meter resolution image base, and the mangrove habitat zone was prepared. Second, an object-based image analysis procedure was performed for mangrove plot basis extraction, including sample collection, image segmentation, image object classification, and post-processing. Third, the detected mangrove plots were checked by interpreters and verified by field work. Finally, the mangrove China dataset (MC2018) was generated and its accuracy was assessed. ISPRS Int. J. Geo-Inf. 2020, 9, x FOR PEER REVIEW 5 of 18 plots were checked by interpreters and verified by field work. Finally, the mangrove China dataset (MC2018) was generated and its accuracy was assessed. Mangrove plots with canopy closure greater than 20% were mapped, but sparsely scattered mangroves and young forests were not included in this inventory. The minimum mapping unit in MC2018 was 100 square meters.

OBIA Procedure for Mangrove Delineation
After the 2-meter resolution imagery and the mangrove habitat zone was prepared, we used an object-oriented image analysis method to extract mangrove distribution features. First, we conducted a sample collection. The preliminary samples were obtained from the Jia2015 dataset [33] by random sampling, and then these samples were manually checked with reference of Google Earth TM submeter images. A total of about 10,000 mangrove sample points and 30,000 non-mangrove sample points were obtained. Non-mangrove samples were image patches including open water, urban area, land vegetation, bare tidal flats, and Spartina alterniflora beach.
Because of the large area of the study region, the east-west span was 1600 km and the north-south span was 1200 km; it was unrealistic to process the whole area image at the same time. We adopted the strategy of tile processing in the second step, cutting the base image with a size of 40 km × 40 km. We obtained 150 tiles in total. The segmentation and classification were processed separately for each image tile, and the final result was a stitch of the tiled export.
The segmentation was performed with a multiresolution segmentation algorithm [51], which was widely used for image analysis. An image is partitioned into object primitives, which have small internal spectral heterogeneity with respect to their neighboring Mangrove plots with canopy closure greater than 20% were mapped, but sparsely scattered mangroves and young forests were not included in this inventory. The minimum mapping unit in MC2018 was 100 square meters.

OBIA Procedure for Mangrove Delineation
After the 2-meter resolution imagery and the mangrove habitat zone was prepared, we used an object-oriented image analysis method to extract mangrove distribution features. First, we conducted a sample collection. The preliminary samples were obtained from the Jia2015 dataset [33] by random sampling, and then these samples were manually checked with reference of Google Earth TM submeter images. A total of about 10,000 mangrove sample points and 30,000 non-mangrove sample points were obtained. Non-mangrove samples were image patches including open water, urban area, land vegetation, bare tidal flats, and Spartina alterniflora beach.
Because of the large area of the study region, the east-west span was 1600 km and the north-south span was 1200 km; it was unrealistic to process the whole area image at the same time. We adopted the strategy of tile processing in the second step, cutting the base image with a size of 40 km × 40 km. We obtained 150 tiles in total. The segmentation and classification were processed separately for each image tile, and the final result was a stitch of the tiled export.
The segmentation was performed with a multiresolution segmentation algorithm [51], which was widely used for image analysis. An image is partitioned into object primitives, which have small internal spectral heterogeneity with respect to their neighboring objects. This algorithm requires three user-defined parameters to control the size and shape of the generated image objects: (1) scale, (2) shape weight, and (3) compactness weight. We prepared different parameter combinations to segment the test images, and we used visual inspection to select images with better segmentation results. From this we determined the segmentation parameters for the mangrove image analysis, which were scale = 15, shape weight = 0.3, and compactness weight = 0.2.
The support vector machine (SVM) algorithm [52] was selected for object-oriented classification, and we used the Gaussian radial basis function (RBF) as the kernel. The SVM model has two important parameters: C and gamma. C is the penalty coefficient. If C is too large or too small, the generalization ability will deteriorate. Gamma is a parameter followed with the selection of the RBF function, and it also affects the generalization ability of SVM models. First, we conducted a parameter tuning process on a test region. We used the grid search optimization strategy to determine parameter settings in order to achieve better classification models. Results showed that setting C = 200 and gamma = 0.1 would produce a relatively strong SVM classification. Then, for each tile, an SVM model was trained, and all image object primitives were classified with the corresponding classifier. Finally, a stitch process was conducted to merge classification results of all tiles, and merged image objects were re-shaped with post-processing procedures, which included merging segments of the same class, removing objects with very small areas, and smoothing object borders. Through object-oriented image analysis and processing, we obtained the mangrove dataset V1 (MC2018V1).

Interpreter Checking
In order to improve the thematic accuracy of mangrove data, we conducted interpreter checking to correct errors in the extraction results and to add some missing patches by reference of the 2-meter base imagery. In the cross-checking step, the MC2018V1 were cross-checked by three operators. Each patch was marked with an attribute of "suspect" or "certain" according to the judgement of the three operators. These inspectors also reinterpreted the imagery, added some possible missing patches, and modified the unreasonable boundaries. Suspect patches would be focused on inspection during the field work. After interpreter cross-checking, we obtained the mangrove dataset V2 (MC2018V2).

Field Verification
The research team carried out the field survey of China's coastal mangrove forests in April and May 2019. The field work focused on verifying mangrove patches with low degrees of certainty. Mangrove patches marked as suspect in the MC2018V2 dataset were all selected for verification. For patches marked as certain, we selected 50% for verification in the field process. During the field work, some missing patches (omission errors) were recorded and manually added to the updated map.
The field survey was conducted along the Chinese mainland coastline, from the border in Guangxi to Taizhou in Zhejiang ( Figure 3). The length of the field survey route was approximately 9500 km. A total number of 2650 mangrove patches were verified during the field work. After field verification, the MC2018V2 dataset was modified and improved, and the final version MC2018 dataset was obtained.
ISPRS Int. J. Geo-Inf. 2020, 9, x FOR PEER REVIEW 7 of Figure 3. Illustration of field verification information of (a) field points and field routes overlapped on the 2-meter imagery; field photo of (b) typical mangrove species Rhizophora stylosa; (c) typical mangrove species Bruguiera gymnorrhiza; and (d) hypocotyl of Bruguiera gymnorrhiza.

Accuracy Assessment
There were two sources of ground truth points for accuracy assessment in this stud First, in June 2019, 223 points were obtained in Guangdong, Guangxi, Hainan, and Fuji by field survey. Second, 2077 mangrove verification points were collected by field inve tigators from submeter imagery of Google Earth TM and Bing Maps TM . We conducted field survey in Taiwan, and the verification points in Taiwan were all collected fro Google Earth TM and Bing Maps TM . The ground truth points fall into two categories: ma groves and non-mangroves. Non-mangrove classes included open water, urban area, lan vegetation, bare tidal flats, and Spartina alterniflora beach. The confusion matrix metho was used for accuracy assessment, and the producer's accuracy (PA), the user's accura (UA), the overall accuracy (OA), and the kappa coefficient were calculated.

Inter-Comparison with Other Available Mangrove Forest Maps or Datasets
The MC2018 dataset [46] was compared with other available mangrove data, inclu ing: (1) [21,22]); and (7) the mangrove forest change of China dataset by Zheng et (MFC2018 [19]). Information about image sources, image resolutions, mapping method and obtained mangrove areas were collected and compared for these datasets. In additio MC2018, Jia2015, Chen2015, MFW2000, and GMW2016 were selected for visual compa son on the two test sites of Qinzhou Bay in Guangxi and Dongzhai Port in Hainan. . Illustration of field verification information of (a) field points and field routes overlapped on the 2-meter imagery; field photo of (b) typical mangrove species Rhizophora stylosa; (c) typical mangrove species Bruguiera gymnorrhiza; and (d) hypocotyl of Bruguiera gymnorrhiza.

Accuracy Assessment
There were two sources of ground truth points for accuracy assessment in this study. First, in June 2019, 223 points were obtained in Guangdong, Guangxi, Hainan, and Fujian by field survey. Second, 2077 mangrove verification points were collected by field investigators from submeter imagery of Google Earth TM and Bing Maps TM . We conducted no field survey in Taiwan, and the verification points in Taiwan were all collected from Google Earth TM and Bing Maps TM . The ground truth points fall into two categories: mangroves and non-mangroves. Non-mangrove classes included open water, urban area, land vegetation, bare tidal flats, and Spartina alterniflora beach. The confusion matrix method was used for accuracy assessment, and the producer's accuracy (PA), the user's accuracy (UA), the overall accuracy (OA), and the kappa coefficient were calculated.

Inter-Comparison with Other Available Mangrove Forest Maps or Datasets
The MC2018 dataset [46] was compared with other available mangrove data, including: (1) [19]). Information about image sources, image resolutions, mapping methods, and obtained mangrove areas were collected and compared for these datasets. In addition, MC2018, Jia2015, Chen2015, MFW2000, and GMW2016 were selected for visual comparison on the two test sites of Qinzhou Bay in Guangxi and Dongzhai Port in Hainan.

Spatial and Area Distribution of Mangroves in China in 2018 3.1.1. Distribution of Mangroves in Different Provinces
In 2018, the total area of mangrove forests in China was 25,683.88 hectares ( Table 2). The distribution of mangrove forests varied greatly among different provinces. Guangdong Province held the largest distribution area of mangroves, which was 10,330.74 hectares, accounting for 40.22% of the total national mangrove area. Guangxi and Hainan followed with distribution areas of 8449.0 hectares and 4676.71 hectares, respectively. The total mangrove area in the top three provinces accounted for 91.33% of the national mangrove area.  Figure 4 shows the spatial distribution of mangrove forest areas along the coasts. Mangrove areas were mainly distributed in Guangxi, eastern Guangdong, and northern Hainan Province, and they were less distributed in Fujian, Zhejiang, and Taiwan. ISPRS Int. J. Geo-Inf. 2020, 9, x FOR PEER REVIEW 8 of 18

Distribution of Mangroves in Different Provinces
In 2018, the total area of mangrove forests in China was 25,683.88 hectares ( Table 2). The distribution of mangrove forests varied greatly among different provinces. Guangdong Province held the largest distribution area of mangroves, which was 10,330.74 hectares, accounting for 40.22% of the total national mangrove area. Guangxi and Hainan followed with distribution areas of 8449.0 hectares and 4676.71 hectares, respectively. The total mangrove area in the top three provinces accounted for 91.33% of the national mangrove area. Figure 4 shows the spatial distribution of mangrove forest areas along the coasts. Mangrove areas were mainly distributed in Guangxi, eastern Guangdong, and northern Hainan Province, and they were less distributed in Fujian, Zhejiang, and Taiwan.   We cut the mangrove patches with a 0.01° fishnet in the latitudinal and longitudinal directions and summarized the area distribution of mangroves in the longitudinal and latitudinal directions. The results are shown in Figure 5. From the longitudinal direction, mangroves were distributed between 108.0° E and 122° E. Mangrove forests in China can be divided into two parts by the longitude line of 114.0° E (near the Pearl River estuary). Most mangroves were distributed to the west of the Pearl River estuary (108.0° E to 114.0° E), and the area of mangroves in this region reached 22,953.40 hectares, accounting for 89.4% of the total mangrove area. The area of mangroves east of the Pearl River estuary (114.0° E to 122.0° E) was 2730.48 hectares, accounting for only 10.6% of the total mangrove area.
From the latitudinal direction, Chinese mangroves were distributed between 18° N and 28.9° N, showing characteristics of high area in the middle and low area toward the ends. Most mangroves were distributed between 19° N latitude and 23° N latitude, with a total area of 21,329.72 hectares, accounting for 83% of the total mangrove area. In regions north of 25° N, only a small area of mangroves was scatter distributed.

Distribution of Mangroves within the Proximity of Urban Areas
In order to obtain the relationship between the distribution of mangroves and the development of coastal cities, we used the buffer and overlay analysis method to calculate the area of mangroves within the proximity of different buffer distances of urban boundaries. We used the global urban boundaries (GUB_Global_2018) dataset produced by Li et al. [53]. We gradually increased the buffer distance in 1 km steps and calculated the area of mangroves in buffer zone segments with buffer distances from 1 km to 35 km.
The results are shown in Figure 6. About 10% of the mangrove area was distributed within 1 km of the urban boundary, and the farthest distance from the mangrove forest to the urban area was 35 km. Figure 6b shows that mangrove area grows rapidly as the distance from the urban boundary increases. About 80% of the mangroves were distributed within a 10 km proximity of the city area, and mangrove area within 15 km from the city boundary accounted for 95%. From the latitudinal direction, Chinese mangroves were distributed between 18 • N and 28.9 • N, showing characteristics of high area in the middle and low area toward the ends. Most mangroves were distributed between 19 • N latitude and 23 • N latitude, with a total area of 21,329.72 hectares, accounting for 83% of the total mangrove area. In regions north of 25 • N, only a small area of mangroves was scatter distributed.

Distribution of Mangroves within the Proximity of Urban Areas
In order to obtain the relationship between the distribution of mangroves and the development of coastal cities, we used the buffer and overlay analysis method to calculate the area of mangroves within the proximity of different buffer distances of urban boundaries. We used the global urban boundaries (GUB_Global_2018) dataset produced by Li et al. [53]. We gradually increased the buffer distance in 1 km steps and calculated the area of mangroves in buffer zone segments with buffer distances from 1 km to 35 km.
The results are shown in Figure 6. About 10% of the mangrove area was distributed within 1 km of the urban boundary, and the farthest distance from the mangrove forest to the urban area was 35 km. Figure 6b shows that mangrove area grows rapidly as the distance from the urban boundary increases. About 80% of the mangroves were distributed within a 10 km proximity of the city area, and mangrove area within 15 km from the city boundary accounted for 95%.

Accuracy of MC2018
The accuracy evaluation results are shown in Table 3. The overall accuracy of the MC2018 dataset, the kappa coefficient, the mapping accuracy, and the user accuracy, were 99.3%, 0.985, 99.3%, and 99.0%, respectively.

Inter-Comparison of Mangrove Maps among Multi-Source Datasets
The MC2018 dataset [46] was compared with the other seven available mangrove maps, including SFA2000 [23], Jia2015 [34], Chen2015 [35], WAM2000 [10], MFW2000 [1], GMW2016 [21], and MFC2018 [19]. Detailed information is listed in Appendix A Table  A1. In these datasets, the mangrove area in China varied greatly. The largest area was 25,683.88 hectares, obtained in this study, and the smallest area was 15,869 hectares, obtained from the GMW2016 dataset. Both the MC2018 dataset and the MFC2018 dataset were mangrove maps of China in 2018, and the mangrove area in the MC2018 dataset was 1,081.43 hectares more than the MFC dataset. The difference may exist because there were many small mangrove patches in China and because the higher resolution images and richer field data used by MC2018 had a better potential to capture these small areas of fragmented mangroves. In addition, some of the area differences should be due to the differences in time (year) of image acquisition and the different methodologies employed for mangrove mapping. Furthermore, from the time sequence of the different mangrove datasets in Table A1, the mangrove area in China had been increasing since 2000, which was consistent with the result from Jia [33].
We selected two test regions of Qinzhou Bay in Guangxi and Dongzhai Port in Hainan, and visually compared the MC2018 and the other four datasets (Jia2015, Chen2015,

Accuracy of MC2018
The accuracy evaluation results are shown in Table 3. The overall accuracy of the MC2018 dataset, the kappa coefficient, the mapping accuracy, and the user accuracy, were 99.3%, 0.985, 99.3%, and 99.0%, respectively.

Inter-Comparison of Mangrove Maps among Multi-Source Datasets
The MC2018 dataset [46] was compared with the other seven available mangrove maps, including SFA2000 [23], Jia2015 [34], Chen2015 [35], WAM2000 [10], MFW2000 [1], GMW2016 [21], and MFC2018 [19]. Detailed information is listed in Appendix A Table A1. In these datasets, the mangrove area in China varied greatly. The largest area was 25,683.88 hectares, obtained in this study, and the smallest area was 15,869 hectares, obtained from the GMW2016 dataset. Both the MC2018 dataset and the MFC2018 dataset were mangrove maps of China in 2018, and the mangrove area in the MC2018 dataset was 1081.43 hectares more than the MFC dataset. The difference may exist because there were many small mangrove patches in China and because the higher resolution images and richer field data used by MC2018 had a better potential to capture these small areas of fragmented mangroves. In addition, some of the area differences should be due to the differences in time (year) of image acquisition and the different methodologies employed for mangrove mapping. Furthermore, from the time sequence of the different mangrove datasets in Table A1, the mangrove area in China had been increasing since 2000, which was consistent with the result from Jia [33].
We selected two test regions of Qinzhou Bay in Guangxi and Dongzhai Port in Hainan, and visually compared the MC2018 and the other four datasets (Jia2015, Chen2015, GMW2016 and MFW2000) in these two regions. Figure 7 shows the results of the Guangxi Qinzhou test area. In this area, the MC2018 dataset performed best and mangrove patches fit the reference image well. The Chen2015 and Jia2015 datasets performed second; they correctly extracted about 50% of the mangrove range, with some missing error. The GMW2016 and MFW_USGS did not perform well enough in the region; they missed a significant amount of the mangrove area. Some of the missing errors were due to the year differences between the mangrove maps and the reference images.
GMW2016 and MFW2000) in these two regions. Figure 7 shows the results of the Guangxi Qinzhou test area. In this area, the MC2018 dataset performed best and mangrove patches fit the reference image well. The Chen2015 and Jia2015 datasets performed second; they correctly extracted about 50% of the mangrove range, with some missing error. The GMW2016 and MFW_USGS did not perform well enough in the region; they missed a significant amount of the mangrove area. Some of the missing errors were due to the year differences between the mangrove maps and the reference images.
In the Hainan Dongzhai Port test area (Figure 8), mangroves were concentrated in large areas, and they have changed little in recent decades. In general, the five datasets have correctly extracted large mangrove areas in this region. The overall accuracies of MC2018, Jia2015, Chen2015 were relatively high, but there were relatively more errors and omissions in the MFW2000 and GMW2016 datasets. The MC2018 dataset was mapped from 2-meter high-resolution imagery, the edge of the patch is smooth, and the structure of the small patch is complete. The mangrove patches in the remaining four datasets all exhibit severe jagged edge effects and the small-area mangrove patches were not sufficiently accurate.  In the Hainan Dongzhai Port test area (Figure 8), mangroves were concentrated in large areas, and they have changed little in recent decades. In general, the five datasets have correctly extracted large mangrove areas in this region. The overall accuracies of MC2018, Jia2015, Chen2015 were relatively high, but there were relatively more errors and omissions in the MFW2000 and GMW2016 datasets. The MC2018 dataset was mapped from 2-meter high-resolution imagery, the edge of the patch is smooth, and the structure of the small patch is complete. The mangrove patches in the remaining four datasets all exhibit severe jagged edge effects and the small-area mangrove patches were not sufficiently accurate.

The Status of Mangrove Protection Areas in China
According to the list of China's nature reserves from 1980 to the present, China has established 35 nature reserves related to mangrove protection, including 7 national-level nature reserves, 7 provincial-level nature reserves, and 21 county-level and municipallevel nature reserves. In addition, since 2001, China has launched a mangrove protection project, continued to carry out afforestation activities, and achieved good results. The results of remote sensing monitoring have shown that the area of mangrove forests in China has gradually resumed growth since 2000 [33].
We overlaid the mangrove forest distribution data with the nature reserve range and found that 64% (16,388.9 hectares) of mangroves were already in the mangrove protection area. This shows that, in recent years, the awareness of ecological protection has been continuously strengthened, and the protection of mangrove wetland ecosystems has continued to improve in China.

The Status of Mangrove Protection Areas in China
According to the list of China's nature reserves from 1980 to the present, China has established 35 nature reserves related to mangrove protection, including 7 national-level nature reserves, 7 provincial-level nature reserves, and 21 county-level and municipal-level nature reserves. In addition, since 2001, China has launched a mangrove protection project, continued to carry out afforestation activities, and achieved good results. The results of remote sensing monitoring have shown that the area of mangrove forests in China has gradually resumed growth since 2000 [33].
We overlaid the mangrove forest distribution data with the nature reserve range and found that 64% (16,388.9 hectares) of mangroves were already in the mangrove protection area. This shows that, in recent years, the awareness of ecological protection has been continuously strengthened, and the protection of mangrove wetland ecosystems has continued to improve in China.

Advantages of 2-Meter Resolution Imagery
Most of the mangrove distribution datasets currently released used Landsat data at 30-meter resolution or PALSAR (Phased Array type L-band Synthetic Aperture Radar) data at a 25-meter resolution. On these moderate spatial resolution images, the vegetation texture and spatial details were missing, and the accuracy of mangrove identification was defective. In addition, small and irregularly shaped mangrove areas were difficult to identify and extract accurately.
In this study, satellite remote sensing images with a resolution of 2 m were used for mangrove mapping. The finer spatial resolution makes it possible to delineate scattered, distributed mangrove plots with small areas and to map narrow strips of mangrove plots distributed along the shoreline. This is particularly important for the mapping of mangroves in China because most mangroves patches (more than 10,000 hectares in total) were less than 10 hectares, accounting for about 40% of the total mangrove area in China. Using high spatial resolution remote sensing images, mangrove forest distributions at a finer scale can be extracted and fine changes in information can be quickly detected. The detailed information of the distribution and changes in these mangroves is highly significant for local management agencies to carry out mangrove ecological protection and restoration and to coordinate coastal economic development and ecological protection.

Confusions between Mangroves and Other Vegetation
Along the southeast coast of China, there exist other types of vegetation growing on the tidal flats where mangrove vegetation grows, and these vegetation types are confused with mangroves in image interpretation. Some vegetation grow in the same habitat as mangroves and form a competitive relationship with mangroves. Such vegetation is typically Spartina alterniflora and Phragmites communis (reed) ( Figure A1). Spartina angustifolia was first introduced from the United States in 1979 and was mainly used for wave and beach protection early on. However, because of the lack of local natural enemies, Spartina alterniflora multiplied quickly and invaded the living environment of local species, especially mangroves. Spartina alterniflora or reeds could often be found around mangrove plots. The spectral characteristics of these vegetations are relatively similar, which increases the complexity of mangrove mapping from imagery in these areas.
In addition, the two types of woody vegetation that are easily confused with mangroves in China are Metasequoia glyptostroboides and Casuarina equisetifolia. Figure A2 shows the field photos of Metasequoia and Casuarina. In the southeast coast of China, most of the Metasequoia glyptostroboides trees were planted artificially, and it was an important tree species for coastal protection. Casuarina equisetifolia trees in the coastal areas were either naturally grown or artificially cultivated, and it was also the main tree species for coastal protection. Similar to mangrove vegetation, Metasequoia and Casuarinas grow on the tidal flats or around the nearby areas; sometimes they even grow spatially connected with mangroves. When interpreting mangroves, these types of vegetation need to be distinguished, and comprehensive judgments need to be made in combination with information from field surveys.

The Importance of Field Work in Mangrove Mapping
Mapping mangroves from remote sensing imagery faces two important difficulties: (1) it is difficult to distinguish mangrove vegetation from adjacent non-mangrove vegetation and (2) it is difficult to distinguish between mangrove and tidal flat herbaceous vegetation (Spartina alterniflora). Using image analysis methods (such as OBIA) alone can hardly generate mangrove maps with high accuracy in large regions. In the OBIA extraction steps in this study, the accuracy of mangrove mapping only reached about 80%. We believe that, in order to produce fine-scale and high-precision mangrove maps, manual editing and strict field verification work are necessary.

Limitations and Caveats
The mangrove inventory only included mangrove plots with a canopy density of 20% or more. Young plants that were newly planted but not forested were not included in this dataset. In addition, in high-tide images, some low-lying mangroves will be submerged by water bodies, making mangroves difficult to identify and extract. Although most of the images we used were in low tide periods, there may be a small amount of missing mangrove area in some low tide beaches. Furthermore, in the process of the field work, we also found sparsely distributed mangroves with canopy densities less than 20%, and these sparse forests were not included in the MC2018 dataset either. Despite the high spatial resolution of Gaofen-1 and Ziyuan-3 sensors, the spectral resolutions of these sensors are limited and consisted in only the RGB and NIR spectrums. Thus, the use of multispectral and hyperspectral wavelengths and multi-modality, multi-sensor data fusion would show strong potential for mangrove mapping in the future.

Conclusions
Remote sensing techniques have become an important method to quickly obtain data about the distribution and change of mangroves. However, the traditional mediumresolution images like Landsat can only obtain relatively coarse mangrove distribution data globally or regionally. With the rapid development of high spatial resolution remote sensing technology, it is necessary to produce fine-scale mangrove data over a large area using higher spatial resolution remote sensing images.
In this research, we used 2-meter spatial resolution remote sensing imagery obtained from GF-1 and ZY-3 satellites, and we employed a hybrid method of object-based image analysis (OBIA), interpreter editing, and field verification. We produced the fine-scale mangrove extent dataset of China in 2018 (MC2018), with an overall accuracy greater than 98%. This dataset can provide detailed information about the mangrove ecosystem status and can support local authorities for mangrove forest planning and management. This study demonstrated the strong potential of the 2-meter resolution satellite imagery of Ziyuan-3 and Gaofen-1 in the annual monitoring of resources and environmental parameters over a large region. The hybrid method used in this study can also provide a reference for fine-scale mangrove monitoring and mapping in other regions of the world.