Extraction of 10 m Resolution Global Mangrove in 2022

Abstract

With the intensification of global climate change, there is an increasing emphasis on protecting natural resources. Mangrove forests, critical to tropical and subtropical intertidal ecosystems, have garnered considerable attention in recent years for their strong carbon sink capacity, rich species diversity, and abundant natural resources. This study utilizes the 2020 global mangrove vector data as a baseline to construct a reasonable buffer zone by calculating the increase in mangrove crown width. The Google Earth Engine (GEE) platform and its Sentinel-2 data from 2022 are employed to acquire synthetic images across all regions using the mosaic algorithm. Then, mangrove forests are extracted using the Otsu algorithm, and a map depicting the global spatial distribution of mangrove forests in 2022 is obtained. The average overall accuracy of the extracted mangrove forests in this study reaches 92.4%, and it is determined that the global mangrove forest area expanded by 4920.6 km² between 2020 and 2022, This study provides crucial data support for the global monitoring of mangrove changes and holds significant importance for protecting and restoring mangrove ecosystems.

1. Introduction

Mangroves are located at the junction of land and sea, comprising woody vegetation distributed along the coastlines of 118 countries and regions in tropical and subtropical areas [1]. Studies have shown that despite their location in submerged saline environments, where hypoxic conditions often inhibit the growth of most terrestrial species, mangrove ecosystems continue to be one of the most productive aquatic ecosystems on Earth, with a primary productivity of 2.5 g C/(m² day) [2]. Globally, mangroves store approximately 4.19 Pg of carbon in vegetation and 1 m of surface soil, exhibiting substantial carbon sink capacity [3]. Therefore, understanding the changes in mangroves on a global scale is crucial.

However, Wang et al. [4] reported that the global mangrove coverage is approximately 170,000 km². From 1980 to 2005, this area declined from 188,000 to 152,000 km², a total reduction of 36,000 km². Specifically, from 1961 to 1996, the mangrove forests in southern Thailand decreased from 3679 to 1905 km², representing a decrease of 48%, mainly due to their conversion into aquaculture land. Within 35 years, 50% of the mangrove forests in the Philippines were transformed into aquaculture sites [5]. Overall, there has been a decreasing trend in global mangrove areas over the past four decades. Hence, studying the distribution and area changes of global mangroves is greatly important for maintaining the ecological balance in tropical and subtropical regions, preserving global biodiversity, and mitigating global warming [6].

There are well-established methods and systems for studying mangroves, and the extraction of mangroves on a global scale has been ongoing for over two decades. The initial global mangrove change statistics were compiled by Hamilton et al. [3], combining the 2000 baseline mangrove data from Giri et al. [1] with forest area loss maps from 2000 to 2012 by Hansen et al. [3]. Hamilton et al. [3] estimated a loss of 1646 km² of mangrove forests between 2000 and 2012. Despite the technological limitations at the time that significantly restricted data accuracy, subsequent studies improved the data resolution. In 2020, Jia et al. [7] extracted a global mangrove dataset with a 10 m resolution based on Sentinel-2 images and analyzed the proportion of mangrove forest areas across various continents and in some mangrove nature reserves or national parks. In the same year, Bunting et al. [8] employed JAXA’s SAR data to produce a long-term series of global mangrove change data from 1996 to 2020, estimating a loss of mangroves over the past 24 years to be 3.4%, with an accuracy rate of 87.4%. In addition, Liao et al. [9] conducted a 30 m resolution temporal statistical analysis of the global mangrove area from 2000 to 2020.

In recent years, advancements in science and technology and related algorithms have led to significant breakthroughs in mangrove research. In 2013, Alsaaideh [10] utilized Landsat 7 data combined with digital elevation models to draw mangrove distribution maps for six islands in southern Japan. In 2016, Monzon et al. [11] combined optical data with SAR data for the ongoing monitoring and mapping of mangrove forests in the Philippines. At the same time, the Otsu algorithm, with its efficient and automatic image binarization processing, has become increasingly prevalent in binary classification research. In addition, the Google Earth Engine (GEE) has been widely adopted in remote sensing and other fields, enabling users to independently write code and process remote sensing data efficiently. In 2016, Reddy et al. [12] applied the Otsu algorithm for threshold segmentation to extract mangrove forests. In 2018, Xia et al. [13] analyzed GF-1 data in Yulin City, Guangxi, producing a mangrove distribution map with a spatial resolution of 2 m. In 2019, Liu et al. [14] distinguished different tree species within mangrove ecosystems using the Otsu algorithm based on significant differences in some sensitive band combinations. In 2020, Zhang et al. [15] monitored the dynamic changes in mangroves from 2016 to 2020 and conducted a change analysis using the Otsu algorithm. In 2021, Pratiwia et al. [16] also employed the Otsu algorithm to extract mangroves, while Wang et al. [17] implemented this algorithm through the GEE platform for the same purpose. In 2021, Cheng et al. [18] developed an automatic, rapid, and highly precise method for classifying intertidal wetlands in the Zhangjiang Estuary Mangrove Nature Reserve, Fujian Province, using the GEE platform, achieving an overall accuracy of 96.5% and a Kappa coefficient of 0.95. Wen et al. [19] utilized multi-source remote sensing data, including LiDAR, Sentinel-2, and GF-2, to extract mangroves’ spatial distributions and estimate the mangrove biomass. Zhao et al. [20] applied the random forest algorithm to classify Sentinel-1 and Sentinel-2 images hosted on the GEE platform, addressing challenges in sampling mangrove distribution areas. In 2023, Shi et al. [21] overcame the limitations of traditional remote sensing data processing methods, such as slow speeds and extensive time requirements, by extracting mangrove forests using the GEE cloud platform and Landsat TM/OLI data. This study aims to distinguish between mangroves and non-mangroves, addressing the binary classification problem. Due to the automatic and fast threshold calculation function of the Otsu algorithm in binary classification problems and its reliability in classification accuracy, which has been proven in previous studies, the Otsu algorithm is employed to extract mangroves in this study.

Although most current research on mangrove extraction is regional, global studies predominantly rely on products with a resolution of 30 m, such as those by Liao et al. [9]. However, the Otsu algorithm has not yet been applied to global mangrove extraction research. In addition, Jia et al. [7] constructed a 1 km buffer zone for mangrove extraction in their study without clarifying the rationale for this specific buffer zone dimension. Thus, this study addresses the challenge of determining an appropriate research area based on the existing mangrove vector map to achieve accurate and efficient mangrove extraction. The primary focus of this study is to propose an automatic and efficient global mangrove extraction scheme with a resolution of 10 m using the Otsu algorithm. In addition, based on the extracted mangrove vector map, this research compiles statistics from various continents and countries with large mangrove areas to analyze the global distribution and trends, providing vital data to support the development and conservation of mangrove ecosystems.

2. Data Sources and Pre-Processing

The data source for this study includes the open and freely available global mangrove vector data from 2020, released by Jia et al. [7] (https://doi.org/10.7910/DVN/PKAN93 (accessed on 29 December 2022)) and Sentinel-2 imaging from the GEE platform. Based on the 2020 global mangrove distribution map, this study utilized Sentinel-2 data to extract mangroves at a resolution of 10 m in 2022.

Due to the extensive volume of data in the 2020 global mangrove vector map, it is difficult to buffer or subdivide it into smaller manageable areas for processing by GEE in a single operation. Hence, the global mangrove forests were initially segmented by continent, and the mangrove forests within each continent were systematically numbered. This method ensures error-free and systematic numbering. Initially, the 2020 global mangrove vector data underwent preliminary pruning. Given the absence of mangrove distribution in Europe and Antarctica, the study divided the vector data into five regions by continent: Asia, Africa, Oceania, North America, and South America. Due to the complexities of precise continental division, especially near the boundaries between North and South America and between Asia and Oceania, there can be a situation where the same mangrove forest in a small area is divided into two continents. Hence, the principle of proximity is adopted at the boundary of continents, where adjacent elements are divided into the same continent.

Given that mangroves naturally expand or are destroyed by natural disasters, it was deemed necessary to construct a buffer zone larger than the original vector map’s range to serve as the study area along with the original range. Remote sensing images have shown increases in tree area, reflected by increases in crown width or by trees dumped due to natural disasters. Based on a crown height model developed by Josicleda et al. in 2016, it was calculated that mangrove tree heights increased by 0.35 m and crown widths by 0.03 m over 3.5 years [22]. In 2016, Lagomasino et al. [23] measured the average crown height of mangrove canopies using field measurements, LiDAR, and VHR CHM methods, recording heights of 10.1, 10.76, and 10.95 m, respectively, with a maximum recorded average height of 15.25 m, none exceeding 20 m. At the same time, considering that the two factors of tree dumping and buffer zone should be greater than or equal to the increase in crown width, it can be assumed that the extension of mangrove growth from inside to outside within two years will not exceed 20 m. Hence, this study defines the potential change zone as the original vector map and its 20 m buffer zone. It employed ArcGIS to create a 20 m buffer zone for each trimmed and numbered mangrove forest. Various levels of Sentinel-2 image products integrated into the GEE platform were utilized for image pre-processing, which included cropping, cloud removal, and tiling. The red and near-infrared bands of Sentinel-2 satellite data were used to produce a 10 m resolution mangrove vector map globally in 2022, and accuracy verification and change analysis were performed.

When loading and pre-processing Sentinel-2 images on the GEE platform, this study initially loaded the vector map of the study area and then utilized the GEE platform’s ee.ImageCollection() method to load raster data corresponding to the date, range, and cloud specifications. The mosaic algorithm was employed to merge the image collection into a single image. During this process, cloud spots were removed using the prewritten cloud removal function using the .map() method. Cloud removal was necessary using a 60 m resolution B10 band, making this method only effective for larger clouds. The impact of smaller clouds on the research was considered negligible and, therefore, overlooked. For this study, the cloud cover threshold was set to 5%. Finally, the vector image was employed to crop and synthesize the resulting raster image to produce the image of the study area.

3. Methodology

3.1. Otsu Algorithm

The Otsu method, a commonly utilized image threshold segmentation technique, has garnered increasing attention. Several improved versions of the Otsu method have been proposed, including the recursive Otsu method, two-dimensional Otsu method, and two-stage multi-threshold Otsu method [24]. Given that this study aims to classify remote sensing images into two categories, mangroves and non-mangroves, it employs binary classification. Therefore, this research employs the Otsu algorithm to extract mangroves, implementing the algorithm on the GEE platform.

The Otsu algorithm constructs a function for a certain band or combination of bands of raster images, with the threshold t being the independent variable and the interclass variance as the dependent variable. The threshold t, which corresponds to the maximum interclass variance, denoted as t*, is identified as the optimal threshold [25]. In this study, the threshold t corresponding to the maximum interclass variance of a correlation index between mangrove and non-mangrove categories more accurately represents the difference between these two categories. After completing the classification based on a specific spectral feature, if further classification within one of the categories is desired, the Otsu algorithm can be reapplied to the selected category, utilizing either a different class of spectral feature or the original spectral feature but targeting a different characteristic distinguished by the newly computed thresholds. In addition, this study relies on the data from Jia et al. [7]; therefore, a substantial portion of the study area consists of mangroves, effectively excluding regular objects such as farmland and buildings. Considering mangroves’ irregular and evolving boundaries, the study employed a pixel-based classification approach rather than an object-oriented classification approach. Since the spectral index and its threshold must be calculated, and other features such as shape do not need processing, the Otsu algorithm satisfies the research requirements.

When processing Sentinel-2 images, a histogram representing the gray frequency distribution of a specific spectral index of the image is inputted through the subfunction otsu1, and then the calculated optimal threshold is returned. Once the image is inputted by the parent function otsu, the return value of the subfunction is called, and the segmented image is directly outputted.

3.2. Extraction of 10 m Resolution Global Mangrove Zoning in 2022

After the image pre-processing was completed, the Otsu algorithm was utilized to extract mangroves. Based on the research by Cheng et al. [18], vegetation, including mangroves and non-vegetation, can be distinguished by the spectral index NDVI, with the threshold determined using the Otsu algorithm. Subsequently, combined with the spectral differences between mangroves and Spartina alterniflora, the two can be distinguished [18].

After calculating the NDVI grayscale images for each small area and generating the histograms, the Otsu algorithm was applied twice consecutively to the NDVI grayscale images of the area to determine the threshold for distinguishing between mangrove and non-mangrove forests. This threshold, generally ranging from 0.41 to 0.45, is influenced by various factors such as region and weather. After visualizing the pixels exceeding this threshold on the NDVI grayscale image and the original buffer within the GEE platform, the comparison image between the mangrove forests and the original buffer was obtained (Figure 1).

Figure 1 indicates that the vegetation non-vegetation threshold calculated for the area is 0.185, and the mangrove non-mangrove threshold is 0.435, consistent with the approximate range of thresholds, 0.410–0.450, derived by Cheng et al. [18] The mangrove non-mangrove threshold in this study is utilized only to differentiate between mangrove and alterniflora in intertidal mangrove ecosystems, not between mangroves and other terrestrial vegetation. Hence, the buffer zone is maintained at 20 m and is not expanded to 1 km or more. In fact, using the same image from the same region for extraction and the same batch of samples for accuracy verification, when the buffer is 50 m, the overall accuracy is 93.4%, which is lower than 94.5% when the buffer is 20 m. When the buffer reaches 1 km, the accuracy falls below 88.0%. In order to improve the algorithm’s efficiency, NDVI band raster images after threshold segmentation are directly visualized as mangrove raster images in Figure 1, with the band color designated as green. Thus, the green areas will display dark and light green shades. This method eliminates the step of converting grids to vectors during visualization. The depth of green represents the NDVI value of the pixel, representing the growth status of the mangrove forests represented by the pixel.

3.3. Synthesis of 10 m Resolution Global Mangrove Vector Map in 2022

In Section 3.2, vector maps of mangroves within each small region were obtained by calculating the spectral indices and threshold segmentation for each small region. However, they are scattered and contain some errors, which need correction and synthesis into a vector graph. Before synthesizing, it is necessary to address the errors present in the vector maps. Due to the network or other uncertainties, when exporting vector maps through the GEE platform, a large number of geometric errors can occur if the geometry of the vector map is complex, and these errors cannot be repaired through the functions available in ArcGIS 10. 8 and other software. Therefore, it is necessary to combine the 2020 vector maps with raster images for manual interaction. The principles are as follows:

(1)

If the error area is only a small part of the study area, the program will not lag significantly, so it can be fixed by deleting excess arcs or nodes.

(2)

If the error area covers the entire study area and causes the program to lag significantly, and if the errors are merely redundant arcs between nodes, then the elements containing the errors can be selected and deleted and filled in with vector elements from the same area in 2020 to be merged as the mangrove vector map for that area. Although this produces an error, the error arises only in the edge areas of the features and not within them. This study selected some regions for statistical analysis during the experimental process. For example, a mangrove vector map of an African area that did not produce errors yielded a circumference of 147,855 m and an area of 2,203,089 m². If the largest element with the largest area was chosen, the area is 927,719 m². Using the same elements from 2020 for substitution, the difference in total area is less than 5%, and the difference in total perimeter is less than 3%, so the error generated can be ignored. It is impossible to validate using the area that generated errors because the area that generated the error cannot be counted, or the data obtained after counting is inaccurate.

After the errors are fixed, the vector maps of each continent can be synthesized first, and then the mangrove vector maps of each continent can be merged into a global mangrove vector map. The global mangrove vector map with a resolution of 10 m in 2022 is shown in Figure 2.

Due to memory limitations and other factors, the synthesized global mangrove vector map cannot be fully loaded in most software systems. At the global or continental scale, changes in the area of mangroves are not readily apparent; thus, statistical analysis is necessary. In this study, each continent was initially counted, and then these figures were aggregated to determine the total area of global mangrove forests in 2022.

4. Results and Discussion

Due to the extensive coverage of mangrove forests worldwide, the thresholds derived from images across different regions vary significantly, and no global mangrove sample library exists. Hence, this study must select the appropriate sample points through visual interpretation to verify accuracy. In order to mitigate the randomness of the samples and regional differences, the selection considered several factors:

• The samples are located in a mangrove national park or a protected area.
• Previous studies have also categorized these areas as either mangrove or non-mangrove.
• Submeter-level satellite images on the GEE platform confirm the mangrove or non-mangrove status of the samples.

For example, the Sundarbans Park in Bangladesh was selected as the sample distribution area for this study, with references to the research results of Jia et al. [7] and Bunting et al. [8] Sentinel-2 data and submeter-level satellite images were concurrently displayed on the GEE platform map to maximize the accuracy of the sample identification. Furthermore, samples were selected from as many continents as possible to avoid regional disparities among global mangrove forests. In regions lacking mangrove national parks or protected areas, this study relied on factors 2 and 3 above and chose areas with dense mangrove distribution. The primary selected areas include the Shankou Mangrove Natural Resources Protection Area in Guangxi Zhuang Autonomous Region, China (Asia), Bangladesh’s Sundarbans Park (Asia), Abu Dhabi Mangrove National Park (Asia), Senegal’s Salum Delta National Park (Africa), Brazil’s Cape Orange National Park (South America), New Zealand’s Rotorua Mangrove (Oceania), and California’s Mangrove National Park (North America), among others. In addition, this study selected some mangrove sample points from non-protected areas or national parks as secondary sample libraries. Although mangroves cover most areas within national parks, the imagery is often heavily clouded due to the tropical and subtropical climates. For accuracy verification in the Northern Hemisphere, the images chosen for this study were from January to May 2022, corresponding to the dry season and typically feature less cloud cover. In this study, the cloud coverage in the images was limited to less than 5% and de-clouded, but there are still cloud shadows covering areas between 100 and 1000 m².

Figure 3 compares mangrove vector maps and submeter-level satellite images of four selected national parks or nature reserves and their surrounding areas in 2022. Analysis of the four images indicates that mangroves in different regions exhibit different spectral phenomena and are influenced by factors such as the timing of image capture, environmental conditions, and the growth status of the mangroves. Mangrove forests in areas with abundant precipitation appear more vibrant in satellite images, as shown in Figure 3b,c, which are situated in tropical rainforests or tropical monsoon climate zones. In addition, Figure 3 allows for a preliminary analysis of the factors that affect the extraction results of mangroves. Figure 3a depicts a vector map and satellite image of a portion of mangrove forests in Abu Dhabi National Park, United Arab Emirates. The climate in this region is extremely arid, the cloud content in the images is usually low, and there is not much distribution of other vegetation types, resulting in high-quality extraction outcomes. The results in Figure 3a are accurate, except for slender transportation facilities such as edge areas and some roads. In contrast, Figure 3b reveals that the area is part of a tropical rainforest climate. Beyond the intertidal zone, substantial evergreen vegetation is distributed along the coastal areas, diminishing the extraction effectiveness. The horizontal and vertical lines of the rules in the figure were left behind during the segmentation of small areas and did not affect the actual process of mangrove extraction and area calculation. In addition, the small river in Figure 3c, similar to the road in Figure 3a, is inaccurately classified. This misclassification stems from the Sentinel-2 series satellites’ 10 m resolution, where narrow roads or rivers can be misclassified due to mixed pixel issues. In Figure 3b, some mangroves on the coastline are classified into seawater or mudflats. In contrast, in Figure 3d, some seawater or mudflats are also classified as mangroves, which can be caused by rising and falling tides. When the tide rises, some low mangroves are submerged by the tide, leading to the phenomenon that the mangroves in Figure 3b are classified into seawater or mudflats. This situation of falling tide is shown in Figure 3d.

This study employed two methods, support vector machine (SVM) and random forest (RF), for accuracy validation. Before conducting the validation, it was essential for both methods to have adequate mangrove and non-mangrove samples to distinguish them between training and testing samples. The selected samples are illustrated in Figure 1, where the green dots indicate mangrove class sample points and the orange dots represent non-mangrove class sample points. This study randomly allocated samples in a ratio of 70% for training and 30% for testing to prevent data overfitting. In addition, only data from the red, near-infrared, NDVI bands, or spectral features were selected as parameter inputs to enhance computational efficiency. Figure 4 indicates that the comparison between Sundarbans Park and its surrounding bare land, grasslands, buildings, and farmland is evident enough that a visual interpretation can be applied for sample selection. The number of mangrove samples selected from each study area for precision validation varied from 52 to 69, while the number of non-mangrove samples ranged from 50 to 68. The number of samples is related to the size of the study area and the difficulty of sample selection, with more samples typically selected from larger and well-categorized study areas. In order to facilitate the comparison of the actual conditions within and outside the study area based on the submeter-level Google satellite images, the vector map was set to be filled with empty values during visualization. The satellite images clearly show that areas with dense vegetation are mangrove distribution areas. For example, during the accuracy verification on Hainan Island, 56 mangrove and 52 non-mangrove sample points were selected through a visual interpretation for accuracy validation. The overall accuracy achieved by SVM was 93.5%, and by RF, it was 94.4%. In addition, cloud cover size is a significant influencing factor; classification accuracy in areas with large cloud cover is usually lower than in areas with minimal cloud cover. Although the differences in results obtained by SVM and RF are not significant, the SVM computations were considerably faster in some cases. Therefore, the accuracy of this study was determined using SVM. The overall accuracy achieved by SVM in one area of Sundarbans Park was 97.1%. The overall accuracy across each region ranged from 86% to 98%, with an average overall accuracy of 92.4%. Areas with high overall accuracy typically exhibited good image quality and lacked complex vegetation structures around the concentrated mangrove distribution areas. The estimated overall accuracy error in this study was 4.93%. The primary sources of error in this study are as follows:

• Sentinel-2 satellite images of each study area are synthesized from multiple images of varying quality using the mosaic algorithm. Even in areas with significant cloud cover, the enhanced mosaic algorithm cannot completely eliminate the cloud interference when combined with the cloud removal function. Cloud-containing areas in the synthesized image are classified into non-mangrove categories, resulting in the incorrect classification of some mangrove forests. In addition, the contrast among different images varies, and the pixel values of some images remain concentrated at a lower level. Without stretching, this can lead to significant differences in the distribution patterns of the same band across different images, adversely affecting sample training.
• Due to mixed pixel issues, certain smaller features may be incorrectly classified as mangroves. Given the insufficient resolution of satellite images, most of these errors cannot be corrected at the algorithmic level. Future research can explore the method of multi-source data fusion to address these issues. For example, DEM data can be used to classify areas with significantly high elevations as non-mangrove forests directly.
• This study is based on the research results of Jia et al. [7] in 2020 and incorporates Galvincio et al.’s [22] mangrove growth rate model and Lagomasino et al.’s [23] mangrove tree height crown width model to determine the maximum increase in mangrove crown width over two years. Hence, a buffer zone was established to define the study area. Although commission errors in the research results of Jia et al. [7] can be corrected through algorithms or manually in the study, it will inherit the omission errors regarding the mangroves located far from the study areas. In addition, this method of determining the study area considers natural growth conditions and fails to address the increase in mangrove areas caused by abnormal factors. This study inherits the omission errors in mangrove classification from Jia et al. [7] and overlooks some mangrove area increases caused by abnormal factors. However, this is unlikely to significantly impact the results of this study, as Jia et al.’s [7] research demonstrated an overall accuracy of 93.6%. In addition, by comparing satellite images of Hainan Island to Jia et al.’s [7] research results, the concentrated distribution of missing mangroves is identified within a one-square-kilometer area near the east coast of Hainan Island. Due to the fact that this omission was only found in no more than 10 locations within this study, with an area of approximately one square kilometer, it can be ignored in global statistics.

Compared to the results at the same resolution level in recent years, the overall accuracy of the global 10 m resolution mangrove forests in 2020, extracted by Jia et al. using Sentinel-2 satellite data, was 93.6% [7]. However, unlike this study, the results of Jia et al. [7] exhibited a larger error in tropical rainforest and tropical monsoon climate zones. This discrepancy is attributed to the fact that Jia et al. [7] constructed a buffer zone of 1 km in their study, which can lead to confusion between mangrove forests and non-mangrove wetland ecosystems [8]. This study also accounted for the total mangrove area in certain countries. Among these, Indonesia has the largest area at 28,614.5 km², followed by Brazil at 11,229.9 km² and Australia at 10,279.8 km². These are broadly consistent with those reported by Jia et al. [7]. It can be observed that Indonesia’s mangrove area is a cliff ahead of the second- and third-ranked countries, which is more than the total mangrove area of the two countries combined. This is because Indonesia is an archipelagic country located in tropical regions with a long coastline and larger undeveloped areas that lack human interaction. On the other hand, countries in tropical desert climates, such as the United Arab Emirates, have maintained a low level of mangrove forest areas despite conservation efforts.

Regarding the selection of sample points, Jia et al. [7] integrated a publicly available mangrove dataset with corrections based on the visual interpretations of submeter satellite images within GEE. In contrast, this study utilized submeter satellite images provided by GEE to adjust the 2020 mangrove vector map by removing misclassified mangroves and incorporating significantly omitted mangrove forests into the 2022 vector map. Additionally, the vector maps of mangroves on each continent were projected, calculated geometrically, and aggregated to determine the global mangrove area. The initial statistics indicated that the total global mangrove area was 116,689.6 km², with the adjusted area totaling 124,255.2 km². Asia has the largest mangrove area, whereas Oceania has the smallest. North America, Africa, and South America are ranked second, third, and fourth, respectively, which is consistent with Jia et al.’s findings [7]. The proportion of mangrove forest areas on each continent and a bar chart of mangrove forest areas ranking the top seven countries are illustrated in Figure 5. It is important to note that some errors in the statistics arise due to imprecise continental boundaries. For instance, the mangrove forests on New Guinea Island and its surrounding islands in Oceania were classified under Asia, resulting in a smaller reported proportion of mangrove area for Oceania than the actual proportion. However, from a global point of view, the proportion obtained from the statistics is basically consistent with the distribution of mangrove forests in each continent.

The synthesized global mangrove vector map reveals that mangroves in Asia are mainly located in Southeast and South Asia, with fewer distributions in West Asia. North American mangroves are concentrated around the peninsula and archipelago near Central America and the Caribbean Sea. African mangroves are primarily distributed along most tropical coastlines, except for Somalia and on the island of Madagascar, with very little distribution in North Africa. South America’s mangroves are concentrated along the continent’s eastern coast and near Central America. The Australian mainland nearly encompasses all mangrove forests in Oceania, with the remainder primarily in larger tropical archipelago nations such as the Solomon Islands and Papua New Guinea.

The total area of mangroves worldwide in 2020 was approximately 119,334.6 km², as provided by Liao et al. [10]. Based on this, the total area of mangroves increased by 4920.6 km² from 2020 to 2022. In addition, the Sentinel-2 satellite data used in this study, with a resolution of 10 m, offers greater detail compared to the commonly used 30 m resolution of Landsat 8 satellite data, hence providing richer details. For instance, in areas with high concentrations of aquaculture ponds or rivers, smaller water bodies can be misclassified as mangrove forests in the Landsat 8 imagery. This study, by comparing the extracted results with the original data, found instances in the 2020 mangrove vector map where land categories such as rivers, bare land, or buildings were incorrectly classified as mangrove categories (Figure 6).

This study obtained a global mangrove vector map for the year 2022 only; thus, it was necessary to incorporate the findings from previous research for temporal analysis. The studies referenced include those by Goldberg et al. [26], Liao et al. [9], and Ragavan et al. [27]. Goldberg et al. [26] concluded that over the 4-year period from 2016 to 2020, the global mangrove area decreased by a total of 3363 km², accounting for 2.1% of the total mangrove area in 2016. The primary drivers of changes in mangrove areas were human activities and natural factors. Human activities have always been the main factor affecting changes in mangrove areas. According to Goldberg et al. [26], the reduction of mangrove area caused by human activities accounted for 62% of the total decrease during 2016–2020, with activities related to the commodity economy and agricultural deforestation being as high as 47% of the decrease. Indeed, as depicted in Figure 7, the transformation of mangroves into buildings and aquaculture ponds due to human development is visible in satellite images. However, as shown in Figure 4a, the mangrove forest captured in the satellite imagery in 2022 clearly extends beyond the boundaries of the mangrove forest extracted in 2020. On the other hand, although global mangrove forests have shown a decreasing trend according to statistics such as Goldberg, there is also an increase in the local areas where no obvious human activities are observed. In addition, combining this with Liao et al.’s [9] research, it is evident that although the mangrove area in 2022 has significantly decreased compared to Hamilton et al.’s [3] statistics from 2000, the rate of reduction is slowing. Further observation and analyses are necessary to determine whether the area of mangroves will maintain a dynamic equilibrium or continue to recover in the future. Area cannot be used as a measure of the productivity of mangrove ecosystems. It needs to be combined with other biological data such as biomass to accurately reflect the productivity of mangrove ecosystems.

In addition to human activities, natural disasters also play a crucial role in the changes in mangrove areas, including phenomena such as earthquakes and typhoons. Goldberg et al. [26] also analyzed the natural factors contributing to mangrove area changes. They identified two main factors: coastal erosion and extreme weather, which accounted for 27% and 11% of the total area reduction, respectively. However, they did not analyze factors such as volcanic earthquakes in offshore and coastal zones. Although this study used a binary classification approach, which enables rapid and efficient mangrove extraction, it does not allow for an accurate calculation of the transfer and change relationships between mangroves and various land features. Nonetheless, by comparing the extraction results with satellite images, Goldberg et al.’s findings can be supported to some extent [26]. For example, as illustrated in Figure 7, there are visible traces of mangrove forests transforming into bare land or buildings in the past two years.

Ragavan et al. [27] studied the distribution of mangroves across India, identifying that nearly half of these species are either rare or endangered. According to their findings, consistent monitoring of mangrove forests in India is essential. Additionally, this study highlighted a significant conversion of mangrove areas to construction land in South and Southeast Asia, which has a serious impact on rare or endangered species. When comparing the results of mangrove extraction with Jia et al.’s 2020 global mangrove vector map [7], it was found that in some areas of Southeast Asia, such as near the city of Bandar Seri Begawan, the reduction in mangrove forests is more severe. Conversely, the mangrove area near the Irrawaddy River mouth in Myanmar exhibited a slight increase. Despite the establishment of national parks and mangrove reserves by many countries, the area of mangrove forests continues to decline in numerous regions. Clearly, protecting mangroves remains a long-term and arduous task.

Compared with previous investigations, this study draws on or utilizes previous research findings, such as Jia et al. [7] and Bunting et al. [8]. It also combines the growth rates of mangroves and the relationship between mangrove height and crown width to construct a suitably sized buffer zone as the study area, keeping the data size within a reasonable range, which can efficiently and accurately extract mangroves. In addition, this study divided the global research area into thousands of small sections, applying the commonly used Otsu algorithm for small-scale mangrove extraction. This method achieved automatic threshold extraction in each small area, significantly improving computational efficiency. The research also used submeter-level Google satellite imagery to manually correct errors in the extraction results and refine their accuracy. Finally, the employment of Sentinel-2 series satellite imagery as raster data in the research allowed the spatial resolution of the extraction results to increase from 30 m to 10 m. This enhancement enabled the effective distinction of narrow features, such as rivers and roads, which were not identifiable in earlier 30 m resolution datasets, thus improving data quality.

This study identified several minor issues that can be addressed in the future. For example, although the Otsu algorithm efficiently distinguishes between targets and backgrounds based on pixel-level applications, integrating other features for object-oriented classification might enhance the quality of the research outcomes. Comparisons between submeter-level images and vector boundaries in Figure 5 and Figure 7 reveal misclassifications of some buildings and small rivers as mangroves. This issue may be due to mixed pixels or an inadequate selection of features to distinguish between mangrove forests and minor rivers. Addressing mixed pixels at the sensor level and improving classification accuracy through object-oriented classification with sufficient features could overcome these issues. Moreover, the sole reliance on samples from national parks or protected areas limits the verifiability of sample accuracy.

The area of global mangrove forests showed a decreasing trend but started to recover in recent years. In previous years, the area of mangroves decreased significantly. Recently, with the continuous strengthening of human protection and restoration of mangroves, the mangrove area has tended to stabilize and even expand. Human activities have always been a primary factor in the reduction of mangrove areas. In addition, if more accurate statistics on the area of mangrove forests are needed, other observation data, such as DEM, may also be required. Nonetheless, determining whether the global mangrove area will continue to expand or remain dynamically stable requires extensive time-series observational data.

5. Conclusions

The methods and data utilized in this study allow for the following conclusions to be drawn through statistical analysis:

(1)

Firstly, this study explores three methods for image mosaic: mosaic, median, and mean. The mosaic and median methods effectively mitigate the influence of residual clouds on the synthesized results without damaging the original data integrity. However, in areas with heavier cloud coverage, the mean method diminishes the synthesis quality, and the resultant image pixel values are typically non-integers, which can degrade the original data. Hence, the mosaic method was employed for image pre-processing in this study. The Otsu algorithm can automatically calculate thresholds for binary classification problems efficiently. This algorithm facilitates mangrove extraction from medium- and high-resolution satellite data sources represented by Sentinel-2 on a global scale. By selecting suitable spectral indices and using the Otsu algorithm for binary classification, not only mangrove forests but also Spartina alterniflora can be accurately identified, ensuring the reliability of the results. In smaller regions, the overall accuracy can exceed 90%. Applying the Otsu algorithm to global mangrove extraction enhances automation and efficiency, providing insights for resolving global binary classification problems.

(2)

In 2022, the total area of global mangrove forests increased by 4920.6 km² compared to 2020. Previous studies have indicated that the decrease in the area of mangrove forests was primarily influenced by human activities. However, in recent years, extreme weather events have increased due to global climate change caused by human activities, and the change of mangrove areas induced by natural factors has also increased. Hence, monitoring the changes in mangrove areas can be essential for monitoring mangrove area variations and provides a necessary basis for future research on mangrove resource conservation and global climate change.

(3)

The global 10 m resolution mangrove vector map produced by this study supports the monitoring of global mangrove area changes in the forthcoming years and holds significant relevance for guiding efforts to protect and restore mangrove ecosystems. In addition, this study contributes data support for subsequent global mangrove extraction research, which can be applied to the temporal analysis of global mangrove area changes.

In addition, this study also identified and discovered the following issues, which may be addressed in future research:

(1)

Lack of detail in data processing. Due to the sheer volume of data, long hours of manual interaction are required to solve such problems, and the manual interaction does not fully resolve the details. In addition, the synthesis of global vector maps is time-consuming and laborious, and the data memory size exceeds the upper limit of software processing, which makes it difficult to complete the work. Although the difference between the error occurrence area and the actual area after the manual interaction method in the study is very small, it is not representative of the actual area, and the results of the accuracy verification of the error occurrence area may not be completely accurate as a result.

(2)

The number of categories is low. This study only extracted data related to mangrove forests, omitting other categories. This limitation hinders the analysis of the relationship between changes in mangrove forests and other categories, and only changes in global mangrove areas can be statistically obtained. Analyzing the drivers behind mangrove area changes and implementing mangrove development and protection measures thus becomes challenging.

(3)

Lack of measured data points on a global scale. The absence of actual measurement points and other factors, such as data source limitations, introduces uncertainty in the accuracy verification of global mangrove monitoring. Therefore, establishing permanent mangrove monitoring stations globally in each mangrove nature reserve or national park is crucial.

Establishing permanent mangrove monitoring stations worldwide is significant for monitoring global mangroves. The results of this study will continue to support the monitoring of global mangrove areas and aid in the creation of permanent mangrove sites globally, thus contributing to the conservation and restoration of mangrove ecosystems.