*Article* **Spectral Reflectance-Based Mangrove Species Mapping from WorldView-2 Imagery of Karimunjawa and Kemujan Island, Central Java Province, Indonesia**

**Arie Dwika Rahmandhana 1, Muhammad Kamal 2,\* and Pramaditya Wicaksono <sup>2</sup>**


**Abstract:** Mangrove mapping at the species level enables the creation of a detailed inventory of mangrove forest biodiversity and supports coastal ecosystem management. The Karimunjawa National Park in Central Java Province is one of Indonesia's mangrove habitats with high biodiversity, namely, 44 species representing 25 true mangroves and 19 mangrove associates. This study aims to (1) classify and group mangrove species by their spectral reflectance characteristics, (2) map mangrove species by applying their spectral reflectance to WorldView-2 satellite imagery with the spectral angle mapper (SAM), spectral information divergence (SID), and spectral feature fitting (SFF) algorithms, and (3) assess the accuracy of the produced mangrove species mapping of the Karimunjawa and Kemujan Islands. The collected field data included (1) mangrove species identification, (2) coordinate locations of targeted mangrove species, and (3) the spectral reflectance of mangrove species measured with a field spectrometer. Dendrogram analysis was conducted with the Ward linkage method to classify mangrove species based on the distance between the closest clusters of spectral reflectance patterns. The dendrogram showed that the 24 mangrove species found in the field could be grouped into four levels. They consisted of two, four, and five species groups for Levels 1 to 3, respectively, and individual species for Level 4. The mapping results indicated that the SID algorithm had the highest overall accuracy (OA) at 49.72%, 22.60%, and 15.20% for Levels 1 to 3, respectively, while SFF produced the most accurate results for individual species mapping (Level 4) with an OA of 5.08%. The results suggest that the greater the number of classes to be mapped, the lower the mapping accuracy. The results can be used to model the spatial distribution of mangrove species or the composition of mangrove forests and update databases related to coastal management.

**Keywords:** mangrove species; spectrometer; spectral reflectance; WorldView-2; dendrogram

#### **1. Introduction**

Indonesia is a global ecological hotspot, judging from the extent and rich biodiversity of its mangrove ecosystem. Bunting et al. [1] mapped and reported the latest data on the world's mangrove area based on a combined analysis of ALOS PALSAR radar images and optical Landsat 5 TM and Landsat 7 ETM+ images taken between 2009 and 2011. The estimated global mangrove area is approximately 137,600 km2, and Indonesia has the largest mangrove forest, with a total area of 26,890 km2. It accounts for 19.5% of the mangroves worldwide and 50.4% of those in Asia. Indonesia is also estimated to contain 43 of around 75 true mangrove species in the world [2,3] or about 57% of all mangrove species worldwide. The characteristics of mangrove forests generally differ from those of mainland forests. For instance, their habitats are not climate-dependent but are shaped by tides, the extent of seawater-inundated soils, elevation, and the presence of canopy structures. They especially thrive on low-lying land without canopy structures [4]. Illegal

**Citation:** Rahmandhana, A.D.; Kamal, M.; Wicaksono, P. Spectral Reflectance-Based Mangrove Species Mapping from WorldView-2 Imagery of Karimunjawa and Kemujan Island, Central Java Province, Indonesia. *Remote Sens.* **2022**, *14*, 183. https:// doi.org/10.3390/rs14010183

Academic Editors: Brigitte Leblon and Chandra Giri

Received: 15 November 2021 Accepted: 30 December 2021 Published: 1 January 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

logging and the creation of ponds are currently degrading mangrove forests. For example, Ilman et al. [5] reported that the estimated decline in mangrove area in Indonesia due to land clearing for ponds and logging between 1800 and 2012 was 913,000 hectares. Put another way, the average rate of decrease in mangrove area is 4300 hectares per year. The conversion of mangrove forests for aquaculture, tourism, and agricultural purposes has disrupted ecosystem stability and reduced physical and biological mangrove functions, affecting the existence of vulnerable mangrove species that are rare or limited.

The Karimunjawa Islands in Jepara Regency, Central Java, Indonesia, are a mangrove ecosystem with high species diversity. This ecosystem is relatively undisturbed and well-preserved because most of its area lies within the Karimunjawa National Park. Karimunjawa and Kemujan Island (two members of the Karimunjawa archipelago) have 3.964 km2 of mangrove forests under the management and protection of the national park. A total of 25 true mangroves species have been recorded in this archipelago [4]. The inventory of mangrove distribution in the Karimunjawa National Park in 2002 found 44 mangrove species, including both true mangroves and their associates. Due to the widespread land conversion and logging in the mangrove area and its surroundings, mangrove species must be regularly mapped to maintain reliable information about the biodiversity of coastal ecosystems.

Cruising, a necessity for mapping activities, can be more challenging in mangrove areas than terrestrial forests. An alternative to cruising is needed to map and monitor the distribution of mangrove species efficiently and thoroughly. Remote sensing technology is an effective substitute in this context because it can minimize fieldwork and reduce the requisite time, cost, and effort of mapping. It is particularly cost- and time-efficient compared with field sampling, plant identification, and vegetation classification [6]. Furthermore, using remote sensing products, mainly satellite imagery, allows past conditions of the observed area to be recorded, enabling multi-temporal analyses. Satellite imagery also provides a synoptic overview of a large expanse, which is efficient for large-extent studies of the Earth's surface.

Mangroves differ from non-mangrove vegetation. They can be identified from specific tones and colors in remote sensing images and their association with coastal areas [7,8]. Healthy mangroves with green chlorophyll have low spectral reflectance in the blue and red bands, high in the green band, and significantly high in the near-infrared band; the reflectance increases with decreased water content in leaves [7]. Visually, true-color composite images for mangroves show a darker green color than vegetation in general because of the absorption of mangrove substrates in the near-infrared band. In satellite imagery, spectral reflectance allows tropical mangroves to be distinguished at the species level on a laboratory scale [9]. The accurate mapping of mangrove species requires images with high spatial and spectral resolution and the spectral reflectance pattern of each mangrove species to be measured in the field.

The spectral reflectance values are measured in the field using a hand-held spectrometer with high spectral detail that accommodates mangrove species mapping. They are used as the basis for a spectral library of mangrove species that comprises endmembers for mapping with a pixel-based classification algorithm. Classification is necessary to create visual results. Furthermore, the grouping of mangrove species based on spectral characteristics entails a cluster analysis to determine the optimal number of classes to be mapped. This study explores the use of three classification algorithms, spectral angle mapper (SAM), spectral information divergence (SID), and spectral feature fitting (SFF), to map the mangrove distribution on Karimunjawa and Kemujan Island. Specifically, the objectives of this study were to (1) classify and group mangrove species by their spectral reflectance characteristics, (2) map their spatial distribution with field-measured spectral reflectance and by applying the SAM, SFF, and SID classification algorithms to WorldView-2 images, and (3) assess the classification accuracy of each algorithm.

#### **2. Materials and Methods**

#### *2.1. Study Site*

The research area included Karimunjawa and Kemujan Island of Indonesia (5◦49 33– 5◦48 23 S, 110◦24 34–110◦28 37 E). Their mangrove ecosystems comprise highly diverse mangrove species and are conservation areas in the Karimunjawa National Park, making them pristine and well-preserved. These two islands are located in the Karimunjawa District of Jepara Regency, Central Jawa Province, about 85 km north of Java Island (Figure 1). Several environmental factors that influence the growth of mangroves at the research site are the coastal physiography (topography), tides (length, duration, and range), waves and currents, climate (light, rainfall, temperature, and wind), salinity, dissolved oxygen, soil, and nutrients [10]. Based on the Schmidt and Ferguson climate classification, Karimunjawa and Kemujan Island have a C climate type with average rainfall of 3000 mm/year and air temperatures of 30–31 ◦C. Topographically, the Karimunjawa National Park area is located 0–506 m above sea level, spanning from flat coastal areas to undulating lowlands [11]. Hilly regions stretch from the east (highest peak) to the west and from the middle to the south. Karimunjawa experiences mixed tides with a prevailing diurnal pattern, i.e., one high and low tide but sometimes two high and low tides each day. The lowest low water level (LLWL) of the island waters is 20 cm, and the highest high-water level (HHWL) is 138 cm, making the maximum tidal range—the difference between the HHWL and LLWL—118 cm; this is classified as micro-tidal [12]. Karimunjawa and Kemujan Island have dark gray grumusol developed from quartz sandstone, micaceous sandstone, and quartz siltstone, and a few coral fragments included in mixed substrates (sand substrates and mixed gravels), with clay deposits along the coast. The soil physical properties on the two islands are fairly similar [11].

**Figure 1.** Study area map—The Karimunjawa and Kemujan Islands displayed on a Landsat 8 OLI color composite image with RGB 546 band combination.

The Karimunjawa archipelago has five types of ecosystems: coral reefs, mangroves, seagrass beds, lowland tropical rainforests, and coastal forests. Mangrove ecosystems can be found on the coasts of Karimunjawa, Kemujan, Cemara Kecil, Cemara Besar, Krakal Kecil, Krakal Besar, Merican, Menyawakan, and Sintok Island, over a total area of 3.964 km2. There are 25 species (13 families) of true mangroves and nine species (seven families) of mangrove associates within the national park area, and five species (five families) of mangrove associates outside the national park [4]. Most of the mangroves on Karimunjawa and Kemujan Island are under the Karimunjawa National Park's jurisdiction, and the community manages the rest. In general, the park consists of eight zones: core, wilderness, marine protection, marine tourism use, historical-cultural-religious use, rehabilitation, marine use, and traditional fisheries. The research sites are located in the wilderness and marine use zones. The mangrove area on Kemujan Island is used for trekking tours through the forest and bird watching towers.

#### *2.2. Image Data and Processing*

The primary data source in this study was the WorldView-2 (WV-2) image covering parts of Karimunjawa and Kemujan Island. It was acquired on 27 June 2020, with a 2 m spatial resolution (multispectral sensor) and eight multispectral bands as described in Table 1 [13]. The image was selected to match the acquisition time to the fieldwork that was conducted at the end of the rainy season and ensure a cloud-free image. In this study, image pre-processing was required to extract the spectral signatures of the targeted objects. The first part of image pre-processing was a radiometric correction to convert the image pixel value from digital numbers into at-sensor reflectance with Updike and Comp's [14] procedure. The second part involved producing pixel values from the at-surface reflectance with fast line-of-sight atmospheric analysis of hypercubes (FLAASH) [15], an atmospheric correction model for the WV-2 image. The atmospheric visibility parameter was estimated from the moderate-resolution imaging spectroradiometer (MODIS) aerosol product [16].

**Table 1.** Band characteristics of WorldView-2 imagery [13].


Differences in the radiometric correction results at each level are visible in the object's spectral reflectance curve (Z profile), presented in Figure 2. The spectral reflectance values in bands 1 (coastal) and 2 (blue) at the DN level (Figure 2a) decreased after correction at the surface reflectance level (Figure 2c), meaning that the atmospheric effects in the image had been corrected. The highest spectral reflectance values were in the red edge, near-IR1, and near-IR2 bands; in other words, the vegetation objects (mangroves) showed high reflectance values in these three bands. The spectral reflectance of mangroves forms a pattern similar to vegetation in general: low in the visible bands and high in the infrared bands (near-IR1 and near-IR2). This means that the radiometric and atmospheric corrections had been successfully applied to the WV-2 image.

The WV-2 image was visually interpreted to distinguish between mangrove and nonmangrove objects. True-color (RGB; 532) and false-color (near-IR1, red, blue; 752) composite images were used in this process, and mangrove and non-mangrove objects were identified based on their colors, tones, textures, and associations. The false-color composite image (752) depicts any vegetation features in red. However, it does not distinguish between mangroves and non-mangroves in areas formerly used as fish ponds. The true-color image composition was added to assist in this discrimination process. Mangrove objects can also be identified from their associations with surrounding objects, i.e., mangroves are located in coastal areas or border the sea. This mangrove object delineation output was then inputted for image masking to separate mangrove objects from others.

**Figure 2.** Spectral reflectance curves of mangrove objects at three correction levels: (**a**) digital number, (**b**) at-sensor radiance, and (**c**) at-surface reflectance.

#### *2.3. Field Data Collection*

The spectral reflectance values of the targeted mangrove species were collected directly from the sampling points during 5–10 March 2021. The number and locations of field samples were selected purposively based on the Pixel Purity Index (PPI) values [17] and the sites' accessibility. The PPI, in this case, was used to locate the purest mangrove pixels in the WV-2 image for spectral reflectance collection. The PPI image was produced from the minimum noise fraction (MNF) algorithm to remove the noise in the data and reduce the computational requirements for further processing. This research used 10,000 iterations because, according to Plaza and Chang [18], PPI produces pure pixels after 10,000 to 100,000 repetitions. These samples were then plotted onto a map that was later used to read the spectral reflectance of mangrove species in the field.

The field data collection resulted in 201 point samples covering 24 targeted mangrove species (Table 2). The distribution of the field samples is presented in Figure 1. The field samples were divided into two major groups: modeling samples for developing the spectral library (24 samples, Figure 3a) and validation models for the accuracy assessment of the resulting map (177 samples, Figure 3b). The modeling samples were spectral reflectance mangrove species collected in the areas with high species diversity. This research only focused on the true mangrove species, and researchers were assisted by park managers or rangers familiar with the targeted species' location to collect these samples. The validation samples were mangrove species identified in the field. The validation sample locations were selected from "white pixels" (i.e., pure pixels) derived from the PPI calculation. In addition to PPI, the validation samples were also determined with aerial photos to identify a mix of two species (*Avicennia marina* and *Ceriops tagal*) found in the field.



The field spectral reflectance values of mangrove species were measured with a JAZ EL-350 portable spectrometer from Ocean Optics (https://oceanoptics.com/, accessed on 14 November 2021). First, the white and dark references were measured. A white reference produces a standard "white object" spectral reflectance reading to calculate the object's spectral sample, while a dark reference produces a reading of an object with perfect absorption to create a "black body" reference [19]. Second, the spectral reflectance values of mangrove species were measured at the leaf level (Figure 3a), based on the criterion that healthy leaves appeared entirely green or contained chlorophyll.

**Figure 3.** Field data collection: (**a**) spectral reflectance measurement of the targeted mangrove species and (**b**) field validation sample collection.

Some important aspects to consider in collecting spectral reflectance data for mangrove species in the field are (1) the field of view (FOV) of the spectrometer sensor, (2) the distance between the spectrometer and the targeted object, (3) the angle and direction of measurement, and (4) the light conditions at the time of observation [19]. Kamal et al. [20] collected the spectral reflectance of *Rhizophora stylosa* on Karimunjawa Island at 2 cm, 50 cm, 1 m, 2 m, and 5 m distances with ten readings at each distance. According to this study, the spectral reflectance curves recorded at close range to the leaf (i.e., 2 cm) and from the furthest distance (i.e., 5 m) showed the lowest curve variation between readings. Therefore, this study used a distance of 2 cm for leaf spectral measurements to ensure that only the mangrove leaf was read (Figure 3a) and ensure highly consistent spectrometer results across readings [20,21]. The spectrometer's measurement angle was set at 45◦ to the nadir and facing the sun to avoid shadows on the target object. In addition, the number of spectrometer reading repetitions affects the degree of confidence in the reading results. The leaf measurements were repeated six times to ensure the consistency of the readings in similar natural lighting conditions in the field.

The field-measured spectral reflectance profiles were stored in a file in .jaz format. A .jaz file contains reference spectrum data (white reflectance), dark reflectance, and the targeted object spectral reflectance data, which are used to construct the object's spectral reflectance curve. The six replicates used to measure each mangrove species object avoided errors such as saturation in the spectral reflectance measurement, which increased the spectrometer readings' precision. The field measurements produced data on objects' light intensity at a particular wavelength; thus, the reflectance values needed to be calculated. These values were normalized and calculated with the formula described in Kamal et al. [19] and Wicaksono et al. [22]. The normalized and mean values were then used as input to build a spectral library.

#### *2.4. Mangrove Species Clustering Analysis*

White and dark reference readings were collected for each measurement of mangrove species samples using the spectrometer to normalize the spectral reflectance of each mangrove species and create a standard range of spectral reflectance values. This allows the derived spectral reflectance curves of different mangrove species to be compared directly. Mangrove species' spectral reflectance can be analyzed effectively at wavelengths of 350–900 nm with a JAZ EL-350 spectrometer [20]. This wavelength range was selected because, based on the specifications of the spectrometer used, noise is likely to occur below 350 nm and above 900 nm. The mangrove species spectra compiled in the spectral library were then resampled according to the center wavelength of the WV-2 image bands and used to develop an optical dendrogram. An optical dendrogram determines how components (i.e., mangrove species) are grouped spectrally. It is also used to identify similarities and cluster distances between species or groups. The optical dendrogram for this study was created in the IBM SPSS Statistics 24 program based on Wicaksono et al.'s [22] work, which used seagrass as the research object and the Ward linkage method. The derived dendrogram was then used as the basis for the mangrove species classification scheme in remote sensing-based mapping.

The dendrogram was built using the Ward linkage method as described by Wicaksono et al. [22]. It analyzes clusters hierarchically by determining the distance between two clusters expressed as an increase in the "error sum of squares" (ESS). The Ward linkage method selects grouping steps sequentially to minimize the ESS at each step. The dendrogram was not developed from the field-measured spectral reflectance but rather the resampled data; therefore, eight wavelengths were selected to represent the eight WV-2 image bands. The resampled spectral reflectance values were inputted in the dendrogram for further use in pixel-based mapping. This dendrogram identifies similarities and cluster distances between mangrove species, producing several large groups according to the combination of cluster distances.

#### *2.5. Pixel-Based Classification and Accuracy Assessment*

The field-collected spectral samples were extracted and divided into three types: white reference, dark reference, and the object's spectral reference. This extraction was performed to produce the spectral reflectance curves of mangrove species for a spectral library, i.e., a collection of references containing the spectral reflectance values of various objects [19,20]. A normalization process was conducted to obtain the appropriate spectral reflectance curves allowing for easy understanding and high representativeness. The extracted spectral reflectance samples of mangrove species were then collected in one container to be inputted into a spectral library in the ENVI 5.3 program. The compiled spectral library of mangrove species from field measurements was spectrally resampled to align with the spectral resolution of the WV-2 image as the basis of mangrove species mapping.

The mangrove species were mapped in ENVI 5.3 using the spectral library from the resampled spectra as the classification input or endmember. For this purpose, three pixel-based classification algorithms were applied and evaluated for mapping mangrove species, namely the spectral angle mapper (SAM), spectral information divergence (SID), and spectral feature fitting (SFF). The SAM algorithm can classify objects based on their spectral reflectance values by considering the illumination angle reflected by the object. SAM determines the similarity between two spectral reflectance objects by calculating the "spectral angle" created between them and treating them as vectors in a space with dimensionality equal to the number (n) of image bands used [23,24]. Because it only factors in the direction of the spectrum, the spectrum length and other factors are considered equal. It determines the similarity of the unknown spectrum *t* to the reference spectrum *r* using Equation (1), below:

$$\mathfrak{a} = \cos^{-1}(t \cdot r ||t|| \, ||r||) \tag{1}$$

The SID algorithm can determine the target pixel based on differences in the object's spectral reflectance information. It calculates the information difference between the target pixel (*r*1) and the target reference (*r*2) from the sum of the relative difference between *r*<sup>1</sup> and *r*<sup>2</sup> (D(*r*<sup>1</sup> *r*2)), and the relative difference between *r*<sup>2</sup> and *r*<sup>1</sup> (D(*r*<sup>2</sup> *r*1)). This algorithm is presented in Equation (2) below [25]:

$$\text{SID}(r\_1, r\_2) = \left(\text{D}(r\_1||r\_2)\right) + \left(\text{D}(r\_2||r\_1)\right) \tag{2}$$

Finally, the SFF algorithm compares the spectral image with the endmember reference. The reference spectra were scaled to match those of the image after the continuum was removed from both datasets (reference and object spectra) [26].

The spectral reflectance of each mangrove species was measured in the field simultaneously with the validation samples. The validation samples are different datasets from those used to map the mangrove species. They contain mangrove species data and their coordinate locations in the field. However, both modeling and validation samples were determined with the same purposive sampling technique because of high object heterogeneity in the field and low accessibility. Both sets of samples were also selected with the same criteria, i.e., the PPI calculation results.

The accuracy assessment was conducted with a confusion matrix to measure the extent to which the mangrove species classification results from field-collected and postfield-processed data were similar. In addition to percent accuracy, it also evaluated each algorithm's misclassification by observing the presence or absence of a logical error in the species classification. This accuracy assessment resulted in user's accuracy (UA), producer's accuracy (PA), and overall accuracy (OA) values, following Congalton and Green's procedures [27].

#### **3. Results**

#### *3.1. Spectral Reflectance of Mangrove Species*

The spectral library for each species was developed by displaying the spectral reflectance curve of the replicated samples with their averages and observing the corresponding curves of all species. The normalized spectral reflectance of 24 mangrove species (10 of primary data and 14 of secondary data) were combined into one spectral reflectance curve, as shown in Figure 4a. The field-measured spectral reflectance has a very high spectral resolution (1586 bands) while that from the WV-2 image has a low resolution (eight bands). Therefore, spectral resampling was conducted to match the spectral resolution derived from the spectrometer to the WV-2 image bands. The center wavelengths of the WV-2 image obtained from the "Radiometric Use of WorldView-2 Imagery" guidelines by Digital-Globe [13] were used as the targeted spectra in the spectral resampling process. The center wavelength for each WV-2 band is as follows: 427 nm (coastal), 478.3 nm (blue), 545.8 nm (green, 607.7 nm (yellow), 658.8 nm (red), 724.1 nm (red-edge), 832.9 nm (NIR1), and 949.3 nm (NIR2). Besides degrading spectral resolution, spectral resampling also simplified the spectral reflectance of objects measured in the field. In general, both field-collected and resampled data had the same pattern: low in the coastal, blue, and red bands, slightly high in the green band, high in the red-edge band, and very high in the near-IR1 and near-IR2 bands (Figure 4a,b).

**Figure 4.** Spectral reflectance curves of the field-measured data (**a**) and the resampling results based on WorldView-2 bands (**b**) for the 24 mangrove species found in the study site.

#### *3.2. Clustering Analysis of Mangrove Species*

The clustering of mangrove species was conducted based on the resampled mangrove spectra. Indirectly, this clustering also grouped several morphological features of plants, especially their leaves. Klanˇcnik and Gaberšˇcik [28] explained that every leaf shape has structural characteristics with distinctive optical properties. The redundancy analysis showed that the leaves' morphological and biochemical characteristics had particular relationships with leaf spectral reflectance. In general, the most dominant plant part that can be identified by satellite imagery is the leaf canopy. In this study, the spectral reflectance approach used as input in the clustering analysis indirectly captured some of the morphological characteristics of the mangroves, especially the leaves.

Based on the resulting dendrogram (Figure 5), there were four possible classification schemes. The cluster distance for the most distinguishable class was 25, while the cluster distance for the least distinguishable class was 0. The dendrogram showed four levels with different numbers of species groups. Level 1 (two groups), Level 2 (four groups), and Level 3 (five groups) can be used as references in pixel-based mangrove species mapping because they have a relatively large distance between clusters (Table 3). Meanwhile, the Level-4 classification scheme cannot be used for this purpose because the distances between clusters are too small. The results of this species clustering are used to design scenarios for mangrove species mapping using the WV-2 image.

**Table 3.** Mangrove species grouping of Levels 1–4 resulting from the cluster analysis.


**Figure 5.** Dendrogram of mangrove species obtained with the Ward linkage method showing four levels of species grouping.

#### *3.3. Pixel-Based Classification*

The classification mapping was based on the dendrogram analysis results, which produced four levels of mangrove species groups. First, the spectral reflectance curves for each group of mangrove species at each level of the dendrogram were averaged. Then, these results were inputted into the three algorithms of mangrove species mapping. The SAM- and SID-based classifications used default thresholds of 0.1 and 0.05, respectively. Because there was no default threshold for the SFF, the threshold value was set at 10. For each dendrogram level, the threshold values of the classification algorithms were made equal to reduce user intervention and maximize software performance. The mapping results of using these three classification algorithms are presented in Figure 6. Overall, the SAM-based mangrove species classification results at Levels 1, 2, and 3 could not classify all mangrove areas, as was evident from the extensive gray areas in the mangrove delineation results. Some of the mangrove areas were unclassified; this could be due to an inappropriate threshold value for the classification. The front/distal formation bordering the seawater was generally dominated by *Rhizophora* groups and some *Bruguiera* individuals. The SAM algorithm found *Pemphis acidula* in the distal formation, but this species was not entirely classified in the Level-3 mapping. Meanwhile, the Level-1 mapping with the SID and SFF algorithms classified the distal formation as Group B with eight species, three of which were *Bruguiera gymnorrhiza*, *Rhizophora lamarckii*, and *Rhizophora apiculata*. According to

field reports, *Rhizophora* groups and some *Bruguiera* individuals generally dominated the distal formation bordering the sea, meaning that *Rhizophora apiculata* is highly likely to have been among the species classified.

**Figure 6.** Mangrove species mapping results using the SAM, SID, and SFF algorithms at group Levels 1–4.

The Level-2 SID-based mapping showed Group A as the dominant species constituent. Group A consisted of several understory genera, such as *Achantus*, *Acrosticum*, *Aegiceras*, and *Scyphiphora* and some trees, such as *Bruguiera sexangula*, *Bruguiera cylindrica*, *Ceriops tagal*, *Rhizophora mucronata*, and *Rhizophora stylosa*, located in the medial and proximal formations. Meanwhile, the distal formation was dominated by Groups B and D, creating a mixed pattern. In contrast, the Level-2 SFF-based classification showed that the distal formation was dominated by Group C, consisting of *Rhizophora lamarckii* and *Sporomusa ovata*. However, these results were inaccurate as these two species cannot grow large and are rarely present in tidal areas but instead reside in the medial and the proximal areas. Meanwhile, the SID- and SFF-based classifications at Level 3 also showed significant differences in species composition; the former showed Group B predominating and the latter, codominant Groups A, C, and D.

The individual species mapping results (Level 4) differed from the other three classifications. The SAM and SID algorithms classified *Pemphis acidula* as the dominant species in the distal formation, although it is rarely found in tidal areas such as this distal formation, typically growing under the canopy of the proximal formation. This causes a visual classification error and subsequently affects the accuracy of the classification mapping using both algorithms. Meanwhile, SFF classified *Bruguiera sexangula* as the dominant species in the distal formation. It is commonly found in tidal areas and is associated with *Rhizophora* but not as a major species with broad distribution. The SFF-based classification suggested the presence of *Bruguiera sexangular* along the tidal area, another imprecise result that affects the mapping accuracy value.

#### **4. Discussion**

#### *4.1. Mangrove Species Clusters*

Mangrove species clustering can adopt plant morphological approaches commonly used in compiling taxonomies. For example, leaf shape, canopy shape, stem characteristics, and plant habitat characteristics such as salinity level, type of substrate, and length of tidal inundation can be used. However, the spectral reflectance curve of the mangrove species object on the WV-2 image was integrated with the field-measured curve for this study. The result of the spectral reflectance resampling was then inputted into the clustering in dendrograms for pixel-based mapping classification.

The results of the dendrogram analysis showed four levels with varying numbers of species groups (Figure 5 and Table 3). Level 1 (2 groups), Level 2 (4 groups), and Level 3 (5 groups) could be used as references for the pixel-based mapping of mangrove species because the clusters were quite far apart. In contrast, Level 4 could not be used for the same purpose because the clusters were too close together. The Level 1 scheme had a cluster distance of 25, resulting in two large clusters in which Bruguiera, Lumnitzera, Rhizophora, Sonneratia, and Xylocarpus were divided into two groups. In the Level 2 scheme with a cluster distance of 5, the mangrove species were divided into four groups. The species Rhizophora lamarckii and Sonneratia ovata were clustered into one group for the similarity of their spectral reflectance; physiologically, the two species also have more wax coating on the leaf surfaces compared with other species. The Level 3 scheme had a cluster distance of two, which divided the mangrove species into five groups where group A, at a rescaled distance of 5, was split into two new groups while the other three groups remained. Unlike in Levels 1–3, the Level 4 scheme grouped mangroves by single species, without forming new groups.

#### *4.2. Accuracy Assessment of the Resulting Maps*

The mapping accuracy was assessed with a confusion matrix to calculate the overall accuracy (OA), producer's accuracy (PA), and user's accuracy (UA). The validation samples were point data obtained directly from the field survey in 2021 (primary data) and secondary data with the attributes of mangrove species. The primary data were acquired using the average point method at 2 m × 2 m pure pixel points that had been created. Additionally, the samples were incidentally collected when encountering certain minor species that had gone unidentified during the interpretation of the aerial photographs. To assess the accuracy of the mapping results, a total of 177 sample points were analyzed using the "extract values to point" tool in ArcGIS on the model developed previously to compile the matrix.

In this step, the mangrove species maps generated using the three classification algorithms were evaluated for their OA, as presented in Tables 4 and 5. In general, the accuracy of the three classification algorithms at all four levels was low. The SID-based classifications at Levels 1 (with two species groups), 2 (four), and 3 (five) had the highest OA at 49.72%, 22.60%, and 15.25%, respectively. Meanwhile, the SFF-based classification for the Level-4 mapping of 24 species had the highest OA at 5.08%. The results demonstrated that the greater the number of classes to be mapped, the smaller the accuracy value. This

finding corresponds to Andréfouët et al. [29] and Kamal et al. [30], who explained that the percentage of classification accuracy expressed in OA decreased with the increase in the number of classes. In addition, the high heterogeneity of objects with natural properties, such as mangroves and mixed species in one pixel, affected the OA. The characteristics of mapped objects also affect the accuracy because if more pixels are mixed, the objects will be more difficult to classify.





The mapping accuracy assessment also calculated UA and PA for each classification algorithm and each dendrogram level to determine which errors reduced the accuracy. In general, the UA of all classification algorithms at each level was higher than the PA, as not all of the validation samples matched the classification model; many fell into a different class (unclassified or class "0"). With many samples considered unclassified, the PA decreased because, in contrast to UA, the number of correctly classified samples as the divisor increased. At Level 1 classifications with the SAM, SID, and SFF algorithms, there were, respectively, 148, 26, and 43 unclassified points among 177 samples. Furthermore, the SAM, SID, and SFF-based classifications led to 137, 24, and 38 unclassified points at Level 2; 132, 15, and 27 unclassified points at Level 3; and 63, 5, and 13 unclassified points at Level 4. The SAM-based classification had the highest number of unclassified points at all levels, resulting in low PA, UA, and OA.

The SAM algorithm demonstrated the lowest accuracy at all levels because the default threshold of 0.1 was used, leaving the mangrove area partially unclassified. The accuracy points extracted from the SAM model did not overlap because the mangrove area was not classified. The number of unclassified points significantly affected the accuracy of the SAM-based mapping results. Not all accuracy points that overlap with the model show accurate or desirable results. At Levels 1, 2, and 3, the points that did not overlap in the class that should have been the majority were placed in the last class, namely classes B, D, and E, respectively. At Level 4, the inaccurate points were mostly classified as *Pemphis acidula* and *Sonneratia caseolaris*. These issues significantly decreased the OA of SAM-based classification.

Mapping single species with the SID- and SFF-based classifications resulted in overlapping classes with more accuracy points than using SAM. The accuracy points in the SID-based classification overlapped more with *Acrostichum aureum*, *Aegiceras corniculatum*, *Pemphis acidula*, and *Sonneratia caseolaris* than with other species. Meanwhile, in the SFF-based classification, the accuracy points overlapped more with *Aegiceras corniculatum*, *Bruguiera sexangula*, and *Pemphis acidula*. However, according to the classification results, there were only nine overlapped accuracy points with the correct values when using the SFF and none when using the SID. In pixel-based mapping, increasing the classes to be mapped will reduce the accuracy value. In contrast, Kamal et al. [30] mapped *Rhizophora stylosa* and produced an OA of 52%. Hirano et al. [31] conducted a similar study mapping *Rhizophora mangle* that resulted in an OA of 40%. Mixed pixels are one of the problems that arise in remote sensing images because one pixel in the image can consist of two or more types of objects. Likewise, mixed pixels were common in the WV-2 satellite imagery used in this study because of the high heterogeneity and species diversity of the area. Hyperspectral field data used as the input in multispectral images also affects the mapping accuracy. Ideally, this type of study should be based on hyperspectral image data so the spectral resolution would be similar.

#### *4.3. Classification Performance Evaluation*

The three classification algorithms have different characteristics, each with particular advantages and disadvantages. They have dissimilar capacities in recognizing objects even from the same type of data source, i.e., the field-measured spectral reflectance of mangrove species. In the SAM algorithm, the similarity between the two spectra is determined by calculating the "spectral angle" between the image spectrum and the object spectrum and then assuming a vector in the same dimensional space with the same number of bands [32]. Its disadvantages include insensitivity to other known factors (besides angle) and using the same treatment for all illumination because SAM only uses spectrum "direction" and not spectrum "length." SAM only considers the spectral reflectance pattern of the object, not the difference in the intensity of the object's spectral reflectance [33]. The SAM-based classification uses vector directions to distinguish between features' spectral reflectance properties. Features with smaller spectral angles are categorized into the same class. SAM classifies pure classes and leaves the rest unclassified, thus failing to classify entire plant

species with pure pixels in the mangrove area [32,34]. In addition, the factors considered in determining the threshold value affect the classification results, especially at Levels 1, 2, and 3.

In contrast to SAM, the SID algorithm considers each pixel a random variable, using a spectral histogram to obtain a mapped probability. Muhammad and Mirza [34] suggest that problems arising from the SAM classification algorithm can be minimized by using SID. SID was applied to the same endmember to classify unclassified species and impure pixels in SAM-based classification results. The SID algorithm uses the size of the divergence to match pixels to spectral references. A smaller divergence means that the pixel is more likely to be similar to the spectral reference. According to Nidamanuri and Zbell [35], SID-based classification can measure the spectral variability of a single mixed pixel and determine similar spectra. However, according to Shanmugam and Srinivasaperumal [36], SID's weakness is that it is more effective on mixed pixel targets.

The SFF classification algorithm produces a separate scale image and root mean square (RMS) image or a combination of both. SFF is an absorption feature-based algorithm that matches the image spectrum (pixels) with the object spectrum (reference) [37]. In this study, SFF classified areas of high species diversity better than SAM and SID in single species mapping (Level 4). However, according to Muhammad and Mirza [34], this algorithm is the most time-consuming to use and the resulting classification still needs improvements. For the input, SFF requires a reference spectrum of the image or spectral library. The second requirement is removing the continuum from the image and spectral reference (object) before analysis [37], for instance, by resampling that degrades the spectral resolution of the data to that of the image.

#### **5. Conclusions**

This study identified 24 mangrove species on Karimunjawa and Kemujan Island. Based on their spectral reflectance characteristics, there were four dendrogram levels: Level 1 (consisting of two groups), 2 (four groups), 3 (five groups), and 4 (single species). The SAM-based classifications at Levels 1, 2, and 3 did not entirely classify the mangroves. The SAM and SID algorithms successfully mapped *Pemphis acidula* massively in the distal formation at Level 4. Using SID, Group B was found to prevail at Level 1 while Group A was dominant at Level 2. At the same level, the SFF algorithm classified Group C as dominant in the distal formation (*Rhizophora lamarckii* and *Sonneratia ovata*). Meanwhile, the SID- and SFF-based classifications at Level 3 showed Group B prevailing in the former and codominant Groups A, C, and D in the latter. The SFF algorithm classified *Bruguiera sexangula* in the distal formation. The best accuracy for mapping mangrove species distribution was obtained by applying the SID-based classification at Levels 1, 2, and 3, with overall accuracies of 49.72%, 22.60%, and 15.20%, respectively. Meanwhile, the best single-species mapping accuracy (Level 4) was obtained with SFF-based classification, with an overall accuracy of 5.08%. In conclusion, the three classification algorithms offered low mapping accuracy due to the high heterogeneity of species in the field, which resulted in many mixed pixels and limited access to obtain evenly distributed accuracy points. The greater the number of classes to be mapped, the smaller the accuracy. A predefined threshold value that is less than optimal is also a source of low accuracy. Future research can focus on assessing whether the number of mangrove species affects the accuracy of mapping results. This can be achieved by replicating the mapping method in a mangrove environment with low species variation.

**Author Contributions:** Conceptualization and methodology, A.D.R., M.K. and P.W.; validation and formal analysis, A.D.R.; writing—original draft preparation, A.D.R.; writing—review and editing, M.K and P.W.; visualization, A.D.R.; supervision, M.K. and P.W.; funding acquisition, M.K. All authors have read and agreed to the published version of the manuscript.

**Funding:** This research was funded by the 2021 Penelitian Terapan Unggulan Perguruan Tinggi Grant scheme provided by the Ministry of Education, Culture, Research and Technology of the Republic of Indonesia (contract number 1738/UN1/DITLIT/DIT-LIT/PT/2021).

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** The data presented in this study are available from the first author on request.

**Acknowledgments:** The authors would like to thank (1) the Department of Geographic Information Science at the Faculty of Geography, Universitas Gadjah Mada, for providing research facilities and equipment, (2) the Karimunjawa National Park management for granting fieldwork permits, and (3) S.M. Ridha and M.S. Usni for their invaluable assistance during the fieldwork.

**Conflicts of Interest:** The authors declare no conflict of interest. The funders had no role in the design of the study; the collection, analysis, or interpretation of data; the writing of the manuscript; or the decision to publish the results.

#### **References**


## *Article* **Global Mangrove Watch: Updated 2010 Mangrove Forest Extent (v2.5)**

**Pete Bunting 1,\*, Ake Rosenqvist 2, Lammert Hilarides 3, Richard M. Lucas <sup>1</sup> and Nathan Thomas 4,5**


**Abstract:** This study presents an updated global mangrove forest baseline for 2010: Global Mangrove Watch (GMW) v2.5. The previous GMW maps (v2.0) of the mangrove extent are currently considered the most comprehensive available global products, however areas were identified as missing or poorly mapped. Therefore, this study has updated the 2010 baseline map to increase the mapping quality and completeness of the mangrove extent. This revision resulted in an additional 2660 km2 of mangroves being mapped yielding a revised global mangrove extent for 2010 of some 140,260 km2. The overall map accuracy was estimated to be 95.1% with a 95th confidence interval of 93.8–96.5%, as assessed using 50,750 reference points located across 60 globally distributed sites. Of these 60 validation sites, 26 were located in areas that were remapped to produce the v2.5 map and the overall accuracy for these was found to have increased from 82.6% (95th confidence interval: 80.1–84.9) for the v2.0 map to 95.0% (95th confidence interval: 93.7–96.4) for the v2.5 map. Overall, the improved GMW v2.5 map provides a more robust product to support the conservation and sustainable use of mangroves globally.

**Keywords:** mangroves; extent; mapping; sentinel-2; global mangrove watch

#### **1. Introduction**

At the United Nations Framework Convention on Climate Change (UNFCCC) Conference of the Parties 26 (COP26) in 2021, an international agreement was made to end deforestation by 2030. To ensure adherence to this, accurate global scale maps of forested ecosystems will be critical. One such ecosystem is mangrove forests, which have witnessed an elevated rate of loss compared to terrestrial forests over the past decades [1] with regional losses exceeding 3%, driven by anthropogenic disturbances [1–3] such as conversion to aquaculture [4] or agriculture [5], urban expansion [6], oil palm plantations [7], and climate change [8]. Mangrove forests support a large number of ecosystem services [9], such as carbon storage and sequestration [10], coastal protection [11], food production [9], and tourism [12]. The ecosystem services of tidal mangroves and marshes were estimated to be worth 193,843 USD per hectare per year for 2007, equating to 25 trillion USD annually [13]. Accurate baseline maps of extent are therefore essential for a local and global ecosystem service accounting as well as verifying COP26 goals. Indeed, the ambitious goals set by the Global Mangrove Alliance (GMA), to restore 20% of mangrove forests by 2030, require accurate baselines upon which their efforts can be built. Furthermore, baseline maps are the keystone for mapping environmental descriptors that characterise this ecosystem, such as biomass [14], understanding the drivers of land cover change [2], and locating primary regions for potential restoration.

**Citation:** Bunting, P.; Rosenqvist, A.; Hilarides, L.; Lucas, R.M.; Thomas, T. Global Mangrove Watch: Updated 2010 Mangrove Forest Extent (v2.5). *Remote Sens.* **2022**, *14*, 1034. https:// doi.org/10.3390/rs14041034

Academic Editor: Chandra Giri

Received: 19 January 2022 Accepted: 17 February 2022 Published: 21 February 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

The Bunting et al. [15] Global Mangrove Watch (GMW) version 2.0 extent maps have emerged as the primary global dataset for characterising mangrove extent. There are a number of initiatives (e.g., GMA, GMW Portal; https://globalmangrovewatch.org; accessed 8 January 2022) that are aiming to preserve and restore mangroves and wider international objectives such as the UN Sustainable Development Goals (SDGs), for which the existing Global Mangrove Watch (GMW) version 2.0 layers [15,16] are a key dataset and are already used for reporting against. Currently, the GMW v2.0 [16] is the most up-to-date mangrove extent at the highest spatial resolution available. However, all mangrove datasets (i.e., [15–18]) published to date have areas that are missing (i.e., not mapped) or where mapping quality is poor. These limitations are evident globally and are caused by, for example, sensor specific characteristics (e.g., Landsat 7 Enhanced Thematic Mapper (ETM+) scan-line error), limited data availability, excessive cloud cover, or a combination of the above. These limitations degrade the performance of the map to meet the needs of the COP26 and GMA global initiatives by the year 2030.

For the GMW version 2.0, Bunting et al. [15] used two random forest classifiers to classify mangrove extent for the year 2010 from a combination of ALOS PALSAR and Landsat sensor data. As demonstrated by [15,19,20], the L-band radar data used in GMW v2.0 are sensitive to mapping mangrove change, while providing limited capability to classify the mangrove extent. However, optical remotely sensed data, particularly those with a Shortwave Infrared (SWIR) waveband, are well suited to the mapping of mangrove forest extent [21]. More recently, a number of studies (e.g., [21–25]) have made use of Sentinel-2 imagery and have demonstrated typical classification accuracies >90% for mangrove extent using ensemble machine learning classification approaches (e.g., random forests) through the Google Earth Engine platform. However, these studies have typically been undertaken over small spatial extents or for a few countries (e.g., [25]), single countries (e.g., [24]), or particular areas of interest at sub-national scales (e.g., [23]). Alternative approaches to mangrove mapping that have focused on mapping through time have also been proposed, such as [26,27], which have used the COntinuous monitoring of Land Disturbance (COLD) [28] method to provide individual site level time-series maps of mangrove extent. However, these time-series approaches are computationally intensive and therefore difficult to apply at a global scale. Nevertheless, they have demonstrated the ability of advanced machine learning and intensive computational processing for delivering maps at the quality required for international reporting.

The aim of this work was to produce an update to the 2010 Global Mangrove Watch version 2.0 [15] suitable for fully supporting the needs of ambitious global level targets relating to mangrove forest preservation and restoration. Specific regions were identified as missing or of poor quality within the GMW v2.0 product and a new method was, therefore, proposed for achieving vastly improved mapping, combining higher-resolution data at higher imaging cadence with advanced machine learning models. These results were combined with the GMW v2.0 map to create the most complete map of mangrove extent currently available and will form the basis of a subsequent study, updating the estimates of mangrove change.

#### **2. Materials and Methods**

The analysis was undertaken on the SuperComputing Wales (SCW) High Performance Computing (HPC) infrastructure using the Remote Sensing and GIS Library (RSGISLib) of tools [29], the KEA image format [30] and the pbprocesstools [31] workload management library to manage the workflow of tasks on the HPC.

#### *2.1. Areas to Be Mapped*

Through user feedback on the GMW v2.0 maps, 204 regions were identified to be either missing or poorly mapped. As detailed by Bunting et al. [15], the choice of 2010 for the baseline map was driven by the use of ALOS PALSAR data, coverage of which was most complete for the year 2010 [15]. However, this resulted in Landsat-5 TM and Landsat-7 ETM+ temporal composites affected by ETM+ scan-line error artefacts, particularly in areas of high cloud cover (e.g., Niger Delta). These artefacts were, in some cases, present with the GMW v2.0 maps. Areas identified that required (re-)mapping as part of this study are shown in Figure 1.

**Figure 1.** Map of regions identified as in need of updating as part of the Global Mangrove Watch (GMW) v2.5 analysis.

#### *2.2. Mangrove Habitat Mask*

Bunting et al. [15] developed a mangrove habitat mask which was used to limit the classification of mangroves to those where mangroves can be expected to be present. For example, mangroves must generally be close to water and at or close to mean sea level. However, this mask was found to have been too tight in a number of regions (e.g., Florida, USA) and therefore caused under-classification of mangroves. Ahead of the mapping, the habitat mask was therefore revised, primarily through manually digitising regions to be added, but also intersecting the maps of [17,18] to ensure that all regions mapped in those products were fully located within the GMW habitat mask.

#### *2.3. Mangrove Mapping*

Bunting et al. [15] used the most appropriate and available data (i.e., Landsat TM and ETM+ and ALOS PALSAR) for 2010 in the original mapping (v2.0). Therefore, to improve the mapping, an approach that used alternative datasets was required. In this case, Sentinel-2 imagery was used to map the areas outlined in Figure 1. For the analysis, 100 global mangrove/non-mangrove XGBoost classifiers [32] were trained and applied to each Sentinel-2 acquisition. To produce a single unified classification, results from all acquisitions and classifiers were merged to create a probability for each pixel to be mangroves. This probability surface was then thresholded to produce the binary map, which was subject to a manual Quality Assurance (QA) process to produce the Sentinel-2-derived maps. These were then combined with the v2.0 2010 baseline and a change detection using the 2010 ALOS PALSAR data was applied to create a revised GMW v2.5 2010 baseline. The processing stages are outlined in Figure 2.

#### 2.3.1. Sentinel-2 Processing

The Sentinel-2 imagery was downloaded from the Google Cloud public dataset [33] and, for the purpose of this work, processed to a 20-m pixel resolution (i.e., the 10-m resolution Sentinel-2 image bands were resampled to 20 m using averaging) orthorectified standardised surface reflectance product. For the 383 Sentinel-2 granules identified as intersecting with the regions of interest (Figure 1), individual acquisitions were selected for download based on cloud cover. Initially, the 10 acquisitions with the lowest cloud cover (maximum cloud cover of 20%) were identified from the entire Sentinel-2 archive (2015–2020). An iterative process was then followed where, for each granule, the scenes were downloaded and processed to produce a cloud mask. The combined cloud masks were checked to estimate whether data were available for all mangrove regions within the scene, as defined by the mangrove habitat mask. If more acquisitions were required, the thresholds for the number of acquisitions and maximum cloud cover were increased. These were increased to a maximum of 100 scenes and a maximum cloud threshold of 75%. In total, 11,262 Sentinel-2 acquisitions were downloaded and used for this analysis.

**Figure 2.** Flowchart of the methods used to generate the GMW v2.5 map for 2010.

To generate a standardised reflectance product for the classifications, the ARCSI software [34] was employed, as successfully demonstrated in past studies [15,26,27,35]. The ARCSI software uses the 6S model [36] through the Py6S module [37] parameterised using the image header information and an aerosol optical depth estimated from a dark object subtraction [15]. Using the method of Shepherd and Dymond [38], the resulting images were normalised for a sensor view angle and local topography producing a standardised reflectance product. A tropical atmosphere and maritime aerosol profile was used for all scenes.

Cloud masking was undertaken using the product of two approaches. The FMask [39,40] algorithm was applied using the Python-FMask implementation [41]. The s2cloudless [42] LightGBM classifier [43] was also applied to each scene. The resulting s2cloudless classification was further refined using a morphological closing (with a 5 × 5 circular operator) followed by a morphological dilation (with a 7 × 7 circular operator). Finally, any cloud objects were removed if they were less than 10 pixels in size. The final cloud mask was defined as the intersection of the two masks. The cloud shadow mask was derived using the approach implemented within FMask, as described in Zhu et al. [39].

Finally, a 'clear sky' mask was derived for each acquisition, defining the areas of the scene to be used for further analysis. The 'clear-sky' mask aims to identify the larger continuous parts of the image, removing small areas between clouds. The first step buffered the cloud and cloud shadows by 30 km, clumping the remaining non-cloud regions. The non-cloud clumps with an area greater than 3000 pixels were then selected and grown to

the 10-km contour of the cloud and cloud shadow pixels. An example of the 'clear sky' mask is shown in Figure 3c.

NP F NP

**Figure 3.** An example of the clear sky mask for part of a Sentinel-2 scene, where (**a**) is the original scene (false colour: near infrared (NIR), shortwave infrared band 1 (SWIR-1), and red bands) and (**b**) is the resulting cloud and cloud shadow mask, which can been seen to have missed some clouds, and (**c**) is the clear sky mask which has masked the regions around the cloud and cloud shadows.

#### 2.3.2. Building the Classification Models

&ORXGV 6KDGRZV &OHDU6N\

Classification of mangrove extent was undertaken on a scene-by-scene basis rather than through the creation of image composites (i.e., merging multiple scenes using a metric such as the greenest pixel). Image composites, whilst relevant for visualisation, often have artefacts due to prevailing environmental conditions (e.g., wet or dry season or, in the case of mangroves, tidal regimes) at the time of the acquisitions or processing errors (e.g., missed cloud or cloud shadows). These artefacts can then impact the classification result. An alternative is to classify each of the scenes independently and then merge those results to create a single map.

To derive training data for the classification, 10,284 samples were created from the existing GMW v2.0 map. These were manually checked against the Sentinel-2 imagery. For regions not already within the GMW v2.0 product training regions, these were manually defined. Non-mangrove regions were defined as regions outside of the GMW habitat product, with points sampled randomly within this region and through manual selection of regions giving a total of 52,555 sample points for training.

The resulting samples were then intersected with all 11,262 Sentinel-2 acquisitions, with each scene masked to the relevant valid clear sky area. This resulted in 4,421,644 mangrove and 9,830,388 non-mangrove pixel values to train the classifier. Given the volume of sample data available, it was decided to split the training data into 100 sets, each with 400,000 samples (200,000 for mangroves and 200,000 for non-mangroves). Those samples were then split into 3 sets for training (100,000 for each class), testing (50,000 for each class), and validating (50,000 for each class) the model.

The XGBoost [32] binary classification algorithm was used for the analysis given its ability to use large training datasets and allow transfer learning (i.e., further training of an existing model). This method has been shown by John et al. [35] to provide good results for the classification of land cover from Earth observation data. To optimise the hyperparameters of the XGBoost model, a subset of 20% was selected from the training (20,000 per class) and validation (10,000 per class) samples. Bayesian optimisation was used to identify the optimal hyper-parameters for each of the 100 classifiers. The range of values for the parameters optimised is given in Table 1. Following identification of the hyper-parameters, each of the 100 models was trained using the full dataset (i.e., 400,000 samples). The testing accuracies of the models (using the 50,000 samples per class) were between 97–99%.

#### 2.3.3. Applying the Classification Models

To apply the 100 global XGBoost classifiers to the individual Sentinel-2 acquisitions, the models were first further trained using the local training data from the Sentinel-2 acquisition, which was limited to 25,000 samples. This allowed the global classifier to be locally optimised for the individual acquisitions. The classifiers were then applied to all the acquisitions, with this creating 112,620 classifications. To avoid incorrect classification of mangroves in areas where they would not be located (e.g., in mountainous areas), the classification was only applied within an updated version of the mangrove habitat layer of Bunting et al. [15].


**Table 1.** The range of hyper-parameter values from the 100 models.

The individual classifications were then merged in two steps to create a mangrove probability for each pixel. The first step merged the 100 classifications applied to a scene to create a single probability output image for the scene. The probability was calculated as the number of times each pixel had been classified as mangroves (i.e., a value of 1 meant that all 100 classifiers classified the pixel as mangroves, while a value of 0.1 meant that only 10 classifiers classified the pixel as mangroves). The second step calculated the mean probability from all the acquisitions for each pixel, providing a single probability surface for all the areas mapped.

To derive the final binary mask of mangrove extent, a global threshold was applied to the probability surface. The threshold was identified through a sensitivity analysis using the mangrove samples based on the 0.1 increments (from 0.2 to 0.8). A mangrove mask was generated for each threshold where the mask with the best agreement with the mangrove samples used to train and test the XGBoost classifiers selected. A threshold of 0.5 provided the greatest correspondence and was therefore applied to all the regions updated using the Sentinel-2 imagery. For studies focused on specific regions, a further local optimisation could be undertaken by selecting a local threshold. However, for this study, a global threshold was applied as defining local regions would be difficult and could result in boundary artefacts within the resulting maps.

Finally, a visual assessment of the mangrove extent was undertaken where polygons identifying regions as incorrectly classified as mangroves were digitised with reference to the Sentinel-2 imagery and high-resolution Google Earth, Mapbox Satellite, and Bing maps imagery. The areas were then removed from the mangrove extent mask.

#### *2.4. Merging Mangrove Maps and Identify Change to 2010*

In addition to the new map produced from the Sentinel-2 analysis, two other mangrove maps were used to resolve issues for particular areas. For the Sundarbans, in India and Bangladesh, the mapping of Awty-Carroll et al. [26] for the year 2010 was added to the Sentinel-2 maps to be merged with the GMW v2.0 products. The Sundarbans were significantly affected by stripping from the Landsat ETM+ data within GMW v2.0. Additionally, mangrove maps for the French overseas territories, where there was found to be a high prevalence of cloud cover that reduces the availability of useable Sentinel-2 data, generated by the French National Mangrove Observation Network [44], were used to improve the new maps.

Following generation of the revised maps, these were merged with the existing GMW v2.0 baseline for 2010 to create the updated 2010 GMW v2.5 baseline. However, the updated areas had been mapped with data acquired over the period from 2015 to 2020 and a change detection was therefore required to backcast the map for 2010. 2010 ALOS PALSAR data were used for this and therefore the new mapping was resampled (nearest neighbour) onto the same 0.000222 degrees (∼25 m) pixel grid of the GMW v2.0 and ALOS PALSAR data layers.

As demonstrated by Thomas et al. [3,19,20], mangroves produce a high backscatter response in the L-band SAR data while the majority of non-mangrove surfaces (e.g., water bodies and mudflats) have a low L-band backscatter. As a result, there is a change trajectory between mangroves and non-mangroves, which was used by Thomas et al. [20] as the basis for a methodology for mapping mangrove change. This was applied globally to produce the GMW v2.0 change layers [16].

For implementation, a low backscatter mask was created for 2010 and used to remove mangroves that were within the new map but not present in 2010. The mask was defined using a combination of the ALOS PALSAR 2010 layer and the Landsat-based Pekel et al. [45] water occurrence layer generated for the period 1984–2020. The analysis was undertaken on a 1 × 1 degree grid, where the water occurrence layer was used to define areas that could be considered as 'permanent' waterbodies, defined as a water occurrence between >90 and <100. However, if no pixels were identified, then the threshold was lowered to >70. For the pixels associated with 'permanent' waterbodies, the 99th percentile of the SAR backscatter was calculated for both the Horizontal-Horizontal (HH) and Horizontal-Vertical (HV) polarisations. The thresholds for classifying the water extent were then calculated for both polarisations as:

$$\text{SAR threshold} = 99 \text{th percentile} - (0.15 \times 99 \text{th percentile}).\tag{1}$$

If no 'permanent' waterbody pixels were identified, then the SAR thresholds where defined as −14 dB in the HH and −17 dB in the HV polarisations. To produce the low backscatter mask, the SAR backscatter was thresholded with values below those calculated above were used and the water occurrence layer had a value > 5.

The low backscatter mask was then used to mask all the tiles, including areas which have not been remapped, updating the mangrove mask and aligning it with the ALOS PALSAR data for 2010. Finally, a Quality Assurance (QA) process was undertaken where the product was visually assessed against a variety of image sources, including highresolution Google Earth, Mapbox Satellite and Bing Maps imagery, the Sentinel-2 data, 2010 ALOS PALSAR, and 2010 Landsat imagery data. Polygons were manually drawn for regions which should be removed from the map (i.e., not mangroves but areas that had been mapped as mangroves) or added to the map (i.e., mangroves but areas that had not been mapped as such). These QA edits were then rasterised and applied to the map producing the final GMW v2.5 layer.

#### *2.5. Accuracy Assessment*

To assess the accuracy of the new v2.5 layer, 26 sites (Figure 4) where new mapping had occurred were selected, representing a range of different mangrove settings, types, and extents. Additionally, a further 34 sites (Figure 4) were distributed globally for assessing the overall product accuracy. For each site, an area of 0.2 × 0.2 degrees was defined and 1000 random stratified points were defined for each class (mangroves and non-mangroves). If there were less than 1000 mangrove pixels within the 0.2 × 0.2 degree area then all mangrove pixels were defined as points and the number of mangrove reference points was reduced. The 2000 points were then split into 200 point sets (i.e., 100 mangrove and 100 non-mangrove) where the sets were assessed in turn until the 95% confidence interval for the macro F1-score was <5%. A minimum of 3 sets (i.e., 600 points) were assessed for each site, where typically 5 sets were required (1000 points) although 10 sets were used for one site. Points were manually annotated with a reference class through a combination of high-resolution Google Earth, Mapbox Satellite and Bing Maps imagery, the Sentinel-2 and 2010 ALOS PALSAR, and Landsat imagery data. In total, 50,750 points were assessed and

used for the accuracy assessment. For sites where the mapping was updated, the points were also used to assess the improvement in map accuracy achieved through this study.

**Figure 4.** A map of the 60 sites used for the accuracy assessment. The 26 red points are over areas which have been mapped with Sentinel-2 as part of the GMW v2.5 analysis while the 34 blue points are further set of sites used to capture the global accuracy of the GMW v2.5 baseline rather than just the areas updated.

#### **3. Results**

#### *3.1. Remapped Regions Comparison*

To compare the accuracy of the updated v2.5 and v2.0 GMW 2010 baselines, the reference points for the 26 sites where the baseline has been updated were intersected with both layers. Summary statistics calculated were an overall accuracy, cohen kappa, and F1 score (per-class and overall), with summaries are provided in Tables 2 and 3. In addition, upper and lower confidence intervals for all metrics were calculated using bootstrapping. It was not possible to calculate metrics, such as the allocation and quantity disagreement as those metrics require a closed map where all pixels are allocated to a class such that the area of the whole region can be used to normalise. However, for this study we only have a single class of interest (i.e., mangroves).

As shown in Tables 2 and 3, the estimated accuracy of mapping in regions where the quality was identified previously as poor or missed increased from 82.6% (80.1–84.9) to 95.0% (93.7–96.4). The range for the individual site accuracies also decreased from 44.7% to 12.4%, demonstrating that the quality of mapping for these areas remapped resulted in a similar quality of mapping for all regions.

**Table 2.** For the region mapped to create v2.5, this table contains an overview of the accuracy statistics from the GMW v2.0 baseline, which can be compared to the statistics in Table 3.


This improvement in mapping accuracy can be seen visually and is illustrated in Figures 5–7. Figure 5a provides a typical example of a region that was affected by the Landsat ETM+ striping but was remapped to improve the output Figure 5b. Figure 6a illustrates an area in Colombia where some areas of mangroves were missed but have now been mapped in GMW v2.5 (Figure 6b). Figure 7 illustrates an example where the habitat mask was too restricted in GMW v2.0 but has been improved within GMW v2.5 by expanding the habitat mask. In terms of the accuracy statistics (Table 3), the example shown in Figures 6 and 7 represents a region where the accuracy will have significantly improved, while Figure 5 resulted in only a modest statistical improvement but is visually much improved.

**Table 3.** For the region mapped to create v2.5, this table contains an overview of the accuracy statistics from the GMW v2.5 baseline, which can be compared to the statistics in Table 2.


#### *3.2. Overall Accuracy Assessment*

Using all 60 sites, the overall accuracy statistics for the v2.5 map was calculated and presented in Table 4. The global assessment estimated an overall accuracy of 95.1% with a 95th confidence interval (i.e., 95% likelihood that the true value is within the range) of 93.8 and 96.5%. This was similar to those published by Bunting et al. [15] for v2.0, which estimated an overall accuracy of 94.0% with a 99th confidence interval of 93.6 and 94.5%. This is to be expected with only approximately 33% of the map having been remapped (i.e., replaced) and with only minor changes masking low backscatter pixels applied to all regions alongside the overall high estimated accuracy of the v2.0 map. However, as demonstrated in Table 2, the local accuracy of the v2.0 map could be as low as 51% where areas were missed.

**Figure 6.** Comparison of GMW v2.0 (**a**) and v2.5 (**b**) products, illustrated with an example from Colombia, where regions had been omitted in GMW v2.0 but included in GMW v2.5.

**Figure 7.** Comparison of GMW v2.0 (**a**) and v2.5 (**b**) products, illustrated with an example from Florida, USA, where the habitat mask was too restricted when used in the production of GMW v2.0 which has been improved for the GMW v2.5 product.

#### *3.3. Area Statistics*

The global mangrove extent mapped in v2.5 was 140,260 km2, an increase of 2660 km2 (2.5%) over the v2.0 GMW map, which had a global total of 137,600 km2. Table 5 provides a range of example countries, some with significant changes in mangrove extent between v2.5 and 2.0. A full country table of mangrove extents for v2.5 and v2.0 has been provided in Appendix A Table A1.


**Table 4.** The overall GMW v2.5 accuracy assessment summary.

**Table 5.** Example country statistics illustrating the changes between GMW v2.5 and v2.0. A full table has been provided in Appendix A (Table A1).


Within the GMW v2.5 data, there are 121 countries with mangroves. Twelve countries, including Bermuda, were missing from the GMW v2.0, however they (and their areas) have now been added to the GMW v2.5 dataset. These were mostly small island nations where persistent cloud cover limits the acquisition of useable remote sensing data. The observed changes at a national level are variable, with 15 countries (e.g., Mozambique) having mangrove area differences of less than 1% between GMW v2.5 and v2.0, while 50 countries (e.g., Australia) had between 1–5% of change. The small changes between the two maps were attributed to the low backscatter pixel mask that was applied to all tiles. However, for countries with a small area of mangroves, these changes can be significant in percentage terms. For example, the area of mangroves mapped in Bahrain was 28% greater in the GMW v2.0.

Of the remaining 44 countries, 18 had a net change between the GMW v2.5 and v2.0 between 5–10% of their mangrove area, with 11 between 10–20% and 10 between 20–50% and 5 with a net change greater than 50% of the GMW v2.0 area. Many of the countries with the largest change area were those with small mangrove extents (e.g., Bahrain or Mauritius). However, in some regions remapped with Sentinel-2, substantial areas were either removed from the GMW v2.5 map (e.g., Angola had 8377 ha of fewer mangroves; Figure 8) or were added (e.g., Benin had 3307 ha more mangroves; Figure 9), with this improving the mangrove mask accuracy for these regions. In the GMW v2.0 map, a number of areas (e.g., Florida; Figure 7) were omitted because of the restricted GMW habitat mask, which was used to limit areas where mangroves could be classified. Improvements in this

mask along with the remapping effort has allowed new areas to be included within the GMW v2.5 map.

However, some regions were found to have a mixture of substantial omissions and commissions within the v2.0 dataset. For example, Fiji, which was remapped with Sentinel-2, had an overall net change of −2.3% between v2.0 and v2.5. However, there were also significant regions of additional mangrove within Fiji in v2.5 as a processing error in the v2.0 product caused the mangroves in the west of the island nation (i.e., −180–−178) to be missed.

The improvement in mapping through the use of the Sentinel-2 data was significant in areas of high cloud cover and particularly in regions such as French Guiana (18.5%), Papua New Guinea (−6.4%), Nigeria (21.9%), and Colombia (13.4%; Figure 6). These areas had often significant striping artefacts present from the use of Landsat ETM+ data in the GMW v2.0 map (e.g., Figure 5).

These changes in the mapped mangrove area are not due to changes on the ground but rather to better input data (i.e., Sentinel-2) or new knowledge (e.g., improvements to the habitat mask) that have allowed us to generate a more accurate mangrove map for 2010.

#### **4. Discussion**

To meet the requirements of global initiatives to achieve ambitious targets on the preservation and restoration of forested ecosystems, accurate and timely maps of the extent are critical. To date, the GMW has produced the most contemporary and comprehensive maps of global mangrove ecosystems. However, poorly mapped and omitted regions were present in the v2.0 dataset. We successfully identified 204 regions in need of updating or inclusion and proposed a new method to refine the maps in the selected locations. Our approach was able to increase the low accuracy of the map in these regions from 82.6% to 95.0%, bringing them in line with the level of overall accuracy of the global map. This updated map is better suited to meet the needs of the COP26 goal of ending deforestation by 2030 and the GMA goal of restoring 20% of mangroves by 2030. Accurate baselines are critical to measuring the success of such ambitious targets and ensuring accountability in reporting. With the accuracy of the updated regions mapped here increased by approximately 10%, GMW v2.5 is situated as the primary global scale mangrove extent product.

#### *4.1. Data and Methods*

As outlined by Thomas et al. [20], radar data used for the GMW v2.0 are limited in its ability to discern mangrove extent. Here, Sentinel-2 was relied upon for high-resolution high-cadence imagery, with spectral bands suited to wetland vegetation mapping. Sentinel-2 data are acquired as often as once every 5 days and do not suffer from the instrument degradations that impacted the Landsat 7 ETM+ imagery used for generating the GMW v2.0 2010 baseline. This provides a dense stack of imagery from which to derive a baseline map for afflicted regions. As an improvement over GMW v2.0, image composites were not used as these can result in image artefacts and inconsistent imaging conditions between images which can lead to classification confusion. Instead, each image was classified separately and a probabilistic approach was used to determine the mangrove extent. This is a more robust approach as it can provide additional minimum and maximum bounds and thus allows flexibility on definitions of extent and transparency on uncertainties. Despite this, errors do persist in specific locations. While the overall accuracy of the map was 95.1%, the ranges on a site-by-site basis was from 87.4% to 99.8%. In some locations, this was caused by the limitations of using moderate spatial resolution (i.e., 20–30-m resolution Sentinel-2, Landsat, and ALOS PALSAR) imagery to map very fine fringes and fragmented stands (e.g., Figure 10), that often are associated with human disturbance. These areas are challenging to identify and interpret and reliably differentiate from other vegetation types, even using very high spatial resolution imagery (i.e., <3 m). Access to local knowledge and field data is increasingly important for achieving high quality results in these locations.

**Figure 10.** Example of fragmented mangroves from Sulawesi, Indonesia. (**a**) 30-m Landsat 8 imagery from 2016 (false colour: NIR, SWIR-1, and Red bands) and (**b**) 30-m Landsat 8 imagery from 2016 overlain with the GMW v2.5 baseline (green), illustrating the resulting v2.5 map for these fragmented areas of mangroves, which only maps the larger regions for mangroves and not the finer detail.

The use of the XGBoost classifier [32] provided the use of a gradient boosted decision tree, which can take advantage of larger training datasets than alternative methods such as random forests. This has advantages over the random forest algorithm used in GMW v2.0, by improving upon the single model through use of an iterative approach and training ensemble models in succession, with each new model correcting errors in the previous one. This is considered a more robust approach to ensemble learning. Our updated method also used a unique approach to classifier training, by using combinations of both local and global training data in order to derive results that are locally tuned but are also representative of the global mangrove extent.

#### *4.2. Future Baseline Mapping*

Going forward, we advocate that a new global baseline at a 10-m spatial resolution using Sentinel-2 be produced. Such a baseline would align well with other global datasets produced using Sentinel-2 and Landsat imagery. JAXA are also reprocessing the ALOS PALSAR and ALOS-2 PALSAR-2 products such that they align with this spatial baseline. Additionally, using Sentinel-2 data, this study and others (e.g., [21–25]) have demonstrated that optical remotely sensed data with a shortwave infrared channel (e.g., Landsat and Sentinel-2) can provide a reliable classification of mangroves and differentiation from other vegetation types in most regions. Regions where other wet tropic forests share a boundary with mangroves (e.g., Papua New Guinea and French Guiana) are however still challenging for classification. The additional spatial resolution (i.e., 10 m) should further help discriminate small mangrove patches, such as river edges and where disturbance and loss/gain patterns are complex. The spatial registration of the radar data used in GMW v2.0 has been found to have 1 or 2 pixels (i.e., 25–50 m) of mis-alignment with the latest Landsat (Collection 1 and 2) and Sentinel-2 datasets. The creation of a low backscatter mask is thought to align the remapped areas with the GMW v2.0 datasets. However, by using the ALOS PALSAR data alone, discrimination of mangroves from other vegetation types is poor, so the uncertainty in mapping the landward extent of mangroves is greater, with this also being the case for the GMW v2.0 dataset. JAXA are set to reprocess their L-band SAR data in order to alleviate this registration issue. Thus, future mapping efforts should maximise the use of these datasets when available.

#### **5. Conclusions**

The paper presented an updated version of the GMW mangrove extent baseline for 2010, producing version 2.5 of this dataset. The update focused on particular areas that were identified as being poor in quality (e.g., due to Landsat ETM+ stripping) or were missed in the GMW v2.0 product. The analysis demonstrated an increase in overall accuracy for updated regions from 82.61% (95th confidence interval 80.1–84.9%) for the GMW v2.0 product to 95.0% (95th confidence interval 93.7–96.4%) for the GMW v2.5 product. To our knowledge, this renders the GMW v2.5 baseline for 2010 as being the most complete global map of mangrove extent. This baseline will be used as the basis for an update to the GMW v3.0 change product, also extending the period from 1996–2016 to 1996–2020, which will be the subject of a forth-coming publication.

**Author Contributions:** Conceptualization, P.B., A.R., L.H., R.M.L. and N.T.; methodology, P.B., A.R. and N.T.; software, P.B.; validation, P.B., A.R., L.H., R.M.L. and N.T.; writing—original draft preparation, P.B.; writing—review and editing, P.B., A.R., L.H., R.M.L. and N.T. All authors have read and agreed to the published version of the manuscript.

**Funding:** The Global Mangrove Watch is funded by the Oak Foundation, the COmON Foundation, the National Philanthropic Trust, DOB Ecology, and the Dutch Postcode Lottery. This research was also funded by the Natural Environment Research Council (NERC) through the UKRI Newton Fund (NE/P014127/1) and the Japan Aerospace Exploration Agency (JAXA).

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** The scripts for the data analysis are available on GitHub: https:// github.com/globalmangrovewatch/gmw\_gap\_fill\_2020 (accessed: 8 January 2022). The output GIS datasets are available here: https://doi.org/10.5281/zenodo.5828339 (accessed: 8 January 2022).

**Acknowledgments:** We acknowledge the support of the Supercomputing Wales project, which is partly funded by the European Regional Development Fund (ERDF) via the Welsh government, for providing the computing infrastructure for undertaking this study. We would also like to thank IUCN-France for the provision of the mangrove maps over the French overseas territories. We also thank all those who have provided feedback on the GMW v2.0 dataset. Tristram Irvine-Fynn is also thanked for reviewing the manuscript before submission.

**Conflicts of Interest:** The authors declare no conflict of interest.

#### **Appendix A. National Mangrove Extent**


**Table A1.** Country level 2010 mangrove extents for both GMW v2.5 and v2.0.

**Table A1.** *Cont.*



#### **Table A1.** *Cont.*

#### **References**

