1. Introduction
Mangrove forest ecosystems, which are a major component of coastal and estuarine zones in tropical and subtropical regions, provide substantial socio-economic functions, such as supplying nutrients and grounds for fish and shellfish, and producing wood and non-wood goods [
1,
2,
3]. In addition, mangrove forests also provide important environmental functions [
4,
5,
6,
7,
8]. For example, mangrove forests play a crucial role in the global carbon cycle [
9,
10,
11,
12,
13]. Previous studies have found that mangrove forests are one of the most carbon-rich ecosystems, and their average total carbon storage at a specified area unit is five times larger than that of tropical terrestrial, temperate, and boreal forests [
6]. Information about the spatial distribution and extent of mangroves is a central component to quantitatively estimate their ecosystem service value [
14]. Recent studies have shown that, over the past two decades, global mangrove forests have been declining rapidly and are gradually fragmenting into small patches. This can be attributed to anthropogenic activities, such as aquaculture and land reclamation [
15,
16,
17,
18]. If the loss continues at the present rate, mangroves are expected to be extinct within the next 100 years [
19,
20]. As a result, there is an urgent need for advancing the monitoring capabilities of mangrove forests to prevent the extinction of mangroves and assess changes in mangrove ecosystem functions.
Remote sensing imagery is ideal to monitor the spatial distribution and extent of land cover [
21], but applying this to monitoring large-scale mangroves is challenging, due to the widespread existence of scattered mangrove trees, which are generally mixed with other types of plants (e.g., salt marsh), and the increasing number of small mangrove patches caused by human destruction and new mangrove plantation. Although high-resolution (<1 m) imagery (e.g., Worldview, Quickbird, and aerial photographs) can be used to identify individual mangrove trees, large-scale mangrove mapping based on these data is limited by cost, data availability, and volume. Consequently, freely available satellite imagery (e.g., Landsat) with moderate resolution (
) is widely used to extract large-scale mangrove forests. For example, Chen et al. (2017) [
22] mapped mangrove forest extent in China from 2015 based on 30-m Landsat imagery; the Global Mangrove Watch (GMW) [
23] applied ALOS PALSAR/JERS-1 (25 m) and Landsat (30 m) imagery to produce global mangrove forest maps from 1996 to 2010.
However, these products with spatial resolutions larger than 30 m are relatively coarse for accurately characterizing the distribution of mangrove forests [
23]. For example, due to new mangrove growth and damage caused by human beings, there are many mangrove patches smaller than the minimum discriminating area of these products [
23,
24,
25].
Approaches for remote sensing-based mangrove mapping can be grouped into two categories. One is using single-date imagery, and the other is using multi-temporal imagery. The first approach mainly adopts single-date optical, SAR (synthetic aperture radar), or a combination of these two types of imagery. The single-date imagery is difficult to be applied in large-scale mangrove mapping because it is hard to guarantee that all clear images covered study area are obtained at low tide, a period during which the canopy of mangrove is not inundated by water and, hence, can be identified [
26]. Additionally, the spectral similarity between mangroves and other land cover types (such as natural terrestrial forests) also makes it more difficult to accurately identify mangroves in single-date imagery [
27]. Compared to using single-date imagery, using time-series imagery can separate mangrove forests from other land cover types at large scales, since the temporal profile of the spectrum for mangroves is distinctive with influence of tidal variability. For example, for mangroves, the spectral reflectance of near infrared (NIR) and short wave infrared (SWIR) obtained at high tide declines abruptly compared to that obtained at low tide. Spectral-temporal features (such as quantiles) derived from the time-series of remote sensing imagery are usually adopted to delineate spectrum temporal profiles and are applied to extract mangroves. For example, Hu et al. (2018) [
27] applied spectral-temporal features derived from Landsat time-series imagery to monitor mangrove forests in China. However, most of studies used optical imagery, and only a few studies adopted SAR data. For example, Chen et al. (2017) [
22] used information about yearlong and fresh water bodies derived from Sentinel-1 time-series data (2014–2016) to identify mangrove forests, which are periodically inundated with water with the influence of tidal dynamics.
Since 2014, the constellations of Sentinel-1A & 1B and Sentinel-2A & 2B were successively launched [
28]. Sentinel-1 C-band SAR data are insensitive to clouds and may be useful to monitor mangrove forests, which are distributed in regions that are often cloudy. Sentinel-2 MSI (Multispectral Instrument) imagery has 13 spectral bands, including four red edge bands sensitive to biophysical features (e.g., leaf chlorophyll and water content) of vegetation, with a spatial resolution of 10 m and temporal resolution of 5 days. Compared to Landsat imagery (7/8 bands, 30 m, 8/16 days), which is commonly used in large-scale mangrove extraction, Sentinel-2 has a clear improvement in spatial and temporal resolution [
29]. Therefore, Sentinel-1 and Sentinel-2 imagery, which are freely available, may have the potential to advance the mapping of mangrove forests. In addition, Sentinel data can also be processed through the Google Earth Engine (GEE) cloud computing platform, which facilitates challenges related to massive data volumes processed in large-area mapping. Several studies have been conducted for mangrove mapping at a regional scale using Sentinel imagery. For example, Chen et al. (2017) [
30] identified the mangrove forests in Dongzhaigang, China, with the use of Sentinel-2 imagery; Manna et al. (2018) [
31] mapped mangroves in Sundarban based on Sentinel-2 imagery. Nevertheless, no research has been reported for using time-series Sentinel-1 and Sentinel-2 data to produce national-scale mangrove maps.
Therefore, our study applied Sentinel imagery on the Google Earth Engine platform to generate a national-scale mangrove forest map, and further explored how a 10-m mangrove map derived from Sentinel imagery can improve our knowledge of mangroves on existing 30-m products generated from Landsat imagery. Our study focused on the coastal zones of China, whose mangrove ecosystems have a scattered distribution and are mixed with salt marsh. Besides, in China, the number of small mangrove patches is increasing due to anthropogenic activities, such as destruction and successive planting in large areas [
32]. Several studies have evaluated the spatial distribution and extent of mangrove forests in China based on Landsat imagery with 30-m resolution. However, the actual extent and area of mangroves containing those small mangrove patches is still unclear because these small patches are smaller than the minimum discriminating area (1 ha) of Landsat imagery.
In summary, the objectives of this study are: (1) generate a national-scale mangrove forests map with 10-m resolution in China using Sentinel-1 and Sentinel-2 time-series imagery; (2) explore an optimal combination of features from Sentinel-1 and Sentinel-2 time-series imagery for large-scale mangrove extraction, which has the potential to be transferred globally; (3) quantitatively and qualitatively compare the difference between the 10-m mangrove forest map and existing available products.
3. Methods
3.1. Preprocessing
3.1.1. Cloud Mask
Sentinel-2 MSI imagery is easily contaminated by clouds, especially in tropical and subtropical regions where mangrove forests grow. The Level-1 product of Sentinel-2 provides a Quality Assessment (QA) band that labels cloud and cirrus pixels. By using this band, we discriminated and removed cloud and cirrus pixels from each image.
3.1.2. Spectral Indices Calculation
Spectral indices enhance the spectral properties of vegetation and water. Here, we adopted the seven spectral indices listed in
Table 3. Given that Sentinel-2 has both 10-m and 20-m bands, before calculating above spectral indices, we resampled all of the 20-m Sentinel-2 bands to data with the spatial resolution of 10 m.
3.1.3. Texture Information Extraction
Texture information, which is usually extracted based on a statistical method called the Grey-Level Co-Occurrence Matrix (GLCM), has been widely used in discriminating different land cover types with similar spectral features [
41,
42,
43]. GLCM represents the co-occurrence probability for two neighboring pixels, which are separated by distance d with angle a, and with grey levels i and j [
41,
42,
43]. There are several indices derived from second-order statistical of GLCM reflecting texture characteristics. In our study, we adopted three texture indices called Contrast (CON), Entropy (ENT), and Correlation (COR), which have been proven to be effective in identifying heterogeneous objects [
43,
44,
45]. These three texture indices are determined by following equations:
where
is the largest grey level, and
,
and
,
are the standard and mean deviation of GLCM along row x and column y [
41,
42].
The above indices were calculated for 10 original bands and 7 spectral indices of Sentinel-2 imagery, and two original bands of Sentine-1 C-band imagery.
3.2. Spectral/Backscatter-Temporal Variability Features Extraction
Previous studies have proven that the spectral/backscatter-temporal variability, generated by using remotely-sensed time-series data, could be used to effectively separate mangroves from other land cover types (e.g., salt marsh, tidal flats, cropland, and terrestrial vegetation) [
27]. Here, we adopted five quantiles (10%, 25%, 50%, 75%, and 90%) of time-series imagery to quantitatively delineate the temporal profiles of spectral and backscatter signatures. The quantiles were calculated as follows: (1) combining all images during 2013 and 2018 into a multi-dimensional array; (2) for the pixels of each band, sorting all clear values from 2013 to 2018 and extracting quantiles requested. By employing the above five quantiles, it can be seen that mangroves show a remarkable difference from other land cover types. As shown in
Figure 2, the quantile values of NDVI for mangroves are much higher than those for salt marsh and tidal flats, while the quantile values of the SWIR1 band for mangroves are much lower than those for cropland and terrestrial vegetation. Additionally, the Student’s t-test result also showed that the differences between the quantile values for mangroves and those for other land cover types are significant (significance level = 0.05).
The five metrics were generated for 10 spectral bands, 2 backscatter bands (VV, VH), 7 spectral indices bands (NDVI, NDWI, MNDWI, PSRI1, PSRI2, PSRI3) and 36 (3) textural bands derived from 10 spectral bands, 2 backscatter bands, and 7 spectral indices. Finally, we in total identified 275 features per pixel.
3.3. Mapping Mangrove Forests with Random Forest
Random forest, a widely used machine learning algorithm, was applied to identify the mangrove forest, achieving a better accuracy of land cover map than any other algorithm [
27,
46]. This algorithm classifies the data by integrating votes from a mass of decision trees, which are built by fitting the features of training samples subsets generated randomly. There are three parameters governing the Random Forest algorithm: the number of trees (ntree), the minimum number of terminal seeds (nodesize), and the number of features (mtry) [
47]. In our study, ntree was set to 200, since previous studies have found that if the number of decision trees is larger than 120, the accuracies of maps become more stable [
27]. For the other two parameters, we adopted the default values (nodesize: 1; mtry: the square root of the number of all features). Finally, each classification result derived by using Random forest was recalculated using a
spatial window [
27].
To train the Random forest classifier, we collected 19378 training samples. The sample selection process contains two steps. (1) Picking sample sites. In our study, we manually selected sample sites (10 × 10 m, the size of Sentinel-2 pixel), which were evenly distributed and covered different land cover types (such as mangrove forest, water, cropland, terrestrial forest, or impervious surface) within the 10-km buffer zone along coastlines. (2) Interpreting samples as “mangrove forest” and “non-mangrove forest.” “Mangrove forest” was defined as areas in which the mangrove canopy cover is larger than 20%, while “non-mangrove forest” was defined as areas in which the area percentage of other land cover types (such as forests, shrubs, salt marshes, tidal flat, water, and cropland) is greater than 80%. For sample sites not distributed in the intertidal region, we interpreted them as “non-mangrove forest” because mangrove forests only grow in intertidal zone. For sample sites distributed in intertidal regions, we utilized Google Earth high-resolution imagery to determine whether an area was covered by mangrove forest, since, in Google Earth imagery, mangrove forests can be easily distinguished from other land cover types, such as salt marshes and tidal flat.
To explore an optimal combination of features from Sentinel-1 and Sentinel-2 time-series data for extracting mangroves, we conducted three experiments: (1) using Sentinel-1 data (S1 in
Table 4); (2) using Sentinel-2 data (S2 in
Table 4); (3) using the combination of Sentinel-1 and Sentinel-2 data (S3 in
Table 4).
3.4. Accuracy Estimation
The classification accuracy was quantitatively evaluated by applying validation samples. Based on a mangrove dataset integrating all results generated from all scenarios mentioned in
Section 3.3, the locations of validation sample sites were determined by using a stratified random sampling. The interpretation method of these validation samples was the same as the method mentioned in
Section 3.3. Finally, we got 1367 validation samples.
The above samples were applied to estimate producer’s accuracy, user’s accuracy, and F1-score of each mangrove forest map. The producer’s accuracy is the probability that a pixel was correctly classified as a given class [
48]. The user’s accuracy is the probability that a pixel classified as a certain class in the map actually represents that class on the ground [
48]. The F1-score is a useful index to assess class-level accuracy and is calculated from the harmonic mean between producer’s (PA) and user’s (UA) accuracy for mangroves as follows [
49]:
Additionally, we also qualitatively estimated the performance of our maps by overlaying them on high-resolution Google Earth imagery.
3.5. Comparison with Existing 30-m Mangrove Forests Products
Several regional and global mangrove products with 30-m spatial resolution have been generated from Landsat imagery. In this study, we mainly used three freely available products: (1) Mangrove forest map of China in 2015 by Chen et al. (2017) (MFM_2015) [
22], which was mainly generated by using information of greenness, tidal inundation, and canopy coverage derived from time-series Landsat imagery. The map was downloaded from the website [
50]. (2) Mangrove forest product of China in 2015 by Hu et al. (2018) (MFP_2015) [
27], which was produced by adopting spectral-temporal variability metrics derived from time-series of Landsat imagery. This dataset was available from the website [
51]. (3) Global mangrove watch in 2016 by Thomas et al. (2018) (GMW_2016) [
23], which was derived from cloud-free Landsat mosaics and PALSAR data. We got this product from the website [
52]. We aggregated all these maps to a common grid of 0.1° and calculated the area differences of mangrove forests between our 10-m map and the above 30-m products.
6. Conclusions
This study advances mangrove forest mapping in China by using time-series Sentinel-1 C-band SAR and Sentinel-2 MSI imagery based on GEE, and demonstrates that Sentinel imagery with 10-m resolution has the capability of accurately detecting national-scale mangrove forests. Small mangrove patches (<1 ha), which are difficult to extract using Landsat imagery with 30-m resolution, can be easily identified in the 10-m mangrove map. When applying Sentinel imagery to mapping national-scale mangroves, we found that the combination of Sentinel-1/2 time-series data is the optimal approach. In addition, we also found that the four red edge bands in Sentinel-2 make little contribution to mangrove extraction. Generally, the 10-m mangrove map, derived by combining Sentinel-1 and Sentinel-2 data, identified a total of 20,003 ha of mangrove forest in China. Among these, the total area of small mangrove patches (<1 ha) was 1741 ha, occupying 8.7% of the whole area of mangroves in China. At the provincial level, Guangxi, Guangdong, and Hainan account for nearly 93.6% of the total mangrove area in China. The largest area (819 ha) of small mangrove patches is located in Guangdong Province, and in Fujian, the percentage of small mangrove patches in the total mangrove area is the highest (11.4%). The above 10-m mangrove dataset in China can be used to analyze landscape fragmentation, quantify the budget of coastal carbon, and take effective steps to protect mangroves and restore degradation. In addition, the approach developed in our study could feasibly be used to map 10-m mangrove forests at continental and global scales based on the GEE.