**4. Discussion**

#### *4.1. Comparison with Other Mangrove Models*

Combining multi-source remote sensing data and surface observations helps us better understand global distribution of mangrove AGB. The model developed during this study is better able to explain the spatial variability in mangrove AGB (*R*<sup>2</sup> = 0.48) than those of other models. Rovai et al. [33] developed a set of statistical climatic-geophysical models based on the environmental signature hypothesis, which explained only 20% of the variability in mangrove AGB in the Neotropics. Twilley et al.'s [31] latitude-based model explained 7.6% of the variation in mangrove AGB at the global scale, while Hutchison et al.'s [32] climate-based model explained 26.7% of the variation. We used our plot data to test Twilley et al.'s latitude-based model and Hutchison et al.'s [32] climate-based model; the resulting explanatory power of these two models was much lower at 2.2% and 10.5%, respectively. There are three primary reasons. First, the initial mangrove AGB dataset was extremely small in both Twilley et al.'s (n = 34) and Hutchison et al.'s (n = 52) analyses. Insufficient training data cannot be used to create a robust global scale model. These models, therefore, have large uncertainty when validated against our larger, global data set (n = 342).

Second, machine learning methods are more suitable to estimating global mangrove AGB than multi-linear regression methods. Although the climate variables used in our model and Hutchison et al.'s [32] climate-based model were similar, the explanatory power of our model was greater because of the difference in regression methods. Several studies have demonstrated that random forest performs better than the linear regression method for estimating biomass [64].

Finally, structural information provided by GLAS and EVI improved the accuracy of random forest to estimate mangrove AGB biomass (Figure 6). Recent field studies have found that canopy height is strongly related to biomass for many mangrove species [65,66]. However, structure information provided by GLAS does not have the expected effects in this study when compared with other research into national and global forest AGB mapping. This may have been caused by the low-density footprint of GLAS in mangrove areas, limiting its ability to represent the structure variation in different mangrove species. Based on the statistical importance of each variable in our model (Figure 6), climate

factors were more important than other variables. This is similar to Simard et al.'s results in which precipitation, temperature and cyclone frequency explain 74% of the global variation in maximum canopy height [53].

**Figure 6.** The mean importance of variables for AGB estimation using the randomForest model is indicated by the percentage increase of mean-squared error (**a**) and the increase in node purity (**b**) from highest to lowest. Percentage increase in mean square error is calculated by the increase in mean square error when a variable is removed in the model. The increase in node purity is calculated based on the reduction in sum of squared errors whenever a variable is chosen to split.

#### *4.2. Comparison with Previously Published Mangrove AGB Maps*

Our estimated global mangrove AGB (1.52 Pg) was similar to that of two global maps produced using other remote sensing approaches (Table 3). Tang et al. [52] reported that total global mangrove AGB was 1.908 Pg, while Simard et al. [53] estimated it to be 1.75 Pg. Although the estimated mangrove storage was similar between these three remote sensing approaches, mean AGB density (115.23 ± 48.89 Mg/ha) in our study was lower than that of Tang et al. (146.3 Mg/ha) or Simard et al. (129.1 ± 87.2 Mg/ha). These di fferences are mainly caused by uncertainties induced by the allometric equations. Tang et al. [52] and Simard et al. [53] predicted global mangrove biomass using SRTM's tree height and a global mangrove biomass allometry equation. The mean AGB density reported by Tang et al. was the highest of these three estimates. Compared to Saenger and Snedaker's global mangrove height-biomass relationship used by Tang et al. [52], Simard et al. [53] applied 331 in situ plots across a wide variety of mangrove forest ecotypes to fit a global equation between AGB and basal area-weighted height. Our results for global mangrove AGB storage and mean AGB density were similar to that of Simard et al. [53] because both methods utilized spaceborne LiDAR data. Traditionally, mangrove forest aboveground biomass derived using synthetic aperture radar was underestimated due to its limited ability to penetrate the mangrove canopy. Simard et al. [53] used GLAS data to correct the SRTM tree height, thereby overcoming the issue of estimating mangrove AGB from SRTM tree height.

**Table 3.** Comparison of total mangrove AGB and area with previously published results.


The estimated global mangrove AGB storage in our study (1.52 Pg) was significantly lower than those from non-remote sensing approaches (Table 3). Twilley et al. [31] estimated global mangrove AGB at 2.34 Pg based on a latitude model, nearly 54% higher than our result. Hutchison et al. [32] used a climate-based model and predicted that total global mangrove AGB storage was 2.83 Pg, 86% higher than our result. The di fference in the baseline mangrove extent could be a major reason for the variation in these results. Although our study and that of Hutchison et al. [32] both used the mangrove map developed by Spalding et al. [19], the final global mangrove area in our study (13,065,675.00 ha) was 15% smaller than that used by Hutchison et al. [32] (15,314,094 ha). This di fference was caused, in part, by inconsistent land boundaries between our predictor variables, mangrove distribution map, and country extents. These layers have di fferent spatial resolutions and extents, so small mangrove patches along the coast or in the islets were omitted during our analysis. These places are also areas with a large distribution of mangrove [19]. Consequently, the disparity in area led to variations in total mangrove AGB storage between the two results. Part of the variation can also be explained by the models used by Hutchison et al., which may overestimate mangrove AGB [32]. Rovai et al. [33] found that these climate- and latitude-based models overestimated mangrove AGB by 25.3% to 44.4% in the Neotropics region. In addition, the structural data provided by spaceborne LiDAR in this study can provide better information for estimating mangrove AGB at larger geographical scales, thereby reducing uncertainty in estimates of mangrove AGB storage.

Most of the 10 countries with largest total mangrove AGB stock from our study were also reported in other research, such as that of Hutchison et al. (2014) [32] and Simard et al. (2019) [53], but the order in which these countries appear on the list was di fferent. Indonesia has the largest mangrove AGB stock, which is consist in each study, even though the mangrove AGB in Indonesia and Papua New Guinea reported by our study was much lower than those of Hutchison et al. (2014) [32] and Simard et al. (2019) [53] (Figure 7). This phenomenon maybe caused by the model we used to predict biomass. Validation (Figure 5) showed that our model tended to underestimate mangrove AGB density at high values (>125 Mg/ha) since observations are limited in these high biomass areas. The mangrove AGB stock in Mexico, Cuba, and Colombia differed between the three studies. The difference in Mexico and Cuba was induced by a bias in predicted mean mangrove AGB density in the different studies. Adame et al. (2013) reported that the AGB in tall, medium and dwarf mangroves in the Mexican Caribbean were as much as 176.2, 114.2 and 7.1 Mg/ha, respectively [68]. The mean mangrove AGB density of Mexico in our study was 113.30 Mg/ha which is closer to that of the medium mangroves reported by Adame et al. (2013). Simard et al. (2019) [53] reported a mean AGB in Mexico of 37.9 Mg/ha, which is much lower than that of the medium mangroves. This underestimation in Simard et al. (2019) [53] may have been caused by using a global allometric equation to predict biomass. The difference in Colombia was mainly caused by inconsistencies in mangrove extent. The mangrove area in Colombia reported by Simard et al. (2019) [53] is much lower than that in our study and in Hutchison et al. (2014) [32].

**Figure 7.** Comparison of (**a**) mean mangrove AGB density, (**b**) mangrove area, and (**c**) total mangrove AGB stock between our study, Hutchison et al. (2014) and Simard et al. (2019) in ten countries with the highest mangrove AGB.

#### *4.3. Limitations and Future Studies*

Although our model estimates global mangrove aboveground biomass fairly well, there are limitations to this study. Available observation data was limited when compared with other regional and global studies. Field data is fundamental for accurate estimations of global mangrove biomass. In this study, we collected 510 records from a number of sources but more than 30% of them could not be used because of uncertainty in location information. Moreover, geolocation errors in mangrove plots cannot be reduced by the point-radius model used by Su et al. [47] and Hu et al. [46], because randomly shifting mangrove plot locations had a large probability of relocating the plot into the ocean. Furthermore, the mismatch of spatial observation scales between plots and remote sensing data was a problem. Researchers have begun recently to use drone-based LiDAR to retrieve mangrove biomass [69], using it as a bridge to scale AGB from the plot level to the scale of satellite observations [70]. The increase in drone-based LiDAR data in mangrove areas will benefit global mangrove forest biomass mapping efforts in the future. Second, sparse GLAS datapoints within areas of mangrove lose some of the variability in structure during extrapolation. Even though we used GLAS data within a 100-km bu ffer of the coast to increase the number of GLAS datapoints, the explanatory power of the extrapolation models were nearly 10% lower than those for China and global forest mapping [46,47]. Fortunately, the Global Ecosystem Dynamics Investigation (GEDI) project [71] recently started collecting global waveform LiDAR data, which will provide higher density data with a smaller footprint than GLAS. This data will help us better understand variability in mangrove structure and biomass distribution. Third, factors such as salinity [72] and river discharge [73] that specifically control the distribution and production of mangrove forest should be added to the model in the future. With these factors, we can more accurately estimate the biomass and better understand how mangrove AGB varies under di fferent environmental conditions.
