2.1.

Study Area

Recent global mapping of tidal marsh found a third of the global extent to occur within the USA.³⁶ In the CONUS, salt marshes are composed of a variety of halophytic vegetation, including Spartina alterniflora, Salicornia spp., Juncus gerardii, Phragmites australis, Distichlis spicata, and Spartina patens. Salt marshes are concentrated in the back bays of barrier islands, estuarine bays, and deltas in low-energy environments, allowing for sediment accumulation.³⁷ In the United States, salt marshes have experienced significant conversion to other land uses, especially in urban areas such as Boston, which is estimated to have lost 81% of its salt marsh extent.³⁸ Salt marshes can be found across the Atlantic, Pacific, and Gulf coasts of the United States. Salt marsh species vary with tidal amplitude, temperature, salinity, and elevation.

2.2.

AGB Modeling

We used Google Earth Engine (GEE) to process the time series data and the MLR package in R statistical software (3.6.2) to train machine learning models to predict AGB.³⁹ Previous studies found the lack of overlap between Sentinel-1/2 and the in situ data collection limited its usefulness in modeling biomass.¹⁸ To address this issue, we identified stable salt marsh pixels, i.e., those that changed little from the biomass sampling year to Sentinel data collection. Stable sampling areas then had the average Sentinel-1 VV, and VH sampled for the 2 weeks preceding and the 2 weeks after the sampled date across the Sentinel-1 archive. We applied the same method to process Sentinel-2 10 m bands and Normalized Difference Vegetation Index (NDVI). A total of 530 samples were deemed stable. Additional biomass samples from the Georgia Coastal Ecosystem Long Term Ecological Research Reserve in 2017 were retrieved from their database and incorporated into the model for 723 stable training points (Fig. 1). We used this subsample to train a 10-m biomass model. We compared the stable model with models trained with all available training locations ($n=984$). The validation included locations from 2019 for the GCE LTERR [Dataset]⁴⁰ and sites from Plum Island, MA [Dataset].⁴¹

Fig. 1

Model training data and source from either Byrd et al.¹⁸ or GCE LTER Project and Pennings.⁴⁰ Service Layer Credits: Esri, Garmin, GEBCO, NOAA NGDC, and other contributors.

Three machine learning algorithms, SVM, extreme gradient boosting (XGBoost), and random forest, were compared with out-of-box cross-validation and in situ validation data. Previous studies have demonstrated the potential of this training data¹⁸ for regional²⁵ and national monitoring of AGB.²⁶

2.3.

Time Series Stability

We explored the temporal stability of in situ training data locations collected before the operation of Senintel-1 and Sentinel-2. We assessed training data stability with a time series analysis of the Soil-Adjusted Vegetation Index. In situ, continuous monitoring plots demonstrate significant variation in AGB from 343 to $1324\mathrm{g}{\mathrm{m}}^{-2}{\text{year}}^{-1}$ in Plum Island.⁴² The biomass plot size ranges from 0.0625 to $1{\mathrm{m}}^2$, much finer than the satellite imagery (10 to 30 m). While fine-scale seasonal variation is minimized at the pixel scale, shifts in vegetation composition or disturbance could cause changes. This study assesses pixel stability using the Landsat archive. In situ plots collected at the exact location and month were consolidated into a single training point with average biomass and imagery variables. Trend and breakpoints were calculated for the resulting 830 training points. We decomposed the time series using the Prophet package in R, isolating trend and seasonality to determine stability.⁴³ The training data had many more gain points than losses. Therefore, we derived a change threshold of 0.05 SAVI from the 0.05 quantile of loss. Two hundred sixty-one training locations were outside this threshold, and SAVI increased in 84% of these locations. These excluded points were further analyzed with Breaks for Additive Season and Trend to determine if they had experienced a break following biomass collection (Fig. 2). The monitoring period was from 1999 to the beginning of the year of sampling. We found that 100% of points with an absolute trend $>0.05$ also experienced a break. As a final test, we compared models trained with all data and stable training data with two in situ testing sets to verify the performance of this stable training set.

Fig. 2

Stability assessment of a point at Twitchell Island, CA. Time series demonstrates significant divergence from the expectation following biomass field data collection.

Tides affect salt marshes in various ways and have been addressed in remote sensing research.^25,44,45 The Landsat imagery was tidally filtered using the methods detailed in Campbell and Wang²⁵ and adapted to GEE. All values were cloud, quality filtered, and temporally filtered from June 01, 2020, to September 30, 2020. We did not filter the Sentinel-2 imagery; instead, to avoid cloudy data and minimize the impact of the tidal stage, the mean of the 60% to 80% range of data was used to target the high biomass period while minimizing cloud impacts.

2.4.

Salt Marsh Extent

This study updates the NWI maps with a combination of Sentinel-1 and Sentinel-2 and an ensemble machine learning approach to predict the probability of a pixel being an emergent estuarine wetland. We used three machine learning algorithms (rotational forests, SVM, and XGBoost) in an ensemble approach. GEE was used to preprocess all remote sensing inputs.⁴⁶ Variables included Shuttle Radar Topography Mission (SRTM) (resampled to 10m), National Elevation Dataset, average Sentinel-1 VV, Sentinel-1 VH, all 10m Sentinel-2 bands (2, 3, 4, and 8), and NDVI and Normalized Difference Water Index. Mean values were calculated after filtering for cloud, quality, and temporally from June 01, 2020, to September 30, 2020.

Each model predicted the probability that a pixel within 1 km of NWI layers estuarine emergent class was salt marsh. We calculated the standard error of these probabilities and used that to predict a spatial estimate of extent uncertainty. We compared the performance of this uncertainty analysis with confidence intervals derived from the methods of Olofsson et al.⁴⁷

Models were trained for individual watersheds or clusters of similar and spatially proximate watersheds. Approximately 5000 random points were placed within 1 km of salt marsh and designated salt marsh or not by the NWI. Using high-resolution imagery, we examined each salt marsh location to confirm that it was still a salt marsh in 2020. We used R statistical software (3.6.2) to train and classify each watershed area within 1 km of the NWI salt marsh boundary. These classifications were then post-processed, including merging areas smaller than three pixels within the marsh extent and requiring salt marsh to be within 100 m of salt marsh in the NWI. We conducted an accuracy assessment across the CONUS with a stratified random sample of 10,000 points.

2.5.

Drivers of AGB in CONUS Salt Marsh

We explored spatial autocorrelation with Moran’s I and Local Moran’s I. Then, we utilized machine learning (XGBoost) and Shapley values to understand the relationship of AGB to drivers. Shapley values are a game theory approach to determining the variable contribution to a particular modeled outcome (Sundararajan and Najmi, 2020).⁴⁸ AGB was explored relative to potential drivers at a 3 km scale ($n=8347$). Explanatory variables included PRISM climate data (August 2020 total precipitation and mean temperature);⁴⁹ average regional sea-level rise, tidal amplitude, and relative tidal elevation;⁵⁰ August 2020 composite of chlorophyll-a (NOAA/NESDIS/STAR 2022a);⁵¹ August 2020 composite of diffuse attenuation coefficient (NOAA/NESDIS/STAR 2022b);⁵² hurricane landfall and intensity;⁵³ C-CAP LCLU variables;⁵⁴ and water extent and change.⁵⁵ These variables represent likely drivers of biomass across the CONUS that are available at a $<3{\mathrm{km}}^2$ spatial resolution.

3.1.

Model Comparison

The three machine learning algorithms’ internal metrics performed similarly. For example, the out-of-bag (OOB) root mean square errors (RMSEs) were all within $100\mathrm{g}{\mathrm{m}}^{-2}$ (Table 1). The inclusion of Sentinel-1 and 2 had minimal impact on OOB metrics. However, we found a notable improvement in the RMSE of XGBoost over the other algorithms for the Georgia 2019 validation data at the site and vegetation plot scale. The Plum Island sampling extents were polygons (136 to 874 salt marsh pixels) with no exact sampling locations [Dataset];⁴¹ therefore, we calculated a min, mean, and max values within the classified salt marsh extent for the location and compared, i.e., if a location had $<500\mathrm{g}{\mathrm{m}}^{-2}$ of biomass, we compared it with the minimum for that region, biomass between 500 and $1000\mathrm{g}{\mathrm{m}}^{-2}$ was compared with the mean, and areas with greater than $1000\mathrm{g}{\mathrm{m}}^{-2}$ were compared with the maximum of a region. XGBoost was the best-performing algorithm across the board and consistently performed better with the stable training data set. The models trained with only Landsat data performed worse in most metrics and were limited to a 30 m resolution. Therefore, we used XGBoost with Sentinel-1/2 to classify AGB for salt marsh areas across the CONUS. These results demonstrate similar uncertainty to other machine learning approaches in salt marsh environments, e.g., Chen et al.⁵⁶ found an RMSE of $371\mathrm{g}{\mathrm{m}}^2$ when predicting Spartina alterniflora.

Table 1

Machine learning model algorithm performance in both the OOB and validation datasets.

Spatial variables ($x$ and $y$), Landsat band 6, and Sentinel-1 VH polarization were the model’s four most important variables (Fig. 7). Sentinel-2 NIR and NDVI were the ninth and eleventh most important variables, respectively. Although the 10 m models performed better compared with the 30 m results (Table 1), the Landsat inputs comprised 6 of the 10 most important variables, suggesting that relying on the Sentinel data alone would result in a loss of predictive power. The importance of Landsat bands is probably due to the time between in situ and Sentinel data collection.

3.2.

Extent Classification

We estimated a CONUS salt marsh extent of $\mathrm{14,491}{\mathrm{km}}^2$. We found an overall accuracy of 96.34%. Following the methods of Olofsson et al.,⁴⁷ we determined a confidence interval of $3175.6{\mathrm{km}}^2$, slightly less than the confidence interval derived from our machine learning approach ($3473.5{\mathrm{km}}^2$). These results suggest the multiple machine learning algorithm approach to be a reasonable estimate of error and useful for providing the locational uncertainty, which can result in a more precise understanding of AGB. The difference between the spatially derived upper and lower confidence intervals is evident when examined visually (Fig. 3). The most prominent differences between the low and high extents were the inclusion of similar ecosystems such as tidal mudflats and upland areas of possible salt marsh transition. In many areas, these differences were minor but represent 26% of the average estimate of biomass in the CONUS.

Fig. 3

Difference between the upper and longer spatial estimates of salt marsh extent. The upper extent includes all areas within the lower extent. Sentinel-2 imagery (NIR, G, B) in the background for a section of Deal Island, Maryland.

3.3.

AGB Across the CONUS

The patterns of AGB varied significantly across the CONUS, including within individual watersheds. In the CONUS, a max AGB of $1735\mathrm{g}{\mathrm{m}}^{-2}$ was found in HUC6 180701 (Ventura-San Gabriel Coast). In total, the CONUS had 8.32 (7.15 to 9.35) Tg of AGB in 2020, which using the conversion of 0.441 from Byrd et al.¹⁸ would be 3.67 (3.15 to 4.12) Tg C. The upper salt marsh extent increased AGB slightly less than the lower extent decreased it. This reflects the inclusion of more low biomass areas in the upper salt marsh extent such as pixels with a mosaic of vegetation and unvegetated areas or tidal flats (Fig. 3). In comparison the entirety of North America’s grasslands is estimated at 207.72 Tg C and an average of $75\mathrm{g}\mathrm{C}{\mathrm{m}}^{-2}$⁵⁷ compared with an average of $255.7\mathrm{g}\mathrm{C}{\mathrm{m}}^{-2}$ in salt marshes. Our average is significantly lower than global estimates from the literature ($430\mathrm{g}\mathrm{C}{\mathrm{m}}^{-2}$) but very similar to the median value of $240\mathrm{g}\mathrm{C}{\mathrm{m}}^{-22}$. This is due in part to not every ${\mathrm{m}}^2$ within a pixel being vegetated. Recent studies modeling gross primary production across tidal marshes of the CONUS found significantly higher productivity ($4.32\pm 2.45\mathrm{g}{\mathrm{C}/\mathrm{m}}^2/\text{day}$⁵⁸); this measure includes woody vegetation, the inclusion of pixels with high coverage, and coarser spatial resolution. Visually, biomass followed many expected trends when aggregated to $3\times 3\mathrm{km}$ areas, with low marsh areas having lower biomass and high marsh areas having slightly higher biomass (Fig. 4). These patterns were further explored within our explanatory machine learning analysis.

Fig. 4

(a) AGB for the Mid-Atlantic, USA, summed for the maximum extent of salt marsh within a $3\times 3\mathrm{km}$ area. (b) Latitudinal plot of total AGB summed for each 0.1 decimal degree. (c) AGB for the Gulf of Mexico, USA, summed for the maximum extent of salt marsh within a $3\times 3\mathrm{km}$ area. (d) Longitudinal plot of total AGB summed for each 0.1 decimal degree. Figure 8 is a CONUS-wide map.

We compared the annual variation with the extent of uncertainty for a single year. These variables are expected to have a significant impact on AGB estimates, but how great an impact is unclear. We estimate their effect on the Lower Chesapeake Bay Watershed. The extent of uncertainty results in a range of 0.34 to 0.604 Tg of AGB in the watershed. In comparison, if the extent midpoint is used, the range of annual variation (2015 to 2020) was between 0.48 and 0.70 Tg of AGB. If the NWI extent is utilized, a very similar total AGB to the upper extent is estimated (0.607 Tg of AGB). Annual variation had a slightly larger effect on AGB uncertainty than the extent did, and we further explore the environmental variables that determine this variation with our machine learning analysis.

CONUS-wide median biomass in $3\times 3\mathrm{km}$ bins with at least 20 pixels varied between a max of 1227.2 and a min of 38.4 occurring in HUC 6 watershed 020700 and 030902, respectively [Fig. 5(a)]. The Gulf of Mexico, Mid-Atlantic, and San Francisco Bay all had high median estimates of AGB. A pattern of increased AGB as you move inland is noticeable in the Mississippi Delta, suggesting that the increased AGB seen in Fig. 4 is due to both extent and AGB increases [Fig. 4(b)]. In the Chesapeake Bay, an increase in median AGB is evident in the fresher tributaries potentially due to Phragmites australis, which can have significantly more biomass than more salt-tolerant species such as Spartina patens.⁵⁹ The standard deviation was consistent across much of the CONUS with low AGB regions such as Georgia, southern Florida, Maine, Pacific Northwest, and Texas having lower standard deviations [Fig. 5(b)].

Fig. 5

(a) The median biomass across the CONUS and (b) the standard deviation across the CONUS. The $3\times 3\mathrm{km}$ squares had the median and standard deviation of all AGB estimates calculated.

3.3.1.

Biomass drivers

We found significant spatial autocorrelation of AGB using Moran’s I ($I=0.84$, $p<0.001$). AGB clustering is expected due to climatic and tidal effects on AGB. Continuing our analysis of drivers, the four variables with the highest absolute Shapley value were tidal amplitude, relative sea level rise (RSLR), precipitation, and temperature (Fig. 6, Table 2). These impactful variables were split evenly between tidal/elevation and climate drivers. The drivers varied by region; e.g., hurricane landfall/category was the fifth most impactful variable in the Gulf Coast region. The response curves of the elevation drivers fit expectations with low rates of RSLR, which has limited or even negative effects on AGB and increases to a plateau. By contrast, high rates of RSLR ($>5\mathrm{mm}{\text{year}}^{-1}$) had a large negative effect on AGB. The response of AGB to tidal amplitude was interesting, with nontidal systems having lower biomass and a slight increase in microtidal systems. Then, AGB declined as amplitude increased [Fig. 6(d)]. This response is likely due to the greater range of vegetated tidal elevations in these high tidal amplitude systems and the greater likelihood of inundated pixels impacting the analysis. The response of AGB to precipitation fits the expectation, with low monthly precipitation reducing AGB. Drought is a major driver of salt marsh die-off.⁶⁰ By contrast, the relationship to temperature was more complicated, with AGB increasing to an inflection point around 26°C, which corresponds with approximately Cape Fear, South Carolina, coinciding with the upper extent of hurricane impacts in 2020.

Fig. 6

Four most impactful drivers of AGB across the CONUS with examples of their spatial variation. Response curves demonstrate the average response of AGB to the variable at a certain value. (a) Precipitation visualized for a region of the Atlantic coast, (b) temperature for a region of the southern Atlantic coast, (c) RSLR for a section of the eastern shore of Maryland, and (d) tidal amplitude visualized for a region of Chesapeake Bay.

In situ studies have examined climate drivers of AGB in salt marshes, including for a single species,⁶¹ long-term spatio-temporal trends (Bice et al. 2023),⁶² and latitudinal gradients of AGB’s relationship to belowground biomass (BGB).⁶³ River discharge was found to be a major driver of Spartina alterniflora AGB, especially along creek banks.⁶¹ Our study found precipitation to be a major driver, which is the closest analog to discharge in our set of drivers. A spatiotemporal analysis of AGB in a single watershed found temperature, river discharge, drought, sea level, and river nutrient concentrations to drive AGB (Biçe et al.). Precipitation was the only variable that was not found to have a causal link to AGB (Bice et al.). Our study did not have discharge data, river nutrient concentrations, or draught indicators, which likely have high covariance with precipitation. When a latitudinal analysis of the ratio of AGB to BGB was conducted on the Atlantic Coast, trends in the increased allocation of BGB with lower temperatures were observed.⁶³ Our study does observe similar declines in AGB going from South Carolina to Massachusetts. However, Crosby et al.⁶³ only examined Spartina alterniflora’s latitudinal response. These studies lack our study’s large geographic scope but demonstrate similar trends, such as increased biomass as temperature increases.

The most significant variables did vary by coast. For example, the hurricane category was most impactful in the Gulf because that region had direct landfalls. Low precipitation likely explains the lower-than-expected AGB across the Gulf for 2020. The regional exploration of drivers allows for identifying climatic, tidal, and LCLU patterns that affect salt marsh biomass. Understanding current AGB and its drivers is critical for improving ecosystem resilience and change estimates.

AGB is a minor component of the larger salt marsh carbon budget, i.e., the mean carbon density of soil organic carbon (SOC) of salt marshes in the CONUS is $27.0\mathrm{kg}\mathrm{C}{\mathrm{m}}^{-3}$. Still, SOC stock loss is more likely in marshes with low or no AGB.⁶⁴ This product’s 10 m spatial resolution allows for finer scale determination of these loss areas. Repeat extent classification and AGB prediction can enable improved carbon monitoring. Future research directions should include global AGB prediction in tidal marsh ecosystems and assessment of regional trends in AGB at a 30 m spatial resolution.

This dataset has several limitations, including the focus on a singular blue carbon ecosystem, which ignores some of the complexity of coastal ecosystems and the gradient across these landscapes. The tidal filtering is limited in this approach and could be improved by incorporating algorithms with potential applications on Sentinel-2 data, such as Flooding in Landsat Across Tidal Systems.⁶⁵ Applications that require a high temporal resolution, such as monitoring the seasonality, would benefit from products such as Harmonized Landsat Sentinel-2⁶⁶ with little loss of information from the 30 m spatial resolution. AGB represents a small portion of the total carbon within these ecosystems, and individual pixels demonstrate high uncertainty when evaluating the test datasets. Using multiple sensors compounded the potential geolocation error and tidal stage impacts, which is one reason that uncertainty was reduced at the site level.