A review of carbon monitoring in wet carbon systems using remote sensing

Not long ago, mangrove biomass and carbon estimates relied upon the extrapolation of field data, environmental conditions, and partial extent maps [e.g. 31, 41–45]. Giri et al [46] created the first global mangrove map using Landsat imagery. This map and other advances in remote sensing have enabled regional-to-global-scale analyses of mangrove carbon stocks and carbon stock change [21, 36, 38, 40, 47–49]. Mangrove carbon monitoring combining field-based surveys and remote sensing occurs across; local [e.g. 50–56], regional [e.g. 49, 57–62] and global [21, 34, 36, 63–67] scales. Continued advancement, including machine learning, have led to recent studies classifying species [68–71], quantifying height distributions and biomass [36, 72–74], change in extent [40, 47, 49, 75] and stand age [60], and productivity [76, 77].

Passive sensors are used to map mangrove extent, change, and extrapolate C storage with field data [8, 59, 60, 78–81]. Active sensors (e.g. light detection and ranging and radar) can measure mangrove structural attributes, such as canopy height. Simard et al [56] first derived accurate height estimates in the Everglades with Shuttle Radar Topographic Mission (SRTM). Subsequently, canopy height was estimated using satellite stereo images, Synthetic Aperture Radar (SAR) interferometry, and lidar [56, 60, 82–84]. Canopy height enables estimates of aboveground biomass [e.g. 36, 85–88]. Additional active spaceborne sensors (e.g. SRTM, Sentinel-1, TanDEM-X, ICESat, and GEDI) have improved canopy height models [e.g. 82, 84, 89] enabled the identification of change hotspots [39, 40, 49], and the development of mangrove carbon monitoring initiatives [37, 47, 60, 90]. The Japanese Aerospace Exploration Agency L-band SAR sensors (ALOS and ALOS-2) are an important active sensor for mangrove mapping, including the identification of invasive species [58], prediction of aboveground biomass (AGB) [74, 91], and long-term monitoring [39, 92]. Medium resolution sensors have enabled global-scale analysis but can miss small mangrove patches and edges or small-scale restoration efforts.

The recent increase in the resolution and accessibility of satellite imagery has provided fine-scale mangrove data products suitable for MRV. The European Space Agency's (ESAs) Sentinel-1 and Sentinel-2 launched in 2014 and 2015, respectively, increased the spatial resolution of new mangrove maps from 30 m (i.e. Landsat) to 10 and 20 m [53, 93, 94]. Moreover, access to high-resolution satellite, aerial, and unoccupied aerial systems (UAS) imagery has further increased the spatial resolution of mangrove maps (<5 m) [51, 59, 60, 70, 71, 79, 94, 95]. Data fusion with combinations of multispectral, hyperspectral, lidar, radar, and high-resolution data have been applied to increase the spatial and temporal resolution of mangrove carbon storage and flux estimates [60, 88]. The increased temporal resolution also facilitates monitoring of short-term disturbance and recovery [8, 49].

Coarse spatial resolution sensors such as MODIS are also informative and often used with other satellite imagery [96]. The high temporal resolution of MODIS is particularly beneficial when tracking net primary productivity (NPP) [76] or gross primary productivity (GPP) change [55, 77], including due to disturbance events like hurricanes [97] and insect outbreaks [77].

Field and climate studies provided the first global mangrove carbon models [41, 63] and continue to be essential for monitoring mangrove carbon [30, 66, 98, 99]. Mangrove height and biomass models have increased in accuracy, providing improved estimates of aboveground C stocks and change through restoration [100, 101], afforestation and encroachment [50, 51, 59, 102, 103], natural disturbances [40, 49, 57, 58] and local anthropogenic impacts [40, 49, 57, 58]. Anne et al [104] modeled mangrove soil carbon with hyperspectral data, which improved on Landsat-based models. Global mangrove carbon density has been extrapolated from 250 m to 30 m with a combination of machine learning, earth observation, and ancillary data [e.g. 21, 34].

Remote sensing further complicates the quantification of uncertainty in carbon monitoring (figure 2). Simard et al [36, 56] demonstrate that allometric equations can introduce considerable bias (>100%). However, the remote sensing canopy height model error was low with a root mean square error (RMSE) of 2 m). In situ carbon monitoring samples are limited globally. If these samples are not representative, uncertainty will be high and unquantified. Extrapolating carbon stocks and fluxes from relatively few in situ measurements makes the accurate quantification of spatial uncertainty extremely important. For example, Sanderman et al [21] used the existing 250 m SoilGrid data, in situ training data, and Landsat imagery to create a 30 m organic carbon stock (OCS) map. The study resulted in an average uncertainty of 40.4% of the mean OCS [21]. Remote sensing methods can quantify the spatial uncertainty improving stakeholder understanding of regional carbon estimates and accuracy.

Despite comprising only 0.3% of global coastal ocean area, mangroves contribute ∼55% of air-sea CO₂ exchange from the world's wetlands and estuaries, 60% of dissolved inorganic carbon (DIC) and 27% of dissolved organic carbon (DOC) from tropical rivers to the coastal ocean [105–107]. Over half of mangrove carbon production was unaccounted for until recently [45, 108], when mangrove carbon export (particularly DOC and DIC) were quantified [106, 107]. Only 14 TgC yr⁻¹ of mangrove NPP is buried in soils, while export to coastal oceans is approximately an order of magnitude higher (158 TgC yr⁻¹) [107]. Mangroves export an estimated 15 Tg particulate organic carbon (POC) yr⁻¹, 51 Tg DOC yr⁻¹, and 124 Tg DIC yr⁻¹ to coastal oceans [106, 107]. Models of river and tidal flow through mangroves informed by remote sensing have improved estimates of carbon export [126], identifying relationships between environmental conditions (tidal height, river-flow, precipitation, biogeochemical constituents of water) and carbon export associated with tidal pumping [45, 126], particularly of DIC [127]. Furthermore, ocean color techniques can identify the source of organic matter through absorption coefficients [109, 110], allowing for detection of mangrove derived chromophoric dissolved organic matter (CDOM) and DOC [111]. Carbon export from mangroves is spatially and temporally heterogeneous, and remote sensing can help resolve this variability indirectly through characterizing water flow and directly through the identification of CDOM.

Remote sensing has been essential for carbon monitoring of mangroves due to their unique landscape position, structure, and spectral characteristics. These data have enabled relatively precise quantification of mangrove extent, carbon stocks, and carbon fluxes from local-to-global scales (table 1). Mangroves are among the most carbon-dense ecosystems (table 1) [107] and are likely to become increasingly impacted by anthropogenic and natural disturbances [112]. Continued remote sensing carbon monitoring is necessary with a particular focus on climate-related range-shifts associated with sea-level rise (coastal contraction and inland expansion [113, 114]) and poleward range expansion [115–118].

Tidal marsh studies utilized earth observation data to constrain and upscale in situ data, predict biomass and soil organic carbon (SOC) stocks, and model productivity. Land-use change was a primary theme, including migration, invasion, and long-term monitoring. The most common carbon indicators were GPP and biomass. Other indicators included sedimentation, leaf area index (LAI), vegetation fraction, nitrogen, and gas fluxes [104, 139–141]. Temporal dynamics of spectral indicators of biomass, i.e. normalized difference vegetation index, were explored in tidal flats, too [142, 143]. Most tidal system studies (n = 33) pertained to marshes dictating this section's focus.

GPP is a common carbon indicator for tidal systems (n = 8). MODIS combined with eddy covariance towers was used to predict GPP in tidal environments [144–146]. Less common were gas flux chambers and incubation [139, 143]. Feagin et al [147, 148] improved on the MOD17 GPP product with an ecosystem-specific model. Tidal inundation is a source of uncertainty within GPP estimation, and studies addressed the tidal stage with spectral index filtering and tidal modeling [143, 145, 149]. These studies primarily rely on MODIS at a minimum spatial scale of 250 m, biomass, derived from Landsat or other high-medium resolution sensors is often used to track finer scale change.

In the 1980s, tidal marsh AGB was first predicted with in situ spectral measurement and expanded to Landsat imagery [150–152]. Since those foundational studies, researchers have assessed other sensors' capacity to predict AGB, including Worldview-2, Hyperion, UAS, lidar, MODIS, AVIRIS-NG, Planetscope, and data fusion [141, 153–160]. A major limitation of biomass prediction in tidal marshes is the site and species-specific limitations of the modeling results. Studies have sought to address this limitation with model transfer but resulted in inaccurate predictions [157], though regionally trained models have been successful [161–163]. AGB prediction scope and accuracy have increased since the first modeling approaches, but scale and uncertainty limit their applicability to global carbon monitoring.

Tidal marsh change was a frequent research topic, including tracking invasive species [164], determining marsh migration [165], time-series change analysis [166], and multitemporal regional change [167]. Studies frequently upscaled carbon measurements with land-use maps [167–170]. Braun et al [171] determined that geomorphic change can dictate whether and how freshwater coastal wetlands serve as sources or sinks for terrestrial carbon and how carbon stocks can fluctuate on a geologically rapid timescale. A few studies used remote sensing data to constrain in situ sampling with land-use maps [165, 172]. The lack of baseline data availability and a focus on local methods limited regional monitoring applications.

The lack of a global extent map and change estimates limits the use of remote sensing in tidal marsh carbon estimates (table 2). CMS has supported the development of the US coastal wetland greenhouse gas inventory [173]. In the contiguous US between 2006 and 2011, coastal wetlands emitted 10.3 Tg CO₂e yr⁻¹ (1.6–21.3 Tg CO₂e yr⁻¹), and a robust sensitivity analysis demonstrates major sources of uncertainty where remote sensing could improve the model, including coastal salinity classifications—and resulting CH₄ emission categories—and the depth of soil deposits lost to erosion [174, 175]. Improved predictions on the fate of soil carbon following marsh loss events could combine earth observations and additional ocean physical modeling. Carbon stock values are well constrained compared to the uncertainty of methane emissions and loss events [174, 175]. So far, strategies for mapping US coastal wetland soil carbon stocks using nationally available soil and wetland maps have not outperformed simpler strategies of applying a single average value for carbon stocks. Holmquist et al [175, 176] utilized an extensive soil core database to predict tidal marsh soil carbon to 1 m depth (0.72 PgC) within the Contiguous United States (CONUS). The study also showed a way to improve future mapping would be to generate maps based on environmental drivers that differentiate between organic and inorganic soils, differentiated by a threshold of 13% organic matter by dry mass. Elevation relative to the tidal amplitude [177, 178], and long-term rates of relative sea-level rise [179] could be potential predictors of carbon stock. These CMS funded studies demonstrate the need for connecting earth observations and models between land, wetland, and open water; further in situ data collection of environmental driver data such as salinity and tidal elevation; and the development of tidal marsh class and change products that can be applied globally.

Additionally, global carbon export from tidal marshes to estuaries is uncertain. The connection between tidal marshes and coastal waters is a long-standing consideration. Teal [185] identifies outwelling as an important potential component of the system, and its magnitude and role have been debated since [186]. The magnitude of C export is highly variable, with tidal marshes being both a sink and a carbon source to coastal waters [187]. Salt marshes export an estimated 3.3 Tg POC yr⁻¹, 14 Tg DOC yr⁻¹, and 29 Tg DIC to coastal oceans [107]. Remote sensing of ocean color to estimate DOC and CDOM can discern spatial and temporal patterns of tidal marsh export [188]. Gao et al [144] explored the connection between tidal marsh productivity and detritus export using in situ sampling of detritus. Monitoring coastal waters is a difficult remote sensing task (see sections 3.1.3 and 3.3.1). The use of ocean color methods and fine-scale satellite imagery could enhance the capacity to monitor C export from tidal marshes.

Seagrass biomass is below the water's surface; therefore, atmospheric and coastal water conditions influence mapping [190]. Similarly, temporal and spatial variability in water quality and depth hinder seagrass identification e.g. [193–197]. These difficulties can result in misclassification between seagrass and algae [197–201]. Due to the remote sensing challenges, seagrass mapped extent is an order of magnitude less than modeled extents [202, 203]. Consequently, scientists lack a global map of seagrass extent, and recent estimates are uncertain (160 387–266 562 km²) [204]. Seagrass aboveground carbon stocks are even more uncertain due to mapping error and regional, intraspecies, and interspecies variability in biomass [199, 203, 205, 206]. Globally, two-thirds of seagrass living carbon (2.52 ± 0.48 Mg C ha⁻¹) is belowground, and seagrass SOC is ∼65 times greater (165.6 Mg C ha⁻¹) [190].

Novel methods for linking remote sensing and in situ data have improved our understanding of seagrass cover and carbon storage. For example, seagrass cover estimates from UAS and in situ images can bridge the scale differences of AGB samples and remote sensing imagery [194, 196, 197]. Seagrass extent mapped with UAS imagery has been used to scale in situ carbon samples to the landscape by percent cover [207]. Zoffoli et al [208] used a linear model to predict biomass with in situ radiance (RMSE = 5.31 g m⁻²) and applied that to Sentinel-2 imagery, successfully capturing seasonality. Modeling optical properties of seagrass has led to the development of a model to estimate LAI that does not require in situ data [209, 210]. In addition to satellite and aerial platforms, ship-based acoustic sensors can identify species [211] and estimate biomass [205]. Data fusion between ship-based sensors, satellites, and UAS has improved seagrass extent maps [212] and benefit biomass mapping.

Mapping was the primary seagrass research topic reviewed due to the challenges of modeling seagrass carbon and the need to address data gaps in known seagrass extent. These challenges have resulted in high uncertainty in seagrass extent estimates (table 3). For example, high-resolution imagery, informed by a species distribution model, was used to manually digitize seagrass beds within a single bay, resulting in a 44% increase in mapped seagrass extent [213]. Poursanidis et al [214] map change between submerged vegetation and other benthic substrate following the cyclone season. Both Landsat and Sentinel-2 have the capacity for regional to global mapping of seagrass.

Additionally, higher spatial resolution sensors, such as Planetscope, have improved classification accuracy compared to Sentinel-2 [197]. UAS imagery (<5 cm) has shown the capability to map local seagrass extent and carbon [207, 217]. Object-based methods help separate areas of similar seagrass cover, water quality, and depth [212] but do not necessarily improve accuracy [217]. Recent advancements in acoustic measurements of photosynthesis-derived oxygen bubbles [218] and tracking seagrass grazing animals [219] have increased seagrass mapped extent. Furthermore, machine learning has improved seagrass bed identification [195, 220, 221]. Remote sensing methods, including object-based image analysis, machine learning, physics-based modeling, and integration of multiple scales of training data, have improved carbon monitoring of seagrass.

Estimating seagrass carbon fluxes with remote sensing is difficult due to varying light, tides, currents, water quality, [e.g. 192, 197, 203] and biogeochemical process (i.e. carbon fixation and CaCO₃ [201, 222–225]), even with in situ CO₂ flux measurements [226]. Furthermore, the major drivers of sediment carbon changes within regions from autochthonous to allochthonous based on seagrass canopy complexity, turbidity, and wave environment, further complicating carbon flux monitoring [227]. Water depth is an important factor in estimating seagrass carbon storage [227–229]. Thomas et al [230] demonstrate a data fusion approach using ICESat-2 and Sentinel-2 to map bathymetry in shallow, optically clear coastal water addressing a key data gap in most optical seagrass mapping approaches. Carbon fluxes are challenging to monitor, but modeling and remote sensing have improved our understanding of the biogeochemical processes and site characteristics contributing to flux variability.

The carbon impacts of seagrass loss are hard to quantify due to a lack of precise mapping and carbon storage information. Local estimates of seagrass loss range from highs of ∼2.8% yr⁻¹ [191, 231] to lows of 1.2% yr⁻¹ [206], and globally, since 1990, seagrass loss rate is estimated to be ∼7% yr⁻¹ [24]. The major drivers of seagrass loss are direct anthropogenic impacts [232–234] from boats, development, dredging, and marine pollution [191, 235], as well as overgrazing due to alterations to the food web [236]. Marine heat waves due to climate change can exacerbate seagrass loss [192, 237], and temperature increases are likely to drive future losses [206]. Seagrass beds experience multiple stressors associated with water quality, temperature increases, and overgrazing which can shift seagrass beds from stable ecosystems to rapid deterioration [192, 206, 217]. However, both improvements in water quality [207, 231, 238] and planting have successfully restored seagrass and increased carbon storage and ecosystem services [239, 240]. High but uncertain loss rates and the success of restoration necessitate improved remotely sensed and in situ quantification of seagrass baseline and change in extent to facilitate its inclusion into carbon monitoring and offset programs.

Wetland belowground carbon is primarily determined with field-intensive surveys to collect soil core samples, e.g. the National Wetland Condition Assessment in the United States [245]. Remote sensing has been increasingly deployed to upscale field observations from sample points to the plot or study area scale. For example, distribution maps of soil carbon stocks have been created from soil core measurements using satellite imagery [253, 254]. Satellite imagery has also aided measurements of carbon accumulation in sediments [255, 256]. Other fine-scale approaches have used UASs and Ground Penetrating Radar [257–259]. Despite these advances, high uncertainty in soil carbon estimates from remote sensing remain due to a lack of consistent depth measurements, including the depth of the upper horizons where most carbon is stored and can differentiate more mineral soil wetlands from peatlands [245].

The prediction of carbon storage in AGB with remote sensing is well studied, particularly in forested ecosystems. For mineral wetlands, studies have used remote sensing to upscale plot-level data of AGB to wider extents, such as the watershed-scale [260–262] including forested riparian wetlands [263]. Lidar has been used extensively in forest biomass research, and mineral wetland applications are increasing [264]. Studies scale site-level aboveground carbon metrics from estimates of AGB or carbon through land-use maps and spectral indices from Landsat [265], MODIS [266], Hyperion [267], and commercial satellites [268]. Budzynska et al [269] predicted other carbon indicators, e.g. LAI and % soil moisture, with SAR and optical data. Riegel et al [264] estimated aboveground carbon using aerial lidar and aerial imagery. Productivity rates, including both GPP [270, 271] and NPP [261], have been measured and upscaled to local, regional, and global scales.

Carbon gas fluxes, in particular methane (CH₄) emissions, have been of interest in recent research for mineral wetlands [272]. Most of this research has focused on peatlands in northern latitudes with fewer measurements and less a focus on mineral soil mineral wetlands [241]. In terms of scale, CH₄ has been evaluated with remote sensing at the regional or country level by combining satellite imagery with process models [273]. Inundation detection has been a key component to broad-scale CH₄ mapping with many models using the Global Inundation Extent from Multi-Satellites dataset [274, 275]. However, more recent research has used fine-scale, 3 m resolution satellite imagery to map inundation detection at the watershed scale to evaluate CH₄ fluxes [276]. Lu et al [277] used eddy covariance data from flux towers to demonstrate that mineral wetlands are net sinks and identify a need to incorporate remote sensing to predict CO₂ flux spatially.

Global assessment of mineral wetland carbon is limited by in situ carbon measurements and wetland map coverage. Recent assessments of global wetland coverage have utilized coarse-scale inundation mapping downscaled by topographic metrics [257, 278]. However, inundation approaches do not distinguish wetland types, e.g. these maps often include peatlands (section 3.2.2) and mineral wetlands. Thus, the best-estimated extent comes from Lehner and Döll [279] and the Global Lakes and Wetlands Database, which estimated that non-peatland marshes, swamps, and forested wetlands cover 3.7 × 10⁶ km² or ∼2.5% of the terrestrial land surface [13].

Global scale carbon measurements have yet to account for these changes in areal extent estimates. For example, Bridgham et al [11] used an average of two older sources [280, 281] for freshwater mineral soil wetland area (2.315 × 10⁶ km²) to upscale carbon burial, carbon soil stock, and CH₄ flux (table 4). Similarly, Roehm [282] utilized two older sources [283, 284] to combine areal extent estimates of northern and tropical marshes and swamps (3.5 × 10⁶ km²) to upscale NPP and CO₂ flux (table 4). This latter estimate is closer to the Lehner and Döll [279] estimate than the one used in Bridgham et al [11]. Carbon monitoring research interest in CH₄ is high due to its global warming potential. Thus, the global assessment of a CH₄ flux has been parsed by wetland type and separated from peatlands [275].

Tropical peatland carbon indicators included AGB, degradation, subsidence, and canopy height. Southeast Asia (n = 34) was the primary focus of tropical peatland research, with additional studies focused on South America and Africa. South American studies mapped carbon stocks [296–299], extent and degradation [300], and mountain peatland stocks using SAR and multispectral imagery [301, 302]. In Africa, research focused on mapping the extent, depth [302, 303] and estimating carbon stocks [304]. In Southeast Asia, degradation, loss, and recovery were major research topics enabled by lidar, SAR, and multispectral imagery. Studies have used lidar to detect illegal logging and carbon sequestration [305], map peat depth [306], and estimate AGB for tropical peatlands [307]. Minasny et al detail an open data and mapping methodology with the ability to predict peat depth at a lower cost than lidar [308]. SAR particularly useful in tropical peatlands due to cloud and forest canopy penetration and its sensitivity to inundation and biomass [290, 296]. SAR applications included dinSAR to map subsidence across Southeast Asia [309] and predict AGB [310]. Studies have addressed remote sensing limitations by using multiple satellites to expand spatial and temporal coverage of fires [311] and the utilization of lidar to expand training data [310]. Despite the significant research interest, tropical peatlands lack regional and global scale monitoring due in part to data availability, extent uncertainty, and resources.

Historically, temperate peatlands have frequently been managed for fuel, drained for agriculture, or other land-use [288, 312, 313]. Temperate peatland indicators included GPP [313, 314], water table dynamics [315], erosion [316], disturbance [317], peat depth [318], and moisture [313, 314]. Due to the prevalence of past anthropogenic disturbance, restoration and recovery are common research topics [319–322]. High-resolution imagery is common for site-scale studies, including satellite [316, 323, 324], UAS [325], aerial [326], and handheld spectrometers [314]. Aitkenhead and Coull [318] conducted a regional carbon monitoring system creating a national map of peat depth and Scotland's carbon content. The variety of temperate peatland vegetation and the importance of subsurface carbon stocks are challenges for regional and global monitoring.

The boreal and tundra regions (n = 35) are data-poor due to remoteness and the short field season limiting in situ data collection. There are also significant human development pressures in parts of the boreal zone for petroleum exploration, mining, forestry, agriculture, and infrastructure operations. Even low impact disturbances such as seismic lines will increase the fragmentation of wetlands and have ecological impacts [345]. Most degraded peatlands are tropical [288] but boreal peatlands and permafrost will change significantly with warming and changes to precipitation [346].

Optical remote sensing data in boreal environments is limited due to sun angle, cloud cover, and the short growing season [347]. The floristic similarity between peatlands and non-peatland ecotypes makes identifying landform and hydrology with active sensors particularly important. The focus on topography and landform included identifying permafrost peat mound degradation with aerial and high-resolution imagery [348], classifying boreal bogs with microtopographic variation from lidar [349], mapping thermokarst lakes with spectral imagery [350, 351], detecting freeze thaw dynamics with SAR [352], detecting permafrost extent with electromagnetic imaging [353], and mapping lake extent with multispectral imagery [354]. An integration of multi-season SAR and multispectral imagery was complementary in detecting vegetation and hydrologic differences in bogs versus fens in the boreal zone [290, 355]. Carbon monitoring efforts included modeling gas fluxes [272, 328, 356], upscaling in situ emission estimates with land cover maps [350, 329, 330, 357], and peat extent [358]. Major change drivers within the system include increasing temperatures [351, 331, 332, 359] and fire [333, 360]. Boreal systems are critical for understanding the global carbon cycle, and unique challenges to in situ and remote sensing data collection are being addressed by science programs such as the NASA Artic-Boreal Vulnerability Experiment [361].

The global importance of peatland fires in Southeast Asia has long been acknowledged, with peak yearly emissions equaling 13%–40% of the mean annual global carbon emissions from fossil fuels [23]. Earth observation has enabled and verified peatland fires. Page et al [23] primarily used fire extent mapped from Landsat to understand peatland fires carbon emissions (2002). Lidar has been used to map fire scars and burn depth improving emission estimates [362]. Emission estimates and burn area models have used satellite-derived peatland fire data from the Global Fire Emissions Database for verification [334, 335, 363]. SMAP soil moisture data has been used to provide fire warnings, predict burn area [364], and as an input in emission models [365]. Drought can worsen emissions from forest fires within temperate/subtropical peatlands (0.32 PgC) [366]. Thus, earth observation is critical for modeling and verifying this important source of CO₂ emission. CMS has supported several fire mapping efforts of which peatlands are the focus [367] or included in more general fire data [368].

Fires in permafrost regions are also a major climate concern with remote sensing monitoring applications. Remote sensing has identified fires as the most prevalent disturbance in the permafrost region [369], leading to widespread permafrost thawing [370]. SAR has been used to track subsidence following vegetation loss in permafrost regions, including subsidence of 0.5–3 cm yr⁻¹ in deforested areas [371] and the rapidly developed thermokarst following fires with rates of subsidence up to 6.2 cm yr⁻¹ [372]. Studies have used SAR interferometry to model recovery and loss estimating that 4 m of permafrost is lost in a fire event and recovery takes up to 70 years [373]. For wildfire effects, algorithms were developed for assessing peat burn severity (depth) using Landsat-5 [293] and Landsat-8 [374]. Projections suggest rapid thawing will release 60–100 PgC, and gradual thaw regions will release another 200 PgC by 2030 [375]. Peramafrost thawing will release significant mercury into the environment [376].

Globally, peatlands represent a massive SOC stock (table 5) and a remote sensing challenge due to their disparate data needs and global range. Peatland extent is ∼4.0106 × 10⁶ km² and 4.5104 × 10⁴ km² in the northern and southern hemispheres, respectively [287, 290]. Boreal regions of the northern hemisphere are 25%–30% peatland and comprise most of the global extent [290, 377]. Tropical peatland carbon (88.6 Pg) is estimated to be 15% of the global peatland carbon, with boreal and temperate peatland carbon estimated to be 521.4 Pg [343]. Temperate peatland carbon is understudied and as a result has high uncertainty in the carbon estimates [378]. The area of tropical peatlands is uncertain (387 201–657 430 km²), and the largest area (56%) and most of the carbon stock (77%) are in Southeast Asia [327], followed by the Amazon basin [379]. Africa's lowland peatland area is largely unknown except for the Congo Basin [303]. Tropical alpine peatlands are numerous in the Andes, many islands, and Africa [290, 380, 381]. Permafrost peatlands are estimated to contain 277 PgC and are changing rapidly due to global warming and fire [336, 337, 382]. The carbon store within the permafrost region is estimated to be ∼1300 Pg (1100–1500 Pg) with 500 Pg within the active layer [383].

Phytoplankton photosynthesis is the primary process by which carbon dioxide is fixed from the water column and overlying atmosphere. Remote sensing applications to estimate phytoplankton photosynthesis or primary production in the marine environment are numerous (see section 3.4). However due to the spatial variability and optical complexity, applications to freshwater systems are scarce. Advances in remote sensing platforms and algorithm development have allowed for the characterization of phytoplankton abundance and productivity in various freshwater environments [e.g. 392–397]. Remote sensing approaches hold much promise for sampling many lakes on the planet [398] and understanding global trends in phytoplankton [399].

Globally, freshwater lakes exhibit a wide range in size and shape, creating a unique challenge for applying remote sensing methods. Accurate estimates of freshwater phytoplankton biomass require remote sensing data with specific wavelengths associated with spectrally narrow chlorophyll-a absorption features and high signal-to-noise ratios [400]. Satellite sensors with these spectral requirements often target oceans and typically have a coarser spatial resolution (300 m–1000 m), limiting their ability to observe smaller freshwater features. These requirements limit how carbon monitoring of freshwater lakes. The Plankton, Aerosol, Clouds, ocean Ecosystem will address spectral needs of ocean color remote sensing but is still too coarse (<1 km²) to discern small-scale waterbodies [401]. Instead, data fusion using high-resolution imagery and ocean color remote sensing are likely necessary to improve mapping of phytoplankton biomass in lakes and ponds.

Several recent works have used a mechanistic carbon fixation model adapted for use in the Laurentian Great Lakes [396, 397] to estimate carbon fixation in the world's large lakes [402]. Additionally, a simple depth integrated model approach (DIM) was refined to estimate growing season carbon fixation of ∼80 000 freshwater lakes [403]. The DIM approach relies on the light-utilization index, which relates to latitude providing a straightforward method to estimate carbon fixation rates when minimal limnological data is available. The marine standard Vertically Generalized Production Model has also been applied to estimate carbon fixation in large lakes [409]. Remote sensing approaches to estimate freshwater phytoplankton carbon fixation have been developed and applied for small lakes [410, 411].

Lake carbon budgets are highly dependent on carbon fixation rates, yet these rates are unknown for most lakes. McDonald et al [412] estimated that there are over 60 million lakes. Estimating carbon fixation in all these lakes would be impossible with in situ methods. Therefore, the development of remote sensing methods to estimate carbon fixation rates is a current focus. Still, global-scale estimates remain elusive due to the lack of readily available remote sensing products appropriate for optically and spatially complex inland lakes. However, several recent works have estimated global scale freshwater carbon fixation for satellite observable lakes [402, 403]. These works also examined carbon fixation on multiple temporal scales ranging from a single year growing season snapshot for ∼80 000 lakes [403] to a 15 year time-series for the world's eleven largest lakes [402]. The latter work revealed significant changes in carbon fixation for several lakes, likely in response to changes in climate. The use of remote sensing to understand spatial variability across lakes is lacking in many of the existing global carbon monitoring values (table 6).

Additionally, common carbon monitoring measurements in lakes include chlorophyll-a, DOC, CO₂ flux, and GPP. Kuhn et al [413] calculated boreal lake GPP with aerial and satellite imagery and verified the result with in situ measurements. Remote sensing combined with in situ measurements clarified that boreal lakes may have a limited role in carbon mineralization [414]. Lake color in the region has been tracked with the satellite record identifying increased connection with the surrounding landscape [415]. The combination of electromagnetic imaging and satellite imagery has mapped the temporal dynamics of hydrological connectivity between boreal lakes and permafrost [416]. Remote sensing is critical for understanding the effect of climate change on the carbon cycle in permafrost regions.

Methods to estimate chlorophyll-a concentrations range from empirical and semi-analytical approaches and, more recently, machine learning and artificial intelligence-based techniques (Reviews in [417, 418]). Initial work has been done to estimate global freshwater lake chlorophyll concentrations [398]. However, more robust methods to account for the varying optical complexity of lakes should be developed. The GloboLakes initiative has developed a freshwater chlorophyll retrieval algorithm, generated a global time-series of chlorophyll concentrations for ∼1000 lakes, and provided monitoring products for the ESA Climate Change Initiative [419].

DOC and CDOM are also frequently monitored in lake environments with remote sensing. CDOM algorithm comparisons found that remote sensing algorithms did not predict the highs or lows well [420], but continued algorithm refinements are promising [421]. Brezonik et al [422] demonstrated a strong relationship between CDOM and DOC but were cautious about assuming a consistent relationship. Geographically diverse study sites suggest the possibility of applying these methods globally [423].

CO₂ flux has been estimated with partial pressure of CO₂ (pCO₂) in coastal oceans [424]. Simple relationships have also been developed between bio-geochemical properties and freshwater carbon flux through comparisons with Eddy-covariance flux tower measurements [425]. While these methods show promise, applications in freshwater lakes are infrequent. Similarly, very few efforts have been made to fully characterize carbon fractions in freshwater systems, although initial efforts seem promising [426] and are worthy of continued research.

Riverine C export is well constrained using global streamflow discharge and measurements of aqueous carbon concentrations [384, 433]. Stream gauge and water quality data can provide the necessary data for the extrapolation of C export at continental scales [436] and global scales [437, 438]. Mass balance analysis and data initiatives have refined global estimates [15, 439]. Many studies have established empirical relationships between surface organic C concentrations (e.g. POC, DOC, and phytoplankton) and remote sensing data across various aquatic systems, including river reaches [e.g. 440–447].

Remote sensing derived C concentrations have been used to estimate riverine export to estuaries and coasts. For example, Chunhock et al [448] calculated river-to-ocean fluxes with remote sensing derived DOC concentrations near the river mouth in conjunction with discharge data. Liu et al [449] used Landsat-derived POC concentrations and monthly river discharges near the mouth of the Yangtze River to assess long-term patterns of riverine POC fluxes from 2000 to 2017. Successful application of remote sensing methods requires continual monitoring of constituent concentrations and a large enough water body (e.g. river with a width larger than 90 m) to be observed from satellite imagery [450], making it challenging to apply them in small rivers and streams. High-resolution UAS imagery was applied to detect water quality parameters in small rivers/streams, such as chlorophyll-a, Secchi disc depth, and turbidity, with limited success [451, 452]. Therefore, carbon monitoring of rivers and small streams with remote sensing requires further research and technological advancement.

In the past two decades, quantification of regional and global CO₂ outgassing from streams and rivers has made great progress. Richey et al [453] in a pioneering study of regional-scale CO₂ outgassing in the Amazon River basin, used field observations to estimate CO₂ emissions per unit area in mainstem and floodplains. They used JERS-1 L-band SAR to estimate areal coverage and inundation status of rivers and floodplains (>100 m in width) and developed empirical relationships for smaller streams. Since then, multiple studies have continued to use remote sensing to refine estimates of CO₂ emissions from the running waters in the Amazon. For example, Johnson et al [454] constrained their analysis to seasonally inundated areas based on SAR detected high and low water periods. De Fátima et al [455, 456] used a 100 m DEM to improve the surface area calculation. Recently, Sawakuchi et al [431] used model estimates of surface area to estimate outgassing in the lower Amazon River Basin. These advances have resulted in an estimate of ∼0.96 PgC yr⁻¹ CO₂–C outgassing from the rivers and streams in the Amazon [433], nearly double the estimate by Richey et al [453], which included both rivers and floodplains. In another salient example of regional carbon accounting, Butman and Raymond [457] used numerous USGS field observations and the National Hydrography Dataset plus [458] river networks to estimate surface water area. These studies demonstrate the need for high-resolution river networks and water surface area data to monitor CO₂ outgassing from rivers reliably.

At the global scale, Richey et al [453] upscaled their estimates of CO₂ outgassing in the Amazon River Basin to calculate outgassing from rivers and floodplains in the global humid tropics. This was much higher than contemporary estimates not informed by remote sensing [384, 459]. Battin et al [460] analyzed the net heterotrophy (respiration—GPP) of 130 rivers and streams and extrapolated the results to the global scale by multiplying average emissions of streams and rivers by global surface area. Utilizing remote sensing data including hydrological data derived from the SRTM, Raymond et al [430] reported a 1.8 PgC yr⁻¹ of CO₂ outgassing from global streams and rivers. Recently, the global relevance of dry inland waters to the carbon cycle has been identified [461–463], representing unreported CO₂ emissions [430].

Progress has also been made for accounting CH₄ emissions from global freshwaters. Bastviken et al [408] synthesized and calculated average areal field observations of CH₄ fluxes of various types of freshwaters (including lakes, reservoirs, wetlands, and lakes) by different latitudes to estimate a total of 103.3 Tg CH₄ yr⁻¹, of which rivers contributed ∼1.5 Tg CH₄ yr⁻¹ (or ∼10 Tg C (CO₂₋e) yr⁻¹. The scarcity of observed data points and the exclusion of small streams in the river surface area likely contributed to low river flux [464]. Overall, CH₄ emissions from streams and rivers are less studied than CO₂ emissions.

In general, current riverine CO₂ and CH₄ outgassing estimates are subject to large uncertainties due to difficulty in accurately measuring surface water area, partial pressure of CO_2, and gas exchange rates [427]. More field data and high spatial resolution remote sensing are needed to refine surface water area and gas exchange rates. The global studies rely on river networks primarily based on NASA's global SRTM DEM at 90 m. In the US, 10 m DEM-derived high-resolution hydrological data is available [465]. Still, higher resolution lidar derived DEM with (∼2 m) can improve river network delineation [466]. Additionally, SAR, multispectral, and hyperspectral data collected from aerial, satellite, and UAS, have been used to map surface water area and characterize river channel morphology [467–470]. Although those applications achieved noticeable success, upscaling them to continental or global scales face many challenges [471].

The Surface Water and Ocean Topography (SWOT) satellite mission [472], will measure terrestrial water at a spatial resolution of 50 m and provide river vector products that represent reaches with a collection of nodes spanning every 200 m [473, 474]. Researchers estimate gas exchange coefficient with remote sensing derived width and water surface slope measurements, while surface water area can be multiplied with estimated gas emissions per unit area to estimate total C degassing. Note that small streams are difficult to discern at SWOT's spatial resolution; therefore, data fusion of SWOT river vectors with high-resolution DEMs holds promise to provide more accurate data regarding rivers and streams.

Global estimates of aquatic organic C burial are between 0.15 and 1.6 Pg CO₂–C yr⁻¹ [405, 475]. These studies focus on sedimentation in reservoirs, lakes, and wetlands, without explicit global-scale C burial in rivers and streams. Watershed models that explicitly integrate terrestrial and aquatic carbon cycling processes are being developed to quantify the burial of particulate organic carbon (OC) in rivers. For example, Qi et al [476, 477] incorporated OC deposition, resuspension, and diagenesis processes in the soil and water assessment tool and showed that a significant fraction of terrestrially originated POC is deposited on the bed of small streams and further decomposes into CO₂ and CH₄. These results indicate that the inclusion of C burial in rivers and streams would improve the accounting of global C burial in inland waters. High-resolution riverine networks will be critical for updating and improving existing carbon monitoring (table 7). Additionally, certain carbon pathways are relatively unknown, e.g. how much carbon enters wetlands and subsequently enters rivers.

The carbon cycle in coastal shelf seas is very similar to that of the open oceans. Thus, carbon monitoring methods in shelf seas tend to overlap with oceanic approaches. However, the coastal shelf has unique data needs i.e. spatial resolution to discern coastal features, spectral influence of depth and terrestrial hydrology. Thus, oceanic remote sensing methods require alteration for use in coastal shelf regions. NPP monitoring has used methods derived for ocean systems both directly and with slight modifications [504, 554], but the performance of these methods is lowest in coastal shelf seas [504]. One region of particular focus is the Arctic Ocean and its surrounding shelf seas, where many regional algorithms exist [e.g. 514, 565, 566]. Lee et al [512] provide an assessment of 32 Arctic NPP satellite models, finding the models performing relatively well in low-productivity seasons and deep-water regions. However, the algorithms tended to overestimate NPP, but yielded underestimates when a subsurface chlorophyll-a maximum was present.

Given that the shelf sea represents the continuum between terrestrial and ocean ecosystems, there are additional factors to be considered compared to carbon monitoring in ocean systems. NASA CMS studies have focused on improved observation and modeling of lateral transport of terrestrial carbon into the watershed and ultimately to the coasts [567–574]. Other studies have focused on the estimation of DOC and CDOM [188, 575–578]. As described in section 3.3.1, CDOM only makes up a fraction of the total DOC pool, but in coastal systems dominated by terrestrial discharge, CDOM and DOC co-vary. The exact form of the relationship between CDOM and DOC varies both temporally and spatially, driven by terrestrial source characteristics and biogeochemical processes, thus the need for regional approaches. However, Vantrepotte et al [579] demonstrated the potential of a generalized approach in deriving DOC from CDOM in very contrasting coastal environments.

The CMS program has supported carbon monitoring in the northern Gulf of Mexico and the region influenced by the Mississippi River. This included efforts to map pCO₂ and estimate fluxes [580–584], model simulations using a coupled physical-biogeochemical model [585] and satellite-derived estimation of pCO₂ and air-sea flux of CO₂ [424]. Studies have also examined patterns in phytoplankton community composition and potential relationships to carbon dynamics [586, 587]. Other CMS program efforts examined carbon properties in both the Gulf of Mexico and the Atlantic coast [564, 588], and other studies have focused on sedimentation and flux of various carbon forms to the seafloor [554, 589–592].

There remains considerable uncertainty in the estimate of global coastal ocean uptake of CO₂. Some of this is related to differences in the extent to which estuarine and inland waters are included in the inventory. However, estimates based on in situ extrapolations as well as global models have generally converged around −0.1 to −0.3 PgC yr⁻¹ [556, 557, 593].