The role of satellite observations in understanding the impact of El Niño on the carbon cycle: current capabilities and future opportunities

Abstract

The 2015/2016 El Niño was the first major climate variation when there were a range of satellite observations that simultaneously observed land, ocean and atmospheric properties associated with the carbon cycle. These data are beginning to provide new insights into the varied responses of land ecosystems to El Niño, but we are far from fully exploiting the information embodied by these data. Here, we briefly review the atmospheric and terrestrial satellite data that are available to study the carbon cycle. We also outline recommendations for future research, particularly the closer integration of satellite data with forest biometric datasets that provide detailed information about carbon dynamics on a range of timescales.

This article is part of a discussion meeting issue ‘The impact of the 2015/2016 El Niño on the terrestrial tropical carbon cycle: patterns, mechanisms and implications’.

1. Introduction

Global mean atmospheric concentrations of carbon dioxide (CO₂) have increased by 40% from about 277 parts per million (ppm) in the late nineteenth century, prior to the advent of the industrial revolution, to present-day values of more than 400 ppm [1]. Current atmospheric concentrations are now higher than any time at least in the past 800 000 years [2]. This unprecedented atmospheric rate of increase is primarily due to growing human activities: widespread combustion of fossil fuel, cement production to meet growing construction demands and land use change to meet global demands for food and timber. These activities are embedded within a large and active natural biospheric cycle. Current estimates report that approximately 50% of CO₂ emissions are absorbed by the land biosphere and the oceans (e.g. [3,4]), but if this rate of absorption cannot be sustained we will be exposed to a larger atmosphere fraction of the additional CO₂ we emit.

Our current understanding of the land biosphere and how it responds to exogenous drivers is particularly uncertain, limited primarily by sparse ground-based measurements. Figure 1, for example, shows the distribution of continuous and flask samples of atmospheric CO₂ mole fraction collected by the US National Oceanic and Atmospheric Administration (NOAA, https://www.esrl.noaa.gov/gmd/ccgg/trends/global.html). Charles Keeling established a measurement site on Mauna Loa, Hawaii during the 1957 international geophysical year [5], with the objective of understanding atmospheric variations of CO₂ on large spatial and temporal scales. Data collected at this site contributed to the scientific justification for the establishment of a global network. The NOAA network in its current form has large gaps in the tropics, reflecting the difficulty of maintaining a measurement programme over these regions. Nevertheless, the data from this network have since been a key source of information that underpins current understanding of the global carbon cycle [1]. Figure 1 also shows latitude–time Hovmöller diagrams for the years surrounding the 1997/1988 and 2015/2016 El Niño events that illustrate a reduced form of the available information. By averaging over zonal bands we effectively discard longitudinal variations, allowing us to emphasize the latitude variations of CO₂ as a function of time. The atmospheric mole fraction data require the application of atmospheric transport models to infer the location and magnitude of the responsible net CO₂ fluxes (emissions minus uptake). The two Hovmöller diagrams shown in figure 1 illustrate the contrasting atmospheric CO₂ patterns associated with the different responses of the tropical carbon cycle to El Niño events (e.g. [6–10]).

Figure 1.

(a) Tropical and subtropical distribution of continuous and flask samples of CO₂ from the NOAA Global Greenhouse Reference Network (denoted by red circles) and ground-based remote sensing CO₂ column measurements from the Total Carbon Column Observing Network (denoted by green squares). (b) and (c) Hovmöller diagrams of monthly mean NOAA CO₂ mole fraction measurements described in 15° latitude bins for the 1997/1998 and 2015/2016 El Niño events, respectively.

Satellite observations of Earth's atmosphere and land surfaces have been available for decades, but we are now entering a new observational era for the carbon cycle in which we have (or soon will have) the capability to observe simultaneously many of its constituent components. What is particularly exciting is that these space-borne data are beginning to observe the atmosphere and land surface on spatial scales comparable with in situ ecological measurements.

Figure 2 shows the Niño 3.4 sea surface temperature (SST) anomaly index, a common metric used to identify phases of the El Niño Southern Oscillation (ENSO). ENSO is a pan-tropical climate variation driven by gradients of equatorial SSTs over the Pacific. Under neutral conditions trade winds push warmer waters westward, piling it up in the western Pacific; during the La Niña phase of ENSO the neutral case is exaggerated, resulting in a northward shift in the jet stream. During El Niño these trade winds weaken and the east–west SST gradient weakens with less upwelling of cooler, nutrient-rich water over the eastern Pacific that eventually impacts the local fishing industry. Changes in equatorial Pacific SSTs also result in a southward and intensified shift in the jet stream that effectively impacts weather patterns over the tropics. Figure 2 shows the prominence of the 1982/1983, 1997/1998 and 2015/2016 El Niño events [12].

Figure 2.

Monthly area averaged SST anomalies over the Niño 3.4 region (5° S–5° N, 170°–120° W) taken from the HadISST1 dataset [11]. Anomalies are calculated by removing the 1951–2000 monthly mean from individual months. A non-comprehensive list of launch dates for satellite instruments relevant to carbon cycle science is also shown. The reader is referred to §2 and table 1 for further details on individual instruments.

Figure 2 also shows when different satellite data became available. The 2015/2016 El Niño was the first major climate variation that provided a natural showcase for these types of space-borne data. During this period we had access for the first time to simultaneous satellite-retrieved data to characterize photosynthesis (solar-induced fluorescence (SIF)), fire (active fire data, burned area, trace gases), leaf phenology (vegetation indices), hydrology (water storage inferred from gravitational anomalies), land surface temperature (respiration) and atmospheric CO₂ and other trace gases. The main purpose of this perspective is to highlight some of these data that can be and are being used to study the impacts of El Niño on the tropical carbon cycle. Section 2 describes individual types of relevant satellite data. We conclude in §3 with a discussion about how to best use these large volumes of heterogeneous data to further basic knowledge about the tropical carbon cycle in the context of in situ measurements.

2. Satellite observations relevant to El Niño

Earth-observing satellites are growing in number and variety, see e.g. http://www.earthobservations.org. There is an ongoing challenge to exploit the vast amount of information satellite instruments produce every day in order to understand planet Earth. Current satellite observations are beginning to challenge our fundamental knowledge of the carbon cycle, with the next generation of sensors being launched so they can play a role in determining global stock takes of carbon as part of the Paris Agreement. Analysis of these observations is an exemplar of the big data challenge that is now faced by many scientific disciplines.

Some satellite data have been available since the 1980s, e.g. thermal infrared (IR) anomalies to determine fires, land and SST and leaf phenology data such as leaf area index (LAI) [13]. The start of the current century marked the advent of sensors that were sensitive to tropospheric chemistry, or more accurately our repurposing of instruments originally intended for monitoring ozone chemistry in the stratosphere. These have since been replaced by instruments dedicated to studying tropospheric chemistry, including gases and aerosols that are relevant to climate and air quality and by-products of combustion and biogenic emissions (e.g. [14,15]). It is only in the past decade we have begun to see the launch of dedicated instruments that are focused on specific components of the land and ocean surfaces, e.g. soil moisture [16,17] and ocean salinity [17]. Table 1 shows an overview of satellite instruments relevant for understanding the response of tropical carbon cycle to El Niño.

Table 1.

Summary of key current and future satellite data products that are relevant to the tropical carbon cycle. As much as possible we use values taken from the World Meteorological Organization Observing Systems Capability Analysis and Review Tool: https://www.wmo-sat.info/oscar/. SS and G denote sun-synchronous and geostationary orbits, respectively. We only report one overpass time for instruments that require sunlight for their measurements. *The GRACE data are described using a 300 km Gaussian filter, and we anticipate a similar approach being used by GRACE-FO. ISRO, Indian Space Research Organisation; ISS, international space station; NASA, National Aeronautics and Space Administration; SAR, synthetic aperture radar (L-band).

variable satellite instruments	dates	relevant data products	orbit/local equatorial overpass time	nadir dimension of data product
leaf phenology
MODIS (moderate resolution imaging spectroradiometer on terra/aqua)	1999–/2002–	leaf area index, fraction of absorbed photosynthetically active radiation	SS/1030 and 1330	500 m
(+ many instruments)
GOME-2 (global ozone monitoring experiment)	2006–	solar-induced fluorescence (SIF)	SS/0930	80 km × 40 km
GOSAT (greenhouse gases observing satellite)	2009–		SS/1330	10.5 km diameter
OCO-2 (orbiting carbon observatory)	2014–		SS/1330	1.29 km × 2.25 km

forest structure
GLAS (geoscience laser altimeter system)	2003–2010	forest biomass	drifting orbit	66 m
ICESAT-1	2003–2008	vegetation biomass/height	drifting orbit	40 m
PALSAR/PALSAR-2 (phased-array L-band synthetic aperture radar)	2006–2011/2014–	forest biomass	SS/1200 and 0000; 91 day repeat	10 m
LandSat-8 onwards	2013–	forest biomass	SS/1000 and 2200	30 m

hydrology
SMOS (soil moisture and ocean salinity)	2009–	soil moisture (SM), liquid water equivalent anomaly (LWE)	SS/0600 and 1800	15 km (SM)
SMAP (soil moisture active–passive)	2015–		SS/0600 and 1800	10 km (SM)
GRACE (gravity recovery and climate experiment)	2002–2017		SS drifting	1°* (LWE)
GRACE-FO (GRACE-follow on)	2018–		SS drifting	1°* (LWE)

atmospheric CO2
GOSAT	2009–	column CO2, CH4	SS 1300 common to all	10.5 km diameter
OCO-2	2014–	column CO2		1.29 km × 2.25 km
TANSAT (exploratory satellite for atmospheric CO2)	2016–	column CO2		2 km × 2 km

atmospheric chemistry
MOPITT (measurement of pollution in the troposphere)	1999–	CO	SS/1045 and 2245	22 km × 22 km
OMI (ozone monitoring instrument)	2004–	HCHO, NO2	SS/1330	24 km × 13 km
GOME-2	2006–	HCHO, NO2	SS/0930	80 km × 40 km
IASI (infrared atmospheric sounding interferometer)	2006–	CO, CH4	SS/0930	circle diameter 12 km
Sentinel-5P	2017–	HCHO, NO2, CO, CH4	SS/1330	7 km × 3.5 km

fires
VIRS (visible and infrared scanner)	1997–2015	active fire detection (AF), burned area (BA), fire radiative power (FRP)	drifting	2 km (AF)
VIIRS (visible/infrared imager radiometer suite)	2011–		SS/1330	375 m/750 m
MODIS (terra/aqua)	1999–/2002–		SS/1030,2230 and 0130, 1330	0.5 (BA), 1 m (AF,FRP)
Himawari-8	2014–		G over Asia	2 km (AF,FRP)
SEVIRI (spinning enhanced visible infrared imager)	2004–		G over Africa and Europe	3 km (AF,FRP)
GOES (geostationary operational environmental satellite)	2007–		G over Americas	2.4 km × 4 km (AF,FRP)

future missions
atmospheric CO2
GOSAT-2	∼2018	column CO2, CH4	SS	10.5 km
OCO-3	∼2019	column CO2	drifting on ISS	1.29 km × 2.25 km
MicroCarb	∼2021	column CO2	SS	∼5 km
Sentinel 5	∼2021	column CO, CH4	SS	7.5 km × 7.5 km
GeoCarb (geostationary carbon cycle observatory)	∼2022	column CO2, CH4, CO	G over Americas	3 km × 6 km

forest biomass
GEDI LiDAR (global ecosystem dynamics investigation LiDAR)	∼2018	forest biomass/height	drifting on ISS	25 m
ICESAT-2 (ice, cloud and land elevation satellite) LiDAR	∼2018	vegetation biomass and height	drifting orbit	40 m
PALSAR-3	∼2020	forest biomass	SS/1330	5 m–30 m
SAR-L for NASA-ISRO SAR	∼2021	forest biomass/height	SS/0600 and 1800; 12-day repeat	10–20 m
BIOMASS	≥2022	forest biomass/height	SS/0600 and 1800; 6-month repeat	50–60 m

fluorescence
FLEX (fluorescence explorer)	≥2022	SIF	SS/1000	300 m

Figures 3 and 4 show the spatial and temporal distributions of a range of satellite observations relevant to studying the tropical carbon cycle. We show the August 2015 spatial distribution of atmospheric and land surface properties observed by satellite (figure 3) to illustrate some of the data available during the 15/16 El Niño.

Figure 3.

Monthly distribution of GOSAT data, the corresponding a posteriori CO₂ fluxes and correlative data for August 2015. Further details of individual datasets are discussed in §2 and table 1. Data are described on a regular 1°-grid, with the exception of GOME-2 SIF that is described on a regular 0.5°-grid. EVI, enhanced vegetation index.

Figure 4.

Monthly mean satellite observations over northern tropical Asia, 2014—2017, of (a) a posteriori CO₂ fluxes inferred from GOSAT data, (b) EVI from NASA MODIS and SIF from GOME-2, (c) analysed prepitation fields and LWE thickness anomaly from NASA GRACE and (d) dry matter (DM) burned from Global Fire Emission Database (GFED) and HCHO columns from Netherlands Agency for Aerospace Programs/Finnish Meteorological Institute (NIVR/FMI) OMI with fire-free scenes discarded. The geographical region studied is shown in inset of (a). Further details of individual datasets are discussed in §2 and in table 1.

(a). Vegetation cover and phenology

Earth-observing satellites using optical and radar methods can detect land use and land use change. Optical remote sensing methods use the sun as their source, while radar and LiDAR sensors generate their own energy source and use backscatter as their signals. Optical data, with the latest Landsat-8 sensors using wavelengths spanning from the ultra-blue to the thermal, provide an assessment of forest canopy cover that can be used to derive estimates of the density and health of leaves [18]. Optical instruments cannot see through clouds, which are prevalent in the tropics, and cannot see below-canopy changes. Radar satellites use microwaves to penetrate through the canopy to determine forest structure. One application of these microwave data is to determine aboveground biomass using ground-based relationships between biomass from plot and LiDAR metrics [19]. However, the most sensitive L-band sensors saturate at values much lower than the biomass densities commonly found over the tropics (approx. 250 tonnes ha⁻¹) [20], although future missions will use the P-band that will increase the values at which the sensors saturate.

There are several remotely sensed vegetation indices of which the three most commonly used include the normalized difference vegetation index (NDVI), LAI and enhanced vegetation index (EVI). These indices provide a crude estimate of canopy greenness that reflects changes in leaf area, chlorophyll content and canopy structure (e.g. [21]). NDVI takes advantage of the fact that denser vegetation will reflect more at near-IR than visible wavelengths. It is mainly used as a quantitative metric of vegetation density, although recent studies have highlighted a possible role for forest stand age in NDVI variations [22]. LAI is the leaf area per unit ground area and is determined by the fraction of incoming photosynthetic active radiation between 0.4 and 0.7 nm absorbed by the plant canopy. EVI is defined similarly to NDVI but is more sensitive to higher biomass regions that are found in the tropics and the contributing wavelengths take into account atmospheric influences such as aerosol scattering. Figure 3c shows the monthly mean distribution of EVI for August 2015 from the NASA moderate resolution imaging spectroradiometer (MODIS) aboard the NASA Aqua satellite. These data show the extent and density of vegetation across tropical ecosystems.

(b). Hydrology

The ESA soil moisture ocean salinity (SMOS, [17]) and Jet Propulson Laboratory (JPL) soil moisture active passive (SMAP, [16]) both use L-band passive microwave remote sensing to estimate the amount of water in the top 5 cm of soil. Variations of soil moisture observed by SMOS or SMAP correspond closely to inputs from precipitation. Maximum rooting depths of vegetation are commonly used to characterize their susceptibility to drought (e.g. [23]). Tropical ecosystems, including grassland and savannah species, can develop rooting depths of 1–10 s metres to reach the water table in response to the length of the dry season they experience. Consequently, changes in soil moisture in the top 5 cm rarely describe the limits of ecosystem access to groundwater. The gravity recovery and climate experiment (GRACE), launched in 2002, was originally intended to study variations in gravity [24]. GRACE accurately measures the distance between its two satellites that are flown in tandem. A successful data product from GRACE has been the liquid water equivalent (LWE) thickness anomaly [25], which has been related to changes in the terrestrial water storage, an integrated measure of the water column, with an accuracy of 38 mm (15 mm) at a spatial scale of 500 km (1000 km) [26]. Figure 3e shows the monthly distribution of GRACE LWE for August 2015. The data show regions with positive water column anomalies over western Africa and negative water column anomalies over central Africa, over much of Brazil and over much of Southeast Asia and northern Australia. The negative water anomaly over Southeast Asia appears to have exacerbated the ongoing agricultural practices of draining peatlands over Indonesia by increasing the susceptibility of this fuel to combustion [7].

(c). Fires

Landscape fire is an integral part of many tropical ecosystems. The most established satellite data product associated with fire is thermal IR anomalies. These data, measured since the 1980s, provide information on the location, extent and to a lesser extent duration of actively burning landscape fires but nothing about the strength of a fire or about the fuel being burned. Newer products include fire radiative power [27], which can be derived from more sophisticated sensors operating in the thermal IR, particularly those sensitive to the 3–5 µm spectral region. This product can be used to determine biomass combustion rates and totals, and to help determine pyroconvective injection heights that describe the maximum altitude reached by the emitted smoke due to intense surface heating (e.g. [28,29]). Further developments include routine global burned area maps, which take advantage of the rapid changes in land surface reflectance (visible to short-wave IR) associated with the burning of vegetation and the surface deposition of charcoal and ash to identify daily updates to global burned area [30].

The characteristics of the gases and aerosols emitted by fire depend on a number of factors [31], e.g. fuel loading, fuel moisture and combustion phase (e.g. smouldering or flaming). Common trace gases that are used as markers for biomass burning include carbon monoxide (CO) and formaldehyde (HCHO). Measuring CO has the advantage that it has an atmospheric lifetime long enough that it is easily measured but short enough that surface emissions can be identified above a global background. The measurement of pollution in the troposphere (MOPITT), launched in 1999, has since provided the science community with CO columns measured at thermal IR wavelengths that are most sensitive to the free troposphere, with near IR wavelengths providing sensitivity to the lower troposphere [32–34]. Past work has shown that biomass burning emissions rarely possess sufficient energy to be deposited directly into the free troposphere [35]. Consequently, CO ascends with larger-scale weather systems that result in a diluted signal away from the fire. Use of HCHO columns has the advantage that its short lifetime of a few hours means that it resides mostly in the lower troposphere close to the emission source [36]. HCHO has a primary emission source from combustion but also has a secondary source from the oxidation of a range of volatile organic compounds co-emitted by the fire, and by other processes—most notably from biogenic emissions of isoprene (e.g. [37,38]). Separating the pyrogenic signal for HCHO requires additional data, e.g. thermal IR anomalies [36]. However, HCHO is only a weak absorber at UV wavelengths that are used to retrieve HCHO columns [39], and for less energetic fires HCHO will be chemically lost before it can be effectively observed. Figure 3f shows monthly HCHO column distributions from the Dutch/Finnish ozone monitoring instrument (OMI) [40] aboard the NASA Aura for August 2015. As a (crude) effort to identify elevated HCHO columns from pyrogenic emissions we have retained columns that correspond to the location of active fires identified by thermal IR anomalies. The location of elevated columns indicates the location of fires, and the column values are a function of the characteristics of the fuel being burned. For instance, we expect HCHO columns to be lower over Central Africa where the main fuel is grassland savannas and savanna woodlands that generally have lower calorific content and are less densely packed than vegetation over tropical South America. There is also moderate burning over North Sumatra during August, but the spatial extent increases dramatically later in the El Niño event [7].

(d). Solar-induced fluorescence

Absorption of incoming solar radiation by plant pigments drives photochemical reactions that eventually produce glucose as part of photosynthesis [41]. Any excess energy is dissipated as heat or as fluorescence. Of the solar radiation absorbed, 20% is eventually dissipated as heat and 2% is emitted by SIF between 685–690 nm and 730–740 nm. The result is a small offset (typically less than 1–2%) to the reflected sunlight that can be observed by satellite remote sensing in the 750 nm spectral range. Linking SIF with gross primary productivity is a key step that is being developed through biome-specific empirical relationships with data collected at flux tower sites (e.g. [42]), building on mechanistic studies (see review by Frankenberg & Berry [41]). SIF products are available from a range of imaging and spectroscopic instruments (e.g. GOME-2 [43], GOSAT [44], OCO-2 [45]). Formal integration of these data with models is key to extracting the information, which is the subject of ongoing work (e.g. [46]). Figure 3d shows the monthly distribution of SIF for August 2015 from the GOME-2B satellite aboard the ESA/EUMETSAT MetOp satellite. It has an equatorial overpass time of 09.30 local time, by which time vegetation will have received approximately 3 h of sunlight. Values will change with, for example, vegetation type, density and health. They could also be affected by residual cloud contamination by virtue of their pixel size (40 km × 80 km), an effect that is less prominent with instruments with finer spatial resolution, e.g. GOSAT and OCO-2 (table 1).

(e). Atmospheric observations

The global atmosphere is a nearly passive component of the carbon cycle, with global atmospheric CO₂ mass mainly responding to changes in land and ocean surface fluxes; in practice, there is also a small, diffuse CO₂ source from the oxidation of reduced carbon [47]. Atmospheric mixing results in information, associated with these atmospheric signals, being irreversibly lost so that atmospheric transport cannot be inverted deterministically. This places limits on our ability to estimate geographical distributions of CO₂ fluxes from observed atmospheric variations of CO₂. With such imperfect knowledge of the atmosphere, Bayesian inference methods are typically used that take advantage of a priori information from flux models (describing static inventories and dynamical land surface models) that are informed by, for example, by field campaigns.

It is because of its mixing properties that the atmosphere is an effective integrator. The original placement of ground-based instruments to measure atmospheric CO₂, indeed, relied on those atmospheric properties. But even these properties cannot overcome the measurement gaps over the tropics (figure 1). Satellite observations are ideal to fill in those gaps but substantive progress has only been made in the past decade with the launch of the Japanese Greenhouse gases Observing SATellite (GOSAT) in 2009 [48] and the NASA Orbiting Carbon Observatory (OCO-2) in 2014 [49]. Both instruments reside in a sun-synchronous orbit with an early afternoon equatorial overpass time of 13.30. They observe atmospheric CO₂ using three modes: nadir mode that measures the column in the local nadir; sunglint mode that takes advantage of the high signal-to-noise from specular reflection off ocean and (to a lesser extent) land surfaces; and target mode that locks onto a point on the surface and tracks it while flying overhead. The challenge has been to develop the short-wave IR sensor technology that is sufficiently precise to observe the small (typically less than one per cent) changes in the CO₂ column from surface fluxes. Thermal IR retrieval of atmospheric CO₂ has been available for a lot longer, but it is most sensitive to changes in CO₂ in the free troposphere. Relating these thermal IR data to surface fluxes relies more on atmospheric transport models (particularly vertical motion) than observations at short-wave IR wavelengths that are more sensitive to changes in CO₂ in the lower troposphere. Figure 3a shows the monthly distribution of CO₂ dry-air mole fraction observations from GOSAT for August 2015. values over the tropics for this month range by 6 ppm (1.5% of the 400 ppm background) illustrating the demanding precision requirement of this measurement. Figure 3b also shows the a posteriori fluxes that have been inferred for August 2015 using observations from that month and later months, which clearly show large coherent regions of CO₂ emissions and uptake. These fluxes appear smoothed because the data density supports independent flux estimates on spatial scales of O (500–1000 km [50]).

These data are not a panacea for carbon cycle science. The main advantage of satellite data over the ground-based data is that they have global coverage, including coverage over tropical ecosystems that have largely been unobserved. A disadvantage is that individual data retrievals are relatively noisy compared to the ground-based measurements. This is because our signal, i.e. fresh surface fluxes, represent (at most) a variation of a few per cent atop a large column abundance that places demanding requirements on the measurement precision. One advantage of the column nature of these satellite data is that they are less sensitive than vertically resolved measurements to model errors associated with vertical mixing [51]. The disadvantage is that a column is generally a superposition of different geographical CO₂ fluxes atop of background that has slow and fast modes of variability due to atmospheric growth and weather [52].

Substantial efforts have been made to minimize systematic errors, particularly on spatial scales between 100 and 1000 km, which will alias themselves as erroneous CO₂ fluxes. For example, the total carbon column observing network (TCCON, [53]), comprising upward-viewing Fourier Transform Infrared (FTIR) instruments that are calibrated to a common standard (figure 1), was established to help provide anchor points for satellite data. Satellite column retrievals of CO₂ are now moving to the scientific forefront (e.g. [9]) as challenges associated with measurement uncertainties and regional biases are beginning to be addressed.

Understanding the contribution of land biosphere emissions to observed variations in atmospheric CO₂ remains a key science challenge (see [54] for a discussion of the counter focus of isolating anthropogenic CO₂ contributions). There is no single effective approach to link these variations to individual biophysical processes. From an atmospheric perspective, there are several potential reactive trace gases (atmospheric lifetimes a few months) that could be used to help separate or at least help improve understanding of the relative importance of biomass burning, anthropogenic activities and land biosphere fluxes. Observed atmospheric variations of these reactive gases over the tropics are interesting in their own right, but together can help identify combustion sources from biomass burning (e.g. CO, HCHO) and anthropogenic activity (nitrogen dioxide, ethane). By virtue of their absence, these gases can help isolate land biosphere fluxes. Observed variations of other gases such as carbonyl sulfide could also point to photosynthesis [55]. Clearly, a multi-species analysis is needed to disentangle anthropogenic, pyrogenic and biospheric contributions to atmospheric CO₂. This is an area that has not yet been seriously addressed because of a lack of available data and expertise, but this situation is beginning to change (e.g. [56]).

3. Future opportunities

We live in an era in which many environmental variables are now observed by satellite sensors. These data include a wide range of atmospheric trace gases and land surface properties relevant to the carbon cycle. However, interpretation of these data is arguably still in its infancy, and, with a few notable exceptions, we have only achieved piecemeal analysis of individual datasets. Considerable information resides in the integration of different data, allowing us to make further inroads into answering the biggest scientific questions associated with the tropical carbon cycle.

In the previous section, we showed examples of individual satellite datasets for August 2015 (figure 3) as they were introduced. As discussed above, inferring CO₂ fluxes from satellite data over the tropics is an important first step towards understanding the impact of climate variation on tropical ecosystems and how they can feedback to climate. Although the datasets shown in figure 3 are only a subset of the available information (consider additional variables and the time dimension of these data), we find many similarities in the spatial distributions of these variables, as expected. Figure 4 shows an example of temporal variations of land surface properties and a posteriori CO₂ fluxes over northern tropical Asia. These variations illustrate how different factors that influence the tropical terrestrial carbon cycle (e.g. leaf phenology, photosynthesis, fire, drought) covary. Collectively, these data have the potential to answer fundamental questions about the tropical carbon cycle. What is the impact of drought on photosynthesis? Can we distinguish between the role of fire and drought ecosystem-level CO₂ fluxes? Can we distinguish between carbon emissions from vegetation and soil in anomalously dry and hot environments?

To reach beyond the simple comparisons shown in figures 3 and 4 we have to formally integrate different data using computational models. Models can range from simple falsifiable hypotheses to link a few model components to more comprehensive large-scale atmosphere-bio-physical models. Scientific insights have been gained already from both approaches. Linking models to the data is a key step. Not all land biosphere models are sufficiently developed to take full advantage of available satellite data, but rapid progress is being made (e.g. [46]). Similarly, atmospheric transport models have in the past used simplified flux inventories or simply used output from land biosphere models, but this is also changing. Once atmosphere–biosphere models have been developed so that they can be confronted by land surface and atmospheric satellite remote sensing data, the next step is to statistically fit the model to the data, taking into account model and measurement uncertainties. This statistical approach essentially describes a weighted least-squares fit of the model to the data.

For the purpose of readability of this perspective, we will ignore the details of the various fitting approaches but suffice to say techniques exist and are well established. Previous studies have mostly inferred time-dependent surface fluxes of CO₂ (e.g. figure 3b) but this places limits on our furthering knowledge of the land biosphere, i.e. it certainly does not improve our predictive capability. However, this approach serves as an intermediary objective. The ultimate objective must be to estimate model parameters (e.g. light sensitivity) that describe flux variations that subsequently drive atmospheric CO₂ variations. Model parameter estimation has been attempted by various studies but mostly with ground-based flux tower data (e.g. [57]). Reducing the large and heterogeneous data volumes available to us from satellites introduces its own challenges, but we anticipate that the richness of these data will support a larger number of estimated model parameters than from using only ground-based data. Nevertheless, we argue that model development should include the maximum (minimum) number of (un)falsifiable parameters, determined by the quality and volume of data, necessary to describe the key variations of a system [58], an approach that readily allows identification of model error.

Satellite data alone will not address the major uncertainties associated with the tropical carbon cycle and how it responds to climate variations. For decades, terrestrial ecologists have painstakingly collected a range of pan-tropical forest biometric data (e.g. mortality, respiration, soil carbon, leaf litter) from sample plots across the major continents (e.g. [59,60]). These data are sparse but accurate and provide insights into biomass dynamics on relatively small spatial and long temporal scales. The impact of the 2015/2016 El Niño will likely result in mortality rates and reduced ecosystem functioning that will not be fully realized for many years to come. It will be difficult to assess these impacts using current satellite data. Consequently, to develop a more comprehensive understanding of the tropical carbon cycle it is critical that all available data be used. Harmonizing these biometric data, some of which describe changes on decadal timescales, with the satellite data represents a scientific challenge but also an opportunity to build collaborations between terrestrial ecologists, satellite remote sensing scientists and atmospheric scientists.

Looking to the future, there are a number of satellite instruments that will eventually replace GOSAT and OCO-2 using similar orbits (table 1). GOSAT-2 is due for launch in ∼2018 and the French-UK bilateral MicroCarb is due to be launched in ∼2021, and the CO₂ component to the Copernicus service (currently being defined) will likely include multiple satellites and be launched after 2025. Major innovations will come from the adopted orbit and technologies. OCO-3 is due to become an instrument aboard the international space station (ISS) sometime in 2019. The ISS is in an inclined orbit so it precesses between ±51.6° latitude, which could lead to more clear-sky scenes over the tropics [61]. The NASA GeoCarb, scheduled for launch in 2022, will be launched in a geostationary orbit above the Americas with a vantage point of ±50° latitude that covers much of North and South America; similar complementary concepts have been proposed for Asia and Africa. Concepts that use active remote sensing to measure atmosphere CO₂ (e.g. NASA ASCENDS) are not subject to the limits of available reflected sunlight and will provide data during day and night. Active remote sensing concepts will at least double the volume of atmospheric CO₂ data over the tropics. These data will have a spatial resolution much smaller than current atmospheric transport models, so they will likely need to be averaged to model grid scales to ensure a meaningful comparison. Recall that columns are a superposition of geographical fluxes from various times, so that measuring CO₂ columns over the Amazon basin during night, for example, will reflect values from earlier times of the day and from far upwind. BIOMASS and FLEX are two ESA flagship Earth Explorer missions. Both build on our observing capabilities with current satellites. BIOMASS will employ the P-band to determine the amount of biomass and carbon stored in forests. As discussed earlier, using the P-band avoids the signal saturating at biomass values that are much lower than found over tropical ecosystems. BIOMASS will be launched, at the earliest, in 2022. Fluorescence explorer (FLEX) mission, should also be launched by 2022, at the earliest, and includes sensors to measure fluorescence, hyperspectral reflectance and canopy temperature with a focus on addressing spatial and temporal scaling issues associated with comparing measurements collected at towers and by satellites. FLEX will fly in tandem with the Sentinel 3 satellite that also has complementary optical and thermal sensors, allowing a more integrated assessment of plant functioning.

We may very well be at the start of a golden age of observing the tropical carbon cycle using satellite instruments. Major advances in scientific understanding will come from integrating interrelated information collected from different satellite instruments, and more broadly from strengthening links between terrestrial ecologists, satellite remote sensing experts and atmospheric scientists.

Data accessibility

Thanks to Liang Feng for providing the a posteriori CO₂ flux estimates corresponding to the GOSAT data. We gratefully acknowledge Gonzalo González Abad (Harvard-Smithsonian Center for Astrophysics), Joanna Joiner (NASA GSFC) and the science teams of MODIS, GRACE, GOME-2 and GOSAT for providing their data. We also acknowledge NOAA ESRL/GMD and all the contributing measurement teams for the CO₂ mole fraction data plotted in figure 1 (https://www.esrl.noaa.gov/gmd/dv/data/). The NASA OMI HCHO data (OMHCHOv003) are available from the NASA Data and Information Services Center; the GOME-2 SIF data are available from the NASA Aura Validation Data Center; the GRACE data are available from JPL Physical Oceanography Distributed Active Archive Data Center; MODIS EVI data are available from the Land Processes Distributed Active Archive Center; and the GOSAT v. 7.1 X_CO2 full-physics retrievals are available from the University of Leicester.

Competing interests

I declare I have no competing interests.

Funding

P.I.P. is funded by the NERC National Centre for Earth Observation (grant no. PR140015).