Earth's drylands are home to more than two billion people, provide key ecosystem services, and exert a large influence on the trends and variability
in Earth's carbon cycle. However, modeling dryland carbon and water fluxes with remote sensing suffers from unique challenges not typically
encountered in mesic systems, particularly in capturing soil moisture stress. Here, we develop and evaluate an approach for the joint modeling of
dryland gross primary production (GPP), net ecosystem exchange (NEE), and evapotranspiration (ET) in the western United States (US) using a suite of
AmeriFlux eddy covariance sites spanning major functional types and aridity regimes. We use artificial neural networks (ANNs) to predict dryland
ecosystem fluxes by fusing optical vegetation indices, multitemporal thermal observations, and microwave soil moisture and temperature retrievals from
the Soil Moisture Active Passive (SMAP) sensor. Our new dryland ANN (DrylANNd) carbon and water flux model explains more than 70 % of monthly
variance in GPP and ET, improving upon existing MODIS GPP and ET estimates at most dryland eddy covariance sites. DrylANNd predictions of NEE were
considerably worse than its predictions of GPP and ET likely because soil and plant respiratory processes are largely invisible to satellite
sensors. Optical vegetation indices, particularly the normalized difference vegetation index (NDVI) and near-infrared reflectance of vegetation
(NIR
Earth's drylands are critically important to society yet exceptionally vulnerable to climate change. Drylands are home to more than two billion people and make up more than 40 % of Earth's land surface (Reynolds et al., 2007). Primary production of dryland vegetation supports many rare and endemic species, as well as extensive rangelands and croplands (Bestelmeyer et al., 2015). Dryland ecosystems are also important regulators of global trends and interannual variability in Earth's carbon cycle (Humphrey et al., 2018; Ahlström et al., 2015; Poulter et al., 2014) due to both their large spatial extent and high climate sensitivity (Biederman et al., 2016; Zhang et al., 2022). Hotter and atmospherically drier conditions associated with anthropogenic climate change will likely increase water limitation (Cayan et al., 2010; Cook et al., 2015; Williams et al., 2020; Ault, 2020; Cook et al., 2020), possibly leading to the expansion and degradation of drylands (Huang et al., 2016, 2017). There is therefore a pressing need for satellite-based monitoring of dryland carbon and water cycling at large scales.
While many remote sensing techniques were originally developed and tested in drylands (e.g., Huete, 1988; Huete and Jackson, 1987; Rouse et al.,
1974), satellite-based modeling of dryland carbon and water fluxes has been a long-standing challenge. For example, early validation studies of the
Moderate Resolution Imaging Spectroradiometer (MODIS) science products noted tendencies to overestimate mean dryland productivity (Heinsch et al.,
2006; Turner et al., 2005, 2006a, b) and to miss important features of the seasonal cycle (Heinsch et al., 2006; Turner et al., 2006b). Recent work
has also shown that both satellite models (Biederman et al., 2017; Stocker et al., 2019) and process-based models (MacBean et al., 2021) dramatically
underestimate the interannual variability in dryland carbon and water fluxes while also frequently failing to capture the “flashy” and multi-modal
seasonal dynamics of dryland carbon cycling (Barnes et al., 2021). For instance, the widely used MODIS gross primary production (GPP) and
evapotranspiration (ET) products substantially underestimate the variability in carbon and water fluxes of the western US, capturing only
Several issues make drylands uniquely difficult to monitor and model with remote sensing (Smith et al., 2019). First, ecosystem carbon and water exchange are more tightly coupled to soil moisture in drylands than in wetter, more mesic systems where moisture tends to be more plentiful (Novick et al., 2016; Stocker et al., 2018), but most existing satellite-based models do not explicitly represent soil moisture stress (Song et al., 2013). Instead, light-use efficiency models often represent moisture stress using vapor pressure deficit (Running et al., 2004; Zhang et al., 2016). While vapor pressure deficit is well suited as a water stress indicator for mesic regions, it often does not fully capture water stress in drylands, where soil moisture plays a particularly important role in regulating surface conductance and carbon and water fluxes (Novick et al., 2016; Stocker et al., 2018; Dannenberg et al., 2022a). Soil moisture therefore needs to be incorporated into satellite-based carbon and water models to represent temporal variability in dryland water limitation (Stocker et al., 2018, 2019; Smith et al., 2019).
Second, dryland plants have physiological responses to water limitation (and precipitation variability more generally) that are not necessarily captured by standard remote sensing approaches. Many dryland plants have drought adaptations that allow them to remain green even while being functionally inactive under extreme moisture stress (Yan et al., 2019; Smith et al., 2019), making it difficult to resolve temporal variation in dryland plant function. Therefore, plant physiological responses to periods of moisture stress are not necessarily reflected in optical vegetation indices (VIs) (Yan et al., 2019; Wang et al., 2022; Smith et al., 2018). The normalized difference vegetation index (NDVI) is the most widely used vegetation index, but it sometimes fails to capture temporal dynamics of carbon and water fluxes in drylands (Yan et al., 2019; Smith et al., 2019; Wang et al., 2022). While other optical vegetation indices overcome some of the weaknesses of NDVI, combining different types of remotely sensed observations – such as those from microwave, thermal, and visible wavelengths – can capture complementary information about plant and ecosystem stress that is unattainable from optical VIs alone (Smith et al., 2019; Stavros et al., 2017; Guan et al., 2017). Land surface temperature (LST) from thermal imaging, for example, is an important determinant of carbon and water fluxes because, among other reasons, both photosynthesis and respiration involve temperature-dependent enzymatic reactions (Farquhar et al., 1980; Atkin and Tjoelker, 2003) and because it is a key indicator of latent heat flux, which cools leaves and land surfaces (Bateni and Entekhabi, 2012). The integration of multi-source satellite remote sensing could therefore improve the representation of plant physiological responses to periodic moisture stress in drylands as compared with optical VIs alone.
Third, drylands tend to be more spatially heterogeneous than many other ecosystems, consisting of complex mixtures of vegetation structural, morphological, functional, and physiological characteristics that vary over relatively short distances. These mixtures of vegetation types within moderate- to coarse-resolution imagery can contribute to significant error in GPP estimates (Turner et al., 2002; Heinsch et al., 2006). Many large-scale remote-sensing-based carbon and water models assume a single vegetation type for each coarse pixel rather than representing the land surface as a continuous mixture of different cover types. In the open canopies typical of dryland ecosystems, optical VIs are also particularly sensitive to soil background reflectance and the presence of senesced vegetation or standing litter (Huete and Jackson, 1987). High spatial heterogeneity in dryland vegetation, in combination with complex terrain in some areas, leads to diverse ecosystem seasonalities. Drylands in the western US often have one or more annual growing seasons occurring in spring and/or summer (Biederman et al., 2017; Dannenberg et al., 2020), and the timing, length, and productivity of those growing seasons can vary substantially from year to year in response to ocean–atmosphere teleconnections (Dannenberg et al., 2015, 2021). Moreover, carbon and water fluxes in drylands depend on intermittent and highly variable “pulses” of precipitation that are less seasonally and spatially uniform than limiting resources (e.g., temperature and light) in more mesic or temperate ecosystems (Huxman et al., 2004; Roby et al., 2020). This combination of high spatial and temporal heterogeneity in dryland ecosystem structure and function leads to highly “unique” patterns in carbon and water fluxes, meaning that models perform poorly when used to predict fluxes at sites on which they are not trained (Haughton et al., 2018), yet the flux tower networks typically used to train remote-sensing-based models have notably low representation of dryland sites relative to their global prevalence (Smith et al., 2019). The spatial and temporal heterogeneity of limiting resource availability, the “uniqueness” of dryland fluxes to their specific location (i.e., low predictive power of models for sites on which they were not trained), and the relatively sparse dryland observation networks combine to increase uncertainty in carbon and water cycling estimates from models primarily calibrated for other regions.
With new sensors, new vegetation indices, and expanded global ground networks, many of these issues are now at least partly addressable. Recent research has focused on using different combinations of remote sensing data, including the integration of soil moisture (Stocker et al., 2019; Jones et al., 2017) and multispectral (Barnes et al., 2021) and thermal infrared (Sims et al., 2008; Anderson et al., 2012) observations in models ranging in complexity from purely empirical to semi-empirical or process-based. For example, dryland-specific GPP estimates based on machine learning of meteorological reanalysis data and optical remote sensing observations outperform globally trained models at capturing seasonal to interannual variability in dryland GPP (Barnes et al., 2021). New satellite microwave missions also allow more direct sensing of soil moisture than previously available (Song et al., 2013; Jones et al., 2017; Smith et al., 2019), which could address one of the biggest contributors to model error in dryland ecosystems: the tight coupling between plant activity and soil moisture that is not well-captured by vapor pressure deficit (Novick et al., 2016; Stocker et al., 2018; Heinsch et al., 2006) or remotely sensed greenness (Yan et al., 2019).
Here, we aim to improve the estimation of dryland GPP, net ecosystem exchange (NEE), and evapotranspiration using an extensive network of eddy covariance
observations and multi-source satellite remote sensing. We specifically develop and test an approach for the data-driven prediction of a full suite of
carbon and water fluxes that are specifically adapted for drylands using a machine learning fusion of multispectral, thermal, and microwave remote
sensing. We use an ensemble of artificial neural networks (ANNs) to jointly predict the key ecosystem carbon and water fluxes – GPP, NEE, and ET –
at monthly, 0.05
We developed and tested DrylANNd across 28 AmeriFlux eddy covariance sites (Fig. 1; Table S1 in the Supplement), each overlapping the SMAP record (2015–present) by at least 1 full year and consisting predominantly of natural vegetation. Based
on the 1981–2010 TerraClimate annual precipitation (
We used a spike detection method to filter out sudden but temporary changes in half-hourly NEE, which can arise either from biophysical effects (e.g.,
sudden changes in turbulence) or from instrument error (Papale et al., 2006). Using REddyProc (Wutzler et al., 2018, 2020) in the R statistical
computing environment (R Core Team, 2021), we then excluded half-hourly NEE observations that occurred during periods of low turbulence based on a
seasonal friction velocity (
AmeriFlux (response) and remote sensing (predictor) variables used in the DrylANNd carbon and water flux model.
For each site, we obtained daily 500
We also obtained 1
To capture soil moisture and soil temperature, we used daily (00:00 UTC retrieval) 9
Since the relationships between vegetation indices and ecosystem function can vary among different vegetation types (Wang et al., 2022), we used the
2020 fractional cover of annual and perennial grasses and forbs, trees, shrubs, litter, and bare ground from Rangeland Analysis version 3 (Jones
et al., 2018; Allred et al., 2021) as static predictors. Rangeland Analysis fractional cover fuses Landsat and MODIS surface reflectance at
From the daily MODIS and SMAP observations, we developed monthly composites of each variable. For the optical vegetation “greenness” indices (NDVI,
kNDVI, EVI, and
Artificial neural network (ANN) architecture conceptual diagram (Olden et al., 2008). Each “neuron” in a given layer represents a weighted combination of the neurons in the previous layer. All remotely sensed predictor variables in the input layer are also listed in Table 1.
We used feed-forward artificial neural networks (ANNs; Fig. 2) to jointly predict monthly GPP, NEE, and ET. ANNs are effective at finding underlying relationships within multidimensional and multi-source datasets, including nonlinear relationships and interactions among predictor variables (Olden et al., 2008). They are particularly useful for estimating biophysical parameters because they support nonlinearity, adaptivity to changes in the environment, and decision confidence (Mas and Flores, 2008; Jensen et al., 2009). Synthetic “neurons,” in which each neuron is a mathematical function, connect the neural network's input and output layers, often through “hidden” layers of intermediary functions. Importantly, ANNs are appropriate for multi-output regression problems, where a single model simultaneously produces predictions of multiple variables (e.g., Atkinson and Tatnall, 1997). Because the multi-neuron output layer of the neural network allows joint prediction of response variables, the ANN framework therefore implicitly preserves some biophysical connections between GPP and NEE, where GPP is the carbon input into the ecosystem, and between GPP and ET, which are coupled via plant stomata.
The DrylANNd model consists of an ensemble of ANNs, each with one input layer of 20 “neurons” (i.e., seven optical VIs, LST observations from four
different times per day, three SMAP variables, and six static fractional cover classes), two hidden layers, and one output layer with three neurons
(GPP, NEE, and ET). The sizes of the two hidden layers (
Each ANN in the ensemble (Sect. 2.4 below) was initiated with randomly assigned weights and biases based on the Nguyen–Widrow method (Nguyen and Widrow, 1990) and with different random subsets of observations for model training (75 %) and validation (25 %), with the precise number of data points used for each individual ANN varying slightly depending on the length of the withheld site's data record. We trained the ANNs using the Levenberg–Marquardt algorithm, which performed faster than and at least as well as other training algorithms in early tests.
Our DrylANNd model consists of an ensemble of 560 individual ANNs, in which each ANN in the ensemble was trained with a different combination of sites
(always withholding data from one site to use for independent evaluation) and initialized with different weights and biases that connect one layer to
the next. Specifically, we withheld each AmeriFlux site from model development for 20 of the ensemble members (28 sites
We evaluated model skill based on the coefficient of determination (
The “black box” nature of many machine learning methods (including ANNs) typically makes it challenging to examine the effect of any given input
variable on model predictions. Here, we examined the importance of the MODIS and SMAP predictor variables for model skill in two ways. First, using
the same leave-one-site-out calibration and evaluation procedure described above, we ran models based on each of the three classes of variables (MODIS
VIs, MODIS LST, and SMAP soil moisture and temperature) individually and in all possible combinations and compared the predictive ability (
Overall performance of DrylANNd GPP
At the monthly scale, the DrylANNd model explained more than 70 % of the combined spatial and temporal variation in GPP (Fig. 3a) and ET (Fig. 3e)
for sites withheld from model training but only about 35 % of the variation in NEE (Fig. 3c). For GPP, the model performed best at shrubland
sites (
Spatial patterns and spatial validation of mean warm-season (April–October) GPP
DrylANNd also effectively captured spatial variation in warm-season carbon and water fluxes across western US drylands (Fig. 4). The model simulates
realistic spatial gradients of GPP (Fig. 4a), NEE (Fig. 4c), and ET (Fig. 4e), with the highest productivity and ET in the subhumid east and in high-elevation “sky islands,” where cooler temperatures and more abundant precipitation provide a more favorable environment than the surrounding desert
lowlands. Across the 16 AmeriFlux sites that completely overlap the SMAP observational period, DrylANNd captured 75 %–80 % of the spatial
variation in warm-season GPP and ET (Fig. 4b and f, respectively) with minimal bias (i.e., with predictions all falling along the
Mean (
Mean (
At the 16 AmeriFlux sites that cover the full SMAP period (2015–2020), DrylANNd effectively captured the mean seasonality (i.e., mean monthly fluxes)
of both GPP and ET across most sites, with
While the seasonality of NEE was mostly well captured by DrylANNd (
Annual and interannual variability in warm season (April–October)
DrylANNd captured roughly 70 % of the variability in annual warm-season GPP (Fig. 7a) and 66 % of the variability in warm-season ET (Fig. 7c)
with MAEs of
Coefficient of determination (
For all three response variables (GPP, NEE, and ET), models that included all three subsets of predictor variables (optical VIs, LST, and SMAP)
performed best overall (Fig. 8a–c), though a combination of optical vegetation indices with SMAP soil moisture and temperature performed nearly as well
for both GPP (Fig. 8a) and ET (Fig. 8c). Models based on VIs and/or LST performed worse than SMAP-based models but still achieved overall
Compared to models based solely on optical VIs, the addition of SMAP soil moisture and temperature generally made the largest difference for model
performance in grasslands and shrublands, while including LST estimates from MODIS thermal infrared made the largest difference for model performance
in evergreen needleleaf forests (Fig. 8). In shrublands, the
Site-level (filled circles, colored by vegetation type) and median (vertical black line) change in MAE (
While the LST-only models performed the worst overall (Fig. 8), the models based on all variables assigned high leverage to LST for all three response
variables, wherein a random permutation of LST led to large increases in MAE and decreases in
Here, we developed and evaluated a data-driven, machine-learning-based approach for estimating monthly carbon (GPP, NEE) and water (ET) fluxes in US drylands using multi-source satellite remote sensing. Our DrylANNd model incorporated information from the optical, thermal, and microwave domains, including newer optical VIs that have shown promise in drylands (i.e., NIRv; Wang et al., 2022), daily land surface temperature observations from multiple times per day, and estimates of surface and rootzone soil moisture and soil temperature. DrylANNd performed particularly well at monthly and seasonal (i.e., mean monthly) timescales, representing a considerable improvement over MODIS GPP and ET estimates across most eddy covariance sites and all vegetation types (Fig. 3).
DrylANNd particularly excelled at capturing monthly (Fig. 3e and f), seasonal (Fig. 6), and spatial (Fig. 4e and f) variation in ET. Given the
importance of ET for linking the carbon, water, and energy cycles (Fisher et al., 2017), accurate ET estimates are critical for understanding and
monitoring global ecosystem functions, especially in drylands where remote sensing of ET is particularly challenging (Smith et al., 2019; Fisher
et al., 2017). By contrast, NEE proved more challenging to estimate than either GPP or ET (Figs. 3c, d and 4c, d) likely because many processes
involved in ecosystem respiration cannot be easily represented with satellite data. While heterotrophic and autotrophic respiration rates are strongly
dependent on temperature (Atkin and Tjoelker, 2003) and soil moisture (Moyano et al., 2013), which can be captured by MODIS LST and SMAP soil
moisture and temperature, they also depend on microbial community composition, substrate availability, and root biomass that are not visible to satellite
sensors. DrylANNd performed moderately well at capturing the seasonality of (Fig. S3) and spatial variation in (Fig. 4c and d) NEE but tended to
systematically overestimate the magnitude of net carbon uptake in US drylands (Figs. 4d and S3), particularly in evergreen needleleaf forests. Many
dryland sites have a net carbon balance near zero and can flip between being sources of and sinks for CO
While DrylANNd captured monthly, seasonal (i.e., mean monthly), and spatial variation in GPP and ET with fidelity, it struggled to predict interannual variability (Fig. 7). This is a common issue for satellite-based models applied in dryland ecosystems (Biederman et al., 2017; Smith et al., 2019; Stocker et al., 2019; Barnes et al., 2021) partly due to the prevalence of “hot moments” (i.e., short periods of high biogeochemical activity) that are disproportionately important to time-averaged carbon and water fluxes in drylands (Kannenberg et al., 2020). While DrylANNd has relatively little systematic bias at capturing low extremes in monthly GPP (Fig. 3a) and ET (Fig. 3e), it tended to underestimate the high extremes. DrylANNd's monthly resolution may smooth the intensity of these short but impactful “hot moments,” leading to the systematic underestimation of monthly high extremes which also propagates to longer timescales, with DrylANNd clearly underestimating the high extremes in interannual variability in warm season GPP (Fig. 7a) and ET (Fig. 7c). Improving estimates of interannual variability in dryland systems may therefore require models that operate at finer temporal resolutions (e.g., daily) to adequately represent short, intense periods of pulse-driven dryland vegetation activity.
Despite the challenges in capturing interannual variability, the ANN machine learning approach used here has several key benefits. First, because it is a data-driven model based solely on remote sensing products with short latencies, DrylANNd would be relatively easy to operationalize at a large scale and in near real time. Second, the ensemble approach allows for intuitive estimates of uncertainty, which are critical for many applications (e.g., ecological forecasting) but which are rarely provided (Dietze et al., 2018). Finally, neural networks allow joint modeling of multiple response variables, providing the means both to efficiently generate multiple indicators of ecosystem activity and to partially preserve the physical connections between GPP and NEE and between GPP and ET, which is relatively rare for remote-sensing-based models. The MODIS and SMAP carbon products, for example, provide joint estimates of GPP and net primary production (Running et al., 2004) and GPP and NEE (Jones et al., 2017), respectively, but neither provides estimates of ET that are coupled to GPP. Zhang et al. (2016), on the other hand, provide coupled estimates of GPP and ET using static, biome-specific water-use efficiencies, but this approach does not provide estimates of downstream plant or ecosystem carbon balances, nor does the model allow for the dynamic changes in water-use efficiency that can occur in response to pulses of rainfall or variation in vapor pressure deficit (Roby et al., 2020).
Previous work has highlighted the potential for combining multiple remote sensing proxies to improve the representation of vegetation dynamics
(Stavros et al., 2017; Smith et al., 2019), and our results support this conclusion and provide further guidance on which remotely sensed variables
contribute most to model improvement in drylands. All three classes of remote sensing variables (optical, thermal, and microwave) contributed
positively to model skill (Figs. 8 and 9). In particular, the inclusion of SMAP soil moisture and temperature resulted in large gains in model skill
(Fig. 8), with the VI
While the LST-only models usually performed worst of all model subsets, and the monthly VI
Somewhat surprisingly, NDVI held higher leverage over model predictions than most other predictor variables (Fig. 9) despite previous research
documenting significant flaws in its ability to track dryland GPP (Wang et al., 2022; Yan et al., 2019). The relatively new
Given the challenges of mitigating and adapting to a changing climate, high-quality remotely sensed carbon and water flux estimates are needed for large-scale monitoring of changes in global ecosystem functions and ecosystem services, especially in dryland regions that are warming more rapidly than many other regions (Huang et al., 2017). Ecosystem production estimates provide the means to monitor and forecast rangeland and cropland productivity (e.g., Hartman et al., 2020) and to track changes in the terrestrial carbon cycle (Xiao et al., 2019). Evapotranspiration estimates are needed for monitoring drought and plant water use and water stress (Fisher et al., 2017), which in turn affect both fire risk (Rao et al., 2022) and mortality risk (McDowell et al., 2022).
Our DrylANNd approach has significant potential to provide these capabilities in the western US. Despite its short calibration and validation period,
DrylANNd's training data encompass much of the climate variability experienced by the western US, including both anomalously wet and dry years that
may serve as analogues when running the model forward in time as new MODIS and SMAP data are released. However, it is possible that the historically
atypical “megadrought” conditions (Williams et al., 2020, 2022; Dannenberg et al., 2022a) under which the model was trained may impose limitations on
the model's predictive capability. Some of the model's limitations in capturing interannual variability could perhaps be ameliorated by incorporating
additional remote sensing data that capture other aspects of dryland ecosystem functions. Solar-induced fluorescence, for example, effectively tracks
vegetation activity in drylands (Smith et al., 2018; Wang et al., 2022), particularly in dry evergreen needleleaf forests where reflectance-based
optical VIs tend to perform poorly (Magney et al., 2019; Wang et al., 2022). However, satellite-based solar-induced chlorophyll fluorescence (SIF) estimates suffer from coarse spatial and
temporal resolutions (e.g., GOME-2; Joiner et al., 2013), discontinuous spatial coverage (e.g., OCO-2 and OCO-3; Sun et al., 2018), or an even shorter
period of record than SMAP (e.g., TROPOMI; Köhler et al., 2018). Fusions of satellite SIF data with MODIS surface reflectance (e.g., Zhang et al.,
2018) overcome some of these limitations but would likely inherit many of the same flaws as reflectance-based optical VIs since they are based on the
same surface reflectance data. As SIF temporal and spatial resolution improves, it will likely become increasingly useful for dryland carbon and water
modeling. Gravimetric estimates of total terrestrial water storage (Andersen et al., 2005; Humphrey et al., 2018) could also improve the
representation of deeper moisture, which can be an essential water source for deep-rooted trees in semiarid systems (Rempe and Dietrich, 2018;
McCormick et al., 2021). However, like the longer-term SIF measurements, estimates of total water storage are limited to very coarse (0.5
Applying the DrylANNd approach at a global scale would require expanding the eddy covariance training sites beyond those used here, which are limited
solely to western US AmeriFlux sites. Drylands are generally defined as regions where annual precipitation is insufficient to meet evaporative demand
(e.g.,
Here, we developed and evaluated a machine learning approach (DrylANNd) for the joint modeling of key carbon and water fluxes (GPP, NEE, and ET)
specifically for drylands of the western US using a combination of satellite optical vegetation indices, multitemporal thermal infrared, and
microwave-based soil moisture and soil temperature. Long-standing challenges in current multispectral satellite-based estimation of dryland carbon and
water fluxes are the result of several interacting issues, including poor representation of soil moisture stress, decoupling between “greenness” and
plant physiology, high soil background reflectance in open canopies, and limited representation of dryland calibration and validation sites available
for model training and testing. Our approach partially addresses these limitations of previous satellite carbon and water flux estimates in drylands.
Soil moisture is explicitly included in the model rather than relying on the covariance between vapor pressure deficit and soil moisture or water-sensitive
vegetation indices as proxies of moisture stress. The model includes new vegetation indices (e.g., The model is trained specifically for dryland ecosystems based on an extensive network of 28 eddy covariance sites spanning a large latitudinal
and arid-to-subhumid gradient in the western US.
We found that this approach effectively captures monthly, seasonal, and spatial variation in GPP and, especially, ET through both space and time, though it still underestimates the magnitude of interannual variability in carbon and water fluxes. DrylANNd was less effective at capturing NEE than GPP or ET likely because respiratory processes are largely invisible to satellite sensors, with the magnitude of dryland carbon sinks overestimated particularly at evergreen needleleaf sites. Compared to models based solely on optical vegetation indices, the inclusion of SMAP soil moisture and temperature was crucial for improving estimates of both the magnitudes and temporal variabilities in all three fluxes, especially in dry grasslands and shrublands of the western US. On the other hand, the addition of multitemporal thermal observations improved flux estimates in evergreen needleleaf forests, where optical vegetation indices have traditionally struggled to capture GPP dynamics. Drylands play important roles both in the global carbon cycle (Poulter et al., 2014; Ahlström et al., 2015) and in ecosystem services supporting a large human population (Reynolds et al., 2007; Bestelmeyer et al., 2015), and DrylANNd significantly improves our ability to quantify carbon and water fluxes in these ecosystems.
The code for all modeling and analysis is available at
The supplement related to this article is available online at:
MPD, MLB, WKS, RLS, and JAB conceptualized the study. MPD, MLB, MRJ, and SKM developed the methodology. MPD and XW conducted analysis. MPD wrote the original draft (with contributions from MLB, MRJ, and SKM), and all authors contributed to review and editing of the manuscript.
The contact author has declared that none of the authors has any competing interests.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
We thank AmeriFlux and the tower PIs for making eddy covariance data publicly available. Any use of firm, product, or trade names is for descriptive purposes only and does not imply endorsement by the U.S. Government. USDA is an equal-opportunity employer and provider.
Matthew P. Dannenberg, Mallory L. Barnes, William K. Smith, and Miriam R. Johnston were supported by the NASA SMAP Science Team (grant number 80NSSC20K1805), and Xian Wang was supported by the NASA FINESST program (grant number 80NSSC19K1335).
This paper was edited by Paul Stoy and reviewed by Andrew Feldman and one anonymous referee.