Semi-arid regions and drylands together cover more than 40% of the globe. In spite of the low fractional cover of vegetation and minimal annual precipitation, the contribution of semi-arid biomes to the global uptake of carbon dioxide from the atmosphere can be significant on annual time scales [1
]. The inter-annual variability of precipitation in these ecosystems can result in both rapid carbon stock accumulation, and subsequent turnover due to frequent disturbance (e.g., drought, insect outbreak, fire), but how these relationships alter the net storage of carbon is poorly understood (e.g., [1
]). The climate in the southwestern USA has recently experienced both increased regional air temperature, and decreased precipitation, both trends which are projected to continue in the coming decades [3
]. The combined effects of these changes in climate have increased drought severity, exacerbated ecosystem stress, and ultimately triggered widespread forest mortality across the region [8
]. Given the rapid state transitions of vegetation in these climatically sensitive biomes, and the projected changes in climate and associated mortality for this region [11
], accurate monitoring of carbon uptake dynamics in these semi-arid ecosystems is an essential part of constraining uncertainties associated with regional carbon balance.
Monitoring ecosystem carbon uptake over large geographic extents requires the use of remote sensing. Typically, remote sensing driven models of ecosystem function predict Gross Primary Productivity (GPP), the total atmospheric carbon taken up via photosynthesis (e.g., [12
]). Satellite remote sensing provides the means to predict GPP by employing light use efficiency-based algorithms [14
] that rely on vegetation indices (VIs), such as the normalized difference vegetation index (NDVI, [16
]). This approach has been used at a variety of sites and scales, both by employing empirical (e.g., [17
]) and light use efficiency/process driven techniques (e.g., [19
]), generally using high temporal resolution, coarse spatial resolution (≥250 m) remote sensing data from sensors, such as the Moderate Resolution Imaging Spectrometer (MODIS) (e.g., [12
]). To this end, the success of most space borne estimates of GPP hinge on a robust relationship between satellite estimated light interception and photosynthesis, via light use efficiency models, which generally perform well relative to in situ
measurements from flux tower networks [20
However, in semi-arid regions, vegetation function is constrained by water availability for the majority of the year, with periods of drought and heat interrupted briefly by spring snow melt or episodic pulses of precipitation. Consequently, in these biomes, changes in NDVI are constrained primarily by the availability of water, rather than light or temperature [21
]. This results in NDVI and GPP often being decoupled in semi-arid regions due to low soil moisture, particularly where evergreen plants are present [22
]. Other characteristics unique to semi-arid ecosystems that challenge the use of NDVI for characterizing changes in GPP in these biomes are highly variable precipitation patterns which trigger high inter-annual variability in GPP due to seasonal water limitations, but low variability in LAI and/or chlorophyll concentration (subsequently low variability in NDVI). Further, the low LAI (<1.5 mean LAI across landscapes, e.g., [23
]) and spatially heterogeneous plant canopies typical of these systems can result in further uncertainties due to high reflectance by the soil background, thus confounding spectral signals relating to plant function (e.g., [24
One of the most universal responses to leaf stress is increasing visible reflectance [27
], due to a combination of stressors which ultimately reduce chloropyll a + b, and consequently reduce the absorption of incident light [27
]. Chlorophyll a + b strongly absorb in the red portion of the visible spectrum, resulting in saturation of the red band at low Chlorophyll a + b, and reducing its potential to track initial chlorophyll loss at the onset of stress. However, the red-edge has been shown to have a more linear response to a wide range of chlorophyll concentrations, increasing its potential as a stress indicator in vegetated systems over conventional red-NIR combinations [26
]. Given the increasing availability of the red-edge waveband in commercial (e.g., RapidEye, WorldView-2, WorldView-3) sensors, and freely available data (Sentinel-2, launched June 2015), testing the potential of the red-edge waveband to improve modeled estimates of GPP in semi-arid ecosystems is becoming a feasible task.
Here, we test the ability of spectral VI’s other than NDVI to model GPP in semi-arid ecosystems. We focus particularly on piñon-juniper woodlands for several reasons. First, because it is the largest biome in the Southwestern US, covering 18 million ha in New Mexico, Arizona, Colorado and Utah. Second, the changes in climate in this region have triggered a significant amount of mortality in this biome [10
], and thus, quantifying the extent of this disturbance on productivity throughout the region is crucial to understanding how this extensive mortality has impacted both current and future carbon dynamics. Finally, we take advantage of an existing experimental manipulation in a Piñon-Juniper woodland in central New Mexico that was girdled in 2009 to simulate the widespread piñon mortality observed throughout the region. The advantage of using this experimental manipulation is that since the girdling in 2009, changes in ecosystem productivity triggered by this mortality have been continuously monitored using eddy covariance, and compared to similar measurements in a nearby intact PJ woodland that serves as a control.
Recent research [34
] conducted in our experimental manipulation suggested that the red-edge employing normalized difference red-edge index (NDRE) [35
] was more sensitive than NDVI to the observed initial decrease in leaf chlorophyll concentration triggered by piñon girdling in this PJ woodland, adding evidence to its potential as a stress monitoring component in GPP modeling efforts. The NDRE was determined to be more sensitive to changes in greenup of the low LAI herbaceous vegetation following mortality in this same system [22
]. Based on these findings, the goal of this study was to test three specific hypotheses. The first hypothesis being that the observed variability in NDRE will allow more accurate estimation of GPP in both the disturbed and undisturbed PJ woodland in this experimental manipulation. Secondly, due to the inherent dependence of productivity on water availability in this biome, adding the normalized difference wetness index (NDWI), a VI that is sensitive to changes in foliar water content, will significantly improve the model fit by constraining the model error during the dry season. Finally, given the low vegetation cover and highly heterogeneous nature of PJ woodlands, that VIs generated from higher spatial resolution (5 m) data will provide more accurate estimates of GPP relative to traditional, moderate resolution (30 m) remote sensing data.
Our results are promising in that we can use simple linear models to estimate GPP in both disturbed and undisturbed PJ woodlands driven by remotely sensed datasets. While structurally sensitive, NDVI is more informative to models of GPP than NDRE except during periods of extreme stress or disturbance. Similarly, we only saw a significant improvement in model performance using NDWI at the girdled site, during the manipulation event that took place in the Fall of 2009. Finally, the use of the RapidEye data did slightly improve estimates of GPP in both the control and girdled sites relative to Landsat ETM+, however this was only true when we reduced the variability in scene to scene sensor view angle in our RapidEye time series. This apparent advantage of the RapidEye data may be due to a combination of factors, including spatial resolution (5 m pixels vs. 30 m pixels) and spectral sensitivity of the sensor. While this may not play a strong role in more homogeneous, closed canopy systems, sensor view angle in this study often imposed more variability on NDVI than natural seasonal variability. Consequently, we recommend that remote sensing efforts to model VI sensitive processes in heterogeneous, low fractional cover systems, place high constraints on acquisition angles for time series, or bin analyses by viewing angle to minimize the potential confounding effects.
We recognize that the temporally resolved RapidEye data set we utilized for this study is not a common commodity and currently carries with it a large cost. Using red-edge data added sensor and illumination geometry complexity, but did improve estimates of GPP during periods of ecosystem stress despite it. Our results suggest high resolution, red-edge employing platforms will potentially be very useful for resolving changes in canopy function during periods of rapid disturbance or recovery where LAI may be changing slowly in relation to chlorophyll content. The recently launched Sentinel-2 satellite mission will allow this to be tested on a broader scale by providing greater spatial and temporal resolution than Landsat, as well as the ability to calculate NDRE and NDWI, and be freely available. Secondly, the upcoming soil moisture active passive sensor (SMAP) may provide either direct measurements, or modeled estimates of soil moisture, providing further predictive power to estimate carbon uptake rates in semi-arid ecosystems.