Water-limited ecosystems cover approximately 40% of Earth’s terrestrial surface and, despite lower primary productivity than forested systems on a per unit land area basis, exert significant controls on global water, energy, and biogeochemical cycles (e.g., [1
]). Some drylands are among the most biodiverse areas in the world, and dryland biodiversity conservation is critical to sustainable development and food and water security [3
]. Drylands are particularly susceptible to degradation through disturbances (acting independently or in conjunction) such as fire, invasive species (e.g., [4
]), grazing (e.g., [5
]), and climate change (e.g., [7
]). Spatiotemporal patterns of vegetation in water-limited ecosystems are complex, and arise as a function of interacting abiotic and biotic processes that exert influence across a large range of spatial and temporal dimensions. Spatial variability in terrestrial vegetation at hillslope scales (e.g., 10 s to 100 s of m) in these ecosystems is both influenced by and can reveal important patterns in surface water, energy, erosion, and biogeochemical cycling (e.g., [8
]). Patterns in vegetation can reflect and conform to gradients in abiotic controls like solar radiation, and topographic convergence, and soil water (e.g., [10
]). A number of biotic factors can also influence patterns of terrestrial vegetation at these spatial scales. Herbivory has been shown to influence the distribution and abundance of vegetation at the relatively fine spatial scales such as hillslopes [5
]. At the same time, biomass in water-limited ecosystems (or proxies such as remotely sensed greenness) can exhibit relatively rapid temporal variability controlled, among other things, by phenology; disturbances such as fire [13
], water stress, and insect infestation [14
]; and atmospheric teleconnections that can produce rare but large precipitation events [15
The complexity, global significance, and sensitivity of dryland ecosystems motivates a need to develop improved capacities to monitor variability in terrestrial vegetation in reasonably fine spatial and temporal detail. The ability to accurately characterize vegetation properties across large regions of space and decades in time at spatial resolutions approaching hillslopes and temporal resolutions of weeks or days would provide the ability to understand the relative influence of various abiotic and biotic drivers of vegetation dynamics in space and time, and significantly enhance the ability to develop, parameterize, and verify spatial models of dryland ecosystems.
Of particular value to the ecohydrology community is the potential to produce historical datasets characterizing spatiotemporal variation in vegetation conditions in sufficiently fine detail to parameterize ecological and hydrologic models. These datasets are of particular interest for ecohydrologic process investigations because spatiotemporal patterns of vegetation are, in effect, an integrative and macroscopically observable manifestation of local patterns of water, energy, and nutrient cycling. Further, these datasets could also be used to condition models of vegetation dynamics, potentially improving the accuracy of model-predicted water, energy, and nutrient cycling and uncertainty quantification and propagation. Models with dynamic vegetation components, particularly those that prognostically simulate aboveground biomass or other complementary elements of phenology (e.g., [16
]) require multitemporal vegetation remote sensing datasets in order to constrain values of model parameters and their uncertainty (e.g., [20
]). In addition, while the vegetation remote sensing record has already led to significant advancement of these models, the ability to obtain reliable spatiotemporal datasets characterizing important attributes of terrestrial vegetation (i.e., Normalized Difference Vegetation Index, NDVI) at much higher spatial resolutions could significantly advance: (1) the ability to quantify spatiotemporal patterns at resolutions not resolved by global land models, (2) the development of parameterizations of sub-pixel resolution ecological processes within these global land models where needed, and (3) application of similar dynamic vegetation models at resolutions approaching individual hillslopes (e.g., [22
Satellites provide global observations of the Earth surface at predictable temporal revisit intervals and play a critical role in monitoring terrestrial ecosystems. Since the 1970s, for instance, the Landsat program has provided insight into global vegetation patterns at a 30 m resolution (e.g., [24
]). The MODerate-resolution Imaging Spectroradiometer (MODIS) sensors on NASA’s Terra and Aqua satellites provide global vegetation products at spatial resolutions of between 250–1000 m [25
]. Until the recent launch of Sentinel-2, the configuration of these two platforms demonstrated the significant tradeoffs for remote sensing of dryland terrestrial ecosystems. Specifically, the higher resolution Landsat products are associated with a revisit of 16 days at best, while some MODIS vegetation data products are available at daily intervals. Both platforms suffer from the risk of clouds contaminating individual images because they rely on the visible and infrared portions of the electromagnetic spectrum. Recognizing the complementarity of the Landsat and MODIS platforms, however, previous studies have sought to use data fusion techniques to leverage the high spatial resolution of Landsat and fine temporal revisit of MODIS (e.g., [27
]). These data fusion frameworks are particularly attractive for dryland ecosystems because remote sensing techniques in drylands rely on phenological changes to elucidate native from non-native vegetation (e.g., [31
]). In addition, the heterogeneity and high soil albedo in drylands can require higher spatial resolution to properly identify vegetation community types [32
]. The Sentinel-2 satellite with global coverage every 5 days (with two satellites) and multispectral bands similar to Landsat 8 OLI at 10 to 20 m (and three additional bands at 60 m) also have the potential to fill this data gap in dryland ecosystems.
In this study, we evaluate the ability of one such technique and software package, the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM) [27
], to accurately capture spatiotemporal variations in vegetation patterns in a water-limited ecosystem with steep environmental gradients. The STARFM is a freely-available data-driven algorithm that characterizes spatial patterns of atmospherically corrected radiance data from sequences of Landsat imagery and then synthesizes imagery at the resolution of Landsat (i.e., 30 m) by downscaling the MODIS imagery of that same landscape during the intervening periods of time. Like [33
], we opted to use the STARFM algorithm rather than more complex variations such as the Enhanced STARFM (ESTARFM; [34
]), which requires at least two pairs of training images (fine scale spatial resolution such as Landsat and coarse scale resolution such as MODIS) [34
], or a modified version of ESTARFM referred to as mESTARFM [35
] because testing is limited and some results are inconclusive about the use of ESTARFM in contrasting landscapes [36
]. The usefulness of STARFM in heterogeneous landscapes and semi-arid environments has been demonstrated in several studies [29
]. The [34
] study found that ESTARFM improves the accuracy of predicted reflectance in three reflectance bands for heterogeneous forested landscapes, while the [29
] study randomly sampled vegetation pixels for reflectance and determined that STARFM was feasible for time series compositing in dryland systems. The [38
] study was conducted in a dryland ecosystem and evaluated the performance of red, NIR and NDVI vegetation pixels stratified by covertype and focused on comparing results among bands and vegetation type as well as the influence of image base pair selection and lag time across a growing season on model performance. The [37
] study evaluated the use of STARFM for monitoring forest disturbance and regrowth using multi-date LiDAR for validation. The present study builds on these previous investigations by evaluating the influences of vegetation type, timing, topography and snow on STARFM performance with the integration of LiDAR and in the context of ecohydrologic modeling applications. Specifically, we focus on the accuracy of 30 m NDVI images synthesized with STARFM in a semiarid watershed that exhibits significant variation in environmental and physiographic conditions. The work is designed to address the following research questions: (1) How accurately can 30-m NDVI images obtained by fusing Landsat and MODIS NDVI retrievals using the STARFM algorithm reproduce corresponding, withheld, Landsat retrievals? (2) How do errors in STARFM-synthesized NDVI images vary with: (a) time of year, (b) vegetation functional type as captured by LiDAR canopy height models (CHM), (c) LiDAR-derived topographic aspect, and (d) presence/absence of snow as captured by the Normalized Difference Snow Index (NDSI)?
The remainder of the paper is organized as follows: Section 2
overviews the methods, including a description of the study area, experimental setup, and error metrics. Section 3
details the results of the analyses in the context of the research questions posed above. Section 4
discusses these results, focusing on how errors in STARFM predictions could conceivably propagate to ecohydrologic simulations of water, energy, and nutrient cycling. Section 5
presents conclusions of the study and highlights new potential avenues of inquiry identified by this work.
Comparisons between STARFM predictions of NDVI using two pairs of training data (Table 2
, Figure 9
) and the null model (Table 4
) suggest the Landsat-MODIS data fusions led to information gain for the five scenes from 13 May 2007 to 1 August 2007, where errors from STARFM predictions were lower than the corresponding errors in the null model. This is particularly evident in the May scene, where the null model was associated with an
equal to 0.410 and the STARFM predictions were associated with an
equal to 0.838. In this case STARFM is seemingly able to accurately apportion changes in NDVI captured by MODIS in space during a period characterized by rapid changes in phenology during spring green-up. By contrast, prediction differences in the null model were lower than corresponding differences obtained via the STARFM algorithm for the 27 April 2007, 17 August 2007, and 2 September 2007 scenes. Differences between the null model and the STARFM algorithm were large in April (
of 0.586 and 0.369, respectively), and minor in August (
of 0.982 and 0.943, respectively) and September (
of 0.982 and 0.968, respectively). The large discrepancy in the April scene can be partially attributed to snow covered pixels, which is consistent with similar studies (e.g., [53
]). When taking the NSE error metric into consideration, all synthetic images in the two-pair model contained new information NSE > 0, which is consistent with studies that found STARFM can capture phenological timing more precisely, even in areas where there are patterns of relatively fine-grained spatial heterogeneity [38
]. It is possible that application of the STARFM algorithm to DCEW (and water-limited, heterogeneous regions more generally) could be improved, especially in the “shoulder months” of the growing season, by using a MODIS product with higher temporal resolution than the 16-day NBAR composite. A study by [29
] tested differences in MODIS daily surface reflectance, 8-day composite and 16-day NBAR on STARFM performance in drylands and found that the 16-day NBAR was the optimal imagery for fusion with Landsat-5 TM; however, study observations highlighted the inherent temporal constraints of composite datasets (i.e., 8-day composite of 16-day NBAR) during times of rapid phenological change, such as green-up or senescence.
With the exception of the April scene, STARFM prediction error was consistently greatest in the NIR band compared to red and green bands. To date, several studies have compared STARFM performance as it relates to differences between predicted and observed values in the NIR and visible bands, with mixed results [29
]. Studies by [53
] found that errors in STARFM-derived predictions in the NIR may be attributable to atmospheric contamination at shorter wavelengths, which has been reported to affect the prediction accuracy for other fusion techniques [28
]. In the study by [28
] the intercept of the relationship between observed and predicted images was positive in all cases, which can be interpreted as a noise signal likely due to atmospheric and BRDF effects. Prediction differences in the study by [54
] could also be related to atmospheric influence, as the Landsat images were corrected to top of atmosphere reflectance not apparent reflectance [29
]. Similar to our study, both [29
] and [38
], which were conducted in dryland systems, found lower correlations in the NIR. Reference [38
] findings differed from previous studies that attributed better NIR [28
] and shortwave infrared [55
] predictions to greater influence of atmospheric contamination on shorter wavelengths. Importantly, that study was set in semi-arid dryland forests and yielded similar results to [29
]. Reference [38
] recommend that all synthetic reflectance products be evaluated prior to use as the influence of site specific conditions might be variable across the visible and NIR portions of the spectrum. Using a modified version of the STARFM algorithm, reference [34
] found that the ESTARFM algorithm performed slightly better than STARFM in a region of relatively homogeneous vegetation, but significantly better in a region of heterogeneous vegetation.
The above discussion reveals the difficulty in comparing across studies using STARFM due to the complexities in the region being studied, the data sets being fused, and the details of the satellite platforms from which those data are collected. In our study domain, absolute errors in STARFM predictions of NDVI tended to be largest in the tree cover type, followed by shrubs, and then grasses. At the same time, however, this conclusion does not hold in its entirety when considering the prediction errors scaled to average NDVI (). The relative error (RMSE in NDVI scaled by ) tends to be largest in the grass cover type, and this impact is greatest in mid-August, when RMSE in the grass regions is approximately 40% of the average NDVI. This suggests that a degree of caution be exercised in applying STARFM, and potentially other data fusion algorithms, in savanna-like ecosystems with significant grass coverage. This result suggests that particular attention is warranted in grass-dominated regions within heterogeneous ecosystems where errors in STARFM-predicted NDVI may be small relative to other regions within a study area in an absolute sense, but significant relative to the average NDVI of the grass-dominated region.
Correspondingly, there do not appear to be significant, generalizable conclusions about the role of topographic aspect in STARFM-derived predictions of NDVI across all vegetation functional types and time periods. Related to the above discussion of the magnitude of higher relative error in grass-dominated regions, this effect is particularly pronounced in south- and southeast-facing hillslopes in DCEW. We caution against over-interpreting this conclusion, however, because south- and southeast-facing slopes found in the middle elevations of DCEW tend to be dominated by grasses. Hence, the observed trend in Figure 5
seems to largely echo the previously discussed conclusions about the role of vegetation functional type as a predictor of STARFM performance. Interestingly, though, when we look at the magnitude of the relative error in NDVI in grass-covered regions we find that south- and southeast-facing pixels demonstrate an increase in relative error in the 1 August 2007 prediction. Relative error in NDVI in grass-dominated regions is again larger in the 17 August 2007 prediction on south- and southeast-facing pixels, but there is also an increase in relative error in east- and southwest-facing pixels relative to the previous 1 August 2007 prediction. By the 2 September 2007 prediction, relative error has decreased on east-, southeast-, south-, and southwest-facing pixels, but risen significantly on north-facing pixels in grass-dominated regions. Thus, topographic aspect may not play as large an influence as vegetation type in defining the magnitude of STARFM prediction errors in water-limited ecosystems, but it may impact the timing of errors when STARFM is being used to synthesize a time series of images.
Results from this study have important implications for the use of STARFM and other data fusion algorithms for creating value-added spatiotemporal vegetation remote sensing datasets that may be used to inform land models. These implications are related to the previous points we have discussed above. Specifically, results suggest that STARFM and other data fusion algorithms can serve to compensate for the tradeoffs between spatial and temporal resolution common to many remote sensing platforms used to characterize terrestrial vegetation, even in water-limited ecosystems with significant spatial heterogeneity in vegetation functional types. This suggests that it is possible to create historical reconstructions of spatiotemporal variation in variables like NDVI, which provide an important window into local dynamics of water, energy, and biogeochemical cycling, with reasonable confidence, over extended periods of time, and in relatively high spatial and temporal detail. When STARFM, specifically, can be applied in a way that makes use of Landsat images bracketing the dates of interest, it is possible to create accurate imputations of NDVI and other variables at the spatial resolution of Landsat and temporal resolution of MODIS. Historical reconstructions with these characteristics would be of significant value for constraining key land model parameters related to the vegetation canopy (e.g., LAI, albedo, etc.). Hence, data fusion algorithms like STARFM may have an important role to play in development, application, calibration, and verification of land models in water-limited ecosystems.
Along with previous studies, this work highlights the importance of continuity in remote sensing datasets with complementary characteristics. Algorithms like STARFM are useful only because of the simultaneous existence and availability of Landsat and MODIS data. STARFM and algorithms like it are sufficiently generic that other complementary pairs of remote sensing platforms could be used to develop historical reconstructions of spatiotemporal vegetation characteristics. While some studies have already begun to explore application of STARFM and other data fusion algorithms to constrain other satellite datasets to each other, additional work on this topic is warranted (e.g., [2
]). Landsat and the Advanced High-Resolution Radiometer (AVHRR) have an even longer period of operational overlap. In addition, although AVHRR is associated with a much coarser resolution than MODIS, it may be worthwhile to explore the performance of AVHRR-Landsat data fusion through STARFM in a variety of ecosystems of interest. The Visible Infrared Imaging Radiometer Suite (VIIRS) is a potential asset with observational characteristics similar to MODIS and complementary to Landsat. With the potential to derive fine spatial and temporal scales, data fusion of Landsat and Sentinel-2 could provide significant advances in monitoring the phenophases of dryland ecosystems.