Rangelands comprise between 26 to 36 percent of the terrestrial land surface [1
]. They consist of many ecosystems, including shrublands, savannas, tundra, deserts, alpine communities, coastal marshes, meadows, and grasslands [3
]. Historically, grasslands covered 300 million ha in the United States and 50 million ha in Canada [4
]. The mixed prairie, one grassland type, is located in the brown and dark brown soil regions of southern Alberta and Saskatchewan, Canada, and extends southward into northern Texas [5
]. Approximately 7 percent of the area of Alberta is covered by the dry mixed-grass prairie subregion, of which 43 percent remains [6
]. Climate is the determining factor for production and health in this ecosystem [1
], and water availability largely controls productivity [1
]. Climate change is expected to increase the amount and duration of growing season drought, as well as alter the frequency and intensity of growing season precipitation [1
]. Knapp and Smith [10
] demonstrated that prairie ecosystems have high production potential and substantial interannual variability, and are likely to be one of the most responsive biomes to future climate change [10
]. Human interest in these areas is usually focused on yearly productivity [12
] or biomass yields for economic and ecological monitoring purposes. The most common land use in the mixed prairie is ranching, as much of it is too dry to accommodate crop production without irrigation. Consequently, the economy in these regions is dominated by livestock ranching and supporting businesses. An emerging interest in prairie regions is yield assessment for agricultural insurance, which for prairies is limited by the inability to produce a direct, measurable yield; unlike traditional agriculture, a harvested crop yield is not the end product [13
]. Prairie vegetation functions to prevent soil erosion and resist landscape degradation, facilitate groundwater recharge, sequester large amounts of carbon dioxide, and support plant and animal diversity [14
], which are often positively related to productivity [16
]. Due to the potential for biospheric carbon sequestration and other co-benefits, in a period with deepening concerns about anthropogenically-induced climate change, there is also an emerging possibility of targeted carbon markets and economic incentives for proper land stewardship in these areas. These economic and ecological issues require accurate methods of evaluating prairie biomass yield in response to interannual variations in precipitation, especially given the current climate variability and predictions for future climate change.
Clearly, accurate monitoring methods are necessary for proper management of the mixed prairie. Traditionally, range management has been accomplished through the subjective evaluation and monitoring of large areas by skilled professionals relying on accumulated judgment and experience [18
]. This methodology has limitations for widespread application. Quantitative estimates of prairie productivity have also often been conducted through biomass harvests, which are expensive and time-consuming [18
]. Biomass harvests are destructive, so repeated sampling of a single plot is not possible, limiting the temporal practicality of the method. The time-consuming nature of manual field sampling, combined with the expansive size of these ecosystems, further limits the utility of biomass harvests for prairie management [19
]. Harvests require measurements to be performed in a representative number of places during a short time period and extrapolated to useful spatial and temporal scales. This often entails measurement errors, inaccuracies inherent in interpolation techniques, and delays in evaluating results that impair effective management.
The requirement for a sampling method that limits error introduced by inadequate sample size and distribution often exacerbated by personal bias [18
] has led to the exploration of remote sensing techniques for rangeland monitoring. Spectral reflectance enables non-destructive sampling over a wide range of spatial and temporal scales. Remote sensing data from multiple satellite platforms have been available since the 1970s [18
]. Many publicly funded satellites monitor the Earth, including the NASA Terra and Aqua satellites, which carry the MODIS (Moderate Resolution Imaging Spectroradiometer) sensors. These satellite sensors provide 250 m spatial resolution from 36 spectral bands, with daily coverage [18
]. MODIS products are created from quality-controlled multiband surface radiance and reflectance, and are freely available through websites developed by U.S. government agencies. The availability of quality-controlled satellite data has led to their widespread use and increased practicality for management and science.
Remote sensing data are also readily available for purchase from commercial satellite vendors. For example, the Système Pour l’Observation de la Terre SPOT 5 satellite produces 10 m spatial resolution multispectral imagery (SPOT 4 can also produce 10 m spatial resolution by co-registering a 10 m panchromatic and 20 m colour image) from four bands, two in the visible wavelengths, one in the near infrared and one in the shortwave infrared [23
]. The increased spatial resolution limits the possible temporal resolution such that SPOT samples a position on the Earth every 26 days, although the sensor can be pointed off nadir at targets not directly below the satellite, increasing the temporal resolution [18
]. Commercial satellites offer several potential advantages over MODIS; the finer spatial resolution of commercial satellites offer the potential of providing data products that more closely match the scales of small management units [24
]. However, images from commercial systems are often prohibitively expensive for practical application for range management [18
]. Commercial satellites also require tasking requests far in advance, and data gaps can occur with excessive cloud cover or otherwise reduced image quality during the tasking period [3
]. Despite these limitations, private satellites represent alternatives to government-funded programs and are also being explored for rangeland monitoring.
Past studies have demonstrated the ability of remote sensing to accurately estimate plant biomass or yield in grasslands [18
]. Satellite remote sensing also allows repetitive sampling, i.e., the creation of time series for evaluating growing season phenology and interannual variability in production. These methods enable productivity monitoring in near real time, relatively inexpensively, and with multiple scales of coverage. Many vegetation indices have been established from remotely sensed data. These indices provide proxies for vegetation biophysical properties, and can be used to diagnose rangeland conditions and trends [27
]. The most commonly used vegetation index is the Normalized Difference Vegetation Index (NDVI) [21
]. This index compares the low reflectance of photosynthetic materials in the red wavelengths to the high reflectance in the near-infrared to produce an estimation of plant biophysical parameters, including biomass, leaf area index (LAI) and the fraction of absorbed photosynthetically active radiation (fAPAR
). Numerous studies have demonstrated that NDVI strongly correlates with green biomass and leaf area index, particularly for grassland ecosystems [19
]. In grassland ecosystems, NDVI also provides a strong indicator of photosynthetic carbon uptake, typically measured via gas exchange or eddy covariance [16
], indicating its potential as a useful metric for carbon sequestration. Furthermore, in prairies, NDVI often scales with metrics of species richness and evenness [16
], so may also provide a useful indicator of biodiversity or other ecosystem co-benefits.
Due to the potential economic benefits of accurate prairie ecosystem monitoring, there is renewed commercial interest in evaluating cost-effective methods of assessing mixed prairie biomass. For example, the agricultural insurance industry has been exploring satellite remote-sensing-based monitoring programs and is interested in evaluating the accuracy of yield estimation in an effort to enhance customer confidence [34
]. This interest influenced the goals of this research, which were to evaluate how effectively NDVI from different common satellite sensors can estimate spatial and interannual patterns of grassland yield, and how satellite measurements compare to biomass harvests and NDVI from ground measurement. A common practice in assessing the utility of remote sensing for rangeland management is to conduct studies over a single growing season, such that evaluation of the impacts of year-to-year variability in precipitation is limited [29
]. Many other studies forego any intensive ground sampling with field spectrometers and rely only on data derived from broad-band satellite sensors [20
], leaving issues of satellite validation unresolved. Furthermore, the lack of a single, universal NDVI formulation due to the many band and sampling configurations of existing satellite sensors requires independent calibrations be conducted. Satellite calibration and validation (“CalVal”) efforts have often focused on a limited number of core sites [35
], and have been less concerned with evaluating the management applications of such sensors in the context of environmental change, leaving many aspects of validation to the scientific or end-user community. Consequently, there is an ongoing need to evaluate satellite sensors by comparing their output with field measurements, particularly when accurate data are needed for management regimes, carbon markets, or insurance purposes. Validation efforts comparing satellite measurements and traditional field sampling must, by definition, compare data collected on vastly different scales, necessitating some kind of scaling regime [36
], adding to the challenges of quantitative yield estimates from satellites.
In this study we analyzed data from the SPOT, MODIS Terra and Aqua platforms, and ground spectrometry to calculate NDVI. We then compared the resulting NDVI values to harvested biomass from southern Albertan mixed-grass prairies over two consecutive summers (2009 and 2010) having contrasting rainfall regimes. The objectives of the study were to (1) determine whether NDVI time-series from satellite systems that are currently utilized for estimating productivity are comparable to measurements from a ground spectrometer and harvested biomass; (2) determine whether variations in productivity as a result of contrasting weather conditions over two consecutive summers are clearly detectable, which could indicate the utility of remote sensing for monitoring long-term or year-to-year climate variability in prairie systems; and (3) determine the practical utility of NDVI for estimating above-ground biomass (“yield”) in the mixed prairie. Addressing the last objective is an essential foundation for cost-effective satellite-based rangeland insurance or carbon management programs. Since accuracy is crucial for carbon markets and insurance programs, we also considered the relative performance of two satellite sensors, SPOT and MODIS (Aqua and Terra), using ground NDVI measurements as a reference.
The weather conditions encountered during this study provided a natural experiment that provided a good test of the utility of remote sensing for monitoring interannual variation in yield. The precipitation differences in the two summers resulted in large differences in NDVI and biomass measurements between the two years for the two sites. Relative to single-year measurements, this interannual weather variability allowed a deeper understanding of the growth response of vegetation to varying moisture conditions, and allowed us to evaluate the ability of remote sensing to provide an accurate yield indicator across a wide range of growth responses typical of dry and wet years. Our findings of a large rainfall effect are consistent with the reports of Knapp and Smith (2001) and Camberlin et al. (2007), who observed that grasslands, such as the mixed prairie, have the capacity for large production response to uncommonly high precipitation [10
]. Clearly, remote sensing methods provide an effective means for tracking year-to-year variation in above-ground green biomass as a result of different weather conditions in this prairie biome.
The time series of the NDVI calculated from the ground spectrometer, SPOT and the MODIS platforms all show trends that are similar to the standing biomass, indicating that there is a clear effect of seasonal and interannual weather patterns on NDVI and biomass, which can be monitored with different remote sensing methods. While the general NDVI trends were similar for the different sampling methods, the exact shapes were not the same, indicating some sensor differences. The two MODIS sensors yielded similar patterns, yet the data from Aqua demonstrate depressed NDVI values in July 2009 and 2010, in comparison to the Terra data. This is more clearly demonstrated by the correlations with ground NDVI (NDVIG
), where the offsets for the two platforms were very similar, yet the correlation was slightly stronger for Terra than Aqua. This slight difference between the two MODIS NDVI responses could be due to a variety of reasons, including specific calibration and instrument or sampling characteristics for each platform. Deering et al. (1992) found that at the FIFE prairie site sensor viewing angle and solar zenith angle had a large effect on vegetation indices [39
]. Different viewing and solar zenith angles, in this case, would result from directly comparing data captured at different times of day, such that shadows and effects of surface anisotropy would affect the angular dependent signal due to the geometric configuration of the sun, sensor and target [40
]. Independent field measurements [31
] have shown a clear diurnal pattern for NDVI for prairie grassland, and this could have partly explained these differences between the two MODIS sensors having different overpass times.
Both MODIS sensors yielded NDVI values that were slightly higher than ground-measured NDVI, particularly at low NDVI values (Figure 5
and Figure 9
). This offset did not appear to be a result of sensor band differences, because it was maintained in the comparison to simulated NDVI data that were designed to eliminate possible band configuration differences (Figure 9
). The tendency of MODIS to produce values that are higher than the corresponding NDVIG
values calculated from ground spectrometers, has been previously reported for a chaparral site [41
]. These authors speculated that this could be an effect of the data processing such as the atmospheric correction, which is applied to all MODIS data, or could be a specific characteristic of the sensor itself [41
]; however, further studies need to be conducted to definitively explain this difference.
In comparison to the MODIS time series, the SPOT time series exhibited more scatter and a different offset when compared to ground-simulated NDVI (Figure 9
). In July, the SPOT NDVI (NDVIS
) overestimated ground NDVI in 2009 and underestimated in 2010, particularly for Sounding Creek (Figure 5
), which is made clearer by comparing the regression with the 1:1 line (Figure 9
). This greater scatter in the SPOT NDVI than in the MODIS NDVI would make simple correction for the offset much more difficult and suggests SPOT would be less reliable than MODIS for monitoring variation in rangeland yield. In our study, the utility of the SPOT platform for monitoring year-to-year variability was limited by the inadequate time series capability resulting from restricted temporal coverage along with periodic cloud contamination. ATIC (the local SPOT data distributor) did not distribute SPOT images with greater than 10 percent cloud cover within the image; consequently, cloud-free SPOT imagery was unavailable for Sounding Creek in both June and August 2010 (Figure 5
), making extensive, direct comparisons with MODIS and ground NDVI difficult. This reduced availability of SPOT data, combined with its high cost, precluded a fully equivalent sensor comparison.
Some studies [24
] have suggested that SPOT is superior to MODIS for rangeland monitoring due to its higher spatial resolution. SPOT satellite imagery has a spatial resolution of 30 m, which resulted in decreased temporal resolution relative to MODIS, as the smaller field of view requires more orbits of the Earth to cover the same position [42
]. In comparison to the daily coverage by the MODIS platforms, the 26-day temporal resolution of SPOT appeared to be inadequate to ensure repeat coverage of an area with data quality comparable to MODIS. While MODIS sensors have lower spatial resolution, the daily coverage allows creation of eight-day composites (employing maximum value compositing and bidirectional reflectance distribution function correction [37
]), so errors due to clouds and other effects are reduced. The SPOT sensor can be directed off-nadir to increase the temporal coverage [24
]; however, this introduces error due to altered sensor angle [42
]. Based on the results of our study, we conclude that the MODIS sensor may actually be more effective than SPOT for mixed prairie yield monitoring and management, particularly when continuous time series are required over large areas.
A number of factors can introduce variability into the NDVI–biomass relationship. For example, scatter in the NDVI–green biomass relationship can be a result of different canopy architecture, which has an effect on NDVI [43
], and this may lead to seasonal progression in the NDVI-biomass relationship (Figure S1
). In May, the variation between NDVI and biomass is partly due to vegetation carryover from the previous year masking new growth. At the beginning of the growing season there had been little or no new green growth, with the initial growth likely to be hidden by the last season’s biomass (the carryover error). The greater variability in August is likely due to senescence, which causes a decline in chlorophyll content (which would affect NDVI); this gradually changing greenness was not fully considered by our harvesting method, which distinguished between green and brown vegetation but did not consider the varying degrees of green associated with changing pigment content at different growth and senescence stages. The effects of dead biomass on reducing the NDVI correlations with biomass are likely to be strongest at the beginning of the season (due to the “carryover” of dead standing biomass from the previous year) and towards the end of the season, when vegetation senescence has begun [44
]. Our seasonal results could also have been influenced by the different species composition in the two townships, with Pinhorn having greater amounts of drought-resistant C4
graminoids than the C3
species dominant in Sounding Creek [46
]. Together, these effects undoubtedly contributed to scatter in the NDVI–green biomass relationship when all seasons and sites are combined (R2
= 0.85, Figure 6
D). Focusing on the midsummer peak (July) values, as suggested by Butterfield and Malmstroem (2009) [45
], effectively removed the early-season carryover error and the late-season senescence phase, and resulted in more accurate biomass estimation (R2
= 0.97, Figure 8
). Interestingly, despite the seasonal NDVI–green biomass shifts, there were strong correlations between peak-season (July) and end-of-season (August) values for NDVIG
= 0.99, not shown) and log green standing biomass (R2
= 0.99, not shown), suggesting a potential to estimate end-of-season biomass from peak-season measurements or vice versa.
Another potential source of error in NDVI measurements can be the background contamination from soil, litter, snow, surface wetness, or low-growing vegetation, which can confound satellite-based assessments of prairie productivity and ground cover [24
]. For instance, Hall-Beyer and Gwyn (1996) determined that the presence of Selaginella densa
, a mat-forming member of the fern phylum that is present in dry mixed-grass prairie, can also lead to errors in biomass estimates from optical remote sensing [48
]. Selaginella densa
mimics sparse grasses in NDVI measurements, yet it does not grow tall enough to be harvested with the methodology used here, so it was not included in our biomass data. Accounting for this effect could reduce scatter in the NDVI–green biomass relationship. By adapting the harvest sampling method to include very low-lying green vegetation in biomass estimates, this “background” green vegetation could be explicitly included in the analysis. Error resulting from this phenomenon would vary with rainfall and productivity. For example, drought could decrease the standing component of the vegetation and could allow more reflectance from lower canopy layers. On the other hand, early in the growing season during a drought year there can be more masking of the slowly emerging green biomass by the litter that is present from previous growing seasons (the carryover effect). This would decrease the amount of green vegetation visible to the spectrometer.
Both spatial aggregation and season affected the NDVI–biomass relationships (Figure S1
), and these findings indicate that careful consideration of sample timing and methods of data aggregation is warranted when developing a biomass monitoring program from satellite data. Selection of peak-season data aggregated at the township scale (approximately 93 km2
) largely removed these complications, and yielded a very strong agreement between NDVI and standing green biomass (Figure 9
), allowing a clear comparison of year-to-year productivity differences (Figure 7
). With the exception of SPOT NDVI, interannual changes in NDVI and standing green biomass (yield) were very similar for the different sensors (Figure 7
), which indicates that NDVI can be an effective proxy for the standing green biomass in the dry mixed-grass prairie. The different shapes of these seasonal trends for NDVI and standing green biomass can be largely attributable to the non-linear relationship between NDVI and biomass (Figure 6
and Figure 8
), which has been reported before [26
The curvilinear relationship between green biomass and NDVI creates an apparent “saturation problem,” where over a certain value NDVI becomes almost invariant to changes in vegetation amount and/or condition [26
]. This saturation, a natural result of the shape of the NDVI-biomass relationship [26
], can largely be explained by the exponential extinction of radiation in vegetation canopies [50
]. This relationship can be readily linearized by taking the log of biomass (Figure 8
), revealing the true nature of the relationship between NDVI and biomass. From the perspective of determining absorbed radiation (fAPAR
), a driving variable for production on this grassland ecosystem, there is less of a saturating problem; near-linear NDVI-fAPAR
relationships are often obtained for grassland ecosystems, demonstrating that NDVI also provides an effective input for productivity models (e.g., the light-use efficiency model [31
]). While our results indicate that NDVI saturation is not a serious constraint for biomass estimation in this mixed-grass prairie, this may not always be the case in more productive ecosystems.
The economic and ecological importance of the mixed prairie creates a need for accurate and cost-effective biomass monitoring. This study has demonstrated that remote sensing can accurately estimate biomass in these areas and that altered productivity due to interannual weather variability can be detected. These results are important for increasing the confidence of data users in the context of economic applications such as insurance payment programs and carbon markets. Ranchers’ perceptions of the accuracy of sampling methods are important for effective adoption and maintenance of remote sensing-based land management programs. Carbon markets would also require accurate monitoring in the mixed prairie, as varying climate and management regimes can undoubtedly influence sequestration. For such purposes, additional sampling (e.g., below-ground sampling) would also be needed. While further work would be required to demonstrate the full utility of NDVI for this purpose, this study strongly supports the use of satellite monitoring for estimating the above-ground component of productivity. Climate change is likely to result in more infrequent, intense precipitation in the areas being studied [1
]. This study indicates that satellite methods can accurately monitor above-ground yield changes that are likely to occur with changing precipitation regimes due to interannual variability or as a result of changing climate.