1. Introduction
Native grasslands are one of the largest ecosystems in the world with an estimated cover area of 40 to 50 million square kilometers [
1,
2]. They are defined as natural ecosystems dominated by naturally occurring grasses and other herbaceous species with the possible presence of woody species, used mainly for grazing by livestock and wildlife [
3].
As population increases, grasslands are becoming important contributors to human food supply (meat and milk), while providing other important ecosystem services, such as carbon sequestration, genetic material storage, water quality, and soils conservation [
4,
5].
The Río de la Plata grasslands (located between 28° S to 38° S) are among the world´s largest temperate-subtropical grasslands, covering the central-eastern part of Argentina, all of Uruguay, and southern Brazil. These grasslands are subdivided into “pampas” and “campos” [
6]: The former are temperate treeless grasslands located in eastern and central Argentina on flat and fertile plains, humid to arid climate, with warm summers and mild winters [
3]. “Campos” are grasslands consisting mainly of grasses, along with other herbaceous species, small shrubs, and occasional trees. They occur on undulating and hilly landscapes, with variable soil fertility, in sub-tropical humid climate, warm in summer and mild in winter, found in Uruguay, southern Brazil, and north-eastern Argentina [
3].
In Uruguay, livestock production is mainly based on extensive grazing of native grasslands, which represent over 65% of the country´s total area. Uruguay has a wide diversity of soil types [
7] and, as a consequence, its grasslands are heterogeneous including tens of species per m
2 [
8,
9] and a total of more than 350 registered species [
10]. These grasslands have a predominance of summer growing species (C4) with an increase in frequency of cold-season species (C3) during autumn and winter [
7]. Considering that warm-season species are responsible of the highest forage production (in spring and summer), and that this is the season with highest rainfall variability, the risks related to drought are very high [
7].
In environments with large variation in herbage production, due to seasonal and interannual variability in rainfall and temperature, the optimal stocking rate needed to reach a specific performance target varies widely among seasons and years [
11,
12,
13]. The control of grazing intensity through the management of stocking rate is an important tool to regulate the amount of solar energy captured and converted into beef production. Within this context, herbage allowance (HA), defined as kilograms of herbage dry matter per kilogram of animal body weight [
3,
14], may be more useful than stocking rate for managing the grazing process [
15]. Managing HA requires herbage mass estimation, or a proxy like herbage height [
16].
Information related to structural characteristics (herbage height or biomass) of native grasslands is essential to support management decisions, not only at the farm level but also at national and regional levels, to inform policy making. This results in demands on researchers to generate “low cost, appropriate and timely information that can be provided to farmers to support their decision-making” [
17]. Bearing in mind that existing field methods are labor-intensive and time-consuming, it is difficult to extend HA control to large areas. In this context, remote sensing monitoring is a promising option for quantifying large areas in a relative short time at a comparatively low cost and offering the possibility of analyzing historical data series.
Considerable research has been conducted to monitor indicators of the vegetation condition [
18] and, more specifically, to characterize grasslands functioning. Thus, remote sensing has been used to estimate above ground net primary production (ANPP), with the advanced very high-resolution radiometer (AVHRR-NOAA) and the moderate-resolution imaging spectroradiometer (MODIS) [
19,
20,
21,
22,
23,
24]. Annual grassland biomass has also been estimated using different satellite sensors (MODIS, SPOT, and AVHRR) with good results in Northern China [
25] and in the Sahel [
26].
Other ecosystem’s structure and function attributes have also been related to vegetation indexes coming from Earth observation. In the Patagonia steppes, Gaitan et al. [
27] assessed the relationship between cover and species richness with nine different vegetation indicators. The authors showed that NDVI explained 30%–40% of the total variability found in these ecosystem attributes. In the arid southwest grasslands of North America, the total herbaceous vegetation cover (green and senescent), height, and biomass estimated using the soil adjusted total vegetation index (SATVI) for cover and the near infrared band from Landsat for height and biomass, was highly correlated with observed information [
28].
Grasslands biophysical parameters, mostly considered in a season or in annual base, have been retrieved from Earth observation for many years. Recently, the modelling approaches are evolving to more complex, robust, and efficient ones [
29]. Also, some studies included higher-resolution sensors such as Sentinel 2 or modern technologies such as radar (active sensor) [
30,
31,
32]. However, more research is needed to accurately estimate the intra-annual performance of some biophysical variables as biomass or height.
In this work, we tried to estimate observed grasslands behavior as a “photograph” of what was available in terms of forage height at different time steps along the year. This approach differs from estimating the height as the annual or seasonal accumulation of biomass. Our approach seems to be particularly important for improving management of livestock production systems. We also analyzed intra paddock variability at different spatial scales.
In this context, the objective of this study was to contribute to improve grazing management by evaluating the ability of remote sensing information to estimate forage height (as an estimator of available biomass) at paddock scale.
4. Discussion
In agreement with previous studies that described native grassland variability [
40], field measurements were extremely variable within the paddocks for each sampling date, with coefficients of variation around 75%. In most paddocks, this heterogeneity was even higher as the height increased, as shown in
Figure 2, where the size of the boxes is larger when the values of the median are higher. This is probably due to the small-scale botanical and structural heterogeneity of this environment, but also because of variation in livestock management and in climate conditions. The amount of variability also showed differences between dates, probably associated with seasonal species composition. On the other hand, as satellite information provides an average value at a pixel resolution (500 m, 250 m, 30 m), it is expected to have much less variability but, differing to what we anticipated, showed no strong differences between these spatial resolutions. Thus, Landsat 8 (with spatial resolution of 30 m), could not represent the variability of native grassland more accurately than MODIS. Therefore, it can be expected that using other sensors with a resolution that is slightly higher, such as Sentinel 2, would not result in a better characterization of this variability either. Although it could be worth to check this fact, we probably must appeal to sensors of much higher resolution (1 m or centimeters) or non-optical ones (radar).
When we compared these extremely variable field measurements with MODIS satellite information, poor correlations were found. MODIS composite or daily variables seem to be not sensitive to grassland height variations. Considering the median height of every date and paddock, the minimum value was 1.4 cm and the maximum was 22 cm (
Figure 3 boxplots), while satellite band reflectance values from GA product vary only from 0.17–0.37 for MIR, 0.20–0.39 for NIR, and 0.03–0.1 for the red band. Regarding the vegetation indices, NDVI varied from 0.43 to 0.83 and the NDWI from −0.15 to 0.28. This is consistent with results found in the semi-arid Sahel by Olsen et al. [
26], who concluded that an increase in NDVI over time cannot always represent an increase in herbaceous biomass. This could be due to the fact that NDVI saturates at high biomass or leaf area index.
In contrast to what was found on monospecific pastures of alfalfa and grass (tall fescue), where good correlations are reported between height and several vegetation indices [
41], our study showed that no daily or composite MODIS satellite information could explain height observed behavior in the natural grasslands of every paddock and date. This could probably be due to the hundreds of species present in native grasslands that result in such a heterogeneous environment, and/or to the presence of non-photosynthetically active plant material that could influence the signal captured by remote sensing sensors. For example, in paddocks 4, 17, and 18 (
Table 2), the correlation between height and FPAR was negative, opposing what we expected. A similar situation occurred in paddock 18 with a positive correlation between height and red band. Careful consideration of the results at individual paddocks revealed that in all cases mentioned above, we found one very influential point with a high value of height and low value of FPAR (and high value of red band). This could probably be due to a situation with high biomass and a large proportion of senescent material. In general, MODIS composite information (MOD13Q1 and MCD15A3H) did not show good results either, and only MIR appeared as a variable worthy of further exploration in future work.
Analyzing paddocks individually, red band was the most promising variable (
Table 4). Some paddocks had high and significant correlation values with several satellite variables (NIR, MIR, red band, NDVI, NDWI). Paddock 9, 11, and 19 showed significant correlations in more than 9 (out of a total of 13) daily remote sensing variables; paddock 13 a total of 8 with daily delta value; and paddock 9 and 19 more than 4 (from a total of 7) with composite images. Other paddocks had no correlation with any variables (paddock 1, 7, 12, and 20). Taking into consideration only the significant values, the sign (+ or −) of Pearson correlation coefficient values result as we expected for all the plots (positive relation between height and NIR, NDVI, and NDWI; and negative relation with MIR and red band).
On the other hand, changes in height and in satellite variable values from one date to the next (delta values) only showed better results than single date values when NIR (GA) was used (
Table 5). There were some paddocks that had relatively good correlations (Person correlation values >0.6) between height and one or more of the different satellite variables such as MIR and NDWI of the GA MODIS product and MIR, red band, and NDVI of Nbar MODIS product, but these cases were isolated and not always with the same paddock involved. As an example, paddock 15 only showed high correlation between delta height and delta MIR (−0.603) but this satellite variable had very poor correlation values in other paddocks (
Table 5).
It is worth considering that in the analyzed period, large variability on weather conditions was observed. During the first three years (2012–2014), weather conditions were relatively favorable, which resulted in NDVI values above or close to the average conditions of a 30 years series, while in April–July 2015, an intense drought period occurred [
42]. Hence, the low correlation observed in our research cannot be attributed to a lack of variability in the observed values.
During the analysis process, we sought for possible common field characteristics in paddocks, such as size of the paddocks (
Table 1), location (
Figure 1), or field data CV (
Figure 3 boxplots) with the same response to a specific satellite information signal, but no explanatory co-variable was found.
Considering paddocks 9, 11, 12, and 13 (all paddocks from the same farm) and taking into account characteristics that could be detected by remote sensing, the first two paddocks had relatively homogeneous conditions related to soil types, elevation, vegetation, and historical management, and the other two had very different conditions, being more heterogeneous in relation to soils, elevation, water sources, vegetation, and size. These could be the reasons that field measurements in paddocks 9 and 11 have strong correlation with different satellite variables while paddocks 12 and 13 did not (
Table 4). On the other hand, paddock 19, the third paddock with significant correlation values with several satellite variables, had no homogeneous conditions and, therefore, this statement cannot be generalized.
Furthermore, the results of Cimbelli and Vitale [
30] suggested that higher grass had a bigger component of the red band, but our results could not be explained by that either. Paddocks with higher values of median height (average or median of sampling dates, and maximum observed value) had different spectral response.
As it was shown, the spatial variability (heterogeneity) observed in native grasslands under grazing conditions is extreme, and this makes it difficult to manage, plan, and characterize at the paddock scale with a single average or even median field measurement value. Considering this, it is even more difficult to try to do so based on earth observation information that provides a value for each pixel, no matter how good the information is.
Additionally, more field information, such as density or vegetation cover, water content, chlorophyll level, or percentage of senescent material, needs to be analyzed and monitored in order to explain differences found between spectral information responses in different paddocks.
Most of the studies, including this one, have used a single sensor to analyze a very complex and heterogeneous ecosystem and this could probably be a limitation. Bearing this in mind, Wachendorf et al. [
17] propose developing a system with complementary sensors to overcame these limitations and provide better estimations of different grassland characteristics. In addition to this, the integration of different sources of information (remote sensing, field data, air photos, and street-level imagery) to monitor grasslands, also seems to be an auspicious methodology [
43,
44]. SAR data (radar remote sensing) calculated from X- or C-band were explored too, with promising results [
45,
46]. On the other hand, some authors propose hyperspectral and high-resolution images as an option to overcome the difficulties at a paddock scale grasslands monitor [
47]. Finally, drones and other unmanned aerial vehicles offer an opportunity for new applications, proving higher spatial resolution and customized spectral and temporal resolution [
17,
48]. It is worth mentioning that as spatial resolution increases, it is more difficult to scale the analysis to a regional or national level. Moreover, a decision support system needs to be simple to be used by different stakeholders.
5. Conclusions
As it was expected, height of native grasslands is extremely variable within paddocks for each sampling date and between dates (seasonal variability). This variability is what we must deal with when we analyze native grassland forage availability and satellite information, and, at least at these spatial resolutions (500 m, 250 m, and 30 m pixel), the estimation of pasture height variability cannot be represented accurately.
We did not find high correlations between field measurements of height and MODIS composite/daily variable when we analyzed all the paddocks considered together or paddock by paddock. However, some areas of future work seem to be justified. The daily red band of Nbar MODIS product seems to be a promising variable to explore, with relatively good correlation values in 41% of the paddocks. When composite MODIS images were considered, MIR had the best performance with 29% of the paddocks showing negative correlation values higher than 0.45.
This work aims to contribute to manage the grazing process on livestock production systems, based on Earth observation information. Our results showed that no MODIS composite/daily variable was able to predict robustly the native grassland height behavior, but some satellite information came out as promising.
Work is needed in order to find remote sensing methods that can be used to monitor the "instantaneous" condition of grasslands (height or available biomass), and this research has evidenced some of the related difficulties and opportunities.