Comparison of Laboratory and Field Remote Sensing Methods to Measure Forage Quality

Recent research in range ecology has emphasized the importance of forage quality as a key indicator of rangeland condition. However, we lack tools to evaluate forage quality at scales appropriate for management. Using canopy reflectance data to measure forage quality has been conducted at both laboratory and field levels separately, but little work has been conducted to evaluate these methods simultaneously. The objective of this study is to find a reliable way of assessing grassland quality through measuring forage chemistry with reflectance. We studied a mixed grass ecosystem in Grasslands National Park of Canada and surrounding pastures, located in southern Saskatchewan. Spectral reflectance was collected at both in-situ field level and in the laboratory. Vegetation samples were collected at each site, sorted into the green grass portion, and then sent to a chemical company for measuring forage quality variables, including protein, lignin, ash, moisture at 135 °C, Neutral Detergent Fiber (NDF), Acid Detergent Fiber (ADF), Total Digestible, Digestible Energy, Net Energy for Lactation, Net Energy for Maintenance, and Net Energy for Gain. Reflectance data were processed with the first derivative transformation and continuum removal method. Correlation analysis was conducted on spectral and forage quality variables. A regression model was further built to investigate the possibility of using canopy spectral measurements to predict the grassland quality. Results indicated that field level prediction of protein of mixed grass species was possible (r2 = 0.63). However, the relationship between canopy reflectance and the other forage quality variables was not strong.

about the application of hyperspectral remote sensing remains due to the fact that in situ canopy reflectance may be heavily influenced by atmospheric variation [34], soil background and leaf orientation and distribution [35]. Such a concern may be even bigger in northern semi-arid mixed grasslands, which are characterized by a large amount of bare soil and dead material [36,37]. Despite the concern, few studies have compared the estimation of forage chemical composition using in situ hyperspectral canopy reflectance measured in the field and ex situ hyperspectral reflectance data measured for dried grass in the laboratory. In addition, the application of hyperspectral remote sensing is also influenced by the mathematical methods used to establish the relationship between canopy reflectance and forage quality [23,33], for example, Mutanga et al. [16] found the continuum absorption approach is better than band width on predicting forage chemical composition.
Grazing could affect the nutritive value of the forage [38,39] and the effects would change as the grazing density change, which were concluded from an experiment in a moist grassland in Czech Republic [40] and a southern mixed grass prairie in the USA [41]. But little research has been focused on the effects of light to moderate grazing on forage quality in northern semi-arid mixed grass prairie. Therefore, the objectives of our study are three-fold: (1) to evaluate hyperspectral measurements for estimating the forage chemical composition of northern semi-arid mixed grass prairie in the field and laboratory, (2) to test the reliability of a field level grass quality prediction method, and (3) to compare the effects of grazing on grassland quality estimation.

Study Sites
The study was conducted in Grasslands National Park of Canada (GNP) (West Block) and surrounding pastures owned by the federal and provincial government and private ranchers in southern Saskatchewan, Canada. Dominant vegetation in this community includes needle-and-thread (Hesperostipa comata Trin. & Rupr.), blue grama (Bouteloua gracilis (HBK) Lang. ex Steud.) and western wheatgrass (Pascopyrum smithii Rydb.). Spikemoss (Selaginella densa Beauv.) and Junegrass (Koeleria macrantha (Ledeb) J.A. Schultes f.) are also frequently observed [42]. Characterized by the semi-arid mixed prairie ecosystem, this region receives approximately 340 mm annual precipitation primarily falling in the growing season (May-September). The mean annual temperature is 3.4 º C with the maximum mean daily temperature of 20 º C in July and the minimum of 15 º C in January [43]. Variation in grassland quality due to long-term changes in productivity or elevated atmospheric CO 2 are of a concern to land managers in this region.

Data Collection
Data were collected in the field and laboratory. Fieldwork was conducted in June-July of 2003. Thirty sites were randomly selected within upland grasslands of the study area. At each site, two 100-m transects perpendicular to each other oriented in the cardinal directions were surveyed. Field level canopy reflectance was collected at 5 m intervals along each transect (40 readings at each site) using an ASD FR Pro spectroradiometer (Analytical Spectral Devices, Inc., USA) to capture within-site variation. The wavelength measurement range was 350-2,500 nm, and the spectral resolution was 3 nm at 700 nm and 10 nm at 1,400 and 2,100 nm. A 25º field of view probe was used pointing down to the canopy at approximately 1 m above ground, yielding a view area of the surface of about 0.15 m 2 . Measurements were taken between 1,000 h and 1,400 h local time under cloud-free conditions. Calibration was made using a white reflectance panel (Labsphere, USA) at approximately 10 minute intervals to minimize solar variation due to the changing sun angle over the measurement period. Reflectance (R) in a given waveband was calculated by dividing the canopy spectra by the white reference spectra. After the canopy reflectance was measured, above-ground vegetation was clipped within a 20 × 50 cm quadrat at 20 m intervals along each transect, which yields 12 clipped samples at each site. Vegetation samples were sorted into green grass, forb, shrub, and dead material shortly after clipping and then dried in an oven at 65 º C for 24 hours. Dry green grass samples randomly selected from 360 samples collected on 30 sites were cut into 1 cm lengths and spread out to 1 cm thickness on a black surface for measuring indoor spectral reflectance [44]. Indoor spectral reflectance was then measured in the laboratory, using the ASD Pro Lamp for illumination, which is specifically designed for indoor lab reflectance measurements over the region of 350-2,500 nm. Therefore, spectral measurements were made on in situ field-canopy samples and ex situ dry green grass in laboratory, and used for developing models for evaluating forage quality. After the indoor spectral measurements, the dry samples were then sent for chemical analysis (ETL ChemSpec Analytical Services Ltd.).

Spectral Data Processing
Hyperspectral remote sensing provides detailed information at small wavelength intervals. But this also generates a great volume of data that can be difficult to interpret. Interpretation is aided by processing the data using algorithms that aggregate regions with known information value. The most commonly used hyperspectral data processing methods are the first derivative transformation and continuum-removed absorption algorithm. First derivative reflectance: frequently, the clearest patterns in reflectance data occur not in the absolute quantity of reflectance, but rather the rate of change of reflectance from one wavelength to another. The first derivative transformation of the reflectance spectrum (R fd ), that calculates the slope values from the reflectance, can be derived from the following equation [45]: (1) where is the first derivative reflectance at a wavelength i midpoint between wavebands j and j + 1. is the reflectance at wavelength j + 1, and Δ λ is the difference in wavelengths between j and j + 1.
Continuum removal: the continuum removal algorithm proposed by Kokaly and Clark [46] is a popular method for absorption feature detection. Spectral absorption regions are selective for chemical composition and density in vegetation because higher concentrations of particular chemicals lower the reflectance in unique spectral regions. Six known chemical absorption regions were selected: two (carotenoids and anthocyanins) in the visible range that are related to chlorophyll absorption and nitrogen concentration (R 470-518 and R 550-750 ), and four in the shortwave infrared region (R 1116-1284 , R 1634-1786 , R 2006-2196 , and R 2222-2378 ), that are the result of lignin, protein and other chemical absorption [16,26,47]. For the six defined absorption regions, a linear continuum was identified from the start to the end points. Differences between the measured value and the continuum were calculated and then summed to represent accumulated absorption in the given region. The other approaches for hyperspectral data processing, such as band depth, band depth ratio, and normalized band depth index, were not tested in this study. The reason is that the research of Mutanga et al. [16] indicated the continuum-removed derivative reflectance (CRDR) is superior to them for estimating forage chemical composition.

Data Analysis
Descriptive statistics [mean, standard deviation (Std), minimum (Min), maximum (Max), and coefficient of variation (CV)] of forage chemical variables and nutrient contents were calculated to understand the general forage quality in the study area. To identify the best indicators for estimating chemical components from laboratory and field conditions, Pearson's correlation analysis was conducted between chemical composition and the raw reflectance of dry samples and canopy reflectance, and the first derivative reflectance as well as the accumulated absorption calculated from the continuum removal approach. The relationships were plotted against wavelength regions for comparisons.
With one-fifth observed protein data and the accumulated absorption data (six each) set aside for validation, the stepwise regression analysis was applied to the other four-fifth data (24) to develop a model for predicting protein using field measured accumulated absorption derived from reflectance. The developed regression model was validated using the jackknife and cross validation approach. This approach operates by withholding the spectral data for one site and building the model functions using the data from the remaining sites. The process of removing one site from the dataset was repeated until all sites had been withheld [48]. Finally, the performance of the prediction model was evaluated by root mean-squared error (RMSE) and average relative error (ARE) by comparing the predicted values to the six observed protein values.
The differences in both chemical components and spectral absorption features between grazed and ungrazed grasslands were investigated through the analysis of variance (ANOVA), to document the effects of grazing on the detection of grassland forage quality.

Vegetation chemical and nutrient contents
Results of the laboratory chemical analysis of the forage quality variables of interest are summarized in Table 1. Most forage quality variables showed a small CV, indicating a relative homogeneity among samples. The only two exceptions are protein and ash content. The relatively high CV of the protein content among samples indicates the diversity of protein content of grass at the field level. Although nutritional requirements of ruminant herbivores vary with physiological state and with body size [49], benchmark values do exist. For a 500 kg mean live weight cow, the requirement of protein content is 3.1% for maintenance and 10.7% for producing 30 kg milk/day. The protein content in the study area is at above the cow maintenance level but is lower than the milk production level.  Figure 1 illustrates the spectral features of the samples in both laboratory and field measures. The upper and lower 95% confident limits based on the 30 samples indicated that the spectral variation is small, which indicates relatively little variation over the sampled areas. Six absorption regions selected for this study were also clearly identifiable. Generally, field level spectral measures were lower than laboratory measures, which is likely due to the influence of bare ground and dead materials. The leaf orientation in the field may also contribute to the difference in the spectral response curve from the measures of dried grass samples in the laboratory which minimize the effects of leaf orientation and distribution.
Relationships between grass quality variables and raw reflectance: Correlation analysis on chemical composition and raw reflectance showed that only lignin, protein and ash were significantly related to reflectance of dried grass in the lab ( Figure 2a) whereas protein, moisture and ash were significantly related to canopy reflectance in the field (Figure 2b).
The visible wavelength region is important: the red region (600-700 nm) correlates with protein and ash, and the blue (400-500 nm) and green regions (500-600 nm) are correlated with lignin concentration. The important red-edge region for estimating protein is consistent with the finding of Kawamura et al. [33] in a sown pasture land in Japan and of Starks et al. [50] in a warm season pasture land in USA, but the blue region (415-460 nm) was also well correlated with protein content in their study. In addition, we found no significant relationship between chemical contents and spectral features beyond the near infrared energy range. This finding agrees with [51] indicating that the visible wavelength region is the most important for grassland biophysical characterization. However, accumulated absorption regions showed significant relationships with the spectral measurements for lignin and protein, indicating a potential of using the continuum removal approach for prediction.  Relationships between quality and first derivative reflectance: By definition, higher first derivative reflectance will occur at steeply sloped regions of the spectral response curve (Figure 1). The most obvious region is at the red edge, popularly used for Nitrogen content estimation of crops and grasses [52,53]. The first derivative of a subset of the red edge region (550-750 nm), showed a dramatic enhancement of the relationship between forage quality variables and spectral measures (Figure 3a). Energy variables showed very similar results, due to similar calculation methods ( Figure 3b). The significant regions in field level measurements are in the blue and red wavelength ranges (Figure 3c). We also compared the effects of the first derivative transformation (550-750 nm) applied to lab and field spectral measurements. In general, the first derivative transformation increased our ability to estimate chemical composition. The relationships between chemical composition and the first derivative reflectance are much more consistent based on spectral reflectance measurements in the lab throughout the wavelength region, even though the relationships were higher for the field measures for most chemical variables. The only exception is lignin content, which is more highly correlated to lab-measured reflectance than that in the field.
Relationships between chemicals and the areas of absorption: For most foliage chemicals, the mid infrared region is important for qualitative detection, but not for quantitative estimation. Considering the limitations of the first derivative method, the continuum removal method was tested. Accumulated absorption was correlated with forage quality variables ( Table 2). All relationships between the forage quality variables and the 470-518 nm, 550-750 nm, 1,116-1,284 nm, and 1,634-1,786 nm wavelength regions were statistically significant. Despite significant correlation, the variation of forage quality variables accounted for by the variation of spectral reflectance varied from 5-66%, indicated by the r 2 values. Interestingly, based on the spectral measurements in the lab, the two mid wavelength regions (2,006-2,196 nm and 2,222-2,378 nm) did not show significant correlations with any chemical components. Nonetheless, based on the canopy reflectance in the field, the wavelength region (2,006-2,196 nm) demonstrates moderate to high correlation with protein, ash, and moisture, and the region of 2,222-2,378 nm is moderately correlated with protein content. In the lab, ash and moisture content only weakly correlated with the 1,116-1,284 nm wavelength region, but in the field both ash and moisture are moderately correlated with two visible wavelength regions, one near-infrared and one mid-infrared wavelength region.
In addition, NDF was only correlated with 470-518 nm wavelength region, but other chemical variables correlated with more than one spectral absorption region in the lab.

Protein prediction model development and validation
When using the six absorption areas derived from field spectral measurements as independent variables in a stepwise regression analysis, we could estimate 63% variation of protein content from the absorption area of the wavelength region of 2,006-2,196 nm with the following equation: (N = 24, r 2 = 0.63, P = 0.000) (2) where A represents accumulated absorption area in the wavelength range indicated in the equation; N is the sample number; and P is the significance value at the 0.05 significance level. This suggests that the model is very good at monitoring protein from in situ field samples ( Figure 4). Higher protein content results in smaller absorption areas due to its strong energy absorption capability in these particular wavelength regions. Therefore, larger absorption areas have smaller protein content. Six observed protein data and the corresponding accumulated absorption areas were set aside and used to evaluate the performance of the prediction model. RMSE and ARE between the observed and predicted protein values were shown in Table 3. RMSE indicates that the predicted protein data series is quite consistent with the observed protein. ARE shows the overall error of the model prediction is 8%, which further indicates that the model (Equation (2)) can accurately predict protein based on the in situ measures of canopy reflectance.  Forage quality variables and the reflectance of the wavelength regions used for calculating the accumulated absorption in grazed and ungrazed grassland were compared in Table 4. ANOVA analysis indicated that canopy reflectance of all the selected wavelength regions is significantly different at the 0.05 significance level in grazed and ungrazed grasslands. Except lignin and protein, the other tested chemical composition, including ADF, Ash, moisture at 135 º C, and NDF, is also significantly different. As for the ADF, ash, moisture, NDF, and Lignin, both the range and CV were higher in grazed sites than those in the ungrazed area, indicating a higher variation in grazed swards. However, both range and CV indicate that the variation of protein is smaller in grazed sward than that in ungrazed sward. In addition, variation of reflectance in the wavelength regions of 1,116-1,284 nm and 1,634-1,786 nm is smaller, but that in other wavelength regions is larger, in grazed sward than ungrazed sward, indicated by CV. With biomass as a covariate, we only found that NDF (F1, 87 = 12.3, P < 0.001) and ADF (F1, 87 = 3.5, P = 0.055) differed significantly among grazing treatments and no quality measure varied significantly with biomass. In both cases, NDF and ADF were higher under the ungrazed than those under the grazed treatments (Table 4). Table 4. Comparisons of chemical composition and canopy reflectance under grazed and ungrazed conditions (N is the sample number; and Mt and Ln stands for moisture at 135 º C and Lignin, respectively).

Discussion
Our data show that the strongest correlation between forage quality and remote sensing data was not from ex situ spectral reflectance measurements of dried grass samples in the laboratory, but from in situ reflectance measures made on the vegetation canopy. Previous research has indicated that water could mask the relationship in fresh samples [25,26]. However, our reflectance values from in situ samples showed consistently higher correlations to chemical composition than from dried samples, despite the field samples being a mixture of green grass, forb, shrub, dead materials, and soil.
Relationships between vegetation quality and spectral measurements differed when measured under laboratory conditions compared to field conditions for a range of chemical variables. Noise was less for indoor spectral measurement, but correlations to chemical composition were also lower. Both the first derivative transformation and absorption continuum removal methods worked very well to improve the relationships between chemical and spectral data. However, each method has its own suitable wavelength regions; the first derivative transformation was for absorption and reflectance transition regions especially the red edge area, while the absorption removal method was mainly for absorption regions that can be extended to mid infrared regions. Furthermore, the first derivative transformation provided more promising results for measurements in a laboratory, which provided smoother spectral response curves. Sixty-six percent variation of protein content could be explained by the total absorption area in the wavelength region of 2,006-2,196 nm. Lignin showed a better relationship with indoor spectral data than with the field measurements. The reason for this result is not fully understood as the sample numbers are lower for lignin compared to other chemical variables because of the high cost for lignin laboratory analysis. Further analysis with increased sample numbers should be investigated. Both first derivative and continuous removal approaches are specifically for hyperspectral remotely sensed data. Fortunately, data from hyperspectral satellite sensors are available currently (e.g., Hyperion on Earth Observing-1 and CHRIS on Proba). Other satellite-based hyperspectral sensors such as Canadian HERO will be launched in the future. At the same time, airborne hyperspectral imagers (e.g., AISA Airborne Hyperspectral Imager of UPM-APSB and HYMAP of Australia) also become available. However, the practical application of hyperspectral remote sensing will not be applicable in the short term due to high cost.
To verify if the remote sensing data at field level could be used for grass quality assessment, we made detailed comparisons on field and dried samples. Except lignin content, all field spectral measurements provided a better indication of chemical composition than spectral measures in the lab. The trend among wavelength regions was clear: higher negative correlations in absorption regions and higher positive correlations in reflectance regions for protein, with significantly positive correlations in near infrared region for moisture content and ash. For the lab measures, only the visible region showed a significant positive relationship with lignin content, which has no significant relationship with field spectral measurements. To measure forage quality using spectral reflectance in such a ecosystem, apparently it is not necessary to dry samples. This is consistent with the findings of Starks et al. [50], Mutanga et al. [16], Mutanga and Skidmore [30][31][32], and Kawamura et al. [33] for forage quality estimation in warm season pasture land in Oklahoma, USA, tropical savanna rangeland in South Africa, and in a sown pasture land in Hokkaido, Japan, respectively.
The interplay between plant moisture, nutritional quality, grazing and reflectance is interesting, which deserves further exploration. It appears that in more mesic vegetation communities, maturational variation in plant quality will be greater, but our ability to detect these changes using in situ reflectance will be greatly impaired. Plant maturation is the factor that largely dictates nutritional quality [54]. Nutritional quality typically declines as maturation increases [55]. Water stress, for its part, typically retards maturation in plants [56]. On average, nutritional quality in plant under water stress will be higher than those in well watered plants [57]. Thus, the rate at which nutritional quality declines with maturation, typically indicated by biomass and growing season, will slow. Unfortunately, in plant canopies containing more moisture, nutritional quality estimated by canopy reflectance is less reliable [16,21,29,58]. So where nutritional quality of swards likely varies as environmental conditions change, our ability to track the variation of nutritional quality will be constrained.
Additionally, forage quality is improved by short-term grazing pressure due likely to the delay of average maturity of plant tissues caused by grazing in the sward [59]. Grazing can also affect reflectance signals by creating more bare ground and reducing the density of non-photosynthetically active material in a canopy [60]. Again, grazing creates a more temporally and spatially dynamic vegetation canopy that, because of what changes are made, make forage quality detection using reflectance measures more difficult. Hence, where moisture and grazing are the dominant effects on vegetation quality, in situ measures of nutritional quality of forage will be challenging. This supposition is consistent with our findings. In the semi-arid grassland, variation in nutritional quality was very low and was only related to grazing treatment (not biomass). Hence our in situ reflectance measures were as effective as or more effective at detecting plant chemical composition than ex situ measures. But these factors also point to why the literature is divided on the effectiveness of in situ measures of forage quality. We should perhaps not expect to find similar results on more mesic grasslands or where grazing is more intense.
This conclusion has interesting implications for tracking large-scale changes in forage quality resulting from increasing atmospheric CO 2 concentrations and climate change. As CO 2 concentrations increase, plant growth rates are expected to increase [61] and communities may shift towards greater dominance by C 3 species [3,[62][63][64][65]. In many ecosystems these changes may be accompanied by changing precipitation patterns that will result in longer periods of little or no precipitation during the growing season [66,67]. Because C 3 species tend to be of higher nutritional quality than C 4 species [68] and decline less in quality with maturation [69], a shift towards C 3 species accompanied by less growing season precipitation suggests less within-season variation in forage quality in temperate grassland ecosystems. Hence, while lower canopy moisture will enhance in situ detection of canopy quality variation that may accompany climate change, there may be less within-season variation to detect.
Landscape level measurement of grassland nutrient content shows a great promise with modern remote sensing tools. Our study supports the idea that grassland quality assessment using remote sensing approaches can be successful in the field. Nevertheless significant challenges, such as teasing out sward moisture effects and accounting for grazing management, remain. The key to overcoming these challenges, we believe, will be the careful comparison of controlled grazing treatments in grassland ecosystems differing in precipitation regimes.