3.2. Spectral Slope Analyses
Figure 1 shows the CR spectra of seven vegetation samples with different percentages of CP. Changes in slope across the different spectral ranges as a function of different CP contents could be seen at: 1748–1764 nm, 1766–1794 nm, 2070–2088 nm, 2278–2286 nm, 2316–2330 nm and 2334–2344 nm (
Figure 1a,b). The spectral assignments for the main absorbance values of the vegetation were taken from Curran [
20] and Schwanninger
et al. [
25].
Figure 1c shows an example of this relationship by zooming in on the 1748–1764 nm spectral range, where the slope is seen to increase with increasing CP content. Increases or decreases in spectral slope as function of CP content were also obtained for the other spectral ranges (
Figure 1b).
Figure 1.
(a) Continuum removal (CR) reflectance spectra of vegetation samples with different percentages of crude protein (CP). Note the variability in the slopes across the different spectral ranges: 1748–1764 nm, 1766–1794 nm, 2070–2088 nm, 2278–2286 nm, 2316–2330 nm, 2334–2344 nm. (b) Visualization of the slope’s tendency to increase or decrease as a function of different CP contents. Up-pointing arrow indicates a slope increase with CP content increase; down-pointing arrow indicates a slope decrease with CP content increase. (c) Zooming in on the 1748–1764 nm spectral range to demonstrate the changes in the slope with changes in CP concentration.
Figure 1.
(a) Continuum removal (CR) reflectance spectra of vegetation samples with different percentages of crude protein (CP). Note the variability in the slopes across the different spectral ranges: 1748–1764 nm, 1766–1794 nm, 2070–2088 nm, 2278–2286 nm, 2316–2330 nm, 2334–2344 nm. (b) Visualization of the slope’s tendency to increase or decrease as a function of different CP contents. Up-pointing arrow indicates a slope increase with CP content increase; down-pointing arrow indicates a slope decrease with CP content increase. (c) Zooming in on the 1748–1764 nm spectral range to demonstrate the changes in the slope with changes in CP concentration.
The distribution of CP, NDF and MEC values is shown in
Figure 2a, and that of the slopes for the six spectral ranges is shown in
Figure 2b. Significant variability in the spectral slopes was obtained, presumably due to differences in chemical composition and the chromophores, which differ for each spectral range.
Figure 2.
Distribution of (a) CP, neutral detergent fiber (NDF), metabolic energy concentration (MEC) and (b) slope values for six spectral ranges.
Figure 2.
Distribution of (a) CP, neutral detergent fiber (NDF), metabolic energy concentration (MEC) and (b) slope values for six spectral ranges.
To gain additional insight into the relationship between spectral slopes and chemical constituents, different ranges of CP conditional on slope ranges were divided into three main classes: slope less than zero (
Figure 3a), between 0 and 0.001 (
Figure 3b), and above 0.001 (
Figure 3c). The negative slope was found to correspond to low CP values (
Figure 3a). Thus, by knowing only the slope’s sign (negative/positive), it is possible to quantitatively assess the amount of CP in a sample. Specifically, slopes in the range of 0–0.001 corresponded to CP values of around 6%–15% (
Figure 3b), whereas slopes above 0.001 corresponded to high CP values (around 15%–30%,
Figure 3c).
In light of these results, the data distribution of CP, NDF and MEC
versus slopes were analyzed and three main categories were found: (1) low, representing about 25% of the data, when CP ≤ 4.5%, NDF ≥ 67% and MEC ≤ 1.6; (2) medium, representing about 50% of the data, when 4.5% < CP < 13%, 53% < NDF < 67% and 1.6 < MEC < 2.5; (3) high, representing about 25% of the data, when CP ≥ 13%, NDF ≤ 53% and MEC ≥ 2.5. For each category of CP, NDF and MEC and for each spectral domain, the range of slope values was identified and defined to assess pasture quality according to the above three categories (summarized in
Table 1). For example, the CP level was classified as low, medium or high when the slope value between 1748 and 1764 nm was ≤−0.0008, between −0.0008 and 0.001 or ≥0.001, respectively (
Table 1).
Figure 3.
Distribution of CP values for slopes (between 1748–1766 nm) of: (a) <0; (b) 0–0.001; (c) >0.001.
Figure 3.
Distribution of CP values for slopes (between 1748–1766 nm) of: (a) <0; (b) 0–0.001; (c) >0.001.
Table 1.
Slope-based criteria for assessment of pasture quality using three levels of crude protein (CP), neutral detergent fiber (NDF) and metabolic energy concentration (MEC).
Table 1.
Slope-based criteria for assessment of pasture quality using three levels of crude protein (CP), neutral detergent fiber (NDF) and metabolic energy concentration (MEC).
%CP |
Spectral Range | Low (≤4.5%) | Medium (4.5%–13%) | High (≥13%) |
1748–1764 nm | Slope ≤ −0.0008 | −0.0008 < Slope< 0.001 | Slope ≥ 0.001 |
1766–1794 nm | Slope ≤ 0.005 | 0.005 < Slope< 0.008 | Slope ≥ 0.008 |
2070–2088 nm | Slope ≤ −0.0065 | −0.0065 < Slope< –0.003 | Slope ≥ −0.003 |
2278–2286 nm | Slope ≥ 0.0058 | 0.0025 < Slope < 0.0058 | Slope ≤ 0.0025 |
2316–2330 nm | Slope ≤ 0.003 | 0.003 < Slope < 0.008 | Slope ≥ 0.008 |
2334–2344 nm | Slope ≥ 0.0048 | 0.0034 < Slope < 0.0048 | Slope ≤ 0.0034 |
%NDF |
Spectral Range | Low (≥67%) | Medium (53%–67%) | High (≤53%) |
1748–1764 nm | Slope ≤ −0.0008 | −0.0008 < Slope < 0.001 | Slope ≥0.001 |
1766–1794 nm | Slope ≤ 0.0052 | 0.0052 < Slope < 0.0076 | Slope ≥ 0.0076 |
2070–2088 nm | Slope ≤ −0.0065 | −0.0065 < Slope < −0.0032 | Slope ≥ −0.0032 |
2278–2286 nm | Slope ≥ 0.0058 | 0.0026 < Slope < 0.0058 | Slope ≤ 0.0026 |
2316–2330 nm | Slope ≤ 0.0032 | 0.0076 > Slope > 0.0032 | Slope ≥ 0.0076 |
2334–2344 nm | Slope ≥ 0.005 | 0.005 > Slope > 0.0033 | Slope ≤ 0.0033 |
| MEC |
Spectral Range | Low (≤1.6) | Medium (1.6–2.5) | High (≥2.5) |
1748–1764 nm | Slope ≤ −0.0005 | −0.0005 < Slope < 0.0016 | Slope ≥ 0.0016 |
1766–1794 nm | Slope ≤ 0.0053 | 0.0053 < Slope < 0.0091 | Slope ≥ 0.0091 |
2070–2088 nm | Slope ≤ −0.0062 | −0.0062 < Slope < −0.0022 | Slope ≥ −0.0022 |
2278–2286 nm | Slope ≥ 0.0057 | 0.0018 < Slope < 0.0057 | Slope ≤ 0.0018 |
2316–2330 nm | Slope ≤ 0.003 | 0.003 < Slope < 0.0093 | Slope ≥ 0.0093 |
2334–2344 nm | Slope ≥ 0.0048 | 0.0031 < Slope < 0.0048 | Slope ≤ 0.0031 |
The slope trends presented in
Figure 2 and
Figure 3 strongly suggest that the spectral reflectance properties of pastures with similar components can provide information regarding the amount of CP, NDF, MEC.
Assessing pasture quality using only slope-based criteria (
Table 2), we found that in general, the total success rates were good for: (1) CP at 72%–80% (low), 50%–74% (medium) and 84%–98% (high); (2) NDF at 6%9–77% (low), 37%–72% (medium) and 79%–94% (high), (3) MEC at 56%–87% (low), 46%–60% (medium) and 79%–89% (high). Rating the overall success rate (“Total” row in
Table 2) for the three categories in each spectral range showed that the best slope criteria for CP were in the 1748–1764 nm and 2278–2286 nm spectral ranges, for NDF in the 1748–1764 nm and 2278–2286 nm ranges, and for MEC in the 2316–2330 nm and 1748–1764 nm ranges. Importantly, the p-value was highly significant between number of identifications per category for all categories (
p < 0.05), indicating that the slopes are highly associated with CP, NDF and MEC contents and their association is not due to random variation.
Table 2.
Success rates of pasture-quality evaluations using slope algorithm. Note that chemical data were used as reference.
Table 2.
Success rates of pasture-quality evaluations using slope algorithm. Note that chemical data were used as reference.
% CP | Total Per Category Based on Chemical Data | Total Per Category Based on Slope Algorithm |
---|
1748–1764 nm | 1764–1794 nm | 2070–2088 nm | 2278–2286 nm | 2316–2330 nm | 2334–2344 nm |
---|
Low (≤4.5%) | 54 | 41 | 76% | 39 | 72% | 41 | 76% | 41 | 76% | 43 | 80% | 43 | 80% |
Medium (4.5%–13%) | 114 | 83 | 73% | 57 | 50% | 84 | 74% | 79 | 69% | 72 | 63% | 68 | 60% |
High (≥13%) | 57 | 55 | 96% | 48 | 84% | 49 | 86% | 56 | 98% | 55 | 96% | 49 | 86% |
Total | 225 | 179 | 80% | 144 | 64% | 174 | 77% | 176 | 78% | 170 | 76% | 160 | 71% |
% NDF | |
Low (≥67%) | 64 | 49 | 77% | 49 | 77% | 44 | 69% | 48 | 75% | 46 | 72% | 45 | 70% |
Medium (53%–67%) | 109 | 79 | 72% | 40 | 37% | 72 | 66% | 71 | 65% | 57 | 52% | 64 | 59% |
High (≤53%) | 62 | 54 | 87% | 51 | 82% | 49 | 79% | 56 | 90% | 58 | 94% | 50 | 81% |
Total | 235 | 182 | 77% | 140 | 60% | 165 | 70% | 175 | 74% | 161 | 69% | 159 | 68% |
MEC | |
Low (≤1.6) | 39 | 30 | 77% | 26 | 67% | 25 | 64% | 30 | 77% | 34 | 87% | 22 | 56% |
Medium (1.6%–2.5) | 80 | 44 | 55% | 41 | 51% | 42 | 53% | 48 | 60% | 47 | 59% | 37 | 46% |
High (≥2.5) | 47 | 42 | 89% | 41 | 87% | 41 | 87% | 40 | 85% | 41 | 87% | 37 | 79% |
Total | 166 | 116 | 70% | 108 | 65% | 108 | 65% | 118 | 71% | 122 | 73% | 96 | 58% |
Overgaard
et al. [
57] showed that one or two years of spectral measurement are insufficient to build fully operational models for cereal property predictions. In this regard, our study employed 10 years (2002–2011) of observations, and should thus adequately represent the expected range of conditions in our study area. Furthermore, Givens
et al. [
27] and Pullanagari [
58] argued that high-precision spectroscopic instruments, spectral resolution and number of wavelengths are crucial for determining pasture quality with high accuracy. However, this study shows that highly accurate estimates of pasture quality using spectral slopes that require only two wavelengths are possible.
The chemometric approach essentially correlates the optical data with “reference” chemical constituents based on different statistical methods, enabling high-quality estimates of vegetation composition [
8,
9,
26,
59,
60]. However, high-precision instruments with high spectral resolution are needed to use these methods, such as the Foss NIRS system model 5000 spectrometer. Furthermore, in HS remote sensing, it might not be possible to obtain such high-precision information due to atmospheric effects, a low signal-to-noise ratio, a varying field of view for every pixel, spectral instability and problems of spectral mixing [
61]. Therefore, the developed method might overcome this limitation. Importantly, four of the six selected spectral regions were completely outside the spectral range of water absorbance (e.g., 1300–1650 nm, 1430–2230 nm). This result is encouraging, because it might expand the possibility of using the developed methodology with airborne HS or even multispectral sensors. Indeed, one needs only two wavelengths to calculate the slope data (e.g., the beginning of the spectral range and its last point), which are then incorporated into the model to assess the chemical constituents.
Next, to investigate the feasibility of our method for airborne HS sensor application, we resampled the spectral data from the Foss NIRS 5000 spectrometer (700 bands) to AISA, an airborne HS sensor with 109 bands between 1100 and 2363 nm. The procedure applied to the resampled data set was the same as that applied to the original data,
i.e., CR spectra for visual inspection of spectral slope change
vs. chemical data. Five spectral ranges (
Table 3) were identified as candidates for the slope method, and a new set of slope criteria was defined (
Table 3). The results are summarized in
Table 4. Four of the five spectral ranges were similar to the four spectral ranges in the original data (
Table 1 and
Table 3) and an additional “AISA-fitted” spectral range (2306–2317 nm) was introduced. In general, the total success rates (
Table 4) of the three categories were as good as those of the original data set (
Table 2) for: (1) CP at 70%–91% (low), 58%–70% (medium) and 86%–100% (high); (2) NDF at 64%–95% (low), 33%–67% (medium) and 79%–92% (high); (3) MEC at 67%–85% (low), 46%–63% (medium) and 83%–94% (high).
Table 3.
Slope-based criteria for assessment of pasture quality using three levels of CP, NDF and MEC, after resampling the spectral region of FOS-5000 (700 bands at 1100–2363 to spectral region of advanced imaging spectrometer for applications (AISA) (109 bands at 970–2500 nm).
Table 3.
Slope-based criteria for assessment of pasture quality using three levels of CP, NDF and MEC, after resampling the spectral region of FOS-5000 (700 bands at 1100–2363 to spectral region of advanced imaging spectrometer for applications (AISA) (109 bands at 970–2500 nm).
% CP |
Spectral Range | Low (≤4.5%) | Medium (4.5%–13%) | High (≥13%) |
1747–1770nm | Slope ≤ 0.00025 | 0.00025 < Slope< 0.003 | Slope ≥ 0.003 |
2061–2096 nm | Slope ≤ −0.014 | −0.014 < Slope< −0.008 | Slope ≥ −0.008 |
2270–2293 nm | Slope ≥ 0.01 | 0.01 > Slope> 0.0045 | Slope ≤ 0.0045 |
2306–2317 nm | Slope ≤ −0.001 | −0.001 < Slope< 0.0025 | Slope ≥ 0.0025 |
2317–2328 nm | Slope ≤ 0.0045 | 0.0045 < Slope < 0.008 | Slope ≥ 0.008 |
% NDF |
Spectral Range | Low (≥67%) | Medium (53%–67%) | High (≤53%) |
1747–1770nm | Slope ≤ 0.0005 | 0.0005 < Slope< 0.003 | Slope ≥ 0.003 |
2061–2096 nm | Slope ≤ −0.014 | −0.014 < Slope< −0.008 | Slope ≥ −0.008 |
2270–2293 nm | Slope ≥ 0.01 | 0.01 > Slope> 0.0035 | Slope ≤ 0.0035 |
2306–2317 nm | Slope ≤ 0.0004 | 0.0004 < Slope< 0.003 | Slope ≥ 0.003 |
2317–2328 nm | Slope ≤ 0.0045 | 0.0045 < Slope < 0.008 | Slope ≥ 0.008 |
MEC |
Spectral Range | Low (≤1) | Medium (1.6–2.5) | High (≥2.5) |
1747–1770nm | Slope ≤ 0.001 | 0.001 < Slope< 0.004 | Slope ≥ 0.004 |
2061–2096 nm | Slope ≤ −0.013 | −0.013 < Slope< −0.006 | Slope ≥ −0.006 |
2270–2293 nm | Slope ≥ 0.01 | 0.01 > Slope> 0.001 | Slope ≤ 0.001 |
2306–2317 nm | Slope ≤ −0.0005 | −0.0005 < Slope< 0.0035 | Slope ≥ 0.0035 |
2317–2328 nm | Slope ≤ 0.0045 | 0.0045 < Slope < 0.009 | Slope ≥ 0.009 |
This “coarse” spectral resolution might be implemented in HS and/or multispectral remote-sensing sensors. In recent years, there has been an increase in the availability of images with high spatial resolution [
1]. In this regard, our study might serve as a basic infrastructure for future directions in the use of this technology to determine pasture quality.
Table 4.
Success rates of pasture-quality evaluations using slope algorithm. Note that chemical data were used as reference, after resampling the spectral region of FOS-5000 (700 bands at 1100–2363 to spectral region of AISA (109 bands at 970–2500 nm).
Table 4.
Success rates of pasture-quality evaluations using slope algorithm. Note that chemical data were used as reference, after resampling the spectral region of FOS-5000 (700 bands at 1100–2363 to spectral region of AISA (109 bands at 970–2500 nm).
% CP | Total Per Category Based on Chemical Data | Total Per Category Based on Slope Algorithm |
---|
1747–1770 nm | 2061–2096 nm | 2270–2293 nm | 2306–2317 nm | 2317–2328 nm |
---|
Low (≤4.5%) | 54 | 48 | 89% | 38 | 70% | 45 | 83% | 49 | 91% | 46 | 85% |
Medium (4.5%–13%) | 114 | 72 | 63% | 80 | 70% | 66 | 58% | 77 | 68% | 66 | 58% |
High (≥13%) | 57 | 55 | 96% | 49 | 86% | 57 | 100% | 55 | 96% | 55 | 96% |
Total | 225 | 175 | 78% | 167 | 74% | 168 | 75% | 181 | 80% | 167 | 74% |
% NDF | | | | | | | | | | | |
Low (≥67%) | 64 | 42 | 66% | 41 | 64% | 51 | 80% | 61 | 95% | 46 | 72% |
Medium (53%–67%) | 109 | 73 | 67% | 63 | 58% | 63 | 58% | 36 | 33% | 57 | 52% |
High (≤53%) | 62 | 49 | 79% | 49 | 79% | 56 | 90% | 53 | 85% | 57 | 92% |
Total | 235 | 164 | 70% | 153 | 65% | 170 | 72% | 150 | 64% | 160 | 68% |
MEC | | | | | | | | | | | |
Low (≤1.6) | 39 | 33 | 85% | 26 | 67% | 28 | 72% | 29 | 74% | 33 | 85% |
Medium (1.6–2.5) | 80 | 39 | 49% | 37 | 46% | 50 | 63% | 40 | 50% | 45 | 56% |
High (≥2.5) | 47 | 44 | 94% | 39 | 83% | 44 | 94% | 42 | 89% | 42 | 89% |
Total | 166 | 116 | 70% | 102 | 61% | 122 | 73% | 111 | 67% | 120 | 72% |
3.3. PLS Analyses
The results of PLS modeling of the CP, NDF and MEC values are presented in
Table 5. Both the whole set of six spectral ranges (53 bands) and each individual spectral range was modeled, and both gave relatively accurate estimates of pasture quality. The best-fit model for the calibration set included all six spectral ranges (
n = 53 wavelengths) in the model with
R2 ranges from 0.83 for MEC to 0.96 for CP. Importantly, when only one spectral range was assessed in the model, the
R2 was relatively high for all chemical constituents, with relatively high slope, and low RMSEP (
Table 5). The best spectral domains for CP assessment were 1748–1764 nm and 1766–1794 nm (9 wavelengths and 15 wavelengths, respectively); for NDF assessment they were 2070–2088 nm and 2278–2286 nm (10 and 4 wavelengths, respectively), and for MEC assessment they were 2316–2330 nm and 2278–2286 nm (8 and 4 wavelengths, respectively).
A score plot of the samples from the PLS modeling demonstrated the excellent correlation between spectra and chemical constituents (CP, NDF and MEC), with an increase from left to right. The score plot for the CP model (
Figure 4) indicated that most of the spectral variations observed in our field study were indeed related to the protein, as predicted by the PLS model (and were not influenced by unknown parameters). Furthermore, for the best CP, NDF and MEC models, the first two LV (latent variables) components in the PLS model explained 100% of the X variance (spectra), and 99%, 86%, and 79% (for CP, NDF and MEC, respectively) of the Y variance (chemical components). This indicated that most of the spectral variation in the six selected spectral domains is related to the CP, NDF and MEC components that were modeled by PLS.
Figure 4.
A score plot for the partial least squares (PLS) CP model using the 1748–1764 nm spectral range.
Figure 4.
A score plot for the partial least squares (PLS) CP model using the 1748–1764 nm spectral range.
A comparison of the PLS and slope method results showed that the best PLS models for assessing CP, NDF and MEC are obtained with the spectral ranges of 1748–1764 nm, 2070–2088 and 2316–2330 nm, respectively, whereas using the slope method, only the 1748–1764 nm spectral range could be used to assess all parameters. Indeed, the PLS is a more accurate quantitative method, but requires chemical analysis to establish a calibration model; in contrast, the slope method provides a qualitative evaluation (three categories) with no need for chemical measurement of the samples.
Table 5.
Partial least squares (PLS) regression model results of the correlations between different spectral regions and percentages of CP, NDF and MEC.
Table 5.
Partial least squares (PLS) regression model results of the correlations between different spectral regions and percentages of CP, NDF and MEC.
Spectral Range (nm) (Number of Bands) | | CP Model’s Statistical Characteristics | CP Best Model | NDF Model’s Statistical Characteristics | NDF Best Model | MEC Model’s Statistical Characteristics | MEC Best Model |
---|
| Prediction | Validation | Prediction | Validation | Prediction | Validation |
---|
1748–1764 (n = 9) | Slope: | 0.934 | 0.933 | 2 | 0.856 | 0.850 | 4 | 0.789 | 0.792 | 5 |
Offset: | 0.6 | 0.606 | 8.508 | 8.874 | 0.429 | 0.423 |
RMSE: | 1.703 | 1.757 | 4.183 | 4.305 | 0.252 | 0.263 |
R2: | 0.934 | 0.931 | 0.856 | 0.848 | 0.789 | 0.769 |
1766–1794 (n = 15) | Slope: | 0.913 | 0.908 | 3 | 0.797 | 0.785 | 5 | 0.794 | 0.782 | 4 |
Offset: | 0.79 | 0.834 | 12.045 | 12.750 | 0.418 | 0.445 |
RMSE: | 1.954 | 2.02 | 4.977 | 5.098 | 0.248 | 0.259 |
R2: | 0.913 | 0.909 | 0.797 | 0.792 | 0.794 | 0.783 |
2070–2088 (n = 10) | Slope: | 0.897 | 0.899 | 5 | 0.865 | 0.859 | 2 | 0.716 | 0.700 | 7 |
Offset: | 0.932 | 0.912 | 8.019 | 8.345 | 0.576 | 0.623 |
RMSE: | 2.122 | 2.204 | 4.062 | 4.133 | 0.291 | 0.324 |
R2: | 0.897 | 0.891 | 0.865 | 0.861 | 0.716 | 0.654 |
2278–2286 (n = 4) | Slope: | 0.903 | 0.898 | 4 | 0.856 | 0.854 | 3 | 0.805 | 0.796 | 3 |
Offset: | 0.871 | 0.9 | 8.499 | 8.677 | 0.395 | 0.419 |
RMSE: | 2.051 | 2.096 | 4.181 | 4.217 | 0.241 | 0.251 |
R2: | 0.904 | 0.902 | 0.856 | 0.855 | 0.805 | 0.795 |
2316–2330 (n = 8) | Slope: | 0.892 | 0.889 | 6 | 0.782 | 0.761 | 7 | 0.812 | 0.802 | 2 |
Offset: | 0.978 | 1.004 | 12.911 | 14.157 | 0.380 | 0.406 |
RMSE: | 2.173 | 2.215 | 5.153 | 5.420 | 0.237 | 0.243 |
R2: | 0.891 | 0.888 | 0.782 | 0.760 | 0.812 | 0.804 |
2334–2344 (n = 6) | Slope: | 0.78 | 0.77 | 7 | 0.787 | 0.785 | 6 | 0.715 | 0.707 | 6 |
Offset: | 1.987 | 2.05 | 12.629 | 12.736 | 0.579 | 0.594 |
RMSE: | 3.09 | 3.17 | 5.097 | 5.219 | 0.292 | 0.302 |
R2: | 0.78 | 0.77 | 0.787 | 0.778 | 0.715 | 0.703 |
All ranges (n = 53) | Slope: | 0.967 | 0.956 | 1 | 0.893 | 0.882 | 1 | 0.832 | 0.823 | 1 |
Offset: | 0.32 | 0.39 | 6.310 | 6.850 | 0.340 | 0.368 |
RMSE: | 1.246 | 1.355 | 3.604 | 3.858 | 0.224 | 0.241 |
R2: | 0.964 | 0.958 | 0.893 | 0.878 | 0.833 | 0.809 |