Figure 1 shows the reflectance spectra collected over turfgrass at three different levels of water stress, specifically turfgrass at 16 days without watering, the intermediate situation at 7 days and at the end of the trial with the saturated cores (0 days without water), which serves as control. The differences across the curves are well evident. The major difference is the increase of reflectance at all wavelengths at 16 days without watering, where LRWC was at about 18% (Fig. 2), with respect to the other two spectral reflectance curves. It is so evident from the three different curves that in the Near-infrared (NIR 750–1,300 nm) and Short-wavelength infrared (SWIR 1,300–2,500 nm) four major absorption troughs are present. These strong reflectance troughs, located approximately in the NIR at 970 and 1,175, in the SWIR at 1,450 and 1,950 nm, are due to the absorption by water^{11}. The troughs around 1,450 and 1,950 nm are less accentuated in the turf with high degree of desiccation (16 days without watering). Also González-Fernández et al.^{47} recommend calculating the band area for 1,450 nm and for 1,950 nm because of its link to equivalent water thickness, thus to estimate vine water status. Rallo et al.^{48} observed typical spectral responses in the SWIR region, where at leaf scale, absorbance bands near 1,450 and 1,900 nm could be related to the leaf water content of an olive grove.

**Figure 2**

Decline in volumetric soil water content (SWC) (%) and leaf relative water content (LRWC) (%) after watering ceased. Each point is the mean of six replications. Bars indicate one standard deviation error.

However, in the regions of 1,350–1,480, 1,800–2,000 and 2,350–2,500 nm measurements of spectral reflectance of crop leaves are not possible in nature, also with fully sun-light conditions, because of the strong atmospheric absorption of light due to water vapor^{14,32,49} and are generally not exploited for landscape level studies. Consequently, to correctly measure these regions of wavelengths, a portable spectroradiometer system with an artificial light source must be chosen^{49}. In fact, in our experiment an artificial light source was used, thus 1,430 and 1,950 can be considered key wavelengths for the measurements under artificial light source.

In the NIR spectral region there is a more commonly exploited troughs around 970 nm and in the region of 1,150–1,260, which are the most studied spectral ranges for estimation of vegetation water content^{14}. It was interesting to note that the troughs of reflectance spectra underwent a gradual reduction in depth as the turfgrass desiccation increased, up to almost disappear in most cases, as showed in the 16 days without water curve. Some of the wavelengths associated with these troughs are, in fact, exploited by the spectral indices used in this study (see Table 1).

Figure 2 shows SWC and LRWC values, averaged over each set of six replicates with one standard deviation error bars, plotted with respect to the number of days without watering. Volumetric SWC declined as the days without watering increased. Starting from a value of 43.78% for the control cores with 0 days without watering, it decreased reaching a much lower value of 5.19% after two weeks without watering. Similarly, also LRWC declined as the number of days without watering increased. LRWC rate of decline was smaller than SWC as the days without watering were 4 or less (LRWC equal to 98.7%, 94.3% and 94.2% for 0, 1 and 4 days without watering, respectively). Then LRWC steeply decreased as the number of days without watering increased above 4. Observing the two parameters it is interesting to note that, with the exception of data collected in cores at 4 days without water, the trend of SWC and LRWC is similar (Fig. 2). In fact, from 1 to 4 days without water, turfgrass leaves try to preserve more water even if the soil water content decreases.

Figure 3 plots bar graphs of the selected indices in Table 1, where the indices are averaged over each set of six replicates of turfgrass at same water stress condition. One standard deviation error bars are also plotted. As is evident, all selected indices correlate with water stress level (Fig. 3).

**Figure 3**

Bar graphs of spectral indices averaged over each set of six replicates at same water stress condition, with one standard deviation error bar. (**a**) NDVI, (**b**) WI, (**c**) NDWI_{2130}, (**d**) NDWI_{1240}, (**e**) WI/NDVI.

A quantitative analysis of these correlations, and specifically with respect to SWC, LRWC and SM, is reported in Table 2, which reports the Pearson product-moment correlation coefficients evaluated among the various parameters and indexes studied in this work.

**Table 2 Pearson product-moment correlation coefficients (r) among volumetric soil water content (%) (SWC) measured using a time domain reflectometry (TDR); leaf relative water content (%) (LRWC); soil moisture (%) (SM) and vegetation indices selected for the study.**

### Volumetric soil water content (SWC)

As expected, SWC was found to be highly correlated with SM (r = 0.98, *p* < 0.001), and among the calculated indices the strongest relationship was with WI/NDVI studied by Peñuelas et al.^{29} (r = 0.94). For this relationship the exponential function proved to be the most suitable mathematical representation of the correlation. Thus, this index presented an exponential decrease when SWC values progressively increased (Fig. 4b). Relating SWC with WI, as also demonstrated by McCall et al.^{3} the correlation coefficient is still high (r = 0.89, *p* < 0.001) (Table 2; Fig. 4a). As the SWC increased to more stressed levels, also the WI and SM increased (Table 2). The range of values is between 4.75% and 47.05% of SWC corresponding respectively to a minimum WI value of 0.94 to a maximum of 1.05 (Fig. 3a). The relationships with NDWI_{1240} (r = 0.87) and NDWI_{2130} (r = 0.88) also show high coefficients (Fig. 4c,d).

**Figure 4**

Relationship between volumetric soil water content (SWC) and (**a**) water index (R_{900}/R_{970}); (**b**) ratio WI normalized difference vegetation index (WI/NDVI); (**c**) normalized difference water index (NDWI_{1240}); (**d**) normalized difference water index (NDWI_{2130}) in Bermudagrass cores. Values represented the 6 replications.

### Leaf relative water content (LRWC)

For correlations (r) among LRWC and the indices selected for the evaluation of water content, significantly high r values were found for NDVI (r = 0.96), WI (900/970) (r = 0.98, *p* < 0.001), WI/NDVI (r = 0.95, *p* < 0.001), NDWI_{1240} (r = 0.94, *p* < 0.001) and NDWI_{2130} (r = 0.95, *p* < 0.001) (Table 2).

As also studied by Jiang et al.^{26} and Johnsen et al.^{22}, NDVI presents a significant correlation coefficient with LRWC (r = 0.96), indicating that factors beyond water availability can impact in turfgrass quality.

As demonstrated by Peñuelas and Inoue^{32}, when evaluating reflectance indices associated with water and pigment contents of peanut and wheat leaves, WI closely track changes in LRWC, but it is frequently influenced by architectural canopy parameters. Similar results were obtained by Steidle Neto et al.^{50}, when assessing water and chlorophyll contents from spectral indices in sunflower plants under drought conditions. WI is effective to represent changes also in sunflower water content.

To minimize these effects, thus maximizing the effect of vegetation water content, Peñuelas and Inoue^{32} also studied the ratio of WI with NDVI, as NDVI is an index that follows color changes in the drying leaves. Moreover, also NDWI_{1240} has registered high r value (r = 0.94, *p* < 0.001) (Fig. 4c), Gao^{30} and Serrano et al.^{37} demonstrated that NDWI together with WI showed high sensitivity to changes in canopy LRWC, better than those formulated using SWIR bands. In our research, also NDWI calculated using R_{2130}, as suggested by Chen et al.^{31}, showed a high correlation coefficient with LRWC (r = 0.95, *p* < 0.001) (Fig. 4d). In fact, in the SWIR, the region of 2,130–2,200 nm is one of the most suitable for measuring optical remote sensing of vegetation water content, together with the NIR wavelengths of 900, 970, and 1,150–1,260 nm band, as we can notice also in the high correlation coefficient between LRWC and NDWI_{1240} (r = 0.94, *p* < 0.001) (Table 2). Regression equations between LRWC (%) and (a) water index (R_{900}/R_{970}); (b) ratio WI normalized difference vegetation index (WI/NDVI); (c) normalized difference water index (NDWI_{1240}) and Normalized difference water index (NDWI_{2130}) in Bermudagrass cores are reported in Fig. 5.

**Figure 5**

Relationship between leaf relative water content (LRWC) and (**a**) water index (R_{900}/R_{970}); (**b**) ratio WI normalized difference vegetation index (WI/NDVI); (**c**) normalized difference water index (NDWI_{1240}); (**d**) normalized difference water index (NDWI_{2130}) in Bermudagrass cores. Values represented the 6 replications.

Regarding the relationship between LRWC and WI (R_{900}/R_{970}) the determination coefficient was r = 0.98. Relative water content increases linearly with increasing WI with valued ranging from 0.94 to 1.05, corresponding to LRWC values ranging from 12.18% to 100% (Fig. 5a). For the relationship between LRWC and WI/NDVI the exponential function proved to be the most suitable mathematical representation of the correlation (r = 0.95). Thus, this index presented an exponential decrease when LRWC values progressively increase (Fig. 5b). In Fig. 5c, d linear regressions between LRWC and NDWI centered at different wavelengths are reported, where in both cases NDWI values increases linearly with increasing relative water content (NDWI_{1240} r = 0.94; NDWI_{2130} r = 0.95). Thus, although Serrano et al.^{37} showed that NIR-based NDWI was more sensitive to changes in canopy LRWC than those SWIR-based, in the present research the results showed that in the relationships studied between LRWC and the two NDWI, in the NIR and in the SWIR regions, the determination coefficients are significantly similar (Fig. 5c, d). Thus, they can both detect turfgrass relative water content.

### Soil moisture (SM)

Regarding the relationship between soil moisture and the vegetation indices studied in the present research, the highest correlation coefficient was found with WI/NDVI (r = 0.92). In this case the exponential function proved to be the most suitable mathematical representation of the correlation. Thus, this index presented an exponential decrease when SM values progressively increase (r = 0.92) (Fig. 6).

**Figure 6**

Relationship between soil moisture (SM) and ratio WI normalized difference vegetation index (WI/NDVI) in Bermudagrass cores. Values represented the 6 replications.

### Potential of multivariate analysis

Figure 7 shows the outcomes of the PCA, specifically as regards the first three PCs. Figure 7a displays spectral plots of the PCA loadings for each of the first three PCs (shown in different colors), plotted with respect to the wavelengths. The closer the loading at a given wavelength is to ± 1, the stronger is that specific wavelength contribution to the given PC. In order to have a better visualization of the impact of each wavelength over the PCs, Fig. 7b, c display the spectral ranges that contributed most to each of the three PCs, plotted as horizontal bars and obtained by thresholding (with two different thresholds) the corresponding loadings. To make the interpretation easier, an example of turfgrass spectral reflectance (and, specifically, the average reflectance for the 7 days without watering condition) is plotted on the same graphs. The figures clearly show that the ranges contributing most to the first PC (in red) are most of the short-wave infrared, with the water absorption bands around 1,450 and 1900 nm providing the strongest contributions. The near infrared band is the range contributing most to the second PC (in green), where the visible range and part of the short-wave infrared provide the strongest contributions to the third PC (in blue). As is evident, the wavelengths used by the spectral indices employed in this study and highlighted with orange arrows in Fig. 7b, c, are included within these spectral ranges. However, there are several other wavelengths that, according to PCA, are strongly “informative” and are worth being exploited for water stress monitoring. For completeness, Fig. 7d plots a bar graph of the cumulative percentage of explained variance by the first three PCs. The first PC by itself explains about 89% of spectra variability and the three PCs together explain more than 99% of it.

**Figure 7**

Outcomes of principal component analysis (PCA) for the first three principal components (PCs). (**a**) Spectral plots of PCA loadings. (**b**,**c**) Spectral ranges more involved in each of the first three PCs, obtained by thresholding the corresponding loadings—threshold equal to ± 0.1 for (**b**) and equal to ± 0.06 for (**c**). An example of turfgrass reflectance curve is superimposed in gray color to make result interpretation easier. The orange arrows indicate the wavelengths employed by the spectral indices used in this study. (**d**) Bar graph of the cumulative explained variance percentage.

Figure 8 plots a scatterplot of the turfgrass reflectance spectra in the three-dimensional space spanned by the first three PCs. Although the reflectance data place themselves in the space following a rather complex data structure, the data are mostly arranged in a sort of ordered fashion with respect to water stress, i.e. data related to similar stress conditions are closer to each other within the data structure, whereas data related to different stress conditions are placed apart to each other.

**Figure 8**

Scatterplot of the spectra of turfgrass at different water stress levels over the three first PCs.

Application of the *k-means* multivariate clustering method, applied with *N* = 3 clusters to the data subsets related to the aforementioned ‘strong’ (16 days without water), ‘medium’ (7 days without water), and ‘absent’ (just watered ) water stress conditions resulted in a correct identification of the three different groups, as shown in Fig. 9a where a scatterplot of the three data subset is shown and the different colors denote the different clusters obtained with *k*-means. By comparing Fig. 9a with Fig. 8, it is evident that the clusters denoted in Fig. 9a with *c*_{A}, *c*_{B}, and *c*_{C}, corresponds respectively to the ‘strong’, ‘absent’, and ‘medium’ stress conditions. However, although having separated correctly the three subsets of data, application of *k*-means by itself does not tell us anything about the water stress condition of each subset of data. Evaluation of the average of both WI (900/970) and WI/NDVI indices (which have revealed to be, among the indices studied here, the more effective indicators of water stress) over the three clusters identified with *k*-means allowed the clusters to be sorted according to a descending order of water stress level (i.e., according to an ascending order of WI or descending order of WI/NDVI, respectively), as shown in Fig. 9b,c.

**Figure 9**

(**a**) Scatterplot of the result of the *k*-means clustering method applied with *N* = 3 clusters. The method succeeded in correctly identifying the three groups of turfgrass. (**b**) Bar graph of WI index averaged over each identified cluster (with one standard deviation error bar). An ascending order of WI indicates a descending order of water stress level. (**c**) Bar graph of WI/NDV index averaged over each identified cluster (with one standard deviation error bar). A descending order of WI/NDVI indicates a descending order of water stress level.

Although performed on a data set limited in terms of sample size, this analysis has shown that multivariate methods have great potential for water stress monitoring in turfgrass and surely deserve further investigations.

Source: Ecology - nature.com