Improving Soil Moisture Assessment of Turfgrass Systems Utilizing Field Radiometry

The need for water conservation continues to increase as global freshwater resources dwindle. Turfgrass mangers are adapting to these concerns by implementing new tools to reduce water consumption. Time-domain reflectometer (TDR) soil moisture sensors can decrease water usage when scheduling irrigation, but nonuniformity across unsampled locations creates irrigation inefficiencies. Remote sensing data have been used to estimate soil moisture stress in turfgrass systems through the normalized difference vegetation index (NDVI). However, numerous stressors other than moisture constraints impact NDVI values. The water band index (WBI) is an alternative index that uses narrowband, near-infrared light reflectance to estimate moisture limitations within the plant canopy. The green-to-red ratio index (GRI) is a vegetation index that has been proposed as a cheaper alternative to WBI as it can be measured using digital values of visible light instead of relying on more costly hyperspectral reflectance measurements. A replicated 2 × 3 factorial experimental design was used to repeatedly measure turf canopy reflectance and soil moisture over time as soils dried. Pots of ‘007’ creeping bentgrass (CBG) and ‘Latitude 36’ hybrid bermudagrass (HBG) were grown on three soil textures: United States Golf Association (USGA) 90:10 sand, loam, and clay. Reflectance data were collected hourly between 07:00 and 19:00 using a hyperspectral radiometer and volumetric water content (VWC) data were collected continuously using an embedded soil moisture sensor from soil saturation until complete turf necrosis by drought stress. The WBI had the strongest relationship to VWC (r = 0.62) compared to GRI (r = 0.56) and NDVI (r = 0.47). The WBI and GRI identified significant moisture stress approximately 28 h earlier than NDVI (p = 0.0010). Those metrics also predicted moisture stress prior to fifty percent visual estimation of wilt (p = 0.0317), with lead times of 12 h (WBI) and 9 h (GRI). By contrast, NDVI provided 2 h of prediction time. Nonlinear regression analysis showed that WBI and GRI can be useful for predicting moisture stress of CBG and HBG grown on three different soil textures in a controlled environment.


Introduction
Water conservation is a global concern for the agricultural sector. Commitments to increase crop production will be required to meet increasing food demand [1]. Reducing water consumption across intensively managed turfgrass systems provides potential to redirect water savings to accomplish this goal. Golf courses within the United States occupy 608,746 hectares of maintained turfgrass, of which 484,978 hectares are irrigated [2]. The Golf Course Superintendents Association of America reported in 2005 that golf courses used 2.934 × 10 9 m 3 of water. This volume was reduced by 21.8% in 2013 by adapting new management methods such as wetting agent applications, hand-watering, enhanced-efficiency irrigation systems, deficit irrigation, and more frequent use of irrigation audits and upgrades [3][4][5].
The turfgrass industry has attained the most substantial water conservation advancement by irrigating based on evapotranspiration (ET) and with the use of handheld soil moisture measurements. Scheduling irrigation based on ET allows turfgrass managers to deliver water applications over large acreage from previously recorded weather data [6,7]. However, soil heterogeneity and variations in microclimates on golf courses require different irrigation needs for optimal turfgrass growth [8]. Time-domain reflectometers (TDR) are hand-held devices that assists turfgrass managers in determining when water availability may be limiting by quantifying soil volumetric water content. These technologies are inadequate when considering large acreage areas because they are labor intensive and time-consuming to use, and often misidentify areas of localized moisture stress. In response, underground soil moisture sensors were developed to integrate the benefits of handheld TDRs and automated irrigation. Grabow et al. [9] found that water usage was reduced by 39% using soil moisture sensors compared to using ET-based irrigation. The use of soil moisture sensors, however, does not account for geospatial variability in edaphic and environmental conditions. To continue the progression of water conservation on golf courses, making soil moisture assessments over large acreage rapidly, non-destructively, and accurately is needed.
There is growing interest in using light reflectance over turfgrass canopies as a nondestructive estimate for assessing plant health, as influenced by various abiotic or biotic stressors (e.g., nutrient issues, diseases, insects, excess UV radiation, water stress, etc.) [10][11][12]. The most widely used reflectance-based assessment for plant performance or stress assessment of crops and maintained turfgrass is the normalized difference vegetation index (NDVI = (R NIR − R VR )/(R NIR + R VR ), where R is light reflectance from a given spectral region, e.g., NIR = near infrared and VR = visible red) [13,14]. The NDVI is strongly associated with turfgrass quality (TQ) and its components (e.g., shoot density, color, and canopy uniformity) [15][16][17][18][19]. Shade stress negatively impacts turfgrass canopies and can be quantified using NDVI [18,20,21]. Other research has shown significant correlation between volumetric water content (VWC) and NDVI (r = 0.22-0.68) [22,23]. Light absorption in visible wavelengths (400-700 nm) is closely associated with overall plant pigment concentrations (r = 0.97) [24,25], but isolating damage by a specific stressor is difficult. Hyperspectral sensors measure energy at narrow bandwidths across larger spectral regions, making them more sensitive to subtle differences in reflected energy compared to traditional multispectral sensors. The expanded energy range of hyperspectral sensors can assist with determining moisture stress without the confounding effects of pigment concentrations impacted by background stressor noises, such as nutrient limitations or pest invasion.
For example, total plant tissue water content influences near-infrared (NIR 700-1100 nm) and short-wave infrared (SWIR = 1100-3000 nm) regions differently. Once NIR light waves penetrate the tissue, they are scattered or transmitted by changes within the turgor pressure of spongy mesophyll cells [26]. Portions of the SWIR region are strongly absorbed by leaf water content once these lower energy wavelengths diffuse inside the leaf [27]. Both regions have key water absorption troughs with centers at 970, 1200, 1450, 1950, and 2250 nm [28]. These water absorption bands are used to estimate the relative water content (RWC) and tissue water content determined by dry mass [29,30]. Water absorption features occurring at 1200 nm and beyond possess higher absorption coefficients and penetrate less into the tissue before being absorbed, whereas the 970 nm has the lowest absorption coefficient [31]. Penuelas et al. [32] found a negative correlation (r ≥ −0.79) between the water band index (WBI = R 900 /R 970 ) and RWC with gerbera (Gerbera jasmonii) plants and saw significant changes with the WBI when RWC was ≥85%. Sims and Gamon [28] demonstrated how both equivalent water thickness (kg m −2 ; r 2 = 0.66) and WBI (r 2 = 0.59) had the highest correlations with water content of thin plant tissue (<5 mm) and a weaker relationship (r 2 ≤ 0.35) when considering total canopy water content.
The WBI has been used for water stress assessment for several plant systems such as rice, forest, chaparral, Sudangrass canopies, and turfgrass systems [33][34][35][36][37]. Within turfgrass systems, Dettman-Kruse et al. [38] identified the capability of predicting creeping bentgrass (CBG) water stress symptoms one day prior to the onset of drought symptoms using the spectral region where the WBI is located (750−1100 nm). McCall et al. [37] found that WBI was highly correlated with soil VWC (0.80 ≤ r ≤ 0.81) for CBG grown in a sand medium. Working in a clayey, urban soil system, Badzmierowski et al. [39] also found that during the most water limited stress periods, WBI was highly correlated to soil VWC (0.69 ≤ r ≤ 0.79). However, WBI requires hyperspectral sensing, which typically require expensive, specialized equipment. For this reason, there is interest in using the green-to-red ratio index (GRI = R 550 /R 670 ) to assess soil moisture status in turfgrass systems [37,39]. The GRI represents a potentially cost-effective approach for detecting moisture stress compared to WBI, as it can be measured using relatively inexpensive sensor types (e.g., multispectral or true-color visible light). Previous work has shown that GRI can have moderate to high correlation with VWC [37,39], though neither of these studies has investigated these relationships with multiple grass species across multiple soil textures. In response, the objectives of this greenhouse study were to assess two common grass species, creeping bentgrass and bermudagrass, grown in three different soil textures to (i) determine whether previously investigated vegetation indices can estimate moisture stress, and (ii) investigate vegetation indices as early predictors of moisture stress compared to visual symptom development. The two grasses represent the largest maintained acreage on golf course fairways, while the different soil textures account for the often highly variable underlying soil characteristics that are often found in these systems. Therefore, the results will have broad applicability for golf course superintendents looking to adopt alternative water management practices.

Sample Preparation
A dry-down greenhouse study was conducted at the Glade Road Research Facility in Blacksburg, VA. A 2 × 3 factorial design was used, in which two grass species, '007' CBG and 'Latitude-36' hybrid bermudagrass (HBG), were planted into three soil textures. Each of six replicates were repeated in time from June to September 2018. Soil textures consisted of a locally manufactured United States Golf Association (USGA) 90 sand: 10  The CBG plugs were harvested in March and June 2018 using a USGA-specified 10.2 cm cup cutter (R&R Products, Inc., Tuscon, AZ, USA). Roots were washed and pruned to 0.64 cm to remove pre-incorporated soil and then planted into the three soil types. The HBG plugs were established from sprigs starting in February 2018 into metal autoclave trays filled with 0.014 cm 3 of USGA 90:10 sand, top-dressed with sand. The HBG plugs were extracted after two months (i.e., in April 2018) using the same methods described for the CBG transfer. Due to their high light requirements, all HBG plugs were grown under lights supplying an average of 600 µmol m −2 s −1 during a 14 h. photoperiod, whereas natural sunlight was sufficient to meet lighting requirements for CBG [40][41][42][43].
All CBG and HBG treatments were transplanted into horticultural pots (Kord STD, Toronto, ON, Canada) with dimensions 15.3 cm × 14.6 cm and an 1835 cm 3 capacity; a minimum of two months of root growth was ensured before using experimental units during dry-down cycles. To ensure both CBG and HBG treatments with the same soil were treated similarly, soils were sieved (4.75 mm, Dual Manufacturing Co., Chicago, IL, USA), oven-dried at 100 • C for 48 h and filled to a specified mark on three pots to determine average bulk densities for each soil type (USGA 90:10 sand-1.03 g/cm 3 , loam-0.97 g/cm 3 , and clay-0.80 g/cm 3 ).
During the grow-in period, plugs were treated with SQM's 28-8-18 Bulldog fertilizer (Atlanta, GA, USA) at 2.5 kg N ha −1 for CBG and 12 kg N ha −1 for HBG on a weekly basis. To prevent disease, fluxapyroxad + pyraclostrobin (Lexicon, BASF) was applied at 0.24 and 0.49 kg ai ha −1 on a 28-day interval, and fluazinam (Secure, Syngenta) at 0.78 kg ai ha −1 on a 14-day interval. All inputs were applied using a CO 2 pressurized spraying system calibrated to deliver 813 L ha −1 through TeeJet TTI11004 flat spray nozzles (Glendale Heights, IL, USA). Plants were maintained at a height of 1.3 cm with Black & Decker garden shears (Towson, MD, USA), with trimming occurring twice per week.
Researching the relationship between turfgrass and soil moisture stress becomes difficult within field conditions due to several factors such as wind speed, direction, fluctuating climatic conditions, topography, etc. These multiple influencing variables necessitates the need for isolated dry-downs because establishing accurate plant and soil water relationships becomes increasingly cumbersome when designing experiments to account for these influencing factors. Thus, all pots were subjected to dry-down to establish plant and soil water relationships by withholding irrigation after bringing all experimental units to field capacity Prior to initiating a dry-down replication, plugs were irrigated with two ten-minute applications of overhead irrigation that applied approximately 1 L min −1 of water and spaced two hours apart to allow gravitational drainage of water and fully saturate all soils. Pots were allowed 12 h. of gravitational drainage after the final irrigation application before carefully inserting 5TM soil moisture sensors (Meter Environment, Pullman, WA, USA) according to manufacturer recommendations to ensure accurate quantification of soil moisture. Because the measurement volume of the sensors was asymmetrical relative to the sensor prongs, probes were oriented so that zone of influence encompassed the soil. Each sensor measured a volume of approximately 1640 cm 3 , therefore providing an accurate representation of the average water content for the entire pot.
A wind tunnel fan (Lasko Products, West Chester, PA, USA) was used to mimic low humidity and high ET associated with rapid plant moisture stress symptom development often observed under field conditions. Pots ran perpendicular to the fan with a row of three pots placed 0.75 m from the fan and a second row of three pots placed at a distance of 0.90 m. Fan speed averaged 11.3 km h −1 (at the closest row of pots) to 9.7 km h −1 (at the second row of pots). Experimental units were randomized for each of the six replications to ensure that each treatment was not in the same location (i.e., distance from fan) more than twice. The fan ran continuously each day from 07:00, representing approximate sunrise, until 19:00. at the last daily data collection. Data collection continued until all turfgrass canopies reached complete necrosis.

Soil Moisture Release Curves
The USGA 90:10 sand, loam, and clay soils were sieved to 4.75 mm, oven-dried, and loosely packed into soil cores with dimensions 2.5 cm × 1.2 cm. The cores were placed onto porous plates and saturated with water for 24 h. ensuring each soil was saturated from capillary rise. Three replications of each soil were placed into soil moisture pressure plate extractors (Soil moisture, Inc., Santa Barbara, CA, USA) at 33, 100, 300, 500, and 1500 kPa. Each pressure chamber had a drainage tube; when water ceased exiting this tube, soil cores were considered to have reached equilibrium with the applied chamber pressure. Once at equilibrium, soil sleeves were removed, weighed with a Mettler Toledo analytical scale (Columbus, OH, USA), oven dried at 105 • C for 48 h, then reweighed. Mass differences were used to calculate gravimetric water content [θg = (M wet − M dry )/M dry ] where M wet = the mass of wet soil and M dry = mass of oven dry soil. Gravimetric values were converted to VWC [θv = θg * β], where θv = VWC and β = the mass of dry solids divided by total core volume. The three replications of each soil were averaged together. Water retention parameters for the van Genuchten model [44] were fit by minimizing the residuals between modeled and average measured VWC for each pressure head.

Specific Soil Sensor Calibration
Volumetric water content from the 5TM soil sensors were calibrated for each soil based on manufacturer recommendations [45]. Four liters of each soil was sieved to 4.75 mm and oven-dried at 105 • C for 48 h in preparation for sensor calibration. The soil was added to a mixing container, and for each data point 100 to 200 mL of water was added. Once mixed, the soil was added to the same horticultural pot used in the dry-down study and packed to the same specific soil bulk density of the plugs when they were transplanted. Raw soil sensor data were collected before 103 cm 3 of the soil was excavated into soil cores, weighed, oven-dried for 48 h at 105 • C, and reweighed to determine gravimetric water content. The bulk densities of soil cores were used to convert gravimetric water content to VWC through the same calculations used for soil retention curves. Soil sensor raw data were plotted against calculated VWC, and the regression analysis that best fit the data were used to adjust the output data from the 5TM sensors. The USGA 90:10 sand was best explained with linear regression (r 2 = 0.99), while the loam and clay soils had the highest relationships with logarithmic regression analysis (r 2 ≥ 0.98). These equations allowed for data corrections to establish continuously adjusted VWC.

Root Length and Biomass
Roots from supplemental plugs of each treatment listed in Section 2.1 were washed, measured, and harvested through four replications. Roots were washed of soil by placing plugs onto a 4.75 mm soil sieve with water slowly running down the side to minimize root destruction. Length (cm) of the three longest roots was measured using a ring stand and all roots were harvested for biomass. Harvested roots were weighed and dried at 70 • C for three days for dry root weight plus any soil still attached to the roots. A 240 V Muffle Furnace (Thermo Scientific, Waltham, MA, USA) was used to ash root samples at 500 • C for 12 h. Total root biomass is reported as dry root weight minus washed weight.

Data Collection
Spectral and visual data were collected each hour from 07:00 to 19:00 each day during dry-down. Spectral reflectance data were collected first using a PSR-1100F spectral radiometer (Spectral Evolution, Lawrence, MA, USA) equipped with a 2.5 cm contact probe. The radiometer measures narrowband reflectance from 512 different wavelengths within a range of 320 to 1100 nm at a 1.4 nm sampling bandwidth. A BaSO 4 panel was used for white reference calibration by placing the contact probe flush with the pad immediately before each hourly data collection to ensure no spectral irregularities from background noise. Other data collected after light reflectance readings were as follows: visual turf quality (TQ), visual estimation of wilt percent (WP), and air temperature. Turf quality was assessed using a 1 to 9 scale (6 = minimum acceptable quality and 9 = highest acceptable quality) according to guidelines established by the National Turfgrass Evaluation Program [46]. Wilt ratings were based on previously established literature and guidelines that were modified to percent wilt coverage of turfgrass canopy. Guidelines for rating WP are described as when leaf firing, dormancy, or no plant recovery occurs [46,47]. Soil VWC was collected continuously using a CR300 datalogger (Campbell Scientific, Linden, NJ, USA) equipped with 5TM soil sensors. Data logger software (PC200W, Campbell Scientific, Linden, NJ, USA) was used to extract raw data to be transformed to calibrated VWC from established equations mentioned in Section 2.2.2.

Data Transformation and Analysis
Light reflectance data were extracted using Spectral Evolution DARWin SP v.1.2.5093 software and converted to comma-separated values file format using SED to CSV v.1.2.0.0 conversion software. Raw reflectance values were used to calculate NDVI, WBI, and GRI using previously described equations [13,32,48]. All data were analyzed using JMP Pro 13 (Cary, NC, USA). Vegetation index and WP data were modeled using a non-linear regression four parameter logistic (4PL) model for each treatment by replication over time: where y = index type, a = growth rate, IP = the inflection point, LA = lower asymptote, UA = upper asymptote, and HAI = hours after initiation of the dry-down cycle. Parameters of each fitted 4PL model were analyzed for variance by grass species, soil type, VI type and their interactions. Means were separated using the students' t-test (α = 0.05) where appropriate. Pearson correlation coefficients were used to examine the relationship between WBI, NDVI, GRI, TQ, WP, and VWC.
The point at which WP reached fifty percent (WP 50 ) data were determined using a custom inverse prediction model within the JMP software based on our 4PL regression models. We define the initial physiological response to drought stress as being when three consecutive measurements occurred more than one standard deviation below the running average of the 4PL upper asymptote. All measurements below this point are defined as drought-stressed turfgrasses. The >1 standard deviation criterion was chosen to provide a conservative estimate for the time at which the VI was undergoing a consistent change as opposed to random noise. For each sample, the difference between WP 50 and time to moisture stress response was quantified to provide the wilt prediction time (hr) for each index. Analysis of variance was used to compare wilt prediction times by grass species, soil type, and their interactions were examined.

Pearson Correlations
An average of 108 collection events were conducted per dry-down replication across six treatments, providing approximately 3900 data points for each parameter (i.e., WBI, NDVI, GRI, TQ, WP, and VWC). All reported coefficients ranged from −0.92 to 0.92, with all relationships being highly significant (p ≤ 0.0001; Table 1). The three indices (GRI, NDVI, and WBI) were all negatively correlated to WP, with WBI having the strongest relationship (r = −0.92). The WBI also had the strongest relationship to VWC (r = 0.62) compared to NDVI (r = 0.47) and GRI (r = 0.56), which follows the same trend observed by McCall et al. and Badzmierowski et al. [37,39]. Soil VWC had a negative relationship with WP (r = −0.66), suggesting that wilt observed was in response to soil moisture stress. All three indices had a moderate relationship to TQ (0.60 ≤ r ≤ 0.66) with WBI having the strongest correlation (r = 0.66) ( Table 1). Wilting percent was negatively correlated with TQ (r = −0.68; data not shown) and was primarily responsible for turfgrass decline.
All indices were strongly correlated to each other. The NDVI and GRI had the strongest correlation (r = 0.92), followed by WBI to GRI (r = 0.89), and the lowest correlation was NDVI to WBI (r = 0.83). Both NDVI values and visual assessments of turf quality are impacted by a variety of stressors (Table 1). The GRI is a measurement of visible light and, therefore, is likely impacted by other stressors as well. Reflectance spectra used for determining WBI are not associated with light absorption within the photosynthetically active region (PAR), rather a specific region of NIR light that is closely associated with water absorption. The WBI is, therefore, a safer index that is specifically associated with water availability in the plant canopy and less related to other stressors. A primary limitation of the WBI is that small spectral bandwidth at longer wavelengths are both more costly and subject to background noise of spectral measurements. However, GRI has a strong relationship with WBI and is therefore potentially useful for moisture stress identification with some additional ground validation measurements.   Figure 1 shows how the 4PL regression analysis, r 2 values, were used to explain the relationship of WBI to soil moisture stress of CBG planted into USGA 90:10 sand across six dry-down replications (r 2 = 0.74). Parameter estimates of upper and lower asymptotes along with calculated IP were derived from the 4PL equation. The IP (blue dot) is the halfway point between no moisture stress indicated as the upper asymptote (black dot) and complete necrosis of turfgrass expressed as the lower asymptote (purple dot). The WBI intersects with WP in close proximity to 50% wilt, which we deemed as significant wilting. The comprehension of the WBI compared to visual 50% wilt and their intersection point occurring in close proximity when considering the hours when WBI's IP occurs is essential for understanding the results discussed throughout this manuscript.

Analysis of Model Parameters and Inflection Points
Grass species and soil type were separated by WBI and GRI for 4PL regression analyses based on analysis of variance effects test of IP (Table 2, Figure 2a-d). NDVI values were excluded from the 4PL regression analyses because it had the least significance when comparing the three indices to VWC (Table 1). Bermudagrass IP was best explained utilizing WBI (r 2 = 0.77) compared to GRI (r 2 = 0.65) (Figure 2a,c). Interestingly, the same relationship was observed with bentgrass between both WBI and GRI (r 2 = 0.70). This is presumed to be caused by the physiological differences both species of turfgrasses display when coping with drought conditions. The WBI and GRI had the weakest relationship to USGA 90:10 sand (r 2 = 0.66). The GRI had the strongest relationship to the loam and clay treatments (r 2 = 0.68 for loam and r 2 = 0.78 for clay) (Figure 2b,d). ± Inflection points (IP) derived from four-parameter logistic regression analysis of the water band index (WBI = R 900 /R 970 ), normalized difference vegetation index (NDVI = (R 760 − R 670 )/(R 760 + R 670 )), and the greento-red ratio index (GRI = R 550 /R 670 ). ψ Calculated from the difference when 50% wilt was observed to the upper data cloud before wilt occurrence. The hours where a presumed physiological change occurred was determined by subtracting one standard deviation from the mean of the upper data cloud of WBI, NDVI, and GRI values using a four-parameter logistic regression analysis.
Analysis of variance for IP revealed highly significant differences for index type, grass species, and soil type (p < 0.0001) ( Table 2). While all these factors were independently significant, none of the interactions were significant (p ≥ 0.25). Therefore, index data were pooled over grass species and soil type, grass species were pooled over index type and soil type, and soil type was pooled over index type and grass species.
The IP of each index was used to determine the time from dry-down initiation to when significant wilt stress was detected due to the strong relationship between each index and observed wilt stress. To corroborate the validity of using the IP as the critical point for detecting significant moisture stress, the IP occurred 9 to 23 h after WP 50 based on 4PL for all treatments (data not shown). Confirming that the IP occurs after the WP 50 across all treatments allows us to use these critical points for temporal analysis knowing significant wilt has already occurred. WBI and GRI performed similarly with regard to hours to IP, whereas the IP associated with NDVI was approximately 27 h later (Table 3).  Table 3. Mean hours of inflection points (IP) for three vegetation indices based on light reflectance data collected from creeping bentgrass and hybrid bermudagrass grown in three soil textures and subjected to drought stress under greenhouse conditions in Blacksburg, VA across six dry-down cycles. Means followed by the same letter are not significantly different based on the students' t-test (α = 0.05). Inflection points of WBI were used to explain the significance observed between grass species and soil type sources (Table 2) because of the strong significance observed between WBI and VWC for our research and congruent literature [37,39]. The GRI and NDVI data were removed because of the strong evidence that supports WBI being most closely associated with VWC and WP (Table 1). Since there were no treatment interactions with that VI (Table 2), the WBI IP data were pooled across soil type to determine differences due to grass species, and then pooled by grass species to determine differences between soil types. On average, the IP occurred 35 h earlier in CBG than HBG (Figure 3a). Likewise, IP occurred approximately 42 h later in the USGA 90:10 sand treatment displayed compared to the finer textured loam and clay soils (Figure 3b). The inflection points of the water band index (WBI = R 900 /R 970 ) and the green-to-red ratio index (GRI = R 550 /R 670 ) were used to identify hours from the initiation of the dry-down cycle until IP. Inflection points were modeled using a non-linear four-parameter logistic regression of spectral reflectance data and pooled across (a) WBI, GRI, and soil types (b) WBI, GRI and grass species.

Index Type IP (Hours)
Root length and biomass were collected to supplement our research findings. Root length and biomass were comparable among soil types but differed by grass species. Before any moisture stress, HBG had 19.5 cm roots and CBG had 13 cm root length. The HBG root biomass (1.79 g) was significantly greater (p < 0.0001) than CBG (0.25 g). Root length and biomass were taken from experimental units not subjected to drought stress but grown in same greenhouse conditions as ones used for dry-down cycles. Figure 4 is a representation of the non-linear 4PL regression model of the WBI for one dry-down cycle with CBG grown in the loam soil. Once three or more WBI data points occurred below the upper data cloud (UDC), the floor (FUDC) and ceiling (CUDC) of the upper data cloud were established. The distribution was analyzed, and one standard deviation (σ) was subtracted from the mean (µ) of all data within the UDC. This value, deemed as the UDC value and indicated by the blue circle in Figure 4, represents the presumed point of initial physiological change in response to drought stress based on the 4PL model's inverse relationship to WP. The hour of wilt prediction (WP 50 ) represents the difference between visually estimated 50% wilt and µUDC − σUDC. There was an interaction between grass species and soil type (p ≤ 0.0191; Table 2). There was no interaction between index type and any other main effect (p ≥ 0.3781; Table 2). All main effects (index, grass species, and soil texture) were significant (p ≤ 0.0093). Indices are presented with all data pooled over grass species and soil type, and individual treatments of grass and soil type are presented separately. Wilt prediction time provided by the WBI and GRI datasets were similar, with respective WP 50 values of 12 and 9 h ( Figure 5). The NDVI predicted significant moisture stress by 2 h, but was not as effective as either WBI or GRI (p = 0.0317). Similar to the correlation analysis (Table 1), these data suggest that NDVI is not as closely associated with drought stress as WBI or GRI. As with the IP results, subsequent data for WP 50 is explained with WBI. While GRI was statistically the same compared to WBI in regard to wilt prediction capabilities, GRI has the potential to be influenced from other stressors because of its relationship with visible light. For these reasons, the interaction between grass species and soil type is reported using WBI and excluding GRI and NDVI.

Inverse Prediction of Drought Stress
The WP 50 of individual treatments of grass species grown on different soil textures were significant (p = <0.0001) ( Figure 6). Bermudagrass grown on sand (BMS) was the only treatment where we were unable to predict wilt, with WP 50 occurring 10 h after BMS reached 50% visible symptom expression. The inverse prediction of drought stress using 4PL parameters were positive for all other treatments. Wilt prediction of all treatments, excluding BMS, compared favorably with each other and had WP 50 of approximately 11 to 26 h.

Discussion
Our results demonstrated how three vegetation indices-the water band index (WBI), green-to-red index (GRI), and normalized vegetation difference index (NDVI)-can be used to estimate and predict moisture stress across a variety of turfgrass canopies prior to substantial visible wilt symptom expression. Overall, the WBI provided the best estimation of moisture stress based on relationship to VWC, wilt percentage, and turf quality. The WBI uses spectral reflectance energy bands outside of electromagnetic regions influenced by endogenous pigment concentrations that allow the index to effectively discern moisture stress from other stressors. The GRI performed similarly to WBI in visual turfgrass estimations, time to IP, and prediction time (9 and 12 h, respectively) to significant moisture stress. This finding is congruent with previous studies that documented these indices are useful to estimate TQ under the conditions tested [17,19,37,39]. The NDVI, which is used most commonly in agricultural settings, was the least reliable at estimating and predicting drought stress. The NDVI hours to IP was approximately 27 h later than WBI and GRI and only provided 2 h of prediction time before extensive drought stress. These results are congruent with the variable and often poor soil moisture stress detection observed in previous studies of NDVI [22,37], compared with the consistent ability of WBI to detect leaf water content independent of degrading pigment concentrations [13,24,28,32,49,50].
In this study GRI provided similar results to WBI in terms of predicting wilt. Therefore, GRI may represent a practical and cost-effective method for estimating moisture stress in turfgrasses suitable low cost. However, since the calculation for GRI uses wavelengths within the visible light spectrum, it is not known if the greenhouse-based results will translate to field conditions due to potential variability within field conditions. For instance, typical reflectance measurements of plant canopies within the visible range are limited because of high light absorption within the photosynthetically active region. Subtle changes in ambient light conditions during remote data collection can significantly alter GRI and other visible-light indices. The controlled conditions here helped avoid other (non-moisture related) stressors such as brown patch (Rhizoctonia solani) and compaction in bermudagrass that can cause shifts from normal spectral reflectance [51,52]. Furthermore, other moisture related factors were accounted for by means of small plot, dry-down research by eliminating factors such as localized areas of varying soil texture, elevation, topography, fluctuating climatic conditions, etc. All of these factors can influence the rate of evapotranspiration, amount of lateral soil water movement, and the amount of water infiltration into the soil profile [53,54]. While extraneous factors were reduced in this study as much as possible, it is highly probable that their influence will impact practical applications as these methods are expanded to larger, field scale research.
When assessing the data at the main factor of grass species, there were differences observed in the ability of the WBI and GRI to detect moisture stress. Bermudagrass is a warm-season plant due to its isolation of the Calvin cycle around high concentrations of carbon dioxide (CO 2 ). In comparison, CBG CO 2 fixation reactions are not isolated and stomates must stay open for longer periods of time to maintain adequate concentrations. During drought stress, to maintain efficient photosynthesis and cool the plant, stomata remain open and results in higher water loss [55,56]. Our results found that the development of wilt symptoms was 35 h earlier in CBG than HBG. This is likely attributed to differing photosynthetic pathways of the grasses tested. Before any irrigation was restricted from treatments, HBG had a significantly greater root system compared to CBG and likely aided in accessing more water deeper within the soil profile to delay moisture stress symptoms. The HBG uses water more efficiently during drought conditions resulting in delaying of wilt conditions by closing their stomates and eliminating the possibility of photorespiration, while CBG rapidly progresses to wilting when soil water content becomes the limiting factor [57]. Furthermore, expanding upon these data could establish moisture stress thresholds and objectively evaluate the drought tolerance of grass species and cultivars. Knowing this information would allow turfgrass managers to make the most optimal grass selection based on their facility's geographical location. This is especially true for transition zone locations, since many facilities, such as the Mid-Atlantic United States, grow both warm-season and cool-season grasses simultaneously. Using these data may be useful for further reducing total water usage through proper turfgrass selection and precise irrigation by grass species and cultivar. Soil texture also influenced the ability of WBI and GRI to detect moisture stress and the time to the inflection point (TTIP). The weakest relationship for both WBI and GRI was with the coarse particle size, USGA 90:10 sand (r 2 = 0.66). The WBI and GRI were most correlated to the loam soil (r 2 = 0.77 and 0.78, respectively). As the particle size decreased, the correlation to WBI and GRI decreased (r 2 = 0.68 and 0.70). It appears that the nonlinear regression analyses seemed best correlated as the particle size distribution became more balanced such as the loam soil. Further work is needed to assess different soil textures in order to confirm this finding. The time to the IP was 42 h later for the USGA 90:10 sand compared to the other two soil treatments. The USGA 90:10 sand likely had increased plant available water and allowed the CBG and HBG treatments to have greater accessibility at lower soil VWC. When comparing soils with similar upper and lower VWC field capacity limits (sand: 5-15%, loam: 10-25%, clay: 25-40%), we know that the permanent wilting point occurs at higher VWC as soil particle size decreases [53,54]. The custom blended USGA 90:10 sand with organic matter content (1.2%) and sand particles (0.15-0.50 mm) provided an optimal environment for plant water accessibility and water retention. While the L and C had larger organic matter contents, 2.1 and 1.7%, respectively, the larger percentages of silt (46, 20%), and clay (18, 41%) for each soil is likely the reason that an earlier expression of wilt symptoms occurred. As these heavier soils dried, the plant accessible water approached the wilting point sooner in regard to time compared to the sand treatments. Furthermore, the USGA 90:10 had a greater root biomass compared to the L and C within each grass species. The more robust root system within the USGA 90:10 sand treatments for each grass species provided greater access to water at lower VWC deeper within the soil profiles compared to the other soils.
There was a significant interaction between grass species and soil texture when spectral data were pooled (p = 0.0007). The hybrid bermudagrass (HBM) grown on sand was the only treatment that was not able to predict wilt ( Figure 6). The difference between this and all other treatments are best explained through the combination of bermudagrass' enhanced physiological capabilities to avoid drought stress and the USGA 90:10 sand having greater plant available water due to the organic matter and custom blend of sand particle size compared to other tested soils. These factors allowed the HBM to avoid drought stress and access most of the VWC until reaching the permanent wilting point which VWC varies by soil type at this point (sand: 5%, loam: 10%, and clay: 20%) [56,57]. Once this point was reached for the HBM on sand treatment, the accelerated wilting was so rapid that light reflectance lagged behind moisture stress.
The ability to use these data for wilt prediction would be a considerable step towards utilizing light reflectance as a tool for rapid moisture stress prediction across larger turfgrass surfaces, such as golf course fairways. The strong relationship between WBI and GRI provide flexibility for choosing sensors for future drought studies collected across data collection platforms. A sensor capable of collecting reflectance values within the nearinfrared range cost significantly more compared to a visible-light camera needed to estimate GRI. The readily available data acquisition of GRI from visible light implies a more direct practical use. However, high absorbance of photosynthetically active light may limit GRI effectiveness under variable solar conditions. Previous reports of WBI detecting moisture stress independent of other stressors using narrowband reflectance of more stable nearinfrared light is encouraging when paired with our results, though current technology and associated costs may limit immediate application.
As the spectral reflectance relationships are better defined with certain grasses and soils, we expect increased precision irrigation management. Using sensors on unmanned aerial vehicles can help improve the speed of data collection across large acreage for expanding the application of these indices. It is important to develop these indices in the field rapidly and non-destructively to identify and predict soil moisture stress. Results from these data may be used to help develop index thresholds that imply early onset wilt occurrence for future research. Applying index thresholds to remotely sensed data might allow for computer automation of wilt detection and linkage to automated irrigation systems. Precise, accurate, and well-timed irrigation across large turfgrass systems will improve efficiency by applying water only when and where it is needed.

Conclusions
Light reflectance measurement of turfgrass canopies is a promising tool to create vegetation indices that can assist in identifying moisture stress without the confounding effects from other variables. Our research contributes to the overall understanding of estimating moisture stress across turfgrass systems using spectral reflectance. We demonstrated how three vegetation indices can be used to estimate and predict moisture stress across a variety of turfgrass canopies prior to visible wilt symptom expression. Two indices, WBI and GRI, were more strongly correlated to visual estimations of wilt and to VWC than the more commonly researched NDVI. Regression analysis of WBI and GRI data shows temporal index changes in response to turfgrasses approaching permanent wilting point. Overall, the WBI is most consistently associated with limited water availability, but the GRI was almost as effective in all metrics compared, with both indices providing an approximate 12 h lead time prior to extensive drought stress. The NDVI, which is used most commonly in agricultural settings, was the least reliable at estimating drought stress in our study. The two grasses used in this study represent the most common species grown on golf course fairways worldwide. Additionally, the soil textures examined represent a wide range of underlying root zones that these grasses may be grown on. The grasses and soils evaluated in this study provide a wide array of real-world turfgrass canopy scenarios. Our results should serve as a fundamental approach for extrapolating turfgrass and soil water relationships utilizing these same indices within intensively managed turfgrass systems. Combining our research and future results with the adoption of unmanned aerial vehicles within the industry, large scale remote sensing will allow turfgrass managers to rapidly assess water management decisions across their entire properties.
Author Contributions: All authors contributed to the work presented in this paper. The manuscript was prepared by T.L.R. and all authors provided re-vision suggestions throughout the writing process. All authors also discussed the results and conclusions throughout the entire production of this manuscript. T.L.R., D.S.M., S.D.A. and R.D.S. all provided an appreciable amount of assistance with data curation. T.L.R., R.D.S., E.H.E. and D.S.M. were heavily involved with the experimental methodology formulation and enhancements. M.J.B. was the main contributor for writing, review, and editing with all other authors contributing as well. All authors have read and agreed to the published version of the manuscript.

Acknowledgments:
The authors would like to thank the Virginia Turfgrass Foundation and Virginia Agricultural Council for their financial support of this research. The authors would like to thank the support of Jamie Hodnett and Valerie Breslow for their contributions in greenhouse maintenance work and data collection. Lastly, the authors would like to provide consideration for the assistance that Brandon Lester provided in developing the code for the data logger to collect and store data properly to allow this project to come to fruition.

Conflicts of Interest:
The authors declare no conflict of interest.