Published February 19, 2016
RESEARCH
Predicting Soybean Relative Maturity and Seed Yield Using Canopy Reflectance Brent S. Christenson, William T. Schapaugh, Jr.,* Nan An, Kevin P. Price, Vara Prasad, and Allan K. Fritz
ABSTRACT Optimized phenotyping, the observable characteristics attributed to the interaction between genotype and the environment, using canopy reflectance measurements may increase the efficiency of cultivar development. The objectives of this study were to: (i) assess canopy reflectance as a tool for predicting soybean maturity and seed yield; (ii) identify specific development stages that contribute to maturity and yield estimation; and (iii) test the stability and utility of maturity and yield estimation models across environments. Canopy reflectance, maturity, and seed yield were measured on 20 maturity group (MG) 3 and 20 MG 4 soybean cultivars released from 1923 to 2010. Measurements were conducted on six irrigated and water-stressed environments in 2011 and 2012. Cultivar, environment, and cultivar by environment sources of variation were all significant for maturity, yield, and reflectance. Maturity estimation models were created using the visible, red edge, and near-infrared spectrum as well as normalized difference vegetation index (NDVI) and water index ratios. Yield estimation models using the red edge, near-infrared, and visible NDVI indices explained much of the variation in yield among genotypes. No significant trends were found for canopy reflectance data collected at specific development stages or in different water regimes contributing to more accurate yield estimation; however, later development stages (R5-R6) were more accurate for maturity estimation due to spectral data identifying senescing vegetation. Performance of canopy reflectance models for maturity and yield accounted for a significant portion of variability among genotypes for maturity in some environments and for seed yield in most environments.
crop science, vol. 56, march– april 2016
B.S. Christenson, W.T. Schapaugh Jr., N. An, V. Prasad, and A.K. Fritz, Dep. of Agronomy, Kansas State Univ., Manhattan, KS 66506; K.P. Price, AgPixel, LLC, 5530 West Parkway, Suite 300, Johnston, IA 50131. Received 10 Apr. 2015. Accepted 15 Oct. 2015. *Corresponding author (
[email protected]).
S
oybean has complex genetics, and therefore classic breeding approaches are still used to assess yield potential and improvement. Crosses are made between parents, and progeny are selected based on attributes favorable to specific target environments. Superior genotypes are moved through the process, with seed yield used as a main selection criterion (Loss and Siddique, 1994). Mechanized harvesters measure seed yield efficiently, but this remains an expensive, laborious, and timeconsuming process. Multiple location-based evaluations of progeny will always be necessary to fully evaluate genotypic yield potentials with mechanical harvesters, but if specific regions of the electromagnetic spectrum can consistently account for a large portion of the genetic variation in seed yield, the technology may serve as a useful indirect selection criterion for breeding programs. Most location-based research conducted on yield estimation models using canopy reflectance (R) has focused on one-, two-, or three-band indices, with results that can be highly variable and inconsistent (Babar et al., 2006a, 2006b). Some of these indices have been useful in estimating yield via other plant characteristics, such as chlorosis (Adams et al., 1999), green cover (Dusek et al., 1985; Daughtry et al., 2000), chlorophyll (Datt, 1999; Daughtry et al., 2000), and photosynthetically active tissue (biomass) (Wiegand et al., 1991). These indices have had varying degrees of success, but none have fully captured the underlying physiological and environmental factors behind consistent phenotypic yield estimation (Hatfield and Prueger, 2010). Studies that propose the use of full-spectrum instruments and models for various parameters, such as nitrogen status, yield, and plant health, have been published in recent years for wheat Published in Crop Sci. 56:625–643 (2016). doi: 10.2135/cropsci2015.04.0237 © Crop Science Society of America | 5585 Guilford Rd., Madison, WI 53711 USA This is an open access article distributed under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). www.crops.org 625
(Hansen et al., 2002; Pimstein et al., 2007), corn (Hong et al., 2001; Weber et al., 2012), rice (Lin et al., 2012), and cotton (Zhao et al., 2007) grown in both optimal and water-stressed environments. Additional studies have characterized and explored how to incorporate these new models into high-efficiency platforms that would be able to evaluate thousands of genotypes with high resolution data (Walter et al., 2012; White et al., 2012). Most of indices correlate plant physiological parameters such as pigment status to grain yield. Indices such as the simple ratio (SR), first used by Jordan (1969) and Rouse et al. (1973) and defined as [R nir/R red], captured the ratio of near-infrared (NIR) reflectance to red (RED) reflectance or other visible regions of the spectrum (VIS). SR has been shown to correlate well with biomass, leaf area index (LAI), fractional photosynthetically active radiation (FPAR), ground cover, and wheat and soybean yield (Hatfield, 1983; Wiegand et al., 1991; Ball and Konzak, 1993; Price and Bausch, 1995; Ma et al., 2001; Royo et al., 2003; Hatfield and Prueger, 2010). The NDVI, defined as [(R nir – R red)/ (R nir + R red)], was derived by Deering (1978) and Tucker (1979) to estimate green biomass and intercept photosynthetically active radiation (PAR); the NDVI has been used to predict yield and other plant functions with many crops using hyperspectral and satellite imagery (Wiegand et al., 1991; Peñuelas et al., 1997; Ma et al., 2001; Shanahan et al., 2001; Royo et al., 2003; Prasad et al., 2007a, 2007b; Marti et al., 2007). Shanahan et al. (2001) successfully predicted yields in corn in Nebraska using the Green NDVI (GNDVI). Researchers found that normalizing the green and NIR relationship was highly correlated with grain yield, explaining 70 to 92% of yield variability at mid-grain fill (Shanahan et al., 2001). Ma et al. (2001) found that defining GNDVI as [(R613 – R559)/(R613 + R559)] explained up to 80% of yield variation among soybean genotypes. Mourtzinis et al. (2014) found that using the red edge together with other variables such as planting date increased the accuracy of predicting soybean yield. The photochemical reflectance index (PRI), defined as [(R 550 – R 531)/(R 550 + R 531)], captures the normalized difference between the major green wavelength reflectance of the plant canopy and leaves, which can be used to quantify radiation use efficiency (Gamon et al., 1992). Trotter et al. (2002) and Garbulsky (2011) used the PRI to assess nitrogen use efficiency and radiation use efficiency to distinguish genotypes that were superior to checks. These indices have been used to estimate yield in various environments. Other indices have focused on biophysical parameters, such as water content and leaf area. Leaf area index (first used in satellite imagery by Tucker and Sellers, 1986) was developed to predict vegetation parameters such as green biomass and green leaf area in wheat (Babar et al., 2006a, 2006b). The LAI measures the potential of the plant to capture PAR in most 626
crop species (Clevers, 1997). Aparicio et al. (2002) determined that LAI played a large part in plant function and correlated with yield prediction. The water index (WI) has also been developed to relate reflectance measurements at 850 and 970 nm with bread and durum wheat yield (Royo et al., 2003; Gutierrez et al., 2010). The WI focuses on the water-absorption bands within the NIR spectrum and capture relative water content based on absorption strength of water within the plant leaves. These indices are easy to calculate and give researchers a way to handle the large datasets associated with such research; however, indices tend to be environment-specific and have low correlation outside of the datasets used to create them (Ma et al., 2001; Gutierrez-Rodriguez et al., 2004; Prasad et al., 2007a; Gutierrez et al., 2010). Therefore, complex models that use more of the light spectrum need to be developed to capture the necessary variation in yield across different environments and crops (Pimstein et al. 2007; Weber et al., 2012). High-yielding cultivars tend to have prominent biophysical and biochemical properties such as increased chlorophyll, plant health characteristics, and increased photosynthetic capacity that make distinguishing these genotypes using spectral indices unreliable (Baret and Guyot, 1991; Aparicio et al., 2002; Pimstein et al., 2007). Because of limitations associated with index models to differentiate genotypes due to physical and biochemical parameters such as biomass saturation, dense canopy, high LAI, and high chlorophyll levels, full-spectrum models have been considered for yield estimation. These models have not been utilized, however, because of overfitting and collinearity concerns from the large number of predictors and the typically small sample sizes associated with spectral research (Pimstein et al., 2007). The main approaches used to meet the challenges of hyperspectral data analysis focus on reducing the correlation between predictor variables. One approach is principal component analysis (PCA). PCA has been used by researchers to build yield prediction models that explain 70 to 90% of the yield variability in maize (Hong et al., 2001; Chang et al., 2003) and 39% of the yield variability among soybean genotypes (Hong et al., 2001). Using six bands, regression explained almost 95% of the variability within maize yield (Hong et al., 2001). Another approach (artificial neural network analysis) was used in soybean to build prediction models that explain 46 to 81% of the yield variability in soybean and 42 to 77% in corn (Kaul et al., 2005). Researchers also determined that resampling the dataset to 10-nm intervals reduced the noise within the spectra associated high collinearity and band to band variation (Lin et al., 2012). Multivariate analysis is new to precision phenotyping research but has the potential to create yield prediction models and explain other important phenotypic parameters. Partial least squares regression (PLS) has been used for data pretreatment and variable reduction before performing multiple linear regression techniques (Wold, 1966). Like PCA, PLS extracts successive linear predictors of y called latent variables or factors. The new variables aim to
www.crops.org
crop science, vol. 56, march– april 2016
explain response variation and predictor variation; models are selected based on balancing these two goals. Hansen et al. (2002) compared three different PLS methods and concluded the nonlinear method was the most consistent for yield and protein content estimation in wheat and barley, which explained up to 75% of the variation in protein content and up to 97% of the variation in yield. Using different water regimes, Weber et al. (2012) similarly found that PLS explained a maximum of 40% of the variability within corn yield, and prediction models explained more variability in water-limited environments than in non-water–limited environments. Lin et al. (2012) used orthogonal projections to determine the latent structures in PLS in rice and distinguished the performance of three cultivars with 90% accuracy. Cultivars in a breeding program are selected based on comparative performance; however, special attention and consideration must be given to cultivar maturity. Breeding trials to measure yield may use maturity checks to group and evaluate experimental lines to reduce yield bias associated with maturity. For example, Rincker et al. (2014) found that genetic gains for yield in soybean maturity group (MG) 3 and MG 4 cultivars increased by 22.8 and 19.5 kg ha -1 yr-1, but when yield genetic gains were adjusted for maturity differences among cultivars, the gains decreased by 3.1 and 1.4 kg ha -1 yr-1 among MG 3 and MG 4, respectively. Maturity also affects the spectral reflectance response curves observed from soybean. Christenson et al. (2014) observed that among MG 3 and MG 4 cultivars, significant relationships were observed between maturity and the VIS, RED (700 to 730 nm), and NIR portions of the spectra. The authors also found that the release year of the cultivar accounted for a significant portion of the variation in seed yield (R2 = 0.82 for MG 3, and R2 = 0.86 for MG 4). They concluded that the relationship between cultivar maturities and the reflectance spectra as well as release year of the cultivars and reflectance spectra were very similar to suggest maturity influences the spectra significantly along with yield. Accounting for crop maturity, height, and lodging first in a regression model dramatically decreased the amount of variation in the reflectance spectra among cultivars accounted for by cultivar release year (Christenson et al., 2014). These findings indicate that maturity may be a major contributing factor to results of both soybean seed yield and canopy spectral reflectance. While much is known about the capability of reflectance measurements to evaluate plant phenotypes, more robust models are needed to expand the use of the technology in applied breeding programs, and estimations of maturity need to be integrated into the yield prediction process. The goal of this study was to develop statistical models for determining soybean maturity and estimating seed yield using canopy reflectance measurements. The specific objectives were to (i) crop science, vol. 56, march– april 2016
assess canopy reflectance as a tool for determining soybean maturity and seed yield, (ii) identify specific development stages that significantly contribute to maturity and yield estimation, and (iii) test the stability and utility of maturity and yield estimation models across varying environments.
MATERIALS AND METHODS Location and Experimental Design This study was conducted on soybean [Glycine max (L.) Merr.] at the Kansas State University research farm south of Manhattan, KS, in 2011 and 2012. In 2011, study plots were established at Location A (39°825 N 96°3746 W; elevation of 322 m) on a coarse-silty, mixed, superactive, nonacid, mesic Typic Udifluvents and coarse-silty, mixed, superactive, mesic Fluventic Hapludolls soil. In 2012, study plots were established at Location B (39°757 N 96°375 W; elevation of 322 m) on a fine-silty, mixed, superactive, mesic Fluventic Hapludolls and clayey over loamy, smectitic, mesic Fluvaquentic Hapludolls soil, and at Location C (39°835 N 96°3746 W; elevation of 322 m) on a coarse-silty, mixed, superactive, nonacid, mesic Typic Udifluvents and coarse-silty, mixed, superactive, mesic Fluventic Hapludolls soil. Twenty soybean cultivars of each MG 3 and MG 4, ranging in release year from 1923 to 2010, were selected out of the Soybean Genetic Gain Study coordinated by Dr. Brian Diers, University of Illinois (Rincker et al., 2014). Selected genotypes were a mixture of private and public cultivars. Replicated plots were planted at Location A on 23 May 2011, at Location B on 16 May 2012, and at Location C on 4 June 2012. Each experimental unit consisted of four rows (3.4 m long; spaced 76 cm apart; 26 seeds m–1). Cultivars were planted in both well-watered (Irr) and water-stressed (Dry) environments that were arranged in four randomized complete blocks (e.g., A-Irr indicates the well-watered growing environment at Location A). However, plots at Location C were later designated as C-Irr-1 and C-Irr-2 after an irrigation system failure caused the dryland plots to receive some irrigation water. In both years, weed pressure and soil fertility within the plots were not limiting factors on soybean development and growth. Flood irrigation was applied in furrows to the wellwatered environments starting at reproductive stage R1 and this irrigation continued weekly until R6 (Fehr and Caviness, 1977). In 2011, no supplemental irrigation was applied to the water-stressed environments. Due to extremely dry conditions, irrigation was applied once in 2012, shortly after R1, to ensure crop development in the water-stressed environments.
Phenotypic Traits Maturity, height, lodging, and seed yield were measured on all plots during both growing seasons. Maturity was recorded as the date when 95% of the pods had reached mature plant color (development stage R8). Height (cm) was measured as the distance from the base of the plant to the top of the main stem. Lodging was scored on a scale of 1 to 5, based on the number of leaning or broken plants. Upright plants (no lean) were scored as 1; other scores were 2 = 20° lean, 3 = 45° lean, 4 = 60° lean, and 5 = flat on the ground. The center two rows of each plot were
www.crops.org 627
mechanically harvested using a two-row plot combine, and seed yield was recorded as kg ha-1 adjusted to 13% moisture.
DATA ANALYSIS
Canopy Reflectance Measurements
The MIXED procedure of ANOVA was performed in SAS 9.2 (SAS Institute, 2008). Cultivar and environment were treated as fixed effects, and replication nested within the environment was treated as a random effect for yield and agronomic traits. For hyperspectral data, development stage observations were treated as a fixed effect, and replication nested within development stage observation was treated as a random effect for environment and full-season analysis. Environment was classified as location and water treatment. Individual development stage observations that failed to detect significant differences in reflectance among cultivars for waveband regions were eliminated from further analysis, and only observation days with significant differences among cultivars for waveband regions were used for least squares means calculations and modeling.
Canopy reflectance was measured using an ASD LocationSpec 3 spectroradiometer (Analytical Spectral Devices, Boulder, CO). Solar radiation reflecting from the plant canopy was captured from 350 to 2500 nm in the electromagnetic spectrum using fiber optics with a 25° field of view. Sampling intervals were of 1.4 nm from 350 to 1050 nm and of 2 nm from 1050 to 2500 nm. Moving averages were calculated automatically to achieve 1-nm-width continuous bands. A Spectralon (Labsphere, Inc., North Sutton, NH) reflectance panel was used to calibrate the spectroradiometer with a dark current and white reference every 40 plots or when needed (this ranged from every plot to 40-plot intervals, depending on location conditions) and to create reflected light percentages. The sensor was mounted on an adjustable monopod pole and held vertically, approximately 1 m above the canopy, to achieve approximately a 50-cm diameter collection area. Two measurements were taken per plot on the second and third rows, excluding the first meter of each plot to eliminate border effect. Each measurement was the average of 10 scans, which was calculated automatically. Spectral data were collected on nearly cloud-free days within two hours of solar noon. Canopy reflectance data were collected weekly from R3 to R6 in 2011 and from R2 to R6 in 2012. In 2011, there were four collection dates (both MGs) in environment A-Irr (28 July, 10 Aug., 18 Aug., 2 Sept.) and three dates in environment A-Dry (28 July, 10 Aug., 18 Aug., 2 Sept.). In 2012, there were four collection dates for MG 3 in environment B-Irr (17 July, 24 July, 9 Aug., 17 Aug.), five collection dates for MG 4 in environment B-Irr (17 July, 24 July, 9 Aug., 17 Aug., 20 Aug.), and four collection dates for MGs 3 and 4 in environment C-Irr-1 (24 July, 8 Aug., 15 Aug., 29 Aug.). Also in 2012, there were three collection dates (both MGs) in environment B-Dry (24 July, 9 Aug., 17 Aug.) and two collection dates in environment C-Irr-2 (8 Aug., 15 Aug.).
Data Pretreatment Spectral reflectance data (350 to 2500 nm) were initially trimmed to 400 to 1310 nm to eliminate the noise from atmospheric absorption regions focused around the water absorption wavebands in the infrared and the atmospheric scatter of blue color in the ultraviolet portion of the spectrum. An outlier control was implemented on the raw data from each observation day before averages and totals were calculated. After outliers were excluded, the data were combined to form 10-nm means to reduce the dataset size and eliminate some of the collinearity associated with hyperspectral reflectance wavebands in close proximity (De Jong, 1993). Combined bands contain less variation from sample to sample than single-band measurements (Lin et al., 2012), and due to the high correlation with wavebands in close proximity, minimal information is thought to be lost due to waveband combination.
628
Analysis of Variance
Partial Least Squares A nonlinear iterative partial least squares algorithm (PROC PLS in SAS 9.2), with ten repeats of random cross-validation and 15 factor levels, was used for outlier identification, pretreatment, and predictor variable selection on the data from 2011 and 2012. The predicted residual sums of squares (PRESS) were used for model selection. This algorithm (NIPALS vs. SIMPLS, De Jong, 1993) was chosen because of the high dimensionality and nonlinear nature of the hyperspectral data. Random 10-fold cross-validation was chosen because of the small sample size compared with the number of predictors and the lack of a true validation dataset for validating built models, and the maximum factor number was set at 15, which is the number allowed by SAS for the dataset used. The optimized factor number was selected when the PRESS statistic was most minimized (Wold, 1966).
Pretreatment Before Modeling Because modeling assumptions of normality were not met, it was necessary to auto-scale spectral data and dependent variables before PLS analysis (Harshman and Lundy, 1994). The data were mean-centered and auto-scaled for each analysis. Mean-centering and scaling eliminates bias within the bands to ensure even weighting of band magnitude, so the analysis is focused on the variance (Hansen et al., 2002). Auto-scaling is calculated as the inverse of the standard deviation for each variable [1/(SD)].
Predictor Variable Selection Hyperspectral band regions from 2-yr means and individual environments that most significantly contributed to maturity and seed yield estimation were determined using partial least squares regression (PLSR). In total, 91 10-nm band regions from 400 to 1310 nm were analyzed for variable importance and selected for further analysis through traditional least
www.crops.org
crop science, vol. 56, march– april 2016
squares multiple linear regression (MLR). This procedure was done to extract exact band regions contributing to maturity and seed yield estimation, which is not possible in PLS models. Variables were determined to be important through variable importance within projection plots (VIP), which were created through a macro in SAS. The criterion limit of 0.8 was used as a cutoff point to indicate variables that contributed significantly to maturity and seed yield estimation (Wold, 1966). Band regions were selected when they were concluded to be highly important and grouped within a strong spike of importance across environments, such as in SNP or QTL analyses. Due to collinearity issues, band regions that significantly contributed to maturity and seed yield estimation and were close to each other were combined for final analysis.
and calculated indices models. Models were validated on individual observations, environments, and maturity group for reliability and robustness.
Table 1. Calculated normalized difference vegetation indices (NDVI) and water indices (WI) for each of the VIS, red edge, and NIR wavebands selected through partial least squares regression. Each index was calculated from reflectance at specific points in the light spectra. For example, BNDVI_1 was calculated from reflectance data at 415 and 915 nm as [(R 915 – R 415 )/(R 915 + R 415 )]. NDVI†
Multiple Linear Regression Models for maturity and yield estimation were built using band regions selected through PLSR using PROC GLM SELECT in SAS 9.2. Cultivar means from both maturity groups for each of the six environments were used to create the training model. Cultivar means from both maturity groups across all environments as well as cultivar means for individual observation days in each environment were used as validation data. Reflectance indices were also created using the selected band regions. All singular visible band regions were used with other singular NIR band regions to create all possible combinations of NDVI, and the three NIR band regions around the water absorption regions (900 to 990 nm) were used to create water indices (Table 1). For variable selection, stepwise with forward and backward elimination was used with 0.1 a for variable entry and exit criterion. Due to the small sample sizes, full n-1 cross-validation ( jack-knife) was utilized instead of training and validation datasets. Model selection was completed when the predicted residual sums of squares (PRESS) statistic and Mallow’s Cp statistic were fully minimized. The selected b parameter estimates were then exported into PROC REG (SAS 9.2) for goodness of fit tests. The variable inflation factor (VIF) was used when modeling the waveband regions to determine multi-collinearity between parameters, and a selection value of greater than 10 was used to determine multicollinearity. If a parameter had a VIF greater than 10, it was eliminated from the possible variables, and the models were re-optimized. The combination modeling of waveband regions and indices were not subjected to collinearity testing because most indices containing similar bands are naturally highly correlated with each other, which causes significant collinearity. Models were created for maturity and yield estimation using environmental (n = 220) least squares means with maturity groups kept together for waveband region models crop science, vol. 56, march– april 2016
Calculation using waveband reflectance nm
BNDVI_1
(915 – 415) / (915 + 415)
BNDVI_2
(940 – 415) / (940 + 415)
BNDVI_3
(990 – 415) / (990 + 415)
BNDVI_4
(1100 – 415) / (1100 + 415)
BNDVI_5
(1140 – 415) / (1140 + 415)
BNDVI_6
(1245 – 415) / (1245 + 415)
GNDVI_1
(915 – 550) / (915 + 550)
GNDVI_2
(940 – 550) / (940 + 550)
GNDVI_3
(990 – 550) / (990 + 550)
GNDVI_4
(1100 – 550) / (1100 + 550)
GNDVI_5
(1140 – 550) / (1140 + 550)
GNDVI_6
(1245 – 550) / (1245 + 550)
RNDVI_1
(915 – 680) / (915 + 680)
RNDVI_2
(940 – 680) / (940 + 680)
RNDVI_3
(990 – 680) / (990 + 680)
RNDVI_4
(1100 – 680) / (1100 + 680)
RNDVI_5
(1140 – 680) / (1140 + 680)
RNDVI_6
(1245 – 680) / (1245 + 680)
RENDVI_1
(915 – 715) / (915 + 715)
RENDVI_2
(940 – 715) / (940 + 715)
RENDVI_3
(990 – 715) / (990 + 715)
RENDVI_4
(1100 – 715) / (1100 + 715)
RENDVI_5
(1140 – 715) / (1140 + 715)
RENDVI_6
(1245 – 715) / (1245 + 715)
WI_1
(940 / 915)
WI_2
(990 / 915)
WI_3
(990 / 940)
† Each prefix describing the NDVI represents a region of the light spectra; B, blue; G, green; R, red; RE, red edge.
www.crops.org 629
RESULTS AND DISCUSSION Genotypic Performance in MG 3 and MG 4 Yield and Spectra A wide range in mean seed yield was observed in both maturity groups (Christenson et al., 2014). At Locations A and B, seed yields were significantly greater in the irrigated treatments than in the water-stressed treatments (Table 2). At Location C, the flood irrigation system at early pod-fill failed to shut down at the appropriate time, and excess water from the irrigated plots inadvertently irrigated the dryland plots, thus the two environments at Location C were designated as C-Irr-1 and C-Irr-2. Because of timely rainfall events from that point on in the growing season, yields in these two irrigated environments were not different. Differences in yield, maturity, lodging, and plant height were observed among cultivar and environment means, and the relative performance of the cultivars for these traits differed across environments in both maturity groups (Table 3). Differences (P < 0.05) were detected in reflectance for selected waveband regions among cultivars for each observation day and development stage in both maturity groups (Table 4). In both MG 3 and MG 4, the VIS (400 to 700 nm) and red edge (701 to 730 nm) had consistent differences among cultivar, whereas the NIR (731 to 1310 nm) was not as consistent in reflectance response within and among genotypes. Reflectance of MG 3 cultivars in environment B-Irr on observation days R3 and R3–R4 was highly inconsistent, with only the 550-nm and red-edge wavebands producing significant differences among cultivars. Table 2. Seed yield by environment and range in yield among soybean MG 3 and MG 4 cultivars grown in environments consisting of two different water regimes at three locations. Env.†
Year
Mean
LSD
Range
––––––––––––– kg ha-1 –––––––––––––
Reflectance of MG 4 cultivars in environment B-Irr on observation days R2–R3 and R4 as well as in environment C-Irr-2 on days R3 and R4 produced no significant differences among cultivars, resulting in the elimination of these development stage observations from maturity and yield models. This lack of difference among cultivars may have been due to severe lodging observed in these environments. Also, the spectral data were influenced by soil such that the radiometer sensed variables other than plant vegetation to produce inconsistent reflectance readings in the red and near infrared ranges. Results from individual environment ANOVA analyses were similar to that of the development stage analyses, with significant differences in spectral readings detected among cultivars for both MGs (Table 5). The most consistent differences among cultivars were observed for the VIS, red edge, and NIR wavebands. In all environments, differences in reflectance readings were noted among observational days for most of the spectral parameters; however, in situations when the Cultivar × Day interaction was significant, cultivar tended to be a larger portion of the variation in reflectance than the Cultivar × Day interaction. This indicated that the relative performance of the cultivars tended to be similar among observation days. The ANOVA results for overall environment means (six environments) produced highly significant differences among cultivars and environment main effects, whereas only a few wavebands had significant Cultivar × Environment interactions (Table 6). These results suggest that most of the phenotypic variation in reflectance was due to cultivar (genotypic) differences. The contribution of the Cultivar × Environment interaction to total phenotypic variation suggests spectral data may be robust across environments, making it a candidate for an indirect selection technique. Reynolds et al. (2009) suggested the interaction between
MG 3 (n = 20) A-Irr
2011
2981
430
2520
A-Dry
2011
1759
604
2011
B-Irr
2012
3637
583
2673
B-Dry
2012
2780
921
2312
C-Irr-1
2012
3993
600
2863
C-Irr-2
2012
3944
593
3331
MG 4 (n = 20) A-Irr
2011
3201
521
3272
A-Dry
2011
1763
553
1813
B-Irr
2012
3593
382
3634
B-Dry
2012
3238
414
2487
C-Irr-1
2012
3772
740
2139
C-Irr-2
2012
3989
511
2678
† Env, environment; A through C, study location of field plots; Irr, well-watered with irrigation; Dry, dryland with no irrigation; 1, first irrigated environment; 2, second irrigated environment.
630
Table 3. F-values for effect of cultivar, environment, and their interaction from ANOVA on agronomic traits based on six environments for MG 3 and MG 4 cultivars. Source†
df
Yield
Maturity
Lodging
Height
MG 3 C
19
120.42**
77.04**
55.99**
131.58**
E
5
49.41**
121.74**
99.84**
15.00**
C×E
95
3.33**
4.59**
3.54**
1.81**
MG 4 C
19
102.12**
41.35**
16.54**
65.36**
E
5
208.94**
485.21**
396.17**
106.48**
C×E
95
4.01**
2.69**
2.69**
4.02**
** Significant at the 0.01 probability level. † C, cultivar; E, environment, consisting of two different water regimes at three study locations.
www.crops.org
crop science, vol. 56, march– april 2016
Table 4. F-values for effect of cultivar by ANOVA on canopy reflectance measured at selected waveband regions measured at different developmental stages for MG 3 and MG 4 soybean cultivars grown in environments consisting of two different water regimes at three study locations. Waveband Env.†
Stage
415
550
680
715
915
940
990
1100
1140
1245
1300
––––––––––––––––––––––––––––––––––––––––––––––––––– nm –––––––––––––––––––––––––––––––––––––––––––––––––––
MG 3 A-Irr
A-Dry
B-Irr
B-Dry C-Irr-1
C-Irr-2
R3-R4
2.22**
2.98**
1.81*
2.46**
NS‡
NS
NS
NS
NS
NS
NS
R4
5.51**
8.25**
4.47**
8.91**
3.29**
2.83**
2.56**
2.51**
2.19*
2.42**
2.46**
R5
3.05**
6.83**
NS
4.01**
3.14**
2.57**
2.03*
NS
NS
NS
NS
R6
2.46**
5.90**
4.68**
3.56**
3.12**
2.74**
2.43**
1.94*
NS
NS
NS
R3-R4
4.48**
7.85**
3.93**
7.14**
NS
NS
NS
NS
NS
NS
NS
R5
7.68**
22.50**
5.75**
18.89**
4.23**
3.70**
4.09**
3.42**
3.00**
3.94**
3.97**
R6
11.22**
11.85**
13.25**
6.24**
6.36**
5.38**
4.45**
3.59**
2.40**
1.95*
1.80*
R2
4.51**
7.63**
2.99**
7.87**
3.05**
2.99**
3.20**
2.62**
2.71**
2.78**
2.77**
R3
NS
2.00*
3.00**
NS
3.33**
NS
NS
NS
NS
NS
NS
R3-R4
NS
3.70**
NS
3.64**
NS
NS
NS
NS
NS
NS
NS
R5-R6
4.13**
6.37**
4.96**
6.27**
1.77*
NS
1.88*
1.83*
2.06*
2.11*
2.15*
R3
3.57**
2.89**
1.94*
4.03**
R4
7.00**
8.09**
4.87**
R2
9.67**
16.04**
R3
1.71*
3.34**
R4-R5
3.40**
R6
4.64**
R2
1.99*
R4-R5
1.77*
1.77*
1.89*
NS
NS
2.01*
2.07*
10.15**
NS
NS
NS
NS
NS
1.92*
2.01*
9.11**
9.87**
NS
NS
NS
NS
NS
NS
2.05*
1.94*
3.01**
NS
NS
NS
NS
NS
NS
NS
4.93**
2.90**
5.59**
NS
NS
NS
NS
NS
1.93*
1.82*
6.01**
3.48**
5.76**
2.13*
1.93*
2.55**
2.23**
2.09*
2.89**
2.92**
2.73**
NS
3.83**
2.10*
1.94*
2.07*
NS
NS
NS
NS
3.97**
6.03**
3.59**
9.39**
3.90**
4.07**
4.59**
3.51**
4.05**
4.56**
4.21**
R3
2.33**
3.31**
1.68*
3.94**
2.33**
NS
NS
NS
NS
NS
R3-R4
3.79**
6.38**
3.88**
4.12**
3.73**
2.93**
2.57**
2.29**
NS
NS
NS
R5
4.75**
6.67**
3.67**
4.24**
10.12**
8.42**
6.74**
5.45**
3.89**
3.17**
2.98**
R5-R6
3.13**
3.22**
2.71**
2.17*
4.30**
3.80**
2.65**
1.92*
NS
NS
NS
R2-R3
5.58**
6.13**
4.84**
4.99**
NS
NS
NS
NS
NS
NS
NS
R3-R4
8.63**
7.67**
5.10**
6.08**
4.90**
4.70**
4.49**
3.86**
3.56**
3.65**
3.61**
R5
5.80**
4.66**
5.25**
5.09**
6.81**
6.26**
4.63**
3.63**
2.76**
2.13*
2.05*
R1-R2
5.01**
7.80**
2.99**
9.29**
4.27**
4.62**
3.64**
2.94**
3.12**
2.54**
2.56**
R2-R3
NS
NS
NS
NS
NS
NS
NS
NS
NS
NS
NS
MG 4 A-Irr
A-Dry
B-Irr
B-Dry
C-Irr-1
R4
NS
NS
NS
NS
NS
NS
NS
NS
NS
NS
NS
R5
4.09**
7.26**
3.92**
6.52**
3.05**
2.86**
2.67**
1.97*
NS
1.78*
NS
R5-R6
3.67**
4.80**
3.06**
3.96**
1.96*
1.87*
1.87*
1.98*
2.01*
2.22*
2.27**
R3
4.16**
2.97**
1.93*
2.94**
1.77*
NS
NS
NS
NS
NS
NS
R4
5.03**
4.89**
3.48**
3.71**
3.61**
3.04**
2.91**
3.15**
2.37**
2.45**
2.38**
R6
4.61**
9.07**
4.08**
7.62**
6.00**
5.45**
4.77**
4.55**
3.93**
3.65**
3.52**
R1-R2
8.50**
13.45**
7.54**
7.98**
NS
NS
1.78*
1.80*
NS
1.91*
1.89*
R2-R3
C-Irr-2
2.43**
NS
2.31*
NS
NS
NS
NS
NS
NS
NS
NS
NS
R4
4.99**
6.72**
3.59**
6.23**
2.44**
2.31**
2.19*
NS
NS
NS
NS
R5-R6
6.12**
5.66**
4.78**
5.31**
2.53**
2.63**
2.54**
2.28**
2.38**
2.29**
2.29**
R3
NS
NS
NS
NS
NS
NS
NS
NS
NS
NS
NS
R4
NS
NS
NS
NS
NS
NS
NS
NS
NS
NS
NS
* Significant at the 0.05 probability level. ** Significant at the 0.01 probability level. † Env, environment; A through C, study location; Irr, well-watered with irrigation; Dry, dryland with no irrigation; 1, first irrigated environment; 2, second irrigated environment. ‡ NS, significance level P > 0.05.
crop science, vol. 56, march– april 2016
www.crops.org 631
Table 5. F-values for effects of cultivar, observation day, and their interaction from ANOVA on reflectance measured at selected waveband regions for MG 3 and MG 4 soybean cultivars grown in environments consisting of two different water regimes at three study locations. Waveband Env.†
Source‡
df
415
550
680
715
915
940
990
1100
1140
1245
1300
––––––––––––––––––––––––––––––––––––––––––––––––––– nm –––––––––––––––––––––––––––––––––––––––––––––––––––
MG 3 A-Irr
A-Dry
B-Irr
B-Dry
C-Irr-1
C-Irr-2
C
19
8.22**
16.14**
7.53**
11.51**
7.00**
5.65**
4.87**
NS§
3.06**
2.81**
2.48**
D
3
15.48**
42.92**
84.22**
32.28**
16.96**
11.58**
12.66**
19.86**
9.57**
16.23**
16.26**
C×D
57
NS
2.09**
2.36**
NS
NS
NS
NS
NS
NS
NS
NS
C
19
18.84**
27.48**
18.84**
19.33**
6.96**
6.02**
5.28**
4.22**
3.15**
3.10**
3.01**
D
2
21.28**
31.48**
98.69**
7.94**
6.33*
5.72*
6.04*
4.91*
5.61*
4.98*
4.78*
C×D
38
3.19**
4.40**
5.61**
2.99**
1.86**
1.72**
1.81**
1.51*
NS
1.59*
1.58*
C
19
7.14**
14.25**
6.58**
15.46**
3.49**
3.56**
3.77**
3.20**
3.55**
3.63**
3.69**
D
3
34.99**
82.83**
36.21**
75.67**
48.66**
40.50**
49.97**
48.60**
43.20**
49.56**
48.97**
C×D
57
1.42*
1.78**
NS
1.42*
NS
NS
NS
NS
NS
NS
NS
C
19
6.94**
6.86**
4.82**
9.73**
2.27**
2.19**
2.40**
2.05**
2.31**
3.00**
3.13**
D
1
21.21**
6.26**
NS
41.53**
21.41**
25.15**
29.06**
14.24**
23.95**
29.36**
28.98**
C×D
19
NS
NS
NS
NS
NS
NS
NS
NS
NS
NS
NS
C
19
13.21**
19.28**
14.41**
17.62**
NS
NS
1.97*
NS
1.89*
2.68**
2.82**
D
3
8.64**
31.32**
6.09*
42.61**
39.80**
31.51**
36.48**
36.03**
36.68**
35.42**
32.75**
C×D
57
2.61**
3.53**
2.51**
2.11**
NS
NS
NS
NS
NS
1.72**
1.76**
C
19
4.44**
6.57**
3.34**
10.44**
3.55**
3.49**
4.00**
3.11**
3.18**
4.11**
4.13**
D
1
32.32**
51.43**
3.05*
122.61**
194.20**
74.86**
217.69**
165.46**
90.38**
228.15**
223.70**
C×D
19
NS
NS
NS
1.72*
2.55**
2.49**
2.54**
2.25**
2.02*
1.77*
1.73*
C
19
8.92**
14.11**
9.28**
9.89**
14.65**
11.96**
8.51**
6.78**
4.32**
3.58**
3.38**
D
3
11.69**
NS
26.96**
4.14*
36.04**
24.34**
36.70**
37.85**
18.78**
36.27**
34.42**
C×D
57
NS
NS
NS
NS
1.93**
1.86**
1.61**
NS
NS
NS
NS
C
19
15.17**
13.02**
9.55**
11.12**
8.51**
8.24**
6.41**
5.81**
4.95**
4.14**
4.04**
D
2
NS
NS
17.78**
NS
NS
NS
4.43*
NS
NS
NS
NS
C×D
38
2.04**
2.39**
2.95**
2.50**
1.87**
1.80**
1.62*
NS
NS
NS
NS
C
19
6.95**
10.58**
4.45**
10.02**
2.79**
2.82**
2.32**
1.97*
1.73*
NS
NS
D
4
31.81**
78.40**
21.23**
100.67**
56.28**
48.13**
59.96**
53.73**
47.04**
66.44**
64.96**
C×D
74
1.56**
1.90**
1.47*
1.80**
1.59**
1.54**
1.52**
1.52**
1.45*
1.46*
1.46*
C
19
9.25**
9.52**
5.36**
7.94**
5.85**
5.05**
4.17**
3.92**
2.67**
2.43**
2.31**
D
2
42.14**
28.23**
10.34**
30.34**
128.35**
83.27**
132.52**
152.26**
81.62**
172.80**
160.52**
C×D
38
2.15**
2.18**
1.70*
1.87**
1.92**
1.88**
1.87**
1.68*
1.72**
1.76**
1.76**
C
19
13.14**
17.63**
9.37**
12.68**
2.21**
2.10**
2.04**
1.72*
NS
NS
NS
D
3
36.00**
55.34**
17.93**
33.27**
17.59**
12.29**
16.47**
16.73**
13.37**
17.02**
16.17**
C×D
57
1.60**
2.42**
1.77**
1.73**
NS
NS
NS
NS
NS
NS
NS
C
19
NS
NS
D
1
39.69**
45.50**
C×D
19
NS
NS
MG 4 A-Irr
A-Dry
B-Irr
B-Dry
C-Irr-1
C-Irr-2
NS 12.82* NS
NS 41.26** NS
1.70* 177.76** NS
NS 88.39** NS
1.69* 161.12** NS
NS 135.75** NS
NS 112.33** NS
NS 119.93** NS
NS 119.92** NS
* Significant at the 0.05 probability level. ** Significant at the 0.01 probability level. † Env, environment; A through C, study location; Irr, well-watered with irrigation; Dry, dryland with no irrigation; 1, first irrigated environment; 2, second irrigated environment. ‡ C, cultivar; D, day of observation. § NS, significance level P > 0.05.
632
www.crops.org
crop science, vol. 56, march– april 2016
genotype and environment stability as a major factor for use in evaluating any new phenotyping technique.
Waveband Selection for Yield and Maturity Using PLSR, portions of the spectra significantly contributed to yield and maturity estimation among MG 3 and MG 4 cultivars (data not shown). The wavebands that contributed significantly to yield estimation also contributed significantly to maturity estimation in individual environments as well as grand means using PLSR. Because of this phenomenon, waveband selections were made based on the yield estimation PLSR models. This was done to decrease overall complexity of the modeling and increase reproducibility to other spectral platforms. In 2011, MG 3 in environment A-Irr had the highest significances for yield estimation from the visible and rededge portions, with a sharp decrease at the beginning of NIR and a significant increase from 775 to 955 nm. For environment A-Dry, variable importance was similar to that of environment A-Irr, but a decrease in the blue (400 to 500 nm) portion of the spectra was observed, and no significant peaks were observed from 1035 to 1095 nm. For MG 3 in environment B-Irr, the VIS and NIR spectra had highly significant portions for yield estimation; the red edge was the most significant, followed by the green and red wavebands. A spike in NIR was observed around 915, 1100, and 1165 nm, and there was a sharp peak in the early blue portion of the spectra. For MG 3 in environment B-Dry, the VIS and NIR had highly significant regions above the 0.8 threshold, with a sharp decrease in importance in the NIR portion up to 1165 to 1305 nm. The red-edge portion was the most significant, as in the A-Irr environment. Reflectance of MG 3 in environment C-Irr had high peaks of significance in the VIR, but these peaks were not as strong as those in locations A and B. There was a sharp increase in the red edge and slight peaks from 915 to 1005 nm and from 1165 to 1305 nm in the NIR. Data from
environment C-Irr-2 mostly mirrors those measured in environment C-Irr-1, with a sharp peak in the green waveband (550 nm). Compared with C-Irr-1, reflectance data from environment C-Irr-2 displayed a sharp increase in the early blue region, and the red edge was the most significant for yield estimation. Reflectance of MG 4 in environment A-Irr showed low significance in the blue region, and a sharp increase in the green region was observed, with 550 nm being the highest significance for yield estimation in the green. The red portion of the spectra was not as significant as the green or as in the MG 3 cultivars; overall, however, the red edge was the most significant region for yield estimation, which was similar to trends for MG 3. Significance at the beginning of the NIR region decreased sharply, but the rest of the plateau region (735 to 910 nm) was significant. MG 4 in environment A-Dry had an increase in significance in the blue region compared to data in A-Irr; but similar to the A-Irr environment, green was the most significant region of the VIS. The red edge was significant, as for MG 3 and MG 4 in the A-Irr environment; however, no significant NIR regions were observed for MG 4 in the A-Dry environment. In MG 4 in the B-Irr and B-Dry environments, meaningful portions of the spectra in the blue, green, red, and red edge were observed, as well as some portions of the NIR. For the B-Irr environment, the early blue, green, red, and the red edge were the most useful for yield estimation, as well as a slight peak in reflectance at 1305 nm. For the B-Dry environment, reflectance in the early blue, green, and red edge wavebands were the most highly significant, similar to the B-Irr environment; however, the red portion of the spectra and the NIR had lower significance levels than the B-Irr environment. In the NIR, significant peaks at 765 to 945 nm and a slight, but significant peak from 1065 to 1105 nm were observed in the B-Dry environment that were absent in the B-Irr environment. For the MG 4 cultivars, no optimized model was found for the C-Irr-2 environment. This may have been the result of
Table 6. F-values for effect of cultivar, environment, and their interaction from ANOVA on reflectance measured at selected waveband regions for MG 3 and MG 4 soybean cultivars grown in six different environments. Waveband Source†
df
415
550
680
715
915
940
990
1100
1140
1245
1300
––––––––––––––––––––––––––––––––––––––––––––––––––– nm –––––––––––––––––––––––––––––––––––––––––––––––––––
MG 3 C
19
15.53**
18.15**
7.75**
19.15**
2.51**
2.37**
2.96**
1.97**
2.94**
4.10**
4.33**
E
5
63.67**
35.55**
5.58**
42.17**
146.21**
110.74**
142.66**
79.28**
69.13**
110.86**
94.38**
C×E
95
NS‡
NS
NS
NS
1.66**
1.58**
1.42**
NS
NS
NS
NS
MG 4 C
19
20.43**
28.86**
13.61**
25.14**
12.00**
10.02**
3.55**
8.71**
6.25**
7.01**
6.77**
E
4
104.39**
125.61**
65.78**
160.62**
338.32**
314.56**
316.51**
243.78**
218.21**
223.47**
205.97**
C×E
76
2.01**
1.68**
1.47**
NS
NS
NS
NS
NS
NS
NS
NS
** Significant at the 0.01 probability level. † C, cultivar; E, environment, consisting of two different water regimes at three study locations. ‡ NS, significance level P > 0.05. crop science, vol. 56, march– april 2016
www.crops.org 633
a limited two readings taken in that environment, but it was most likely caused by a high degree of lodging in the MG 4 cultivars that would have negatively impacted the quality of the reflectance data. The C-Irr-1 environment, however, had significant regions in the VIS and red edge as well as most of the NIR (with the exception of 1005 to 1125 nm wavelengths). Overall, the most significant portions of the spectra across all experiments were the visible and red-edge portions of the spectra. The green region around 550 nm and the red region around 675 to 695 nm were the most significant portions of the visible spectra. In the red edge, the most significant region was from 705 to 715 nm. The NIR was highly influenced by environmental factors and was inconsistent in significance for yield estimation across environments. This result is most likely due to atmospheric scatter and observational day conditions influencing the spectra. Band regions for final yield modeling were selected based on significance to yield estimation through importance in projection values in all environments in both maturity groups. Selected portions of the spectra that were close to each other were combined to form 11 spectral regions used for further yield modeling. The final bands used for modeling were 415 nm (400–430 nm), 550 nm (530–570 nm), 680 nm (670–690 nm), 715 nm (700–730 nm), 915 nm (910–920 nm), 940 nm (930–950 nm), 990 nm (980–1000 nm), 1100 nm (1090–1110 nm), 1140 nm (1120–1160 nm), 1245 nm (1240–1250 nm), and 1300 nm (1290–1310 nm). Other spectral regions exhibited variable importance values over the 0.8 criterion, but were not selected due to fears of high multicollinearity between bands and over-fitting due to more predictor variables than the sample size (n = 20). The region from 405 to 435 nm has been correlated with high chlorophyll a and b content as well as beta-carotene absorption (Chappelle et al., 1992). Lower reflection values in this region are due to high absorption of light by both forms of chlorophyll and beta-carotene. Peñuelas et al. (1995) also created a normalized phaeophytinization index that related senescence to reflectance around 415 and 435 nm. The visible green region from 535 to 565 nm has high reflection due to chlorophyll a and b reflecting green light. Chappelle et al. (1992) found that chlorophyll a and b inside soybean leaves most reflected the 550-nm region, and that this waveband could be used in a ratio with 675 nm to explain 93% of the variation in chlorophyll content between soybean leaves and could be related directly to photosynthetic capacity. Also, 550 nm has been used to explain 92% of yield variability in wheat yields (Royo et al., 2003). Ma et al. (2001) also used 559 nm alone as well as in a ratio with 613 nm to explain 13 to 80% of the variation in yield in soybean genotypes. The middle of the red depression region (675 to 685 nm) has been correlated with chlorophyll absorption in soybean (Chappelle et al., 1992). The lower-yielding varieties had higher reflectance in this region, suggesting a lower 634
amount of chlorophyll or inefficiency of the chlorophyll, which resulted in lower yields in soybean (Ma et al., 2001), corn (Weber et al., 2012), and wheat (Royo et al., 2003). The red-edge inflection point (705 to 745 nm) is the sharp increase in reflection values because of the transition from the visible red region to the high-reflection NIR portion due to cellular scatter. This inflection point has been used to distinguish plant health and yield, with healthier plants having large contrasts between red and NIR (Gitelson et al., 2011). The 915-nm waveband region had a high reflection value in the spectra and has been associated with chlorophyll content measurements. Zhao et al. (2007) found that ratios using 551 and 915 nm as well as 708 and 915 nm accounted for 67 to 76% of the variability within chlorophyll content between cotton genotypes; however, Gitelson et al. (2003 and 2005) found that the best waveband regions to estimate chlorophyll in higher plants were from 525 to 585 nm and 695 to 725 nm. The 915-nm region also has been used for total biomass prediction in bermudagrass, accounting for 29.8 to 44.3% of the biomass variation (Starks et al., 2006). Marti et al. (2007) found total biomass had a strong correlation with wheat yield (r = 0.97). Reflection in the 940-nm region has been used in chlorophyll meters (SPAD, Minolta Osaka Co., Ltd., Japan) to capture nitrogen status and chlorophyll content of crops (Blackmer et al., 1994). Vollmann et al. (2011) found a significant correlation with SPAD-502 readings (ratio between 650 nm and 940 nm) and 1000-seed weight in soybean. High reflection values were also observed from 985 to 995 nm and from 1135 to 1155 nm, with higher-yielding genotypes tending to have higher reflection. Wenjiang et al. (2004) found that wavebands selected through regression techniques for winter wheat total foliar nitrogen content were around regions 1000 to 1140 nm (r = 0.83) and 1200 to 1300 nm (r = 0.51). The 1240-nm region of the spectra has been correlated with water content of the leaf in many crops (Peñuelas et al., 1993; Gao, 1996; Datt et al., 2003; Gutierrez et al., 2010). Moreover, lower values in the 1150–1260-nm region have been associated with higher water content (Sims and Gamon, 2003). Prasad et al. (2007b) concluded that indices using water content bands had higher heritability and could distinguish higher-yielding genotypes more consistently than vegetationbased indices in wheat. Higher values in these regions correlate to higher-yielding varieties in corn, but no physiological characteristics have been associated with these regions (Weber et al., 2012). Researchers have been able to characterize plant water status and stress through the preferential absorption of water, and more specifically the hydroxyl ions within the electromagnetic spectrum (Peñuelas et al. 1993, 1997; Gao, 1996; Serrano et al., 2000). The NIR (730–1300 nm) and middle infrared (1300–2500 nm) regions have correlated well with water status of plants, and wavelengths of 970, 1240, 1400, and 2700 nm have been proposed as absorption bands
www.crops.org
crop science, vol. 56, march– april 2016
indicative of water status (Tucker, 1980; Peñuelas et al., 1993; Gao, 1996; Zarco-Tejada et al., 2001; Gutierrez et al., 2010). Peñuelas et al. (1993) first developed the WI, defined as [R970/R900], to predict water stress, and WI was found to strongly correlate with the relative water content of plants. The normalized difference WI, proposed by Gao (1996), defined as [(R860 – R1240)/(R860 + R1240)], has been used to evaluate water stress in corn and soybean (Jackson et al., 2004) with favorable results. Babar et al. (2006a) proposed normalizing the WI, using the wavelengths 970, 900, and 850 nm, and found it useful for screening wheat genotypes for water stress and water use efficiency (Gutierrez et al., 2010). Prasad et al. (2007a) also normalized the WI, using the wavelengths 970, 920, and 880 nm to screen winter wheat lines under dryland conditions (Gutierrez et al., 2010). These authors propose the normalization of the WI provides added genotypic variation
explanation. However, Gutierrez et al. (2010) found that the normalized WI using wavelengths 970 and 880 nm proposed by Prasad et al. (2007b) was the only index sufficient for predicting water stress within environments.
Correlations between Agronomic Traits and Spectral Reflectance Strong relationships were observed between seed yield, maturity, lodging, height, and selected spectral reflectance wavebands and calculated indices. Among MG 3 cultivars, the VIS, excluding the 415-nm waveband, NIR, GNDVI, RNDVI, RENDVI, and WI_1 and WI_2 have strong relationships (P < 0.01) with seed yield as well as maturity (Fig. 1A). As the VIS reflectance and calculated water indices decreased in value, yield and maturity increased. As the NIR and other calculated vegetation indices increased, seed
Figure 1. Correlation coefficients (r) for the relationship between seed yield, maturity (mat), lodging (lod), plant height (ht) and spectral reflectance parameters for (A) MG 3 and (B) MG 4 soybean based on cultivar means averaged across environments. crop science, vol. 56, march– april 2016
www.crops.org 635
yield and maturity increased. These patterns also have been observed in other research in wheat, corn, and rice (Reynolds et al., 1999; Lin et al., 2012; Weber et al., 2012). Lodging and plant height were observed to increase as reflectance in the VIS and NIR and WI_2 and WI_3 increased in value; while lodging and plant height decreased with increases in BDNVI, GNDVI and RENDVI values. These results could be due to the association between the NIR portions of these indices and biomass accumulation (Weber et al., 2012). The RNDVI had no significant relationship to lodging or height among MG 3 cultivars. As in the analysis of MG 3 cultivars, VIS reflectance and calculated water indices decreased in value among the MG 4 cultivars as yield and maturity increased; and yield increased as the NIR and GNDVI, RNDVI, and RENDVI vegetation indices increased (Fig. 1B). However, the relationships between maturity and spectral, and indices parameters were stronger in MG 3 cultivars than observed in MG 4 cultivars, suggesting maturity played a larger role in spectral reflectance values among MG 3 cultivars. Lodging and plant height also have significant relationships, with spectral parameters in the VIS and calculated indices. However, plant height tended to exhibit stronger relationships with the VIS and calculated indices than lodging, suggesting that plant height may be a more significant influence on spectral reflectance among MG 4 cultivars than lodging. These results are consistent with previous research findings that lower reflection values in the red region of the spectra correlates with higher grain yields (Weber et al., 2012). Chang et al. (2003) found similar patterns in correlations between VIR and NIR bands and corn yield; however, they found that in early sampling dates, the NIR had a negative correlation with yield, whereas positive correlations were observed in this study. They concluded that negative correlations were due to soil reflectance confounding the visible red region of the spectra. Lodging and plant height relationships with spectral parameters were not as significant as those of seed yield and maturity. The VIS, NIR, and water indices had a strong positive relationship (P < 0.05) with lodging, whereas the GNDVI and RENDVI had significant negative correlations with lodging. Height also had a significant positive relationship with the VIS spectral parameters but no relationship with the NIR. As with lodging, the GNDVI as well as RENDVI had strong negative (P < 0.05) relationships with plant height. These results indicated that seed yield and maturity have the strongest relationships with spectral parameters, and these agronomic traits tend to follow the same pattern of relationships throughout the spectral parameters, suggesting that these traits are also linked to each other and could be confounded. The relationship with seed yield for these reflectance indices is consistent with observations made by many researchers, specifically in corn, where GNDVI was highly correlated with grain yield in Nebraska (Shanahan et al., 2001). The relationship also applies to soybean (Ma et al., 2001) as well as radiation use efficiency 636
and nitrogen use efficiency (Gamon et al., 1992; Trotter et al., 2002; Garbulsky 2011).
Maturity Model Development and Validation It is necessary to evaluate relative maturity and group cultivars based on relative maturity to efficiently select the most promising genotypes without potential bias. Christenson et al. (2014) found that maturity has significant influence on spectral reflectance and yield of soybean, accounting for 62% of the yield variability in seed yield among MG 3 cultivars studied. In this study, relative maturity, expressed in days after planting (number of days from planting to R8), was used as a dependent variable, with reflectance parameters used as predictor variables to create models based on spectral reflectance to estimate maturity in breeding programs. Two models were created using only the selected wavebands using PLS and calculated vegetation indices using selected observation day environmental means, based on genotypic differences, in both maturity groups (n = 220). Using multiple linear regression with forward-backward stepwise, cross validation, and information criterion (Cp, AIC, AICC, BIC) selection, statistical models with waveband and calculated indices were created (Fig. 2). The waveband model was selected using reflectance at 550, 680, 715, 915, 990, and 1140 nm. The model fit the data well (P < 0.01) and explained 50% of the variability in maturity among cultivars (R 2 = 0.50), with a root mean square error (rMSE) of 5.19 d. In the model, wavebands of 680 and 715 nm accounted for the most variation, with 550 and 1140 nm accounting for the least variation among cultivars (data not shown). The model created using the calculated indices GNDVI_4 (550 and 1140 nm), RNDVI_3 (680 and 990 nm), and WI_1 (940 nm/915 nm) was not as efficient as the waveband model, accounting for 43% of the variation in maturity among cultivars and an rMSE of 5.51 d; however, the model was still highly significant (P < 0.01). The index RNDVI_3 accounted for the most variation in maturity, and GNDVI_4 accounted for the least within the model (data not shown). Although the models used different predictor variables, the wavebands used in the waveband model also are expressed in index form, with the exception of 715 nm. Comparing observed maturity to estimated maturity shows a greater range in observed maturity. The 550-nm region has been used extensively as a singular waveband, as well as in a GNDVI index to estimate yield, radiation use efficiency, nitrogen use efficiency, chlorophyll status, and overall plant health (Chappelle et al., 1992; Gamon et al., 1992; Ma et al., 2001; Shanahan et al., 2001; Trotter et al., 2002; Royo et al., 2003; Garbulsky 2011). The red (680 nm) and RNDVI have been the most widely used band regions for NDVI calculation, with the red associated with chlorophyll absorption in soybean (Chappelle et al., 1992; Morrison et al., 1999; Ma et al., 2001), corn (Weber et al., 2012), and wheat (Royo et al., 2003), as well as many other
www.crops.org
crop science, vol. 56, march– april 2016
Figure 2. Relationship between the waveband and index maturity estimation models and the observed maturity among soybean cultivars of MG 3 and MG 4 for all environment means. GNDVI_4 was calculated as [(R1100 – R550) / (R1100 + R550)], RNDVI_3 was calculated as [(R990 – R680) / (R990 + R680)], and WI_1 was calculated as [(R940 / R915)] using canopy reflectance measured at selected waveband regions.
crops. The red edge (715 nm) has been characterized for overall plant health, with senescing plants shifting to higher reflection values, closer to the red region of the spectra, in the red edge and decreasing plant function (Gitelson et al., 2011). The regions around 915 and 1140 nm have been correlated with biomass and are good indicators of membrane integrity and cellular stability (Starks et al., 2006; Zhao et al., 2007; Marti et al., 2007; Gutierrez et al., 2010; Weber et al., 2012). The region around 990 nm also has been associated with plant function and status, but no known physiological function has been associated with this waveband region. Validation of the maturity estimation models was performed using the intercept and b parameters for each variable in the final model (Fig. 2) and creating maturity crop science, vol. 56, march– april 2016
estimations for each development stage cultivar mean, environment means, and maturity group mean (n = 20). Estimates were based on the reflectance for each cultivar in that grouping; i.e., maturity was estimated for MG 3 in A-Irr by using the mean spectral (n = 20) reflectance for each MG 3 cultivar in the A-Irr environment (Table 7). Overall, the validation and utility of both maturity estimation models were mixed. For both models within both maturity groups, the estimation models performed better using only data collected in the later reproductive development stages, or environmental averages, rather than data collected from earlier reproductive development stages. These observations are not surprising because later development stages are closer to cultivar maturity, and environment
www.crops.org 637
Table 7. Soybean maturity and yield estimation model validation by development stage, environment, and maturity group. Model‡ Env.†
Stage
WMM
IMM
WYM
IYM
––––––––––––––––––––– R2 ––––––––––––––––––––
MG 3 A-Irr
A-Dry
B-Irr
B-Dry
C-Irr-1
C-Irr-2
R3-R4 R4 R5 R6 All A-Irr R3-R4 R5 R6 All A-Dry R2 R3 R3-R4 R5-R6 All B-Irr R3 R4 All B-Dry R2 R3 R4–5 R6 All C-Irr-1 R2 R4–5 All C-Irr-2
NS§ 0.26* 0.19* 0.55** 0.58** NS 0.25* 0.48** 0.56** 0.16* NS 0.27* 0.20* NS NS 0.27* 0.35** NS NS NS NS 0.17* NS 0.15* NS 0.50**
R3 R3-R4 R5 R5-R6 All A-Irr R2-R3 R3-R4 R5 All A-Dry R1-R2 R2-R3, R4 R4 R5 R5-R6 All B-Irr R3 R4 R6 All B-Dry R1-R2 R2-R3 R4 R5-R6 All C-Irr-1 R3 R4 All C-Irr-2
NS NS NS NS NS NS NS 0.41** 0.28** NS NS NS 0.30** NS 0.21* NS NS NS NS 0.32** NS NS NS NS NS NS NS 0.32**
All MG 3
NS 0.26* NS 0.64** 0.45** 0.17* 0.37** 0.59** 0.43** NS NS 0.22* 0.26* NS 0.22* 0.38** 0.17* NS NS NS NS NS NS 0.19* NS 0.40**
0.58** 0.62** 0.57** 0.54** 0.66** 0.37** 0.49** 0.63** 0.59** 0.35** 0.38** 0.33** 0.55** 0.56** 0.53** 0.61** 0.63** 0.61** NS 0.23* 0.44** 0.68** 0.47** 0.15* 0.51** 0.85**
0.45** 0.53** 0.43** 0.56** 0.61** 0.58** 0.57** 0.35** 0.59** NS NS 0.32** 0.59** 0.46** 0.55** 0.64** 0.52** 0.15* NS 0.60** 0.62** 0.51** 0.34** 0.59** 0.34** 0.80**
NS NS 0.19** NS NS NS NS 0.51** 0.24* NS NS NS 0.39** 0.16* NS NS NS 0.21* NS 0.30** NS NS NS 0.18* NS NS NS 0.18*
0.54** 0.68** 0.69** 0.61** 0.68** 0.50** 0.33** 0.47** 0.47** 0.23* NS NS 0.75** 0.33** 0.56** 0.55** 0.65** 0.49** 0.67** 0.30** 0.19* NS NS 0.36** NS NS NS 0.71**
0.68** 0.69** 0.61** 0.70** 0.81** 0.34** 0.47** 0.58** 0.57** NS NS NS 0.39** 0.38** 0.81** 0.23* 0.44** 0.42** 0.72** 0.17* NS NS NS NS NS NS NS 0.69**
MG 4 A-Irr
A-Dry
B-Irr
B-Dry
C-Irr-1
C-Irr-2
All MG 4
* Significant at the 0.05 probability level. ** Significant at the 0.01 probability level. † Env, environment; A through C, location; Irr, irrigated; Dry, no irrigation; 1, first irrigated environment; 2, second irrigated environment. ‡ Models to predict soybean maturity and yield; WMM, waveband maturity model; IMM, Index maturity model; WYM, Waveband yield model; IYM, Index yield model. § NS, significance level P > 0.05.
638
totals captured the entire reproductive development phase of the plant. There were also no discernible differences in model accuracy between well-watered and water-stressed environments despite the fact that the range in maturity among genotypes was greater in well-watered environments than in the water-stressed environments (37 d vs. 28 d for MG 3 and 34 d vs. 21 d for MG 4). Among MG 3 cultivars, the best validation of the maturity models was observed in 2011 (Table 7). For most observation days and environmental totals, no statistical difference was observed between the waveband and index models; however, the waveband model performed better for maturity estimation for MG means than the index model (R 2 = 0.50 as compared to R 2 = 0.40). The models also performed better among MG 3 cultivars than among MG 4 cultivars, which may be due to the high association of wavebands and indices with maturity among MG 3 cultivars compared with MG 4 cultivars (Fig. 2). Among MG 4 cultivars, most of the observation days showed no association between estimated maturity and observed maturity. Also, as among MG 3 cultivars, the only differences in model performance was observed for the MG means, with the waveband model fitting the data and having better association with observed yield than the index model. Maturity (R8) is highly dependent on genotype, environment, and management. Plant stresses during the later portions of the growing season could change the physiological status of the plant enough that early observation dates are ineffective at estimating relative maturity. Also, spectral observations in this study were not taken later than R6, so monitoring reflectance between R6 and R8 may be necessary to fully evaluate and more accurately classify the relative maturity of cultivars.
Yield Model Development and Validation Selected observation days from the individual environments were combined into a single dataset with both maturity groups and used to build the final training models for yield estimation. Datasets were means of each environment (n = 220) calculated from singular observation day data, excluding all days without significant differences among cultivars. Models were created using only selected wavebands and calculated indices based on these wavebands, resulting in two overall models. The summary statistics and equations for the two yield estimation models are presented in Table 8. The waveband model explained 44% of the variability in seed yield among cultivars using waveband 715 and 915 nm with an rMSE of 814 kg ha–1 and a dependent yield mean of 3208 kg ha–1. The red-edge region at 715 nm explained 15% of the variability in seed yield among cultivars, whereas reflectance at 915 nm explained 29% of yield variability. In the training model, the b parameter estimate for the 715-nm waveband was negative, indicating that cultivars with high values for 715 nm had decreased yields. This result is consistent with
www.crops.org
crop science, vol. 56, march– april 2016
Table 8. Model summary for estimating yield using the waveband-only and index-only models with each parameter in the model, the amount of variation accounted for by each parameter, summary statistics of the model, and the model equation. Model
Variables
Partial R2
Model R2
rMSE
DM
Equation
-1
––––––- kg ha ––––––Waveband
Index
715 nm
0.15**
915 nm
0.29**
BDNVI_4
0.14**
GNDVI_4
0.03*
RNDVI_1
0.35**
RENDVI_3
0.06*
0.44**
814
3208
Yield = –2980 – (R715 × 23900) + (R915 × 16630)
0.58**
703
3208
Yield = 9710 – (BDNVI_4 × 58154) – (GNDVI_4 × 44401) + (RNDVI_1 × 74053) + (RENDVI_3 × 31779)
* Significant at the 0.05 probability level. ** Significant at the 0.01 probability level.
previous research findings that lower reflection values in the red region of the spectra correlate with higher grain yields (Weber et al., 2012). The 915-nm waveband had a positive b parameter, which was consistent with previous research that showed that increases in the 915-nm waveband region resulted in increased yield (Reynolds et al., 1999; Lin et al., 2012; Weber et al., 2012). The index model was more complex than the waveband model and explained 58% of the variation in seed yield among cultivars, with the BNDVI_4 (415 nm and 1140 nm) explaining 14%, GNDVI_4 (550 nm and 1140 nm) explaining 3%, RNDVI_1 (680 nm and 915 nm) explaining 35%, and RENDVI_3 (715 nm and 990 nm) explaining 6% of the variation in seed yield among cultivars (Table 8). The index model had an rMSE of 703 kg ha-1 and a dependent mean of 3208 kg ha-1. Both models fit the data well and had error values under 25% of the dependent means. The blue, green, and red NDVI have been characterized heavily with yield estimation modeling as well as other plant functions in corn (Shanahan et al., 2001; Royo et al., 2003), soybean (Ma et al., 2001), and wheat (Prasad et al., 2007a, 2007b). Red-edge NDVI has not been characterized as well as the others, but Mourtzinis et al. (2014) found that including the red edge into soybean yield modeling, could increase the accuracy of the model. Also, in considering individual components of the index calculation, it is possible that the index is sensing overall plant health and photosynthetic capacity (red edge) while accounting for biomass and water status of the plant (990 nm). These assumptions are made about the 990-nm waveband region based on the position of the water absorption region of the soybean cultivars in this study (data not shown). Validation of the yield estimation models was performed using the intercept and b parameters for each variable in the final model and creating yield estimations for each development-stage cultivar mean, environment mean, and maturity group mean (Table 7). Estimates were based on the reflectance for each cultivar in that grouping; i.e., yield was estimated for MG 3 in A-Irr using the mean crop science, vol. 56, march– april 2016
spectral (n = 20) reflectance for each cultivar in the MG 3 in the A-Irr environment. Overall, performance of the models was mixed. The models were more informative of MG 3 than MG 4 cultivar performance across development stages and environments, mainly due to the inability of the models to estimate yield among MG 4 cultivars in environments B-Irr, C-Irr-1, and C-Irr-2. Examining Tables 4 and 5 shows that analysis of reflectance values frequently failed to distinguish differences between MG 4 cultivars in these environments, especially in the NIR regions. The models performed best when estimating maturity group means, accounting for 85 and 80% of the yield for the waveband and index model among MG 3 cultivars and 71 and 69% for the waveband and index model among MG 4 cultivars, respectively. These results are most likely linked to environmental conditions changing through the growing season and environmental factors affecting the spectral data more on certain observation days, resulting in the environmental means being more stable than individual observations. Results also indicated that no obvious development stage was better for yield estimation than another. Within a breeding program, it is important to group lines and genotypes by relative maturity when selecting promising genotypes to reduce selection bias that may be related to maturity differences. The cultivars in the yield estimation models presented in Fig. 3 have been placed into relative maturity groupings based on maturity estimates using the previously described waveband or index maturity models. Based on grand mean estimated maturities cultivars were grouped into three relative maturity groupings classified as early, medium, or late maturing. Each of these groupings had roughly a 10-d range in grand mean observed maturity. Also, the correlation between observed yield and observed maturity was 0.14 (P > 0.05), –0.36 (P < 0.01), and –0.42 (P < 0.01) for the early, medium, and late maturing groups, respectively. Grouping the cultivars based on estimated maturity revealed that the waveband yield model (Fig. 3A) fitted the yield of the early- and medium-maturing cultivars more precisely (early R 2 = 0.55, medium R 2 = 0.48) than the
www.crops.org 639
Figure 3. Relationship between estimated and observed seed yield for MG 3 and MG 4 cultivars placed into early, medium and late maturity groupings based on: (A) waveband maturity estimation model and waveband yield estimation model, (B) index maturity estimation model and waveband yield estimation model, (C) waveband maturity estimation model and index yield estimation model, and (D) index maturity estimation model and index yield estimation model.
overall model (Table 8, R 2 = 0.44) when cultivars were not grouped by maturity. The waveband yield model, however, did not fit the later-maturing cultivars as well (R 2 = 0.18), possibly because the later-maturing cultivars had a longer seed fill duration after spectral reflectance data collection ceased, resulting in less precise yield estimates or inability of the maturity model to adequately predict maturity in later-maturing cultivars. Coupling the waveband yield model with the index maturity model resulted in yield estimation performance similar to that of the waveband yield and maturity models (Fig. 3B) (R 2 values ranged from 0.22 to 0.55). Combining the index yield model with either maturity estimation model increased precision of estimating yield (R2 values ranged from 0.37 to 0.67) compared with not adjusting 640
for maturity, or using the waveband yield model (Fig. 3C and D). As with the waveband model, the index model combined with the waveband and index maturity models performed well for the early- and medium-maturity groups but decreased in precision in the later-maturity grouping. Overall, the index yield model combined with either maturity model produced promising yield estimation results, and when coupled with the index maturity model, it estimated the yield of the later-maturing cultivars better than the other three model combinations without sacrificing precision in the other maturity groupings. Coupling the yield estimation models with a maturity estimation model increased precision in distinguishing yield of cultivars compared with the training models that did not account for maturity, except in the later-maturity grouping. Spectral data were not taken to senescence, which
www.crops.org
crop science, vol. 56, march– april 2016
likely impacted our ability to adequately characterize the latermaturing cultivars for maturity and yield. The early-maturity grouping consisted of only six cultivars, which could have influenced model performance in predicting yield and maturity in that grouping compared to the later-maturity grouping which consisted of sixteen cultivars.
CONCLUSIONS
PLS variable selection was used to detect important spectral waveband regions contributing to soybean yield and reduce the spectral datasets to a manageable size for regression analysis. Most of the selected spectral waveband regions correlated with key biophysical and biochemical components proven to contribute to yield in many crops. In this study, VIS, red edge, and portions of the NIR as well as calculated vegetation indices based on these wavebands were important portions of the electromagnetic spectra, and creation of maturity and yield estimation models based on these waveband regions and indices accounted for a significant amount of the variability among soybean cultivar maturity groups and seed yield. No real differences were detected for wavebands or index calculations, as both accounted for a significant amount of variability in maturity and seed yield; however, the waveband models tended to be simpler than index models. Variability was high within and among development stages, water regimes, and environments for validation of training models. No trends were observed for specific development stages being best for maturity or seed yield estimation. Environmental factors such as weather, time of day, and others appeared to affect spectral reflectance tremendously and led the training models to perform better or worse on different validation datasets. Measurements may need to be taken longer into the growing season to fully assess the relative maturity of the cultivars. Performance of the canopy reflectance models for maturity and seed yield was not as consistent as hoped, but it accounted for a significant portion of variability among genotypes in maturity in some environments and in seed yield in most environments. The training models also had a low enough rMSE to distinguish the top 25% highest estimated yielding cultivars, but to capture the top 25% highest observed yielding, 50% of the cultivars would need to be selected. This experiment demonstrated that canopy reflectance can be used to predict relative maturity and seed yield using a diverse set of soybean genotypes. These genotypes allowed for significant variation in model training datasets; however, experiments need to be conducted with genotypes that have less diversity to validate the models. Integrating spectral reflectance measurements into a high-throughput platform also is necessary before this technology can be adopted in breeding programs.
crop science, vol. 56, march– april 2016
Acknowledgments This is contribution no. 15-036-J from the Kansas Agricultural Experiment Station, Manhattan. This research was supported in part by a grant from the Kansas Soybean Commission. The authors want to thank Dr. Brian Diers, University of Illinois, and Dr. James Specht, University of Nebraska, for providing seed of the soybean genotypes for this study.
References Adams, M.L., W.D. Philpot, and W.A. Norvell. 1999. Yellowness index: An application of spectral 2nd derivatives to estimate chlorosis of leaves in stressed vegetation. Int. J. Remote Sens. 20:3663–3675. doi:10.1080/014311699211264 Aparicio, N., D. Villegas, J. Araus, J. Casadesus, and C. Royo. 2002. Relationship between growth traits and spectral vegetation indices in durum wheat. Crop Sci. 42(5):1547–1555. doi:10.2135/cropsci2002.1547 Babar, M., M. Reynolds, M. van Ginkel, A. Klatt, W. Raun, and M. Stone. 2006a. Spectral reflectance indices as a potential indirect selection criteria for wheat yield under irrigation. Crop Sci. 46(2):578–588. doi:10.2135/cropsci2005.0059 Babar, M., M. Reynolds, M. Van Ginkel, A. Klatt, W. Raun, and M. Stone. 2006b. Spectral reflectance to estimate genetic variation for in-season biomass, leaf chlorophyll, and canopy temperature in wheat. Crop Sci. 46(3):1046–1057. doi:10.2135/ cropsci2005.0211 Ball, S., and C. Konzak. 1993. Relationship between grain-yield and remotely-sensed data in wheat breeding experiments. Plant Breed. 110(4):277–282. doi:10.1111/j.1439-0523.1993. tb00590.x Baret, F., and G. Guyot. 1991. Potentials and limits of vegetation indices for LAI and APAR assessment. Remote Sens. Environ. 35:161–174. doi:10.1016/0034-4257(91)90009-U Blackmer, T.M., J.S. Schepers, and G.E. Varvel. 1994. Light reflectance compared with other nitrogen stress measurements in corn leaves. Agron. J. 86:934–938. doi:10.2134/agronj1994.0 0021962008600060002x Chang, J., D.E. Clay, K. Dalsted, S. Clay, and M. O’Neill. 2003. Corn (Zea Mays L.) yield prediction using multispectral and multidate reflectance. Agron. J. 95:1447–1453. doi:10.2134/ agronj2003.1447 Chappelle, E.W., M.S. Kim, and J.E. McMurtrey, III. 1992. Ratio analysis of reflectance spectra (RARS): An algorithm for the remote estimation of the concentrations of chlorophyll a, chlorophyll b, and carotenoids in soybean leaves. Remote Sens. Environ. 39:239–247. doi:10.1016/0034-4257(92)90089-3 Christenson, B.S., W.T. Schapaugh, Jr., N. An, K.P. Price, and A.K. Fritz. 2014. Characterizing Changes in Soybean Spectral Response Curves with Breeding Advancements. Crop Sci. 54:1585–1597. doi:10.2135/cropsci2013.08.0575 Clevers, J.G.P.W. 1997. A simplified approach for yield prediction of sugar beet based on optical remote sensing data. Remote Sens. Environ. 61:221–228. doi:10.1016/S0034-4257(97)00004-7 Datt, B. 1999. Remote sensing of water content in Eucalyptus leaves. Aust. J. Bot. 47:909–923. doi:10.1071/BT98042 Datt, B., T.R. McVicar, T.G. Van Niel, D.L.B. Jupp, and J.S. Pearlman. 2003. Pre-processing EO-1 Hyperion hyperspectral data to support the application of agricultural indexes. IEEE Trans. Geosci. Rem. Sens. 41:1246–1259. doi:10.1109/ TGRS.2003.813206
www.crops.org 641
Daughtry, C.S.T., C.L. Walthall, M.S. Kim, E. Brown de Colstoun, and J.E. McMurtrey, III. 2000. Estimating corn leaf chlorophyll content from leaf and canopy reflectance. Remote Sens. Environ. 74:229–239. doi:10.1016/S00344257(00)00113-9 De Jong, S. 1993. SIMPLS: An alternative approach to partial least squares regression. Chemom. Intell. Lab. Syst. 18(3):251–263. doi:10.1016/0169-7439(93)85002-X Deering, D.W. 1978. Rangeland reflectance characteristics measured by aircraft and spacecraft sensors. Ph.D. diss. Texas A&M Univ., College Station. Dusek, D.A., R.D. Jackson, and J.T. Musick. 1985. Winter wheat vegetation indices calculated from combinations of seven spectral bands. Remote Sens. Environ. 18:255–267. doi:10.1016/0034-4257(85)90061-6 Fehr, W.R., and C.E. Caviness. 1977. Stages of Soybean Development, Cooperative Extension Service, Agriculture and
Home Economics Experiment Station Iowa State University, Ames, Iowa. Gamon, J.A., J. Peñuelas, and C.B. Location. 1992. A narrowwaveband spectral index that tracks diurnal changes in photosynthetic efficiency. Remote Sens. Environ. 41:35–44. doi:10.1016/0034-4257(92)90059-S Gao, B.C. 1996. NDWI—a normalized difference water index for remote sensing of vegetation liquid water from space. Remote Sens. Environ. 58:257–266. doi:10.1016/S00344257(96)00067-3 Garbulsky, M. 2011. The photochemical reflectance index (PRI) and the remote sensing of leaf, canopy and ecosystem radiation use efficiencies. Remote Sens. Environ. 115:281–297. doi:10.1016/j.rse.2010.08.023 Gitelson, A., Y. Gritz, and M. Merzlyak. 2003. Relationships between leaf chlorophyll content and spectral reflectance and algorithms for non-destructive chlorophyll assessment in higher plant leaves. J. Plant Physiol. 160:271–282. doi:10.1078/0176-1617-00887 Gitelson, A.A., A. Vina, D.C. Rundquist, V. Ciganda, and T.J. Arkebauer. 2005. Remote estimation of canopy chlorophyll content in crops. Geophys. Res. Lett. 32:108403 10.1029/2005Gl022688. Gitelson, A.A., P.S. Thenkabail, J.G. Lyon, and A. Huete. 2011. Remote
sensing estimation of crop biophysical characteristics at various scales In: P.S. Thenkabail, et al., editors, Hyperspectral remote sensing of vegetation. CRC Press, New York. p. 329–358. Gutierrez, M., M. Reynolds, W. Raun, M. Stone, and A. Klatt. 2010. Spectral water indices for assessing yield in elite bread wheat genotypes under well-irrigated, water-stressed, and high-temperature conditions. Crop Sci. 50:197–214. doi:10.2135/cropsci2009.07.0381 Gutierrez-Rodriguez, M., M.P. Reynolds, J.A. Escalante-Estrada, and M.T. Rodriguez-Gonzalez. 2004. Association between canopy reflectance indices and yield and physiological traits in bread wheat under drought and well-irrigated conditions. Aust. J. Agric. Res. 55:1139–1147. doi:10.1071/AR04214 Hansen, P.M., J.R. Jorgensen, and A. Thomsen. 2002. Predicting grain yield and protein content in winter wheat and spring barley using repeated canopy reflectance measurements and partial least squares regression. J. Agric. Sci. 139:307–318. doi:10.1017/S0021859602002320 Harshman, R., and M. Lundy. 1994. PARAFAC: Parallel factor analysis. Comput. Stat. Data Anal. 18:39–72. doi:10.1016/01679473(94)90132-5
642
Hatfield, J., and J. Prueger. 2010. Value of using different vegetative indices to quantify agricultural crop characteristics at different growth stages under varying management practices. Remote Sens. 2:562–578. doi:10.3390/rs2020562 Hatfield, J.L. 1983. Remote sensing estimators of potential and actual crop yield. Remote Sens. Environ. 13:301–311. doi:10.1016/0034-4257(83)90032-9 Hong, S.Y., K.A. Sudduth, N.R. Kitchen, H.L. Palm, and W.J. Weibold. 2001. Using hyperspectral remote sensing data to quantify within-location spatial variability [CD ROM]. Proc. Int. Conf. on Geospatial Inf. In: Agric. and Forestry, 3rd, Denver, CO. 5–7 Nov. 2001. Altarum, Ann Arbor, MI. Jackson, T.J., D. Chen, M. Cosh, F. Li, M. Anderson, C. Walthall, P. Doriaswamy, E.R. Hunt. 2004. Vegetation water content mapping using Landsat data derived normalized difference water index in corn and soybean. Remote Sens. of Env. 92:4 475–482. doi:10.1016/j.rse.2003.10.021 Jordan, C.F. 1969. Derivation of leaf area index from quality of light on the forest floor. Ecology 50:663–666. doi:10.2307/1936256 Kaul, M., R.L. Hill, and C. Walthall. 2005. Artificial neural networks for corn and soybean yield prediction. Agric. Syst. 85:1–18. doi:10.1016/j.agsy.2004.07.009 Lin, W.S., C.M. Yang, and B.J. Kuo. 2012. Classifying cultivars of rice (Oryza sativa L.) based on corrected canopy reflectance spectra data using the orthogonal projections of latent structures (O-PLS) method. Chemom. Intell. Lab. Syst. 115:25– 36. doi:10.1016/j.chemolab.2012.04.005 Loss, S.P., and K.H.M. Siddique. 1994. Morphological and physiological traits associated with wheat yield increases in Mediterranean environments. Adv. Agron. 52:229–276. doi:10.1016/ S0065-2113(08)60625-2 Ma, B.L., L.M. Dwyer, C. Costa, E.R. Cober, and M.J. Morrison. 2001. Early prediction of soybean yield from canopy reflectance measurements. Agron. J. 93:1227–1234. doi:10.2134/ agronj2001.1227 Marti, J., J. Bort, G. Slafer, and J. Araus. 2007. Can wheat yield be assessed by early measurements of normalized difference vegetation index? Ann. Appl. Biol. 150:253–257. doi:10.1111/ j.1744-7348.2007.00126.x Morrison, M.J., H.D. Voldeng, and E.R. Cober. 1999. Physiological changes from 58 years of genetic improvement of shortseason soybean cultivars in Canada. Agron. J. 91:685–689. doi:10.2134/agronj1999.914685x Mourtzinis, S., S.C. Rowntree, J.J. Suhre, N.H. Weidenbenner, E.W. Wilson, V.M. Davis, S.L. Naeve, S.N. Casteel, B.W. Diers, P.D. Esker, J.E. Specht, and S.P. Conley. 2014. The use of reflectance data for in-season soybean yield prediction. Agron. J. 106:1159–1168. doi:10.2134/agronj13.0577 Peñuelas, J., F. Baret, and I. Filella. 1995. Semi-empirical indices to assess carotenoids/chlorophyll a ratio from leaf spectral reflectance. Photosynthetica 31:221–230. Peñuelas, J., I. Filella, C. Biel, L. Serrano, and R. Save. 1993. The reflectance at the 950–970 mm region as an indicator of plant water status. Int. J. Remote Sens. 14:1887–1905. doi:10.1080/01431169308954010 Peñuelas, J., R. Isla, I. Filella, and J.L. Araus. 1997. Visible and near infrared reflectance assessment of salinity effects on barley. Crop Sci. 37:1400–1405. doi:10.2135/cropsci1997.001118 3X003700010033x Pimstein, A., A. Karnieli, and D. Bonfil. 2007. Wheat and maize monitoring based on ground spectral measurements and
www.crops.org
crop science, vol. 56, march– april 2016
multivariate data analysis. J. Appl. Remote Sens. 1:013530. doi:10.1117/1.2784799 Prasad, B., B. Carver, M. Stone, M. Babar, W. Raun, and A. Klatt. 2007a. Genetic analysis of indirect selection for winter wheat grain yield using spectral reflectance indices. Crop Sci. 47:1416–1425. doi:10.2135/cropsci2006.08.0546 Prasad, B., B. Carver, M. Stone, M. Babar, W. Raun, and A. Klatt. 2007b. Potential use of spectral reflectance indices as a selection tool for grain yield in winter wheat under Great Plains conditions. Crop Sci. 47:1426–1440. doi:10.2135/cropsci2006.07.0492 Price, J.C., and W.C. Bausch. 1995. Leaf area index estimation from visible and near-infrared reflectance data. Remote Sens. Environ. 52:55–65. doi:10.1016/0034-4257(94)00111-Y Reynolds, M., Y. Manes, A. Izanloo, and P. Langridge. 2009. Phenotyping approaches for physiological breeding and gene discovery in wheat. Ann. Appl. Biol. 155(3):309–320. doi:10.1111/j.1744-7348.2009.00351.x Reynolds, M.P., S. Rajaram, and K.D. Sayre. 1999. Physiological and genetic changes of irrigated wheat in the post-green revolution period and approaches for meeting projected global demand. Crop Sci. 39:1611–1621. doi:10.2135/cropsci1999.3961611x Rincker, K., R. Nelson, J. Specht, D. Sleper, T. Cary, S.R. Cianzio, S. Casteel, S. Conley, P. Chen, V. Davis, C. Fox, G. Graef, C. Godsey, D. Holshouser, G.L. Jiang, S.K. Kantartzi, W. Kenworthy, C. Lee, R. Mian, L. McHale, S. Naeve, J. Orf, V. Poysa, W.T. Schapaugh, G. Shannon, R. Uniatowski, D. Wang, and B. Diers. 2014. Genetic improvement of U.S. soybean in maturity groups II, III, and IV. Crop Sci. 54:1–14. doi:10.2135/cropsci2013.10.0665 Rouse, J.W., Jr., R.H. Haas, J.A. Schell, and D.W. Deering. 1973. Monitoring vegetation systems in the Great Plains with ERTS. p. 309–317. In: Proc. Earth Res. Tech. Satellite-1 Symp., Goddard Space Flight Cent., Washington, DC. 10–14 Dec. 1973. Royo, C., N. Aparicio, D. Villegas, J. Casadesus, P. Monneveux, and J. Araus. 2003. Usefulness of spectral reflectance indices as durum wheat yield predictors under contrasting Mediterranean conditions. Int. J. Remote Sens. 24:4403–4419. doi:10.1080/0143116031000150059 SAS Institute. 2008. Version 9.2. SAS Institute Inc., Cary, NC. Serrano, L., I. Filella, and J. Peñuelas. 2000. Remote sensing of biomass and yield of winter wheat under different nitrogen supplies. Crop Sci. 40:723–731. doi:10.2135/cropsci2000.403723x Shanahan, J.F., J.S. Schepers, D.D. Francis, G.E. Varvel, W.W. Wilhelm, J.S. Tringe, M.R. Schlemmer, and D.J. Major. 2001. Use of remote sensing imagery to estimate corn grain yield. Agron. J. 93:583–589. doi:10.2134/agronj2001.933583x Sims, D.A., and J.A. Gamon. 2003. Estimation of vegetation water content and photosynthetic tissue area from spectral reflectance: A comparison of indices based on liquid water and chlorophyll absorption features. Remote Sens. Environ. 84:526–537. doi:10.1016/S0034-4257(02)00151-7 Starks, P.J., D. Zhao, W.A. Phillips, and S.W. Coleman. 2006. Development of canopy reflectance algorithms for real-time prediction of Bermudagrass pasture biomass and nutritive values. Crop Sci. 46:927–934. doi:10.2135/cropsci2005.0258 Trotter, G.M., D. Whitehead, and E.J. Pinkney. 2002. The photochemical reflectance index as a measure of photosynthetic light use efficiency for plants with varying foliar nitrogen contents. Int. J. Remote Sens. 23:1207–1212. doi:10.1080/01431160110106096 Tucker, C.J., and P.J. Sellers. 1986. Satellite remote sensing of
crop science, vol. 56, march– april 2016
primary production. Int. J. Remote Sens. 7:1395–1416. doi:10.1080/01431168608948944 Tucker, C.J. 1979. Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens. Environ. 8:127–150. doi:10.1016/0034-4257(79)90013-0 Tucker, C.J. 1980. Remote sensing of leaf water content in the near infrared. Remote Sens. Environ. 10:23–32. doi:10.1016/00344257(80)90096-6 Vollmann, J., H. Walter, T. Sato, and P. Schweiger. 2011. Digital image analysis and chlorophyll metering for phenotyping the effects of nodulation in soybean. Computers and Electronics in Agriculture 75:190–195. doi:10.1016/j.compag.2010.11.003 Walter, A., B. Studer, and R. Kolliker. 2012. Advanced phenotyping offers opportunities for improved breeding of forage and turf species. Ann. Bot. 110:1271–1279. doi:10.1093/aob/ mcs026 Weber, V.S., J.L. Araus, J.E. Cairns, C. Sanchez, A.E. Melchinger, and E. Orsini. 2012. Prediction of grain yield using reflectance spectra of canopy and leaves in maize plants grown under different water regimes. Field Crops Res. 128:82–90. doi:10.1016/j.fcr.2011.12.016 Wenjiang, H., W. Jihua, W. Zhijie, Z. Jiang, L. Liangyun, and W. Jindi. 2004. Inversion of foliar biochemical parameters at various physiological stages and grain quality indicators of winter wheat with canopy reflectance. Int. J. Remote Sens. 25(12):2409–2419. doi:10.1080/01431160310001618095 White, J.W., P. Andrade-Sanchez, M.A. Gore, K.F. Bronson, T.A. Coffelt, M.M. Conley, K.A. Feldmann, A.N. French, J.T. Heun, D.J. Hunsake, M.A. Jenks, B.A. Kimball, R.L. Roth, R.J. Strand, K.R. Thorp, G.W. Wall, and G. Wang. 2012. Field-based phenomics for plant genetics research. Field Crops Res. 133:101–112. doi:10.1016/j.fcr.2012.04.003 Wiegand, C., A. Richardson, D. Escobar, and A. Gerbermann. 1991. Vegetation indexes in crop assessment. Remote Sens. Environ. 35:105–119. doi:10.1016/0034-4257(91)90004-P Wold, H. 1966. Estimation of principal components and related models by iterative least squares. In: P.R. Krishnaiaah, editor, Multivariate analysis. Academic Press, New York. p. 391–420. Zarco-Tejada, P.J., J.R. Miller, T.L. Noland, G.H. Mohammed, and P.H. Sampson. 2001. Scaling-up and model inversion methods with narrow-band optical indices for chlorophyll content estimation in closed forest canopies with hyperspectral data. IEEE Trans. Geosci. Rem. Sens. 39:1491–1507. doi:10.1109/36.934080 Zhao, D., K.R. Reddy, V.G. Kakani, J.J. Read, and S. Koti. 2007. Canopy reflectance in cotton for growth assessment and lint yield prediction. Eur. J. Agron. 26:335–344. doi:10.1016/j. eja.2006.12.001
www.crops.org 643