Efficacy of different indices derived from spectral reflectance of wheat for nitrogen stress detection

Abstract

The productivity of cereal crops is mainly related to their nitrogen status. It is hypothesized that the spectral reflectance data could be used to predict wheat nitrogen status with spectral indices and that their performance depends on the nature of the interaction of the solar radiation with the crop canopy. A wheat crop was raised with 12 levels of nitrogen treatments: 0, 15, 30, 40, 50, 60, 70, 80, 90, 100, 110, and 120 kg ha⁻¹, with uniform phosphorous and potassium nutrition and uniform water and management practice. The spectral reflectance measurements of the crop canopy were taken at 5 nm intervals, throughout the crop growth period. Different spectral indices, both broadband (ratio as well as orthogonal) and hyperspectral indices were computed throughout the growing season. Canopy Colour Difference (ΔE), an index developed from the entire visible region and hence broader than the spectral indices developed hitherto, was also estimated from the reflectance data. Simple linear relationships developed between spectral indices versus applied nitrogen levels as well as the plant nitrogen content revealed that the hyperspectral indices are less sensitive in comparison to broadband indices. The result was reinforced by a higher correlation between the colour difference, NDVI and Greenness Index with plant nitrogen level/content, as opposed to hyperspectral indices.

Keywords:

• Nitrogen status as nitrogen level/plant nitrogen content
• wheat
• broad band indices
• hyperspectral indices
• colour difference

Introduction

The spectral reflectance from a plant canopy depends on the interaction of electromagnetic radiation with the biophysical and physiological parameters and hence contains large amount of information. However, these data being voluminous and widely varying in different wavelengths, it should be compressed into spectral indices for useful applications. Assessment of nutrient status of plant is traditionally based on destructive sampling and point measurements. However, for large-scale estimates of crop performance, simulation models serve as an efficient tool, the use of which is primarily limited because of the lack of spatial input data pertaining to plant/crop nutrition. Diagnostic methods like remote sensing techniques that take advantage of the optical properties of leaves are another possibility to assess the nitrogen status of plants especially a larger spatial scales (Tarpley et al. 2000, Osborne et al. 2002, Wright et al. 2004, Xue et al. 2004). Remote sensing could provide inexpensive large area estimates of crop growth condition and can be used to monitor nutrient and other stresses, utilizing the information contained in the spectral reflectance data they measure. Several indices have been developed to compress the information contained in the reflectance patterns of plants (Gates et al. 1965, Asrar et al. 1984, Huete et al. 1985, Goward et al. 1985, Sellers, 1987, Baret & Guyot 1991, Galvao et al. 1999, Ceccato et al. 2002, Ma et al. 2001, Osborne et al. 2002). The indices derived from spectral data of more than 60 nm bandwidths are known as broadband indices (Baret et al. 1987). Most of the currently used indices are broadband indices that are related to several plant biophysical processes like photosynthesis through empirical or physical models. On the other hand, there are hyperspectral indices, which are obtained from the spectral reflectance at specific wavelengths. There are studies that indicate that broadband indices could be better predictors of photosynthetic capacity than narrow band indices because the sensitive bands saturate too rapidly (Yoder 1992). On the contrary, broadband indices can be poor indicators of physiological changes at fine temporal scales (intra-seasonal) under stress conditions because most of the diurnal physiological changes are manifested at specific wavelengths of the reflectance spectra (Running & Nemani 1988, Gamon et al. 1992). Even though stronger relationships are reported with newer indices, most of the applied research utilises NDVI and SR (Phillips et al. 2004), which are broad band indices. The interaction of electromagnetic radiation with the plant canopy being a complex phenomena depending on the canopy health, nutrition and structure, it is necessarily required to assess the efficacy of a spectral index to predict a plant physiological condition in a comparative manner. Among plant nutrients, nitrogen is the most important and essential element for crop growth and productivity. Nitrogen is an integral part of chlorophyll, which is the primary absorber of light energy needed for photosynthesis. It is hypothesized that the spectral reflectance data could be used to predict wheat nitrogen status with efficient spectral indices and their performance depends on the direct or indirect effects of plant nutrient status on its physiological regime. Thus nitrogen stress might affect the spectral reflectance in a broader manner than in specific regions. The present study is a comparative assessment of broadband indices vis-à-vis hyperspectral indices for plant nitrogen stress/content prediction. The results developed in this study could assist the development of efficient algorithms for crop simulation models.

Material and methods

The experiment was carried out in the Indian Agricultural Research Institute farm during the period 2000–2001. For this experiment, there was a prerequisite to deplete the soil nitrogen to reduce the soil nitrogen variability as well as to increase the crop response. Maize, which is generally considered as an exhaustive crop was grown (in the summer of 2000) without any fertilizer application followed by flooding the fields (twice) to leach residual nitrogen. Soil samples which were collected after the first flooding showed that there was a reduction of available nitrogen in the top layers and an increase in lower layers. After subsequent floodings, it was found that the initial nitrogen status in the soil was reduced by about 50%. It was assumed that the field was devoid of any nitrogen variability after this stage. A wheat crop (Triticum aestivum L. cv HD2329) was taken up in the following winter (2001) with varying levels of nitrogen fertilization with 12 treatments: 0, 15, 30, 40, 50, 60, 70, 80, 90, 100, 110, and 120 kg N ha⁻¹ of nitrogen and three replications. Recommended doses of P (40 kg ha⁻¹) and K (20 kg ha⁻¹) were applied uniformly to all the treatments. All plant protection measures and irrigation and other cultural practices were followed as recommended. Since irrigation was applied uniformly, any water related constraints were assumed to be negligible. However in field conditions, there are several causes of errors arising out of field variabilities. To minimize the errors, a representative unit area (1m²) was earmarked in each treatment plot keeping in view the uniformity of crop growth where spectral reflectance measurements were taken throughout the growing season.

Spectral observations

The reflected radiation from the wheat canopy and soil were taken with LICOR LI-1800 Portable Spectroradiometer, which scans continuously from 330 nm to 1100 nm at 5 nm intervals. The measurements were made throughout the crop growth period (twice a week) on clear sky days at around local noon, 1 m above the crop canopy with the sensor facing perpendicular to the crop canopy. The incident solar radiation was also taken simultaneously during canopy measurements. While taking the observations, care was taken not to cast any shadow over the area being scanned. The crop spectral reflectance, in each treatment (mean of the three replications) was calculated by taking ratio of the measured canopy reflected radiation to the incident solar radiation. Since the study was mainly focused to understand the nitrogen stress predictability of various spectral indices based on their nature, three different approaches of index development were used: Ratioing, Orthogonolization and Euclidian Distance approach. The reflectance data thus collected in a hyperspectral mode was integrated to obtain reflectance in the bandwidth which corresponds to the Multi Spectral Scanner (MSS) sensor of the Landsat satellite, for calculating some of the ratio as well as orthogonal indices (the broad band ones). This integration of the spectral data and calculation of both broadband and hyperspectral indices was done using a programme developed in C language (Table I). A potential remotely sensed measurement is the point of maximum slope in vegetation reflectance spectra, which occurs between wavelengths of approximately 680–740 nm (Salisbury et al. 1987). This point refers to the red edge and marks the boundary between chlorophyll absorption in red wavelengths and leaf scattering in near infrared wavelengths. The red edge is usually defined as the maximum of the first derivative of the reflectance spectra. Red Edge Slope (RES) represents the rising trend of the reflectance curve from the ‘valley’ (Red region) to the ‘peak’ (infrared region). The differential reflectance plotted against wavelength gives a peak in the 650–740 nm region. This peak (red edge) shifts towards the red region (red shift) for a healthy crop and towards the blue region (blue shift) for stressed vegetation. This RES value was obtained after differentiating and plotting the differential reflectance against the wavelength.

Table I Spectral Indices evaluated for their sensitivity to plant nitrogen status in the present study.

Indices evaluated	Type	Formulae	Author
Simple Ratio (SR)	Broad Band	IR/R	Birth & McVey 1968
Normalized Difference Vegetation Index (NDVI)	Broad Band	(IR-R)/(IR + R)	Rouse et al. 1973
Normalized Pigment Chlorophyll Ratio Index (NPCI)	Hyperspectral	(IR 680 -R 430 )/(IR 680 +R 430 )	Peñuelas et al. 1994
Physiological Reflectance Index (PRI)	Hyperspectral	(IR 550 -R 530 )/(IR 550 +R 530 )	Peñuelas et al. 1994
Water Band Index	Hyperspectral	R 970 /R 900	Peñuelas et al. 1994
Hyperspectral NDVI (hNDVI)	Hyperspectral	(IR 900 -R 685 )/(IR 900 +R 685 )	Rouse et al. 1973
Hyperspectral SR (hSR)	Hyperspectral	R 900 /R 680	Peñuelas & Filella 1998
Red Edge Position REP	Hyperspectral	Slope	This study
Greenness (orthogonal index)	Broad Band	A 2,1 X 1 +A 2,2 X 2 +A 2,3 X 3 +A 2,4 X 4 .	Kauth & Thomas 1976
Colour Difference (ΔE120)	Broad Band		This study

To derive orthogonal Indices like Soil brightness (SBI), Greenness (GVI) and Yellowness (YVI), a sequential orthogonalization was performed by the procedure outlined by Kauth and Thomas (1976) known as KT transformation, to generate the information defined by all MSS bands (only Greenness results are considered in this study). The coefficients for Brightness, Greenness and Yellowness were calculated by the procedure outlined by Jackson using the spectral reflectance from a dry and wet soil as well as vegetation at maximum vegetation stage (Jackson 1983, Jackson et al. 1983). The spectral Indices were calculated throughout the growth period of the crop (expressed as Days After Sowing, DAS) for different nitrogen treatments. Table I shows the summary of the spectral indices calculated and evaluated in this study.

Canopy colour discrimination in uniform colour scales

Colour change/difference is a biophysical parameter which is manifested in almost any natural phenomena, the quantification of which is extremely difficult. It characterizes the visible region of the reflected spectra containing information in blue, green and red making it a ‘broader’ spectral parameter. CIE (Commission Internationale de l'Eclairage) determined standard coefficients that are used worldwide to quantify colour (CIE, 1976).The values used by CIE are called L*, a* and b* and the colour measurement method is called CIELAB. L* represents the difference between light (where L* = 100) and dark (where L* = 0). a* represents the difference between green (−a*) and red (+a*), and b* represents the difference between yellow (+b*) and blue (−b*). This represents a three dimensional space with each colour defined in a coordinate system and a colour difference defined by a Euclidean distance. As the variations in the observed colour data with change in Nitrogen content were very subtle, spectroradiometric data were converted to CIE Lab using the CIE X, Y, Z tristimulus values(CIE, 1931). The following equations were used in converting tristimulus values to CIE L* a* b* colour system:

1

2

3

where Xn, Yn, Zn are the tristimulus specifications of the standard illuminant ie Sun (D65).

X, Y, and Z tristimulus values derived from the reflectance spectra by using a software developed in DOS environment by LICOR. The color difference (ΔE) between the a treatment and healthy sample was then obtained by (Hunter 1948 1958, Zhang & Kamdem 2000)

4

where^, ΔL, Δa, and Δb are the differences in the values of L*, a* and b* of a nitrogen treatment from a datum (120 kg N ha⁻¹ considered as unstressed). Thus the colour difference between the healthy and stressed crop is expressed in terms of ΔE, which is obtained from the spectral reflectance patterns of the crop.

Total plant nitrogen content and other physiological parameters

Biomass (above ground) was harvested from a unit area (1 m²) from each treatment at the six growth stages of the wheat crop. The samples were oven dried and dry biomass was estimated. The dry samples were powdered in an electrical mixer and were used for the estimation of total plant nitrogen using N-autoanalyser expressed as nitrogen in the plant sample per unit area (% Plant Nitrogen x Dry biomass per unit area). Other parameters like Leaf Area Index and Total Chlorophyll were also determined but they are not explicitly described in this paper. However, they are used to augment our hypothesis wherever required.

Results and discussion

The experiment was initiated with the assumption that the nitrogen status of the soil at the start of the experiment was reduced to such a level that a good response to the applied nitrogen would be obtained. Although we took the measurements for each treatment, only the results for the lowest (0 kg N ha⁻¹) and the highest (120 kg N ha⁻¹) level are plotted in the figure to make them more readable. However, the remaining treatments were found to be intermediate between the responses of extreme nitrogen levels.

Time series of total plant nitrogen, spectral reflectance and spectral indices

Variations in the aboveground biomass, plant nitrogen content and pigment concentration were translated into clear changes in the spectral signature during the crop period. The total plant nitrogen per unit area (it has to be noted that the percentage of nitrogen decreased with time) showed an increasing trend with a sigmoid pattern, during the crop period. The highest value was observed in 120 kg N ha⁻¹ (48.68 g m⁻²), and lowest for 0 kg N ha⁻¹ (13.58 g/m²), at about 100–118 DAS. Initially, the plant nitrogen content was similar in all the treatments up to 67 DAS (Figure 1a,b) and thereafter, a sudden increase in total plant nitrogen per unit area (hereafter plant nitrogen status) was observed between late jointing stage and maximum vegetation stage. As reviewed by Bergman (1992), the nutrient concentration in a leaf changes with crop age and accumulated biomass. Plant Nitrogen (%) decreases as plants get older due to the dilution effect caused by increased biomass growth at a given level of nitrogen availability (Smith 1962).

Figure 1.  (a) Relationship between plant nitrogen content and dry matter per unit area. (b) Variation of the plant nitrogen content per unit area for different nitrogen levels during the crop growth.

The spectral reflectance from 330 to 1110 nm at 5 nm interval was computed throughout the crop duration, twice a week. The spectral reflectance of the canopy provides several options for the derivation of their physiologic status by quantifying the patterns in the visible and infrared regions of the electromagnetic spectrum. The reflectance patterns at the five growth stages of wheat for 0 and 120 kg N ha⁻¹ treatments are presented in Figure 2. To generalize, there are two mechanisms which cause difference in the reflectance patterns among the treatments. One is the pigment characteristics and photosynthetic differences (in the visible region) and the second is the plant leaf/canopy structural characteristics (in the NIR region). In the visible region, the higher values in the ‘stressed’ treatments as well as in the initial and final stages of the crop might represent lower photosynthetic capacity (in particular, chlorophyll concentration) due to the low nutrient availability (in stressed) and senescence (as the crop progresses). In the NIR region, the difference between the treatments might be attributed to a structural development (leaf as well as canopy). In general, the spectral reflectance of 120 kg N ha⁻¹ is higher in the NIR region than 0 kg N ha⁻¹. This difference is higher in the maximum vegetation stage than initial and final stages of the crop (Figure 2d). Although the spectral signature showed characteristic features of green vegetation (lower reflectance in the red region), the average reflectance in the visible region was relatively high and the NIR reflectance was relatively low. As the canopy developed, reflectance in the visible region decreased and NIR reflectance increased and showed a maximum contrast at maximum vegetation. Thereafter, the crop was flowering/senescent (Figure 2e,f) and the spectral signature showed a steady decrease across the red edge region. While increase in red reflectance can be attributed to the decrease in chlorophyll content resulting from lower nitrogen supply (Filella et al. 1995), decrease in NIR reflectance is mostly due to decrease in LAI and green biomass, as has been widely reported for wheat crops (Asrar et al. 1984, Jensen et al. 1990, Fernández et al. 1994, Peñuelas et al. 1996).

Figure 2.  Spectral reflectance pattern of the crop canopy at different phenological stages for the wheat crop: (a) Crown Root Initiation Stage, (b) Tillering Stage, (c) late Jointing Stage, (d) Maximum Vegetation Stage, (e) Flowering Stage, (f) Dough Stage for different nitrogen levels (only 120 kg N ha⁻¹ and 0 kg N ha⁻¹ are exhibited). The x -axis for all the graphs shows the wavelength in nm. and y-axis reflectance.

The Simple Ratio (SR) values for all the treatments initially increased, reached a peak value and there after decreased, during the crop growth period. The maximum value reached by 120 kg N ha⁻¹ was 14.96 and by 0 kg N ha⁻¹ was 12.34 (Figure 3a). The Normalized Difference Vegetation Index (NDVI), gave a smoother pattern, compared to SR. NDVI also showed difference with nitrogen levels, but this difference was not prominent between the extreme nitrogen levels (Figure 3b). This could be explained based on the LAI – SR and LAI- NDVI relation. Many researchers have demonstrated that the SR increases linearly with LAI, while NDVI shows a curvilinear response that saturate at LAI > 3. The results we obtained also were in agreement with those obtained in previous wheat nutrition studies based on spectral reflectance (Asrar et al. 1984, Wiegand et al. 1991, Serrano et al. 2000). Thus, we speculate that saturation might be the reason for a lower variability in NDVI than in SR. In other words, the addition of more leaf layers to the canopy does not entail large changes in NDVI (Sellers 1987). The Normalized Pigment Chlorophyll Ratio Index (NPCI) time series showed a decrease and then increase in all the treatments, with the lowest value at around 90 DAS, corresponding to the maximum vegetation stage (Figure 3c). This index has been reported to be highly correlated to variation in the ratio of chlorophyll and total pigments concentration (Peñuelas et al. 1994). This is justified because in most plant species, two types of chlorophyll (a and b) determine the reflectance mainly by absorption of blue and red light, and to a lesser extent green light, provided that leaves and stems are functioning well (de Boer 1993). The time series resulted here may be because of the changes in leaf chlorophyll, which is directly dependent on the plant nitrogen status.

Figure 3.  Time Series for different indices evaluated in the present study throughout the duration of the Wheat crop for different nitrogen levels (only 120 kg N ha⁻¹ and 0 kg N ha⁻¹ are exhibited). The x-axis for all the graphs shows the time expressed as Days After Sowing (DAS). •-•-• – 100 kg N ha⁻¹, x-x-x – 0 kg N ha⁻¹.

Water status of plants can be related to the plant nitrogen status. To explore this the Water Band Index (WBI) trend was analyzed. It was seen that plants with high nitrogen status have lower values for WBI and vice versa. WBI, designed to capture the trough at 970 nm in the reflectance spectrum due to water absorption (Danson et al. 1992, Penuelas et al. 1993), was higher in the initial or later stages or a nitrogen-stressed crop (Figure 3d). WBI for all the treatments followed a decreasing pattern up to the maximum vegetation stage and thereafter increased. There was also a ‘dip’ in the WBI time series in all tratments at 30–40 DAS coinciding with the days following the first irrigation. It points to the fact that WBI is directly related to the plant water content and also an indicator of a probable interaction between nitrogen and water stress.

The Physiological Reflectance Index (PRI) showed an erratic trend (Figure 3e) for all the treatments, phenological stages and the differences among the treatments were also small. In the initial stages of the crop, the PRI was more or less similar for all the treatments, but the trend was erratic at later stages. Peñuelas et al. (1994) have reported that PRI closely follows the diurnal variations in the xanthophyll and other pigments responsible for peak reflectance at the green (530–550 nm) wavelength regions. The index may hence follow diurnal changes in xanthophyll pigment and photosynthetic rates of control in nitrogen stressed levels.

Hyperspectral Vegetation Index (hSR) showed very similar trend for all the treatments but the difference in the values between 0 and 120 N kg ha⁻¹ was higher in the maximum vegetation stage than for the initial and later stages of the crop. It can be noted from Figure 3f that the sensitivity of hSR was higher than SR and it has clearly differentiated the nitrogen treatments. So also, Hyperspectral NDVI (hNDVI) showed a similar trend in comparison to its broadband version, NDVI. The peak (90 DAS) that was not clearly distinct in the NDVI time series was distinct in hNDVI time series (Figure 3g). Like hSR, hNDVI also showed a higher sensitivity than NDVI. The red wavelength region between 670 and 690 nm is the point of maximum chlorophyll absorption, while 900 nm is known to be point where maximum NIR reflectance occurs and with minimal influence of the atmospheric and canopy moisture affects (Penuelas et al.1996).

One efficient method for minimizing the effects of soil background on canopy spectral signatures is the use of high-resolution derivative spectra (Demetriades-Shah et al. 1990, Elvidge & Chen 1995). The derivative in this case is the instantaneous rate of change of the reflectance with respect to wavelength. Peñuelas et al. (1994) identified an increase of slope in reflectance between 500 and 600 nm in nitrogen-stressed sunflower leaves from unstressed leaves and a decrease and shift in slope between 650 and 750 nm. It was found that the peak reflectance shifts towards a shorter wavelength in nitrogen stressed plants (Figure 4). The amplitude of the peak also changed with the nitrogen treatment, where a lesser amplitude was seen in the nitrogen stressed plants. It was seen that the difference in the amplitude between the lowest and highest nitrogen treatments increased up to the peak vegetation stage. The Red Edge Slope (RES) increased sharply and then decreased in all the treatments indicating that the slope of the reflectance curve in the Red Edge is dependent on the plant nitrogen status. The maximum value of the RES was for 120 kg N ha⁻¹ 90.32×10⁻⁴, while the maximum value for 0 kg N ha⁻¹ was 53.6×10⁻⁴. At peak growth stage onwards, the red edge position moves to a shorter wavelength in nitrogen stressed treatments. Figure 4 shows the plot between the first order derivative of reflectance (differential reflectance) and wavelength, where a ‘blue shift’ can be observed in stressed 0 kg N/ha treatment (730 nm), compared to healthy 120 kg N/ha treatment (735 nm) at peak growth stage (90 DAS) with a difference of 5 nm.

Figure 4.  Variation of the first derivative of the reflectance for different nitrogen levels showing the change in Red Edge Position at Maximum Vegetation Stage (90DAS).

The Greenness, which is an orthogonal index, and a function of relative vegetation amount, also followed a trend similar to SR and hSR. The highest maximum was shown by 120 kg N ha⁻¹ (25.563) and the lowest maximum (14.978) was shown by 0 kg N ha⁻¹. The peak value was however reached at an earlier date in all the plots (∼80 DAS, Figure 3h), as opposed to other indices. Other orthogonal indices like Soil Brightness Index (SBI) and Yellowness, although calculated, was not directly used in this study because they contain information on other aspects like soil reflectance more than vegetation, which is beyond the scope of this paper.

The Canopy Colour Difference (ΔE) was computed only for the three stages of the crop viz. Initial, Maximum Vegetation Stage and Senescence and not for the entire the crop period because of computational difficulties. But, this did not affect the comparative study of the predictability of different indices for nitrogen stress because all the indices were compared at 90 DAS, which corresponds to the Maximum Vegetation Stage (the details of which are explained in the next section). The temporal variation of ΔE, derived from the L*, a* and b* values for three stages of crop growth with 120 kg N ha⁻¹ as the reference is presented in Figure 5. In the initial stages of crop growth, the difference in colour between the maximum nitrogen level and the lower levels was of the order of 1.5. As the crop growth progressed, this value reached a maximum of 4.5. At senescence, the difference in colour was reduced from its peak stage to 2.5, however having a value slightly higher and erratic than the initial stage.

Figure 5.  Variation of the Colour Difference of the crop canopy with different nitrogen levels for the three stages of the crop.

Plant nitrogen predictability of the spectral indices

Correlation studies were made between different indices with both nitrogen levels as well as plant nitrogen content. It was noted that the former one was more sensitive to various indices than latter because the difference in the plant nitrogen content among the different treatments was lesser in comparison to the corresponding nitrogen levels. It was observed that the sensitivity of each index varied differently at different stages of the crop, as seen in Figure 6a, where SR is shown to be more sensitive between the 77 DAS, 96 DAS compared to later stages of the crop growth. The steep increase in SR may be because of the rapid increase in LAI during these periods (Figure 6b). Hence, a relationship between the indices and nitrogen status may not be relevant and misleading if a comparison is made overlooking a particular crop stage.

Figure 6.  (a) Response of Band Ratio to plant nitrogen content at different stages of crop growth. (b) Response of Leaf Area Index (LAI) to plant nitrogen content at different stages of crop growth.

To find out the exact day of maximum vegetation, all the datasets were subjected to 31 different types of iterative curve fitting procedures using CurveExpert V 1.3. In most of the cases, the pattern was sinusoidal. The rising and the falling limbs of the SR time series were fitted to two linear equations, and the intersection point of these equations was noted as the maximum vegetation stage. The peak vegetation stage from the data was more or less at 90 DAS (88.16 DAS, to be precise) for all the treatments. Since the difference (spectral reflectance, indices, and color difference) between the treatments was highest at the peak vegetation stage (90 DAS), the spectral indices derived on this day of crop growth was used for carrying out further analyses (Table II).

Table II Determination of the day of Maximum Vegetation from Simple Ratio time series of different treatments.

N.level (kg ha^−1)	Rising limb	r^2	Falling limb	r^2	Day of Max. Veg.	Peak SR (curve)	Peak SR (data)
0	Y = 0.1696× − 2.99	0.979	Y = − 0.248× + 33.54	0.973	87.46	11.83	12.23
15	Y = 0.146× − 2.08	0.989	Y = − 0.217× + 29.57	0.979	86.97	10.64	11.14
30	Y = 0.147× − 1.85	0.993	Y = − 0.228× + 31.24	0.979	88.01	11.15	11.10
40	Y = 0.136× − 1.60	0.986	Y = − 0.231× + 31.57	0.982	90.2	10.71	11.88
50	Y = 0.180× − 2.55	0.987	Y = − 0.245× + 33.74	0.983	85.20	12.83	12.83
60	Y = 0.175× − 3.15	0.992	Y = − 0.231× + 33.59	0.998	89.72	12.86	13.05
70	Y = 0.186× − 3.08	0.995	Y = − 0.257× + 35.91	0.991	87.73	13.31	13.31
80	Y = 0.164× − 2.41	0.989	Y = − 0.244× + 33.89	0.986	88.71	12.21	13.02
90	Y = 0.207× − 4.14	0.988	Y = − 0.296× + 40.38	0.978	88.40	14.18	14.57
100	Y = 0.144× − 2.77	0.980	Y = − 0.227× + 31.18	0.983	91.36	10.44	12.02
110	Y = 0.217× − 3.88	0.988	Y = − 0.299× + 40.39	0.977	85.65	14.75	14.08
120	Y = 0.232× − 4.78	0.980	Y = − 0.318× + 44.09	0.985	88.53	15.78	14.96

The broadband spectral indices like SR, NDVI (ratio indices) at 90 DAS when linearly fitted to both nitrogen levels as well as plant nitrogen content, gave a significant positive correlation (Figure 7). As widely reported, consequent to lower amounts of green biomass, vegetation under nutrient stress shows a decrease in reflectance in the near-infrared band (>750 nm), and increased reflectance in the red band, and a blue shift on the red edge. Our results were in accordance with the previous studies (Gamon et al. 1992, Peñuelas et al. 1997, Wright et al. 2001 2004), which reported similar trends in the spectral signature, which lead to a decrease in either NDVI or SR. But it should be noted that the predictability of these indices varied differently for plant nitrogen content and nitrogen levels. It could be seen from the Figure 7 that they perform better with the nitrogen level than the plant nitrogen status. This could be accounted for the higher statistical variability in the nitrogen level than the plant nitrogen status. However, it should be reconciled that the nitrogen level, although not a direct indicator of plant nutrient condition, it is a commonly used agronomic parameter for crop nutrient application, and hence we tested it in this study. It seems that NDVI (r²=0.7559) is a better predictor of nitrogen level than SR (r²=0.7115) while for plant nitrogen status, NDVI (r²=0.4848) predicted lower values than SR (r²=0.5201). This might be because NDVI saturates beyond a certain level of plant nitrogen status, similar to the NDVI-LAI relationship, as opposed to SR. Our studies are in accordance with the results of Wright et al. (2004).

Figure 7.  Correlation coefficients between different indices evaluated in this study, Nitrogen level and Nitrogen content.

The hyper spectral versions of these indices like hSR and hNDVI on the other hand, gave a modest realationship with both nitrogen level (r²=0.6424 and 0.5487, respectively) or plant nitrogen status (r²=0.5829 and 0.4049, respectively). Serrano et al. (2000) defined SR and NDVI according to our hyperspectral definition and opined that “under nitrogen stress conditions, a relationship between these indices and canopy parameters like LAI and fPAR were affected by plant chlorosis and are unreliable”. We also speculate that a hyperpectral index cannot capture all of the plant conditions in a robust manner when the canopy exhibits other symptoms like chlorosis. As seen from Figure 7, NPCI predicts the plant nitrogen attributes with lesser efficiency than NDVI or SR for both nitrogen level (r²=0.5516) and plant nitrogen status (r²=0.3476). The index is likely to respond to variations in nitrogen content in the plant, as nitrogen is known to encourage the formation of leaves with large amounts of chlorophyll and other photosynthetic components per unit area (Evans 1983, Lawlor et al. 1987 1989). NPCI, defined to assess the carotenoid/chlorophyll ratio was found to be higher in N-limited leaves. Since nitrogen stress results in loss of chlorophyll, Thomas and Oerther (1972) used this approach to estimate nitrogen status. Since the carotenoid pigments have absorption maxima in the 300–500 nm region and the chlorophyll in the red region around 680 nm, we thought that the combination of these two would be useful as an index for nitrogen stress detection, because nitrogen forms an important component of these pigments. However it can be concluded that a hyperspectral index like NCPI, can predict plant nitrogen status only weakly, even though it is strongly dependent on plant chlorophyll content (r²=0.888, relationship not shown in this paper).

Physiological reflectance Index and Water band Index gave only a weak relationship with both nitrogen level (r²=0.1805 and 0.147, respectively) or plant nitrogen status (r²=0.2275 and 0.3451, respectively). However, it is important to note that as opposed to other indices, these indices gave a higher correlation to the plant nitrogen content than nitrogen levels. This might be because of a combined effect of a closer relationship with the plant nitrogen status and other physiological parameters, which vary diurnally. Some workers like Sims and Gamon (2002) suggested a close relationship between these two indices and have demonstrated that when combined, they predict photosynthesis in a robust manner. Demmig-Adams and Adams (1996) pointed out that there could be good correlation between PRI and photosynthetic efficiency, which in turn may be related to nitrogen status, nitrogen being a vital component of the chlorophyll pigments. PRI in diurnal cycle tends to in increase N-limited leaves, and decrease in water stressed leaves, hence giving a positive correlation between the index and light use efficiency (Peñuelas et al. 1994). It seems that these indices are suited for the detection of short-term variations in crop conditions (diurnal) than seasonal ones. Raz et al. (2003) opines that although NIR region includes one of the water absorbance bands (970 nm) between 700 to 1100 nm, it is not sensitive for water stress as a whole because, this region was affected by differences in the leaf size and structure that occur due to nitrogen treatments. In the present investigation, though the spectral observations were made around noon, no specific trend was observed with nitrogen levels or plant nitrogen status. The erratic trend observed could be due to the fluctuating environmental conditions and the resultant canopy productivity (Running & Nemani 1988). If incident Photosynthetically Active Radiation (PAR) could be measured in conjunction with PRI, it might be possible to estimate photosynthetic rates, which in turn could be related to the nitrogen levels/status. The Greenness index, which is an orthogonal index containing information on all the MSS band, gave a better performance than other ratio indices in the sense that it gave a higher relationship with plant nitrogen status (r²=0.6147). This might be because this index is broader than the traditional SR or NDVI because it uses all the MSS bands and uses coefficients to describe the relative contribution from each band. To our knowledge, there is not much work done to relate the KT transformation-derived Greenness to plant nitrogen status based on ground based sensors. We tried to simulate the MSS bands on Landsat by integrating the hyperspectral data.

It can be noted that a linear relationship between ΔE and plant nitrogen status gives a correlation coefficient of 0.8095 and 0.750 for nitrogen level and plant nitrogen status, respectively. There are other studies also wherein the colour theory has been employed for differentiating the plant nutrient status, successfully. The relationship between leaf nitrogen content of pear and CIE L*, a* and b* colour space co-ordinates showed that the N content best correlated with the b* value and the ratio of a* to b* (Khemira et al. 1994). Recently, Graeff and Claupein (2003) have confirmed the fidelity of this index. They had computed the ΔE in artificial light and concluded that there is a need to prove the fidelity of this index under natural conditions. But the uniqueness of this study is the comparative assessment under field conditions and natural light source, with other widely used indices. A note of caution is to be added here about the possibility of deficiencies of other nutrients or stress due to disease or even water may interfere with the colour change estimation. However, in the present study since the factors other than nitrogen were kept uniform, it is possible to ascertain the utility of ΔE for the detection of wheat nitrogen stress.

Most of the currently used remote sensing indices (NDVI) are broad band indices in the sense that they are obtained by the ratio or difference between the reflectance in two broad bands. In other words, the information in the hyperspectral mode is integrated to get the reflectance in these broad bands, which, Yoder (1992) insists, are better predictors, particularly of green biomass, LAI and yield. Albeit there are many studies done to assess nutrient status of plants by spectral reflectance, the present study was aimed at finding out the sensitivity of the already established indices for a field scale application, for a simple canopy structure like wheat. The most striking feature of nitrogen deficiency is the loss of chlorophyll or greenness of crop canopy, the observation by which remedial measures are undertaken after confirming with laboratory analysis of soils and plants. A perusal of Figure 7 shows the comparative assessment of the different broadband and hyperspectral indices as a predictor of nitrogen stress reveals that broadband indices are better predictors of nitrogen stress detection than the hyper spectral indices. There are some studies that indicate that broadband indices could be better predictors of photosynthetic capacity than narrow band indices because the sensitive bands saturate too rapidly (Yoder & Pettigrew-Crosby 1995). The hSR and the hNDVI have given lower correlation than their counterpart broadband indices NDVI and SR. It is suggested that hyperspectral indices may be more useful for estimating the parameters like chlorophyll (NPCI), water stress (WBI) etc. ΔE is a broadband index, wherein the spectral energy is in the visible region of the electromagnetic spectrum is integrated to get the tristimulus values and the primary colour coordinates.

Conclusions

The study was done to understand the relationship between the most commonly used spectral indices and plant nitrogen and to cull out the best predictor of wheat nitrogen status. After studying different spectral indices for their sensitivity to nitrogen stress, it is obvious that the best index to assess nitrogen stress is the colour difference (ΔE). The next best index is found to be NDVI followed by the Greenness index. Hence, it can be concluded that broadband-derived indices are better predictors than hyperspectral indices for wheat nitrogen stress/status. It is speculated that a nitrogen stressed plant canopy interacts with the electromagnetic radiation in a broadband manner and not in a hyperspectral manner, owing to several feedback relationships resulting from the plant physiological regime depending on plant nutrition. However, the interference of solar angle and look angle of the sensor can alter the spectral reflectance and consequently, predictability of indices. Hence it is strongly suggested that for future studies, multiple angle-based measurements also be included to decipher the ‘directional’ behaviour of these indices with regard to solar angle variation with time. Extrapolation of the algorithms developed from ground based measurements to satellite remote sensing needs to be done cautiously, but these could be validated and then used for crop simulation modelling at larger spatial scales, using satellite data. However, for the development of ground level sensors in a precision farming era, our results have several prospects. Further studies are needed to predict accurately the nitrogen content from spectral measurements for other crop types, especially crops with complex canopies and radiative transfer mechanisms. The study helps to develop efficient algorithms for coupling the remotely sensed data to process models to understand plant-environment interaction in a robust and quick manner.