^{1}

^{*}

^{1}

^{2}

^{1}

^{3}

^{1}

This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution license (

The retrieval of nutrient concentration in sugarcane through hyperspectral remote sensing is widely known to be affected by canopy architecture. The goal of this research was to develop an estimation model that could explain the nitrogen variations in sugarcane with combined cultivars. Reflectance spectra were measured over the sugarcane canopy using a field spectroradiometer. The models were calibrated by a vegetation index and multiple linear regression. The original reflectance was transformed into a First-Derivative Spectrum (FDS) and two absorption features. The results indicated that the sensitive spectral wavelengths for quantifying nitrogen content existed mainly in the visible, red edge and far near-infrared regions of the electromagnetic spectrum. Normalized Differential Index (NDI) based on FDS_{(750/700)} and Ratio Spectral Index (RVI) based on FDS_{(724/700)} are best suited for characterizing the nitrogen concentration. The modified estimation model, generated by the Stepwise Multiple Linear Regression (SMLR) technique from FDS centered at 410, 426, 720, 754, and 1,216 nm, yielded the highest correlation coefficient value of 0.86 and Root Mean Square Error of the Estimate (RMSE) value of 0.033%N (n = 90) with nitrogen concentration in sugarcane. The results of this research demonstrated that the estimation model developed by SMLR yielded a higher correlation coefficient with nitrogen content than the model computed by narrow vegetation indices. The strong correlation between measured and estimated nitrogen concentration indicated that the methods proposed in this study could be used for the reliable diagnosis of nitrogen quantity in sugarcane. Finally, the success of the field spectroscopy used for estimating the nutrient quality of sugarcane allowed an additional experiment using the polar orbiting hyperspectral data for the timely determination of crop nutrient status in rangelands without any requirement of prior cultivar information.

Sugarcane (

In general, several approaches to measure nutrient status in the plant have been developed and evaluated. The most common method is performed in the laboratory using leaf samples collected in the field [

However, the first method requires more leaf samples from the field, which is a laborious, lengthy and destructive process [

The aim of this study was to develop an estimation model that could explain the nitrogen variations in sugarcane. The field spectroscopy data were analyzed to investigate whether they contained adequate information for determining the nitrogen concentration in sugarcane with several cultivars. Original reflectance was transformed into a first-derivative spectrum (FDS) and two absorption features. Narrow vegetation indices and Stepwise Multiple Linear Regression (SMLR) were applied to compute the estimation models. Subsequently, the relationships between the measured nitrogen concentration and the spectral reflectance were explored. In addition, effects of canopy architecture on the spectral signature and the model precision were analyzed. This research will be expanded to include an additional experiment using polar orbiting hyperspectral data for the timely determination of crop nutrient status without any requirement of prior cultivar information.

The two study sites are located in the Watthana Nakhon district (plot 1) and Mueang Sakaeo district (plot 2), Sakaeo Province, in the eastern region of Thailand (102°15′E, 13°45′N) (

In December 2010, the canopy spectral reflectance was measured using a Fieldspec® 3 spectroradiometer (Analytical Spectral Devices, Boulder, CO, USA) [

The first and second fully expanded leaves from the top of two random shoots for one tiller were collected. The midrib was removed from the leaf blade because the presence of the midrib resulted in a decreased concentration of nitrogen [

FDS was calculated and used as a variable in the estimation model. FDS is commonly used to enhance absorption features that might be masked by interfering background absorptions and canopy background effects [_{λ(}_{j}_{)} is the reflectance at the _{λ(}_{j}_{+ 1)} is the reflectance at the

Nitrogen exhibits absorption features in the visible, near-infrared and shortwave-infrared regions [_{(λi)} is obtained by dividing the reflectance value R_{(λi)} of each waveband _{c(λi)} at the corresponding wavelength

The first and last spectral data values are on the hull; therefore, the first and last values of the continuum-removed spectrum are equal to 1. The output curves have values between 0 and 1, where the absorption pits are enhanced. In this study, only three regions of the wavelength were used, including R420–R530, R550–R750 and R1116–R1284, which are known as the pigment and water content absorption zones. Two variables proposed in [

Continuum-Removed Derivative Reflectance (CRDR) was calculated by applying a first-derivative transformation to the continuum-removed reflectance spectrum R′.

The band depth (BD) was calculated by subtracting the continuum-removed reflectance at wavelength

Narrow vegetation indices were computed from all possible two-wavelength combinations involving the 1,025 wavelengths using FDS, CRDR and BD. These 1,025 discrete bands allowed a calculation of N*N (1,050,625 indices). The two most widely used vegetation indices in estimating agricultural and ecological variables are the Ratio Spectral Index (RSI) and the Normalized Differential Index (NDI) [_{1}: FDS, CRDR and BD between 400–2,500 nm., and λ_{2}: FDS, CRDR and BD between 400–2,500 nm.

SMLR was used to estimate the nitrogen concentration in sugarcane from the measured reflectance spectra. The idea was to identify the spectral wavebands that provide the best correlation with different chemical compounds present in the leaf or canopy [

Two approaches were adopted to evaluate the predictive accuracy: validation by (1) an independent data set and (2) a 10-fold cross technique. Performances of the estimation models were summarized and reported in terms of the coefficient of determination (^{2}_{i}_{i}

The canopy spectral reflectance values of the three cultivars, shown in

The correlation coefficients between all possible wavelength pairs with ^{2}_{(750/700)} and the RVI based on the FDS_{(724/700)} yielded the highest accuracy, with ^{2}_{cv} values of 0.044 and 0.043% N and RE values of 3.34 and 3.27%N for the validation data set and the pooled data set, respectively. The regression equations are Y = 0.37x + 1.39 and Y = 0.14x + 1.04.

The scatter plots in

The number of selected spectral wavelengths, used to estimate the nitrogen concentration by SMLR, ranges from two to five. Using the FDS as an independent variable, SMLR selected five sensitive wavelengths, centered at the visible, red edge and far near-infrared regions of the electromagnetic spectrum. The best model yielded ^{2}_{1} – 471.9x_{2} – 31.98x_{3} + 133.78x_{4} – 115.86x_{5} + 1.4. The selected wavelengths for each data set are summarized in ^{2}

Results from this research indicate that field spectroscopy data contains adequate information for determining the nitrogen concentration in sugarcane with combined cultivars. The univariate method has illustrated that the red edge region contains the sensitive wavelengths for explaining the nitrogen variations in sugarcane canopy. This region has been shown to be insensitive to atmospheric and background effects [

The estimation model with the highest predictive accuracy was developed by SMLR technique. The selected wavelengths, including the spectral bands centered at 410, 430, 720, 754 and 1216 nm, were used to predict the nutrient quality in sugarcane. Wavelengths in the visible region (λ = 410 nm and 430 nm), selected for nitrogen concentration estimation, are related to pigment absorption [

The structure of the plant canopy has a significant bearing on its spectral signature [

Most of the previous publications developed models for estimating the nitrogen variations in paddy rice, wheat, cotton or grass, but only in a few cases in sugarcane [

The optimal goal of this paper was to develop an estimation model that could explain the nitrogen variations in sugarcane with combined cultivars. Reflectance spectra were measured over the sugarcane canopy. Derivative spectra and absorption features were used as independent variables in univariate and multivariate approaches. The most important conclusions that could be drawn from this study are as follows:

Stepwise multiple linear regression could explain the nitrogen variations in sugarcane canopy better than a narrow vegetation index. This technique utilizes more than two wavelengths from the entire spectral region (400–2,500 nm) to estimate the dependent variable.

First Derivative Spectrum (FDS) showed a better relationship with canopy nitrogen concentration than Continuum-Removed Derivative Reflectance (CRDR) and Band Depth (BD) when the models were developed with multivariate approaches and validated with a combined cultivar data set.

It was concluded that a First Derivative Spectrum (FDS) has potential when used to estimate the nitrogen content in sugarcane with combined cultivars at the maturity stage (9–12 months).

Visible, red edge and far near-infrared regions contain more information on canopy nitrogen concentration of combined sugarcane cultivars compared to other parts of the electromagnetic spectrum.

Canopy architecture directly influences the spectral response and the predictive precision. Canopy structure, therefore, should be taken into consideration when mapping sugarcane nutrient quality in rangelands with combined cultivars.

In the case of a known cultivar, partitioning the data into cultivars could increase the estimation capability of the method applied in this research.

The modified estimation model, generated by SMLR technique from FDS centered at 410, 426, 720, 754, and 1,216 nm, yields the highest correlation coefficient value of 0.86 and RMSE value of 0.033%N (n = 90) with nitrogen concentration in sugarcane. This result is much better than those of the previous studies.

Overall, since the field spectroscopy data used in this study was measured over the sugarcane canopy under natural atmospheric conditions, the success of field spectroscopy used for estimating nutrient quality in sugarcane allows an additional experiment using the polar orbiting hyperspectral data for the timely determination of crop nutrient status in rangelands without any requirement of the prior cultivar information.

This research was supported by the National Center for Genetic Engineering and Biotechnology (BIOTEC), Ministry of Science and Technology, Thailand. We would also like to acknowledge the kind assistance of the ES Research and Development Co. LTD., the Geo-informatics and Space Technology Development Agency (GISTDA) and the Agricultural System and Engineering Laboratory, Asian Institute of Technology. We thank the anonymous reviewers whose comments substantially improved this paper.

The study area in Sakaeo Province, in the eastern region of Thailand.

Comparison of the mean canopy reflectance spectra of sugarcane (N = 60 for each cultivar) between the three different cultivars: (

Contour plots showing the regions with high correlation coefficients (^{2}

Measured versus estimated nitrogen concentration in sugarcane with combined cultivars, using narrow vegetation indices (n = 90): (_{750}_{700}_{748}, CRDR_{700}); (_{724}, FDS_{700}); and (_{748}, CRDR_{630}).

Measured versus estimated sugarcane nitrogen concentration of combined cultivars; the model was developed by a Stepwise Multiple Linear Regression (SMLR) approach with two independent variables (n = 90): (

Characteristics of the experimental plots used in this study.

Location | 13.866°N 102.287°E | 13.710°N 102.214°E |

Soil texture | loamy sand | loamy clay |

Annual precipitation, mm (2010) | 1,268.2 | 1,296.6 |

Annual air temperature, °C (2010) | 22.91/35.14 (Min/Max) | 22.8/34.08 (Min/Max) |

Plot size | 36 m × 76 m | 36 m × 76 m |

Characteristics of the sensor, wavelength covering and spectral resolution used in this study.

Fieldspec^{®} 3 spectroradiometer |
400–1,050 | 1.4 | 465 |

1,050–2,500 |
2 | 560 |

The spectral regions between 1355–1450 nm, 1800–1950 nm and 2420–2500 nm, which are associated with the water absorption, were excluded from the analysis.

Descriptive statistics of the nitrogen concentration measured in the laboratory.

Calibration | 90 | 1.142 | 1.483 | 1.313 | 0.098 |

Validation | 90 | 1.148 | 1.457 | 1.319 | 0.084 |

Pooled | 180 | 1.142 | 1.483 | 1.316 | 0.091 |

| |||||

By Cultivar | |||||

| |||||

LK92-11 | 60 | 1.251 | 1.415 | 1.333 | 0.046 |

KK-3 | 60 | 1.275 | 1.483 | 1.395 | 0.062 |

K84-200 | 60 | 1.142 | 1.322 | 1.221 | 0.051 |

Performances of narrow vegetation indices calculated from different independent variables for estimating nitrogen concentration in sugarcane with the combined cultivars.

_{1}/λ_{2} (nm) |
||||||||
---|---|---|---|---|---|---|---|---|

| ||||||||

^{2}_{c} |
_{c} |
^{2}_{v} |
_{v} |
^{2}_{cv} |
_{cv} | |||

NDI | FDS | 750/700 | 0.82 | 0.041 | 0.73 | 0.044 | 0.78 | 0.043 |

CRDR | 748/690 | 0.82 | 0.041 | 0.73 | 0.043 | 0.78 | 0.043 | |

BD | 748/680 | 0.81 | 0.042 | 0.70 | 0.047 | 0.76 | 0.045 | |

| ||||||||

RVI | FDS | 724/700 | 0.80 | 0.043 | 0.73 | 0.044 | 0.78 | 0.043 |

CRDR | 748/630 | 0.83 | 0.041 | 0.73 | 0.044 | 0.78 | 0.043 | |

BD | 748/670 | 0.82 | 0.042 | 0.70 | 0.048 | 0.76 | 0.045 |

λ: Selected wavelength in nm; RMSE_{c}, RMSE_{v} and RMSE_{cv}: root mean square error of calibration, validation and relative cross-validation with a 10-fold cross technique, respectively, expressed as %N; ^{2}_{cv}: relative cross-validated coefficient of determination with a 10-fold cross technique; fit between estimated and observed values at the 0.01 level was considered highly significant.

Performances of narrow vegetation indices calculated from different independent variables for estimating nitrogen concentration in sugarcane with the separated cultivar.

| ||||||||||
---|---|---|---|---|---|---|---|---|---|---|

_{1} |
_{2} |
^{2}_{v} |
_{1} |
_{2} |
^{2}_{v} |
_{1} |
_{2} |
^{2}_{v} | ||

NDI | FDS | 748 | 728 | 0.81 | 718 | 710 | 0.91 | 712 | 706 | 0.80 |

CRDR | 570 | 748 | 0.80 | 736 | 712 | 0.90 | 712 | 700 | 0.79 | |

BD | 748 | 584 | 0.78 | 716 | 710 | 0.92 | 746 | 680 | 0.75 | |

| ||||||||||

RVI | FDS | 748 | 728 | 0.79 | 718 | 710 | 0.90 | 724 | 692 | 0.82 |

CRDR | 748 | 510 | 0.78 | 712 | 574 | 0.91 | 712 | 700 | 0.82 | |

BD | 748 | 586 | 0.78 | 716 | 710 | 0.91 | 712 | 678 | 0.76 |

λ: Selected wavelength in nm; ^{2}_{v}: root mean square error of calibration, validation; fit between estimated and observed values at the 0.01 level was considered highly significant.

List of regression models developed by the vegetation indices using the different data sets.

_{wavelength} |
^{2}_{v} | |||
---|---|---|---|---|

Combined cultivars | NDI | FDS_{750,700} |
Y = 0.37x + 1.39 | 0.73 |

CRDR_{748,690} |
Y = 0.26x + 1.38 | 0.73 | ||

BD_{748,680} |
Y = 9.64× + 10.77 | 0.70 | ||

| ||||

RVI | FDS_{724,700} |
Y = 0.14x + 1.04 | 0.73 | |

CRDR_{748,630} |
Y = −0.09x + 1.15 | 0.73 | ||

BD_{748,670} |
Y = 18.79x + 1.12 | 0.70 | ||

| ||||

Separated cultivar - LK92-11 | NDI | FDS_{748,728} |
Y = 0.95x + 1.71 | 0.79 |

CRDR_{570,748} |
Y = 0.11x + 1.12 | 0.80 | ||

BD_{748,584} |
Y = 4.58x + 5.72 | 0.78 | ||

| ||||

RVI | FDS_{748,728} |
Y = 0.93x + 0.93 | 0.79 | |

CRDR_{748,510} |
Y = 0.24x + 1.21 | 0.78 | ||

BD_{748,586} |
Y = 9.05x + 1.14 | 0.78 | ||

| ||||

Separated cultivar - KK-3 | NDI | FDS_{718,710} |
Y = 1.59x + 1.09 | 0.91 |

CRDR_{736,712} |
Y = 0.44x + 1.39 | 0.90 | ||

BD_{716,710} |
Y = 4.18x + 1.73 | 0.92 | ||

| ||||

RVI | FDS_{718,710} |
Y = 0.53x + 0.61 | 0.90 | |

CRDR_{712,574} |
Y = 0.61x + 2.05 | 0.91 | ||

BD_{716,710} |
Y = 2.45x – 0.7 | 0.91 | ||

| ||||

Separated cultivar - K84-200 | NDI | FDS_{712,706} |
Y = 1.45x + 1.18 | 0.80 |

CRDR_{712,700} |
Y = 0.75x + 1.23 | 0.79 | ||

BD_{746,680} |
Y = 3.47x + 4.59 | 0.75 | ||

| ||||

RVI | FDS_{724,692} |
Y = 0.08x + 1.07 | 0.82 | |

CRDR_{712,700} |
Y = 0.39x + 0.84 | 0.82 | ||

BD_{712,678} |
Y = 0.63x + 0.86 | 0.76 |

Performance of stepwise multiple linear regression in estimating the nitrogen concentration in sugarcane with combined cultivar.

^{2}_{c} |
_{c} |
^{2}_{v} |
_{v} |
^{2}_{cv} |
_{cv} | ||
---|---|---|---|---|---|---|---|

FDS | 410, 430, 720, 754,1216 | 0.90 | 0.030 | 0.80 | 0.038 | 0.86 | 0.033 |

CRDR | 748, 1158, 1184, 1216, 1276 | 0.89 | 0.033 | 0.64 | 0.053 | 0.78 | 0.043 |

BD | 748, 1262 | 0.83 | 0.040 | 0.70 | 0.047 | 0.77 | 0.044 |

λ: Selected wavelength in nm; RMSE_{c}, RMSE_{v} and RMSE_{cv}: root mean square error of calibration, validation and relative cross-validation with a 10-fold cross technique, respectively, expressed as %N; ^{2}_{cv}: relative cross-validated coefficient of determination with a 10-fold cross technique; fit between estimated and observed values at the 0.01 level was considered highly significant.

Selected wavelengths and coefficients of determination between the observed nitrogen concentration and the spectral reflectance of separated cultivar.

^{2}_{cv} |
^{2}_{cv} |
^{2}_{cv} | ||||
---|---|---|---|---|---|---|

FDS | 670, 754, 1,266, 1,494, 2,313 | 0.86 | 750, 1,104, 1,572, 1,586, 2,153 | 0.93 | 552, 1,032, 1,284, 1,604, 2,359 | 0.80 |

CRDR | 656, 704, 1,266 | 0.79 | 598, 674, 740, 1,228 | 0.92 | 678, 746, 1,128 | 0.80 |

BD | 684, 748, 1,228 | 0.82 | 552, 736 | 0.92 | 746 | 0.78 |

^{2}_{cv}: relative cross-validated coefficient of determination with a 10-fold cross technique; fit between estimated and observed values at the 0.01 level was considered highly significant.

List of regression models developed by SMLR from the different data sets.

^{2}_{v} | |||
---|---|---|---|

Combined cultivars | FDS | Y = 212.76x_{1} – 471.9x_{2} – 31.98x_{3} + 133.78x _{4} – 115.86x_{5} + 1.4 |
0.80 |

CRDR | Y = 55.77x_{1} + 19.0x_{2} + 61.92x_{3} – 73.5x _{4} – 34.25x_{5} + 1.26 |
0.64 | |

BD | Y = 25.26x_{1} – 3.7x_{2} + 1.19 |
0.70 | |

| |||

Separated cultivar -LK92-11 | FDS | Y = 520.59x_{1} + 14.794x_{2} – 97.004x_{3} + 95.08x _{4} – 63.49x_{5} + 1.347 |
0.86 |

CRDR | Y = 33.04x_{1} – 18.84x_{2} – 32.51x_{3} + 1.72 |
0.79 | |

BD | Y = −0.47x_{1} + 17.86x_{2} – 0.7845x_{3} + 1.62 |
0.82 | |

| |||

Separated cultivar-KK-3 | FDS | Y = 67.4x_{1} – 24.64x_{2} – 136.2x_{3} – 168.34x_{4} – 42.007x _{5} + 1.4153 |
0.93 |

CRDR | Y = −14.27x_{1} – 105.25x_{2} + 23.31x_{3} + 8.84x_{4} + 1.004 |
0.92 | |

BD | Y = −3.31x_{1} + 2.26x_{2} + 1.17 |
0.92 | |

| |||

Separated cultivar - K84-200 | FDS | Y = −194.8x_{1} + 138.92x_{2} – 94.52x_{3} – 213.97x _{4} – 22.31x_{5} + 1.29 |
0.80 |

CRDR | Y = −64.57x_{1} + 25.36x_{2} + 6.34x_{3} + 28.04x_{4} + 1.08 |
0.80 | |

BD | Y = 7.86x_{1} + 1.12 |
0.78 |

The estimation models for estimating nitrogen concentration in sugarcane from field spectroscopy compared with previous publications (unit: % nitrogen).

^{2} |
||||||
---|---|---|---|---|---|---|

RVI (FDS741, FDS1323) | Y = −0.237x + 0.587 | 0.74 | 0.084 | 25 | Leaf | [ |

RVI (R740, R720) | Y = 0.73x + 0.753 | 0.81 | 0.103 | 37 | Leaf | [ |

NDI (FDS750, FDS700) | Y = 0.37x + 1.39 | 0.78 | 0.043 | 90 | Canopy | This paper |

RVI (FDS724, FDS700) | Y = 0.14x + 1.04 | 0.78 | 0.043 | 90 | Canopy | This paper |

Y = 212.76x_{1} – 471.9x2 – 31.98x_{3} + 133.78x 4 – 115.86x_{5} + 1.4 |
0.86 | 0.033 | 90 | Canopy | This paper |

developed by SMLR technique.