Model for estimation of total nitrogen content in sandalwood leaves based on nonlinear mixed effects and dummy variables using multispectral images

https://doi.org/10.1016/j.chemolab.2019.103874Get rights and content

Abstract

Fertilizer overuse is a common phenomenon in global agroforestry production, and this overuse causes ecological destruction. The ability to accurately estimate the nutrient content of plant leaves in real-time would be a wonderful solution to reduce the degree of environmental damage. In recent years, remote sensing technology has been widely used in the diagnosis of crop nutrition in many countries. Most studies focus on optimal band selection or create new vegetation indices, but these studies have ignored the random impact of natural environmental factors on the estimated results. This paper proposed an estimation model of total nitrogen content (TNC) in sandalwood leaves that takes sampling season and site conditions as the dummy variable and random effect, respectively. Three forestry farms with different locations and site conditions were selected as study areas to enhance the universality of this model. Multispectral images of leaves were obtained using a low-cost five-band camera (RedEdge3, MicaSense, USA), and the experimental results indicate the following: (1) the growth of the tree height, crown width and stem effectively increased under the medium gradient level (N2), whereas a high gradient level (N3) significantly promoted all aspects except tree height; (2) the mean and variance of some image texture features of the G, RE and NIR band were significantly correlated with TNC at the 0.05 and 0.01 levels, and the texture mean value index (TMVI) proposed in this paper can improve the correlation with TNC; and (3) the results obtained using the nonlinear mixed-effects model with dummy variables improved the fitting degree and estimation accuracy compared with results of SVR and BPNN. This study demonstrates the advantages of using the nonlinear mixed-effects model with dummy variables to obtain a more reliable estimation model for the nutritional diagnosis of rare tree species.

Introduction

Sandalwood is a semi-parasitic evergreen tree that is mainly distributed in Southeast Asia, Australia and the Pacific [1]. The wide range of applications of this tree in fragrance production and in pharmaceutical, medicinal and other commercial products makes sandalwood valuable worldwide. However, due to overharvesting and the destruction of its ecological environment, wild sandalwood resources have sharply decreased globally, resulting in a marked decline in exports. Since sandalwood is not native to China, substantial labour and material resources have been invested to study artificial methods of cultivating sandalwood to meet market needs. After decades of research and promotion, people have mastered the basic methods of cultivating sandalwood [2]. Researchers have found that sandalwood cultivation is highly sensitive to water and nutrition resources; these aspects are impacted by poor forest farm management, which can seriously affect sandalwood growth and heartwood formation. The ability to monitor nutritional status in real time is the key to improving the growth rate and quality of sandalwood.

The optimization of nitrogen utilization is a major factor in improving plant yield and quality. Today, many countries apply nitrogen fertilizer to obtain high levels of food production. However, the nitrogen utilization rate decreases when the amount of fertilizer exceeds demand, and the surplus nitrogen will be lost through runoff, leaching, ammonia volatilization and denitrification, which lead to groundwater nitrate pollution, water eutrophication and the greenhouse effect [3,4]. The traditional method of monitoring crop nitrogen content is to collect plant leaves in the field and perform chemical analysis in the laboratory. Although the results are reliable, this method cannot meet the need for rapid and non-destructive monitoring. Therefore, an effective technology to improve the nitrogen utilization efficiency would be to monitor the nitrogen status of plants in real time in an effort to balance the applied nitrogen supply and the plant’s nitrogen utilization [5].

Plant nutrition diagnosis technology has rapidly developed over the past 20 years. The main diagnosis methods are the fertilizer window, leaf colour chart, chlorophyll meter, hand-held spectrometer, hyperspectral remote sensing and digital image processing technology. However, there are some disadvantages associated with these different methods. For example, the fertilizer window cannot determine specific amounts of topdressing, and laboratory chemical analysis is also required [6]. Leaf colour charts cannot explain the reason for chlorosis and are also affected by planting density and crop varieties [7]. SPAD is widely used in different types of research, but its ability to measure leaf area is limited [8]. When plant nitrogen content is close to or higher than the optimal value, the SPAD value cannot accurately characterize the chlorophyll content [9]. Hence, hand-held spectrometers, hyperspectral remote sensing and digital image processing technologies are currently more reliable and popular in plant nutrition diagnosis.

In the early stages of non-destructive plant nutrition diagnosis, digital image processing was widely used because of its convenience and low cost. Due to its flexibility and feasibility, it is only necessary to select suitable carriers for research on different scales. Some results have shown that the changes in plant colour were significantly correlated with chlorophyll content, and the chlorophyll content was strongly affected by the total nitrogen content [10]. José et al. [11] used single-colour channels and double-colour channels from the RGB, HSV, and L*a* b* colour spaces to predict SPAD values in maize. This analysis shows that the correlation of SPAD values and colour data with the concentration of nitrogen is possible. Furthermore, the study shows that when single-colour channel correlations are performed, the channels G, b* and V provide the best modelling accuracy. Miguel et al. [12] developed a method in which principal component analysis is applied to digital images to calculate a greenness index using RGB components of the colour image that yield an estimate of the amount of nitrogen in the plant. The results show that the capacity of this index to predict the nitrogen deficiencies that affect the barley yield was equal to or better than that of the SPAD measurements. Confalonieri et al. [13] estimated leaf and plant nitrogen content using an 18% grey-dark-green colour index (DGCI) method. Compared to the DGCI and the corrected DGCI, the new method is considerably more stable with regard to both trueness and precision. Baresel et al. [14] used a consumer digital camera, SPAD and reflectance spectrometry for the non-destructive detection of the leaf chlorophyll content and the nitrogen nutrition status in wheat. With spectral measurements, the biomass and leaf nitrogen content could not be clearly differentiated; chlorophyll measurements do not reflect the biomass, whereas the described procedure of image analysis permits the consideration of biomass and leaf nitrogen content.

Although image processing technology has many merits, the band of digital images remains limited to visible light and large bandwidths, which results in insensitivity to subtle changes of leaf colour. Therefore, hand-held spectrometer technologies and hyperspectral remote sensing, with their wide band range and rich number of bands, are favoured by researchers. The vegetation indices calculated from the red to near-infrared bands, which can be obtained from wide bands or narrow bands, show a high correlation with the plant nitrogen stress status [15]. Studies that have estimated biophysical and chemical characteristics have suggested that the saturation phenomenon easily occurs in the biophysical parameter modelling of the index obtained by the wide band [16,17] and that the narrow-band index is more sensitive to changes in plant nitrogen content [18]. While hyperspectral sensors describe the plant canopy reflectance in greater detail than described by multispectral sensors, they also suffer from issues with data redundancy and spectral autocorrelation, termed “high dimensional disaster” or the “Hughes” phenomenon. Therefore, to optimize the advantages of narrow bands, methods and approaches must be provided that can overcome data redundancy [19]. Thorp et al. [20] selected 25 relevant bands from 701 narrow bands; the estimation accuracy of durum wheat characteristics, which included growth, nitrogen status, and grain yield, was higher when using the less redundant spectral data. Yu et al. [21] proposed a detection and quantification method for nitrogen concentration in oilseed rape leaf using a deep-learning-based regression model composed of stacked auto-encoders (SAEs) and a fully connected neural network (FNN). The SAE-FNN model achieved a reasonable performance for the nitrogen concentration. Corti et al. [22] assessed the capability of hyperspectral line-scan imaging (400–1000 nm) to estimate crop variables in a greenhouse under combined water and nitrogen stress using multivariate data analysis and two data compression methods: canopy average spectra and hyperspectrogram extraction. Image hyperspectrogram compression, without spatial information loss, produced more encouraging results when canopy structure in crop variables was considered than those results produced by the average canopy spectra method.

In addition to data types, many factors affect the accuracy of the non-destructive diagnosis of nitrogen, such as growth stages and crop varieties. Khanal et al. [23] obtained visual images of a corn canopy from manned aircraft and then established the relationship between vegetation indices and yield. The results showed that image correlation to corn yield increased as the corn growth stage progressed but weakened towards the end of the growing season. Because of the different cultivated varieties, Zhao et al. [24] determined that the most sensitive bands for the nitrogen nutrition index in summer maize was located at 710 nm and 512 nm, whereas Goel et al. [25] reported that the change of nitrogen stress was more obvious at 498 nm and 671 nm. Plant water content is an important factor that affects the estimation of nitrogen content. Therefore, reducing the influence of water stress on colour is key to improving the prediction accuracy. Kusnierek et al. [26] developed a method that can distinguish between nitrogen and water status in spring wheat. The results showed that the first component of processed spectral data was related to the water regime, whereas the second component was related to nitrogen fertilizer.

Most studies focus on improving the correlation between vegetation indices and nitrogen content, but they neglect the importance of the modelling algorithm [27]. At present, the most commonly used modelling methods are multivariate statistics and machine learning algorithms. A comparison of those two methods shows that the nonlinear mapping ability of the machine learning algorithm is stronger than that of multivariate statistics, but the machine learning algorithm is inferior in interpreting the results. However, those two methods considered only the fixed effect of parameters while ignoring the random effects on the estimated results.

Nitrogen content estimation by remote sensing is based on the changes in the reflectance in different bands. However, in addition to the nitrogen content in plants, many other factors can affect the reflectance value, such as sampling season and site condition. In this context, the objects of the current study are as follows: (i) to determine the effects of different nitrogen levels on tree height, crown width and stem of sandalwood; (ii) to discuss the correlation between image features and total nitrogen content in sandalwood leaves, and to select or create best features to represent the nitrogen content; and (iii) to propose an estimation model using nonlinear mixed effects and dummy variables that take other influencing factors into consideration.

Section snippets

Materials and methods

Differences exist in spectral reflectance curves between healthy plants and those under nitrogen stress, allowing vegetation indices to effectively represent the total nitrogen content [[28], [29], [30]]. In addition, several texture features were calculated and applied in variable selection. After the variable set was determined, five kinds of different formulas, including linear combination, reciprocal combination, power function, logarithmic function and exponential function, were tested,

Changes in sandalwood plant growth under different nitrogen levels

The changes in height, crown width and stem growth were calculated for the different levels from the beginning of fertilization to the end of the experiment. In Fig. 6, those three indicators show a rising trend in the different degrees achieved with the increase in nitrogen application. Compared with the control group, the growth of height of N1 to N3 increased by 65%, 301% and 353%. Regarding the crown width, the values are 29%, 237% and 484%, whereas the growth of the stem only increased by

Changes of tree measurement factors

The increase in nitrogen application has a greater impact on tree height and crown width but has only a slight influence on the stem. The changes of those tree-measurement factors were not significant at the N1 level except for the tree height, which indicates that the nutrition was first absorbed by cells in the stem when the nitrogen content was insufficient. When the level of fertilization reached N2, the growth rate of both the height and crown width increased substantially. Under a high

Conclusion

To monitor the nutritional status of sandalwood in real time, we proposed a method based on the nonlinear mixed-effects model with dummy variables to predict the TNC in sandalwood leaves. The results are as follows:

  • (1)

    The VIs calculated by G, RE and NIR bands show great correlation to the TNC. In terms of texture features (including texture mean value and texture variance), nine features calculated by the G, RE and NIR band are related to the TNC at the 0.05 or 0.01 level, which indicate that

Acknowledgment

This study is funded by National Natural Science Foundation of China (grant number “31670642”) and State Forestry Administration of the People’s Republic of China (grant number “[2016] No.11”).

The authors thank the anonymous reviewers sincerely for their useful comments, which contributed to the quality of this study.

References (60)

  • B. Zhao et al.

    Exploring new spectral bands and vegetation indices for estimating nitrogen nutrition index of summer maize

    Eur. J. Agron.

    (2018)
  • P.K. Goel et al.

    Potential of airborne hyperspectral remote sensing to detect nitrogen deficiency and weed infestation in corn

    Comput. Electron. Agric.

    (2003)
  • K. Kusnierek et al.

    Simultaneous identification of spring wheat nitrogen and water status using visible and near infrared spectra and Powered Partial Least Squares Regression

    Comput. Electron. Agric.

    (2015)
  • L. Wang et al.

    Estimation of leaf nitrogen concentration in wheat using the MK-SVR algorithm and satellite remote sensing data

    Comput. Electron. Agric.

    (2017)
  • S.K. Singh et al.

    Assessment of growth, leaf N concentration and chlorophyll content of sweet sorghum using canopy reflectance

    Field Crop. Res.

    (2017)
  • C.J. Tucker

    Red and photographic infrared linear combinations for monitoring vegetation

    Remote Sens. Environ.

    (1979)
  • J. Heiskanen et al.

    Sensitivity of narrowband vegetation indices to boreal forest LAI, reflectance seasonality and species composition

    ISPRS J. Photogrammetry Remote Sens.

    (2013)
  • A.A. Gitelson et al.

    Use of green channel in remote sensing of global vegetation from EOS-MODIS

    Remote Sens. Environ.

    (1996)
  • J.L. Roujean et al.

    Estimating PAR absorbed by vegetation from bidirectional reflectance measurements

    Remote Sens. Environ.

    (1995)
  • A.A. Gitelson et al.

    Relationships between leaf chlorophyll content and spectral reflectance and algorithms for non-destructive chlorophyll assessment in higher plant leaves

    J. Plant Physiol.

    (2003)
  • A.A. Gitelson

    Wide dynamic range vegetation index for remote quantification of biophysical characteristics of vegetation

    J. Plant Physiol.

    (2004)
  • D. Haboudane et al.

    Hyperspectral vegetation indices and novel algorithms for predicting green LAI of crop canopies: modeling and validation in the context of precision agriculture

    Remote Sens. Environ.

    (2004)
  • W. Wang et al.

    Estimating leaf nitrogen concentration with three-band vegetation indices in rice and wheat

    Field Crop. Res.

    (2012)
  • C.S.T. Daughtry et al.

    Estimating corn leaf chlorophyll concentration from leaf and canopy reflectance

    Remote Sens. Environ.

    (2000)
  • A. Huete et al.

    Overview of the radiometric and biophysical performance of the MODIS vegetation indices

    Remote Sens. Environ.

    (2002)
  • J. Chen et al.

    Estimation of total suspended matter concentration from MODIS data using a neural network model in the China eastern coastal zone

    Estuar. Coast Shelf Sci.

    (2015)
  • Y.Y. Lu et al.

    Nitrogen vertical distribution and status estimation using spectral data in maize

    Commun. Soil Sci. Plant Anal.

    (2018)
  • H.S. Kusuma et al.

    The extraction of essential oil from sandalwood (Santalum album) by microwave air-hydrodistillation method

    J. Mater. Environ. Sci.

    (2016)
  • G.H. Ma et al.

    Study on semi-parasitism of sandalwood seedlings

    J. Trop. Subtropical Bot.

    (2005)
  • T. Huang et al.

    Nitrate leaching in a winter wheat-summer maize rotation on a calcareous soil as affected by nitrogen and straw management

    Sci. Rep.

    (2017)
  • Cited by (18)

    • Detection and classification of groundnut leaf nutrient level extraction in RGB images

      2023, Advances in Engineering Software
      Citation Excerpt :

      The implemented methods' results were compared with back propagation neural network (BPNN) and support vector regression (SVR) algorithms. The achieved R2 value by their proposed method is 0.912, which is high when compared to SVR and BPNN [28]. Though the multispectral images provide a greater number of features, they require more processing and computation time compared to RGB images.

    • Rapid identification model based on decision tree algorithm coupling with <sup>1</sup>H NMR and feature analysis by UHPLC-QTOFMS spectrometry for sandalwood

      2020, Journal of Chromatography B: Analytical Technologies in the Biomedical and Life Sciences
      Citation Excerpt :

      Accordingly, new methods that require short analysis time would be essential for the identification of genus Dalbergia, genus Pterocarpus and their counterfeit furniture. Rapid identification model based on mathematics was dominated by cluster analysis and principal component analysis (PCA) [25,26]. They are generally used for processing small data [27,28].

    • A possible fractional order derivative and optimized spectral indices for assessing total nitrogen content in cotton

      2020, Computers and Electronics in Agriculture
      Citation Excerpt :

      This result shows that spectral indices developed by different transformed reflectance showed the same sensitive portions to plant biophysical indicators. Machine learning methods, such as neural networks, random forest, genetic algorithms widely used to quantitative estimation of plant biochemical properties (Chen and Wang, 2019; Sun et al., 2017). However, (Li et al., 2018b; Zhai et al., 2012) indicates that SVR approach can better identify the relationship between nitrogen concentration and hyperspectral parameters, which aims to improve the estimation accuracy of the model.

    View all citing articles on Scopus
    View full text