International Journal of Agronomy

Volume 2015, Article ID 748074, 10 pages

http://dx.doi.org/10.1155/2015/748074

## Indirect Estimations of Lentil Leaf and Plant N by SPAD Chlorophyll Meter

^{1}Department of Soil and Plant Sciences, California State University, Chico, 400 W. 1st Street, Chico, CA 95925, USA^{2}Department of Soil Science, University of Saskatchewan, 51 Campus Drive, Saskatoon, SK, Canada S7N 5A8^{3}Department of Plant Sciences, University of Saskatchewan, 51 Campus Drive, Saskatoon, SK, Canada S7N 5A8^{4}Department of Electrical Engineering, University of Saskatchewan, Saskatoon, SK, Canada S7N 5A9

Received 12 October 2014; Accepted 10 November 2014

Academic Editor: Kent Burkey

Copyright © 2015 Hossein Zakeri et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

#### Abstract

A Soil Plant Analysis Development (SPAD) chlorophyll meter can be used to screen for leaf nitrogen (N) concentration in breeding programs. Lentil (*Lens culinaris* L.) cultivars were grown under varied N regimes, SPAD chlorophyll meter readings (SCMR) were recorded from the cultivars leaves, and leaf N concentration was measured by combustion. Linear regression and the nonlinear Radial Basis Functions (RBF) neural networks models were employed to estimate leaf N concentration (LNC) based on the SCMR values. The closest estimates of LNC were obtained from the multivariate models in which the combination of plant age, leaf thickness, and SCMR was employed as the independent variable. In comparison, SCMR as the single independent variable in both models estimated less than 50% of LNC variations. The results showed significant effects of soil moisture and plant age on the association of LNC –SCMR as well as the relationship of LNC with plant N, grain yield, and days to maturity. However, the effect of cultivar on the measured variables was negligible. Although lentil N can be diagnosed by comparing SCMR values of the crop with those from a well-fertilized (N fixing) plot, the results did not support using SPAD chlorophyll meter for screening lentil LNC.

#### 1. Introduction

The majority of leaf N is accumulated in the chloroplast, where photosynthesis takes place, resulting in a strong association between plant photosynthesis and leaf N status [1]. This association facilitates modeling plant growth and yield via leaf N assessment, because the latter can be rapidly estimated using SPAD chlorophyll meter. This widely used hand-held device measures the ratio of transmitted red (~650 nm) to infrared (~940 nm) electromagnetic radiation from the leaf surface and produces numeric outputs that are related to leaf chlorophyll (*chl*) content (Konica Sensing, Inc., Osaka, Japan). The SPAD Chlorophyll Meter Reading (SCMR) is correlated to leaf N concentration (LNC, e.g., leaf N mass per leaf mass) and can be used to evaluate soil and plant N status, estimate plant N requirement, predict grain yield, and forecast crop maturity [2, 3].

Despite its extensive application, association of SCMR-leaf* chl*/leaf N is often affected by soil and weather conditions, plant age, leaf thickness, leaf area, and leaf position in the canopy [2, 4]. Strategies such as data collection from the canopy apex at mid-day can eliminate daily variations of light and leaf starch concentration and improve the strength of SCMR-leaf N association models [5, 6]. Variations of leaf mass and leaf area can affect LNC (N is diluted in larger mass and area) and leaf spectral reflectance characteristics. Therefore, specific leaf weight (SLW: leaf mass per leaf area) and specific leaf N (SLN: N mass per leaf area) often have cofounding effects on SCMR-LNC models. In rice (*Oryza sativa* L.), LNC had stronger correlation with adjusted SCMR for leaf thickness (SCMR divided by SLW) than it had with SCMR [4]. In tobacco (*Nicotiana tabacum* L.), SCMR produced closer estimates for SLN than it did for LNC [7].

With an assumption of independency of input variables, SCMR (independent variable) is generally fitted against leaf* chl*/leaf N (dependent variable) in a standard linear regression model. When leaf thickness, plant age, soil, and weather affect the model, multivariate and polynomial linear models may result in stronger SCMR-LNC relationships compared to the standard linear model. In three different studies on rice and pigeon pea (*Cajanus cajan* L.), including SLW as the second independent variable improved the prediction power of the model for estimation of LNC based on SCMR values [3, 4, 6]. Similarly, a second-degree polynomial model allowed Castelli and Contillo [7] to interpolate data from two monocot and two dicot species and estimate leaf* chl* based on SCMR. However, combining several independent variables in multivariate linear models () can violate the assumption of variables independency. Under such circumstances, nonlinear regression may provide more reliable fit for SCMR-LNC association than a linear model.

Nonlinear regression continually adjusts the estimated values of the parameters and improves the fit to a minimum satisfactory level [8]. From different nonlinear approaches, artificial neural networks models are widely used to develop association, classification, and prediction models in biology. These models consist of interconnected groups of artificial neurons that pass the information among a series of layers as the weights of observation are changed to achieve the best fit [9]. An advantage of these self-adaptive models is their capability of learning from an existing data set (training), as opposed to entirely relying on theoretical algorithms in linear approach. By employing back propagation neural networks and using plant and soil indexes, Liu et al. [10] estimated rice leaf* chl* concentration by SCMR values with 90% accuracy. In corn (*Zea mayz* L.), actual LNC values were strongly correlated () to the estimated LNC from SCMR by a neural networks model [9]. Despite the complexity in calculation, most neural networks models, such as redial basis function (RBF), are available through statistical packages such as SAS and Matlab and can be employed directly or with minor modifications.

Lentil (*Lens culinaris* L.) is an annual legume plant that produces substantial amounts of leaf, enriched in N. In a field study, 60, 34, and 15% of the above ground biomass and 80, 45, and 13% of plant N content of lentil were accumulated in leaf at flowering, pod, and maturity stages, respectively [11]. We hypothesized that lentil leaf N directly links to plant performance and yield; therefore, plant N content and grain yield of lentil can be rapidly estimated via LNC measurement. To test the hypothesis, we (1) determined the associations of lentil LNC with plant biomass and N content at different growth stages, and with lentil grain yield at maturity, (2) estimated lentil LNC using a SPAD Chlorophyll Meter, and (3) developed linear and nonlinear models for computing lentil LNC via SCMR values.

#### 2. Materials and Methods

##### 2.1. Experiment Setup

Lentil cultivars were grown in Saskatoon (52.05° N and 106.43° W) and Indian Head (50.55° N and 103.65° W), Saskatchewan. In Saskatoon, three N fertility treatments of 50 kg N ha^{−1}, 5.6 kg ha^{−1} granular rhizobia (Nodulator, Becker Underwood, Saskatoon, SK), and a nontreated control were applied on eight lentil cultivars in 2006 and 2007 (N fertility trial). This trial was conducted in two different fields in 2006 and 2007. Compared to 2006, the 2007 field in the N fertility trial had low soil available N and no recorded history of legume crop cultivation and rhizobia inoculation. In the N fertility trial, CDC Greenland, CDC Plato and CDC Sedley (late-maturing group), CDC Milestone and CDC Viceroy (medium-maturing group), and CDC Blaze, CDC Red Rider, and CDC Rouleau (early-maturing group) were grown in both years [12]. The prefix “CDC” represents the Crop Development Centre at University of Saskatchewan, where the cultivars were developed. In both years, lentil was grown in a randomized complete block design (RCBD) with a split-plot arrangement in four replications. The main plots consisted of the three fertility treatments of control, inoculant, and N fertilizer and the subplots consisted of the eight lentil cultivars. In Indian Head, five lentil cultivars were subjected to two different no tillage (NT) durations, one 5 years (short-term NT) and the other 28 years (long-term NT). In this trial, average spring soil available N (NO_{3}-NH_{4}) over the years was 8.9 and 11.3 mg N kg^{−1} soil in the short- and long-term NT plots, respectively. In this trial, CDC Sedley, CDC Vantage and CDC Milestone (medium-maturing group), and CDC Robin and Redcap (early-maturing group) were grown in both years [12]. Here, five lentil cultivars were arranged in a CRBD with three replicates within each NT duration treatment. Varied rainfall and soil available N during the study resulted in four distinct situations: (1) N fertility trial in 2006, where a suitable growing season was terminated by a drought, (2) N fertility trial in 2007, where a severe mid-season drought and low soil N limited N_{2} fixation, N uptake, and grain yield of lentil, (3) NT trial in 2006, where a suitable growing season was terminated by a mild drought, and (4) NT trial in 2007, where a substantial late season rainfall stimulated lentil biomass and N accumulations. More details about the weather, soil N, and lentil performance in these trials are found in [13, 14]. Overall, the N fertility trial, which had 8 lentil cultivars and 3 N fertility treatments, was conducted in Saskatoon in 2006 and 2007. The NT trial with 5 lentil cultivars and 2 no-till duration treatments was conducted in Indian Head in 2006 and 2007.

##### 2.2. Data Collection

Leaf chlorophyll content was estimated using a SPAD Chlorophyll Meter (Model 502 Konica Minolta Sensing, Inc., Japan) at three growth stages of vegetative (up to node 12), first-pod (at least one pod per plant), and late-pod (when the canopy started turning yellow). To eliminate daily variations of light quality and leaf starch concentration, SCMR readings were limited to the uppermost leaves during 10:00 to 12:00 h of day. Three plants per plot were randomly selected and SCMR was recorded from the three fully expanded uppermost leaves of each plant. The average of nine SCMR values in each plot was considered as the plot SCMR value. The leaves were immediately detached and transferred on ice into a refrigerator for further measurements the next day. In the laboratory, leaf surface area was measured using a leaf area meter (Model LI-3100C, LiCor, Lincoln, NE), and leaves were dried at 60°C for 24 hrs, weighed, and ground. Leaf N concentration was measurement by the combustion method, using a Leco carbon-nitrogen determinator (LECO CNS 2000, St. Joseph, MI, USA). Specific leaf weight (SLW), which represents leaf thickness, was the ratio of leaf dry weight (g) to leaf area (m^{−2}), SLN was the ratio of g leaf N (leaf weight × LNC) to leaf area (m^{−2}), and adjusted SCMR for leaf thickness was the ratio of SCMR to SLW. In addition, grain yield and N content of the entire plant biomass (from both trials) and average N content of leaf (referred to entire leaf biomass), stem, and pod from 5 plant plot^{−1} at the three given growth stages (from the N fertility trial) were available.

##### 2.3. Data Analysis

In each trial, the three leaf characteristics (LNC, SLW, and SLN) were analyzed for the effects of treatment and cultivar. In the analysis of variance, the main factor was N fertility treatment (in the N fertility trial) and no-till duration (in the NT trial), and the subfactor was cultivar. Data were analyzed as a year-combined RCBD for each trial-growth stage, with year, treatment, cultivar, and their interactions as fixed variables and block and interaction of block with the fixed factors as random variables [14]. Data were analyzed by the MIXED procedure in SAS, Version 9.2 (SAS Institute, Cary, NC), and differences amongst the means of the fixed effects were tested using Fisher’s protected LSD at . Pearson’s correlation coefficients for LNC with grain yield, harvest index, days to maturity, and biomass, N concentration, and N content (g N plant^{−1}) of leaf, stem, and pod at the three given growth stages were computed for the N fertility trial, using the CORR procedure of SAS.

Following the analysis of variance, the leaf characteristics data from two trials, three growth stages, and two years were pooled to compute the correlation of SCMR and adjusted SCMR with LNC, using the CORR procedure of SAS. This data set consisted of 740 data points for each of the leaf characteristics (LNC, SCMR, SLW, and SLN). The pooled data set was randomly divided into two groups, a training set of 630 data points (85% of the data) and a test set of 110 data points (15% of the data), to develop linear and nonlinear models by the GLM procedure in SAS and the Newrb function in Matlab, respectively. The training set was used in model development and the test data to validate the models. The three growth stages of vegetative, first-pod, and late-pod were arbitrary considered 1, 2, and 3, respectively.

For the nonlinear approach, we tested an RBF neural networks model. This model is linear combinations of radial basis that produces linear outputs based on nonlinear inputs. Using RBF requires specification of the number of hidden unit activation function, the number of processing units, a criterion for modeling a given task, and a training algorithm for finding the parameters of the network. Weight of the model is found through the training process, where network parameters are optimized to fit the network outputs to the given inputs [15]. Four groups of independent variables “SCMR,” “SCMR + SLW,” “SCMR + growth stage,” and “SCMR + SLW + growth stage” were the input independent variables, and LNC was the only dependent variable. Hence, each of the linear and nonlinear approaches resulted in four equations differing in the independent variable (see Table 4 for the linear equations). The developed models were fed by the correspondence independent variable(s) from both the training and the test sets to estimate LNC. The estimated LNC from each set was correlated against the actual measured LNC from the same data set. Pearson’s correlation coefficients of the estimated LNC and actual LNC from the training and test set were considered a measure of model accuracy and the model reliability, respectively.

#### 3. Results

Averaged over the years, trials, treatments, and cultivars, lentil LNC decreased from 4.5% at vegetative to 3.8% at first-pod and 2.7% at late-pod growth stage (Table 1). Average SLN was similar at the vegetative and first-pod growth stages (2.0 g N m^{−2} leaf) and then decreased to 1.4 g N m^{−2} leaf at late-pod. In contrast to leaf N, leaf thickness (SLW) was increased from 42 g m^{−2} leaf at vegetative to 50 g m^{−2} leaf at first-pod and late-pod. Maximum variations of the three leaf properties occurred at the first-pod stage.