Fertilization is considered one of the most effective ways to increase fruit yield and improve its quality. Among the 17 essential nutrients required by plants, nitrogen (N) is one of the key elements for pear tree growth, fruit yield, and quality. Generally, the N demand is relatively large compared with that of other elements. However, the positive response of tree vegetative and reproductive organs to added N triggered a higher application of N fertilizer without paying more attention to actual N requirement of trees, which might lead to an excessive N status in pear trees. The over-application of N not only results in vigorous vegetative growth and decreased sugar-to-acid ratio [1
], but also increases the risk of nitrate leaching from soil, which contributes to the eutrophication of surface waters and degradation of drinking water quality. An excessive soil N availability in summer may delay fruit maturation, negatively impact the total soluble solids (major component is soluble carbohydrate) in fruit, and decrease plant tolerance to pests and diseases, such as psylla (Cacopsylla pyri
L.) on “Bartlett” and post-harvest blue mould (Penicillium expansum
) on “Conference” pears [2
]. Synchronizing N availability with fruit N requirements offers the potential to protect the environment without sacrificing production [3
Tissue testing is one tool that can aid in fertilizer management by evaluating N status and identifying fruit requirements for additional N. The measurement of traditional nutrient status by tissue testing was based on the destructive sampling of leaves, branches and fruits, sometimes even full tree excavation [4
]. The determination of N concentration is labour-intensive and time-consuming in addition to being cost prohibitive (elemental analyser method) and requiring dangerous chemical reagents (Kjeldahl method) which may cause environmental contamination. Recently, with the rapid development and improvement of spectroscopy techniques, monitoring leaf reflectance or the transmittance of radiant energy can be considered a form of rapid, non-destructive tissue testing [6
]. Visible near-infrared spectral reflectance data can be used to evaluate the leaf N concentration and chlorophyll content of cucumber [8
], tomato [9
], oilseed rape [10
] and pear tree [11
] using a specific regression method. However, these studies were generally performed on detached leaves under laboratory conditions without considering variable leaf N status, threshold and diagnosis time, and rarely considering about different models. Moreover, in-field estimating leaf nitrogen using hyperspectral data at canopy scale were mainly found on the trees or crops [12
]. Hence, research is required on developing a similar technology for in-field, non-destructive leaf hyperspectral sensing of pear trees.
Under steady laboratory conditions, spectral measurements are often conducted with integrating sphere because both reflectance and transmittance of leaves can be obtained at the same time. However, it’s too time-costing for the in-field measurement of leaf spectra. For an in-field spectral measurement, method of 25° bare optic fibre was usually used in the situation of whole plant level/the canopy measurement under the solar condition. If the leaf spectra were measured with a plant probe, the black background and the white background were found to be both used and reputable in many studies [14
]. These in-field measurements mentioned above have their own advantages and disadvantages. For one thing, the solar condition is widely used for the in-field spectral measurement owing to the simple measuring device component but the light intensity is always changeful; while an internal and calibrated light of the plant probe is considered to be a steady light source but the measuring time is limited by the battery capacity during the field operation. For another, we are unable to remove the dust on the adaxial surface and the underside of leaves, so that the spectra obtained by using the leaf clip with black and white backgrounds should be compared to improve the signal-to-noise ratio of spectral data. Besides, suitable techniques should be used to make use of hyperspectral data. The techniques must deal efficiently with the strong multi-collinearity present in the spectral data and should not be too sensitive to sensor noise [16
]. The methods such as principal components regression (PCR), partial least squares regression (PLSR), stepwise multiple linear regression (SMLR), and back propagation neural network (BPNN) are found to be good chemometrics methods to dealing with the spectral data [17
]. In addition to these chemometric techniques, the remote sensing community developed over the past four decades a wide range of spectral indicators (e.g., vegetation indices, red-edge indices), responding strongly to the main vegetation biophysical variables, such as leaf area index and leaf pigmentation [19
]. A number of spectrometric studies have been undertaken on plant N content monitoring using the vegetation indices and the red-edge [24
]. However, the four chemometric techniques (PCR, SMLR, PLSR and BPNN) and vegetation indices for modelling the N concentration of pear leaves from spectrometer data collected in the field have not been tested so far.
A non-destructive estimate of yield was mainly conducted by collecting canopy spectra of a satellite platform and airborne hyperspectral scanners. Examples included the prediction of biomass and yield of winter wheat under different nitrogen supplies using spectral indices [27
] and the grain yield prediction using reflectance spectra of canopy [28
]. However, there is a lack of sensors on satellite platforms with optimal spatial resolution to monitor orchard crops at the tree scale [29
]. Therefore, the estimation of fruit yield by the canopy spectra in the orchard often had a low R2
, for example 0.33 in peach [29
] and 0.58 in citrus [30
In the present study, the best method to non-destructively evaluate the leaf nitrogen concentration by in-field visible/near infrared spectroscopy will be established. In addition, the right time for yield estimation by the leaf N status and the reference threshold of leaf N status for fertilization decisions will be find out.
2. Materials and Methods
2.1. Plant Material and Treatments
A two-year field experiment was conducted from 2014 to 2015 in a commercial pear orchard in Gaochun (Jiangsu Province, China; 32°03′ N, 118°46′ E). The experimental site has an annual mean temperature of 15.9 °C and receives 1157 mm of precipitation. The soil has a clay-loamy texture with 69 mg·kg−1
alkali-hydrolysable nitrogen [31
], 43 mg·kg−1
available phosphate by the method of Olsen [32
], 146 mg·kg−1
available potassium by the method of Sommers [33
], 17 g·kg−1
organic matter, and a pH 6.8 in water. Pear trees (Pyrus communis
L.) that had been managed by trellis cultivation in a 3 m × 5 m frame for 12 years were used in the field experiment during the growing period (May to June). Referring to common N fertilization amounts in the region (660 kg·N·ha−1
), six N fertilization treatments consisting of 0 (N0), 165 (N1), 330 (N2), 660 (N3), and 990 (N4) kg·N·ha−1
as urea (46% N) were separately split-applied four times (post-harvest, late autumn, pre-bloom stage and expansion stage) at 20%, 40%, 20%, and 20% of the total amount, respectively. Calcium superphosphate (12% P2
) was applied at 528 kg·ha−1
only in late autumn as base fertilizer. In addition, for the quality of pear fruit, 990 kg·ha−1
of potassium sulphate (45% K2
O) was applied three times (late autumn, pre-bloom stage, and expansion stage) at 40%, 20%, and 40% of the total amount, respectively.
In a completely randomized block design, three replicates of three trees each were arranged in three alternate tree rows. The control comprised of five trees. From late autumn 2013 to the post-harvest stage of 2015, fertilizer was applied to six holes (30 cm wide × 30 cm deep) in a circle around each tree. The standard cultural practices used in local commercial production, including pruning, irrigation, and pest control were employed. Every season, the fruit crop load was manually adjusted by large-scale thinning to ensure high commercial quality.
2.2. Spectra Collection
As reported by Tagliavini et al., the remobilization of nitrogen ceased between petal fall and the beginning of pear fruit development [34
]. Nitrogen was transported into shoot leaves several weeks after bloom; shoot leaves are more dependent than spur leaves on spring N uptake. Therefore, the N concentration in new shoot leaves is sensitive to the N supply and can be used to evaluate the N status of the whole tree. The middle leaves of the year’s spring flush from the external side (east, south, west, and north) of the canopy were collected at 50 days (May) and 80 days after full bloom (80 DAB, June) for the spectral measurement. Eight or twelve leaves facing four directions from each tree were measured to represent the N status of the whole tree. In-field spectral measurements were achieved using a portable field spectrometer FieldSpec 3 (Analytical Spectral Devices, Boulder, CO, USA). The instrument covered wavelengths of 350–1000 nm (with a sampling interval of 1.4 nm and a spectral resolution of 3 nm) and 1000–2500 nm (with a sampling interval of 2 nm and a spectral resolution of 10 nm). The output spectral band number was 2151, and the interval of re-sampling was 1 nm. Before measuring each tree, Teflon white standard (Spectralon, Labsphere Inc., North Dutton, NH, USA) was used to set up the maximum reflectance (99.9%) conditions. The reflectance of each leaf was determined from the measurement of leaf radiance divided by the radiance of the reflective white standard. Three spectral measurement methods were compared: (1) 25° bare fibre under solar conditions, which was restricted by weather (sunny, cloudless days only, from 10:00 a.m. to 3:00 p.m.); (2) a leaf-clip assembly attached to a plant probe with an internal, calibrated light source (the transflectance of each leaf was determined using the white background of the leaf-clip); and (3) the black background of leaf-clip to determine absolute reflectance of each leaf (Figure 1
). Two central symmetrical points on the leaf adaxial surface were designated as the measurement points on all leaves. The scan number at a given position was set to five, and the average value of 10 spectra was used as the final reflectance of the leaf.
2.3. Modelling Methods
Before modelling, preprocessing methods were used on the raw spectra. Wavelengths with very large atmospheric influence were removed for the 25° field of view measurements under solar conditions. Normalization was used on the raw spectra collected by the probe attached with backgrounds using Unscrambler X 10.3 (Camo Software AS, Oslo, Norway). Using the spectral data and N concentration, we investigated the predictive power of principal component regression (PCR), stepwise multiple linear regression (SMLR), partial least squares regression (PLSR), back propagation neural network (BPNN), and vegetation indices (difference vegetation index, ratio vegetation index, and normalized differential vegetation index). Full-spectrum methods as PCR, PLSR, and BPNN use all available wavelengths simultaneously [16
]. In contrast, SMLR selects useful wavelengths from the available spectrum and ignores the remaining wavebands during model application.
The PCR and PLSR methods were analysed comparatively using Unscrambler X 10.3. The model obtained the optimal number of factors by leave-one-out cross-validation and the mean square error in calibration.
The data compression stage of the SMLR consists of selecting a combination of a few spectral bands as regression factors. The regression factor matrix is then used to calculate the biophysical loadings by ordinary least square. In the present study, the wavelengths were first regressed sequentially against the biophysical variables. The wavelength with the highest explained variance was then chosen as the first regression factor. With this first regression factor fixed, the next wavelength was chosen, and so on. Forward and backward elimination were not investigated. SMLR was calculated by SPSS 20.0 (SPSS Inc., Chicago, IL, USA).
The back propagation neural network (BPNN), based on principal component analysis (PCA), was used to develop the models for predicting the leaf N concentration. For the PCA-BPNN models, PCA was performed first to extract information from the whole spectral regions, and some principal components which carried over 80% information of the original dataset were used as the neurons of the network input layer. Several network architectures were tested by varying the number of neurons in the hidden layer with different initial weights (at least 10 times). The optimal parameters of the target error, the training rate, and the iteration were determined by the least prediction error [35
As a baseline method, linear regression models between N concentration and vegetation indices (DVI, RVI, and NDVI) were analysed. For the fast analysis and feature extraction from the observed datasets, we used the report by Yao [36
]. The reduced sampling method was used to construct a difference spectral index (DSI), normalized difference spectral index (NDSI), and ratio spectral index (RSI). In this procedure, the spectral reflectance data were read at intervals of 10 nm within the range of 350–2500 nm. With the duo matrix formulation, all of the possible DSI, NDSI, and RSI based on any two individual bands at the interval of 10 nm from 350–2500 nm were regressed against an N concentration linear equation. According to the changing coefficients of determination (R2
), a contour map of R2
was plotted. From this map, a sensitive spectral range with a relatively high R2
was identified. The vegetation index and BPNN were calculated in MATLAB R2012b (MathWorks, Natick, MA, USA).
The quality of the calibration model was evaluated using the following statistical parameters: coefficient of determination between predicted and measured N concentrations (R2
), root mean square error in calibration (RMSEC), and root mean square error in validation (RMSEV). As mentioned by Saeys [37
], a calibration model with an R2
value greater than 0.91 is considered to be excellent, whereas R2
between 0.82 and 0.90 represents good prediction. A small difference between the RMSEC and RMSEV values is also important to avoid ‘over-fitting’ in the calibration and validation phase [38
2.4. Data Pre-Treatment
No attempts were made to reduce the number of input variables through feature selection in order to compare the modelling accuracy of the seven methods. Before modelling, PCA was used to eliminate the outliers based on the spectral principal component analysis of 720 samples [11
]. We found a large deviation in the spectra of most samples. After elimination, 710 samples were divided into the calibration set and the validation set using a randomly-selected collection for analysis (Table 1
2.5. Leaf Nitrogen Concentration and Yield Estimation
The total N concentration of the leaf samples was measured analytically as a reference for the spectral prediction models. After spectral measurements were taken, the leaves were picked off the tree, sealed in valve bags, and taken to the laboratory for analysis. All of the leaves were first placed in a forced-air oven at 105 °C for 30 min and then dried to a constant weight at 70 °C for 48 h. The samples were finely ground (100 mesh) and analysed for the total nitrogen concentration using an Elementar Vario Macro CHN analyser (Elementar Analysensyteme GmbH, Hanau, Germany). A standard citrus leaf sample (GBW10020) was included to ensure accuracy of N concentration measurements.
The fruit number per tree was counted at both the fruit expansion stage and the maturity stage to ensure the reliability of fruit number. Sixteen pear fruits were collected per tree at maturity and weighed to estimate the average single fruit weight per tree. The yield was calculated by multiplying the single fruit weight by the fruit number as kg per tree. N concentration of each treatment was obtained by an average of three replications’ leaf N concentration. Moreover, we performed polynomial curve fitting between the chemically-measured leaf N concentration and the yield to find the better time estimating the yield.