Introduction
Apricot (Prunus armeniaca L.) is an important fruit crop worldwide, belonging to Rosaceae family. The apricot production was estimated to be about 3,578,412.14 tonnes in the world and 695,632.70 tonnes in Europe, in 2021 (FAOSTAT, 2023). Europe is the second world’s apricot producer after Asia (Figure 1). Italy is the fourth world’s apricot producer, after Turkey, Iran and Uzbekistan, and it is the largest producer in Europe followed by France, Spain and Greece (Figure 1). In 2021, the apricot annual Italian production was 189 570 tonnes on 17 740 ha of cultivated area, in five Italian Regions: Emilia-Romagna (31%), Campania (21%), Basilicata (20%), Puglia (6%) and Sicilia (5%). In the Mediterranean area, a large part of production is primarily fresh fruit, with a smaller amount destined for processing industry for various products such as dried fruit, fruit cans, jam and juices, ice cream, cheese, etc. (Alvisi, Reference Alvisi1997; Chang et al., Reference Chang, Alasalvar and Shahidi2016).
It has been reported that mineral nutrition, particularly nitrogen content (N), plays a key role in plant growth, yield and fruit quality including, firmness, sugar content, phenolic compounds and visual appearance (skin colour) (García-Gomez et al., Reference García-Gomez, Ruiz, Salazar, Rubio, Martínez-García and Martínez-Gomez2020; Zhebentyayeva et al., Reference Zhebentyayeva, Ledbetter, Burgos, Ll´acer, Badenes and Byrne2012; Radi et al., Reference Radi, Mahrouz, Jaouad, Tacchini, Hugues and Amiot1997, Radi et al., Reference Radi, Mahrouz, Jaouad and Amiot2003; Falls and Siegel, Reference Falls, Siegel, Poole, Townshend and Worsfold2005; Macheix et al., Reference Macheix, Fleuriet and Billot2018; Wang et al., Reference Wang, Wang and Wang2007; Asma et al., Reference Asma, Colak, Akca and Genc2007; Dimitrovski and Cvetkovic, Reference Dimitrovski and Cvetkovic1981). Since nitrogen content in plant was directly associated with chlorophyll synthesis (Boussadia et al., Reference Boussadia, Steppe, Zgallai, Ben El Hadj, Braham, Lemeur and Van Labeke2010; Kamnev et al., Reference Kamnev, Sadovnikova and Antonyuk2008), chlorophyll content in leaves is an indicator of the N status (Li et al., Reference Li, Yang, Fei, Song, Li, Ge and Chen2009; Shaahan et al., Reference Shaahan, El-Sayed and Abou El-Nour1999). The proportion of leaf N allocated to the chloroplast is approximately 75% (Hak et al., Reference Hak, Rinderle-Zimmer, Lichtenthaler and Natr1993; Kutik et al., Reference Kutik, Nátr, Demmers-Derks and Lawlor1995). Leaf N is contained not only in chlorophyll but also in enzymes, vitamins and nucleic acids as well as in proteins. (Carranca et al., Reference Carranca, Brunetto and Tagliavini2018; Khasawneh et al., Reference Khasawneh, Alsmairat, Othman, Ayad, Al-Qudah and Leskovar2021; Khasawneh et al., Reference Khasawneh, Alsmairat, Othman, Ayad, Al-Hajaj and Qrunfleh2022; Mratinić et al., Reference Mratinić, Popovski, Milošević and Popovska2011).
Chlorophyll measurements performed by the SPAD (Soil and Plant Analysis Development) sensor (Minolta Corporation, Ltd., Osaka, Japan) (Minolta Camera, Reference Minolta Camera1989) record chlorophyll values in a non-destructive way by acquiring values of the leaf transmittance at red (650 nm) and infrared (940 nm) wavelengths. The standard method for determining chlorophyll content is very accurate but destructive. (Amirruddin et al., Reference Amirruddin, Muharam, Ismail, Ismail, Tan and Karam2020; Wu et al., Reference Wu, Zhang, Zhao, Xie and Hou2023). Compared with the traditional destructive methods (Porra et al., Reference Porra, Thompson and Kriedemann1989), SPAD analyses many leaf samples in small amounts of time, space and resources, leading to an exponential increase of its use in the last decade. (Uddling et al., Reference Uddling, Gelang-Alfredsson, Piikki and Pleijel2007).
Hyperspectral analysis shares the versatility of SPAD, but in a full range of 350–2500 nm. It reads the reflectance with a higher resolution to gather more accurate information (Liu et al., Reference Liu, Li, Zhang, Wang, Guo, Long, Yang, Wang, Li, Hu, Wei and Xiao2020; Tang et al., Reference Tang, Dou, Cui, Liu, Gao, Wang, Li, Lei, Zhao, Zhai and Li2022). These characteristics have been used to assess N content using vegetation indices such as Normalized Difference Nitrogen Index (NDNI) at specific wavelengths (Götze et al., Reference Götze, Jung, Merbach, Wennrich and Gläßer2010; Osborne et al., Reference Osborne, Schepers, Francis and Schlemmer2002) or greater ranges until to the full spectrum between 350–2500 nm (Bruning et al., Reference Bruning, Liu, Brien, Berger, Lewis and Garnett2019; Miao et al., Reference Miao, Mulla, Randall, Vetsch and Vintila2009). To assess N content, Continuum Removal (CR) methodology was carried out (Kokaly and Clark, Reference Kokaly and Clark1999) selecting the ranges where absorption peaks were evident along the spectrum (Curran et al., Reference Curran, Dungan and Peterson2001; Huang et al., Reference Huang, Turner, Dury, Wallis and Foley2004; Van Der Meer, Reference Van Der Meer2004). While SPAD directly measures chlorophyll at certain wavelengths and this value is correlated with nitrogen content, the spectroradiometer works on the range between 350 and 2500, detecting other signals in the short-wave infrared region (SWIR), between 1400 and 2500 nm, correlated with (a) nitrogen content not of chlorophyll but of other compounds as proteins, enzymes, involved in the plant metabolism; (b) other organic compounds as carbohydrates produced during chlorophyll photosynthesis. Thus, it might help to understand more fully the physiological state of the plant.
Leaf nutrient levels in apricots are non-uniform, showing seasonal variations (Leece and van den Ende, Reference Leece and van den Ende1975) and dependence upon cultivar (Bojic et al., Reference Bojic, Milosevic and Rakocevic1999), rootstock (Rosati et al., Reference Rosati, DeJong and Southwick1997, Velemis et al., Reference Velemis, Almaliotis, Bladenopoulou and Karayiannis1999, Jiménez et al., Reference Jiménez, Garín, Gogorcena, Betrán and Moreno2004), interstock (Milosevic Reference Milosevic2006, Milosevic and Milosevic Reference Milosevic and Milosevic2011) and fertilization (Szücs Reference Szücs1986). It has been reported that the organization or relative amounts of photosynthetic components differ between sun and shade leaves (Hikosaka and Terashima, Reference Hikosaka and Terashima1995; Hoel and Solhaug, Reference Hoel and Solhaug1998). Furthermore, an important component for a higher productivity of quality fruit in an orchard is the design (Javaid et al., Reference Javaid, Qureshi, Masoodi, Sharma, Fatima and Saleem2017) for maximum light interception, and trees should be oriented in North-South direction (Javaid et al., Reference Javaid, Qureshi, Masoodi, Sharma, Fatima and Saleem2017). Other authors (Boissard et al., Reference Boissard, Guyot, Jackson, Steven and Clark1990; Leinonen and Jones, Reference Leinonen and Jones2004; Paltineanu et al., Reference Paltineanu, Septar and Moale2013; Zia et al., Reference Zia, Wenyong, Spreer, Spohrer, Xiongkui and Müller2012, Wang et al., Reference Wang, Tuerxun and Zheng2024) have studied how leaf orientation and canopy geometry represented by row orientation, row spacing and plant height interact with environmental factors, and the importance of cardinal point in peach, apple and walnut orchards, but not in apricots.
Based on these considerations, in this work an apricot orchard was studied with the following aims:
-
1) to explore solutions that would make data collection more efficient considering how time-consuming using proximal sensing tools may be. We have studied whether the results of SPAD prediction models might vary depending on the leaf position on the tree. Following studies should evaluate whether some of these positions are more representative of the whole tree, allowing more targeted and efficient data collection. In this preliminary study, the four cardinal points were chosen because they represent easily identifiable standard positions on the apricot tree;
-
2) to assess plant physiological state in an experimental study where synergy of SPAD and hyperspectral sensor was considered. This is a very preliminary attempt to assess how SPAD prediction models may describe metabolic mechanisms involving other substances, such as sugars, proteins, detected by hyperspectral analysis, during chlorophyll photosynthesis.
Prediction models of SPAD measurements were built using hyperspectral data as auxiliary variables along the entire spectrum between 350 and 2500 nm models or in ranges where absorption peaks were evident along the spectrum (Curran et al., Reference Curran, Dungan and Peterson2001; Huang et al., Reference Huang, Turner, Dury, Wallis and Foley2004; Van Der Meer, Reference Van Der Meer2004). These models were built on two types of datasets: (1) with all data collected on tree leaves for each of the two apricot varieties, Farlis and Farbaly; (2) with data grouped according to the four cardinal points on trees of the same apricot varieties.
Materials and methods
Description of site
The study was conducted on apricot orchard in the Apulia Region (Southern Italy) at latitude 40°53’26’’ N and longitude 17°5’5’’ in hilly territory and situated about 20 Km from the Adriatic Sea, in a private farm. The two varieties chosen were Farlis and Farbaly, both were about 8 years old, had a late ripening and a good fruit quality.
The orchard was grown in clay-loam soil and was irrigated using drip system, following the usual agricultural practices of the area. Orchard trees were spaced 3 m within rows and 5.5 m between rows. This study was carried out in July 2021.
Sampling design
Leaves of nine trees of Farlis and Farbaly were analysed in the field. Each tree was divided into 4 cardinal points east, north, south, west. Three leaves were selected on each cardinal point for a total of 108 leaves representing each variety. Three repetitions were measured in situ across the leaf with SPAD and after on the same points with Field Spec 4. Finally, the 324 SPAD and spectral data were averaged to get an overall dataset (OD) of 108 measurements that represented each variety. Furthermore, the data of each cardinal point of varieties were elaborated. For this aim, subsets of the overall dataset (SOD) of 27 SPAD and spectral data represented the cardinal points with each variety.
SPAD
The SPAD-502 (Minolta Corporation, Ltda., Osaka, Japan) (Minolta, 1989) measured the transmittance of red light (650 nm) and infrared radiation (940 nm) through the leaf giving out a SPAD dimensionless reading as an indicator of the amount of chlorophyll in the leaf tissue according to this equation (Naus et al., Reference Naus, Prokopova, Rebicek and Spundova2010) (Eq. 1):
where k is a slope coefficient and C is a confidential offset value.
Spectral data
Hyperspectral analysis
Leaf spectral measurements were performed in the field with ASD Field Spec 4 Portable Spectroradiometer (Analytical Spectral Devices Inc., Boulder, Colorado, USA) on the same points where SPAD analysis was performed. Plant Probe was used to detect a spectral signature in a range of 350–2500 nm. Field Spec 4 provided spectra with 2151 bands having a resolution of 1 nm. The spectral reflectance signatures were averaged over 10 nm to reduce the number of wavelengths from 2151 to 215, smoothing the spectra and keeping down the risk of over-fitting (Shepherd and Walsh, Reference Shepherd and Walsh2002). The calibration was performed by a standard white reference of Plant Probe with a known reflectance of 99% and repeated for each tree, to increase the comparability of measurements.
Spectral transformations
Since the aim of this paper was to study the performance of SPAD prediction models, vegetation indices (VI) associated with chlorophyll and nitrogen content were considered. They were Chlorophyll Index (CI), Normalized Pigment Chlorophyll Index (NPCI) associated with chlorophyll content and Normalized Difference Nitrogen Index (NDNI) with nitrogen content (Table 1).
* Rn represents the reflectance values and n the wavelength.
Furthermore, CR methodology was applied on all the ranges along the spectrum of 350–2500 nm showing absorbance peaks (R ranges) to compute Depth (Table 2) as follows (Eq. 2):
where Rb is the reflectance at the band bottom and Re is the reflectance on the conjunction line called continuum at the same wavelength of Rb so that no local maximum is higher than 1 (Van Der Meer, Reference Van Der Meer2004). This calculation was performed by ViewSpecPro software (Analytical Spectral Devices Inc., Boulder, Colorado, USA).
* R indicate the ranges along the spectrum of 350–2500 nm showing absorbance peaks.
Statistical analysis
Prediction and model estimation
Two types of estimation models were applied on ODs and SODs constituted by SPAD values (dependent variable), and spectral data (independent variables or predictors): (i) ordinary least square regression (OLSR) applied on VI and CR indices; (ii) partial least square regression (PLSR) applied on full spectrum of 350–2500 nm (FS) and R ranges. The two apricot varieties were estimated separately, as well as the cardinal points. The analysis was performed with statistical software package SAS/STAT (release 9.4 SAS ANALYTICS U software).
Cross-validation
Cross-validation was performed as described by Riefolo et al. (Reference Riefolo, Castrignanò, Colombo, Conforti, Ruggieri, Vitti and Buttafuoco2020). This procedure was performed on cardinal points subset too: two-third of the subset (18 samples) as calibration set, and the remaining one-third of the samples (9 samples), as validation set. The performance of prediction models was evaluated by means of three statistics: (i) the coefficient of determination in prediction (R2); (ii) root mean square error of prediction (RMSEP); (iii) residual prediction deviation (RPD) (Bellon-Maurel et al., Reference Bellon-Maurel, Fernandez-Ahumada, Roger and McBratney2010) defined as follows (Eq. 3):
where SD is the standard deviation of the response variable SPAD. It is used to standardize the value of RMSEP with respect to the dispersion of samples enabling to compare the effectiveness of the prediction model as follows: (i) RPD > 2 excellent; (ii) 1.4 ≤ RPD ≤ 2 good; (iii) RPD < 1.4 unreliable (Chang et al., Reference Chang, Laird, Mausbach and Hurburgh2001). After cross-validation, analysis of residuals was performed with Shapiro–Wilk and Kolmogorov–Smirnov tests, to evaluate the normality of distribution. The selection of the best model was based on the following criteria: (i) RPD values ≥ 2.0 with the smaller number of latent variables; (ii) normality of residuals.
Analysis of variance
Analysis of variance of all variables relative to cardinal points was performed to find significative difference among them. The normal distribution of variables was verified by Shapiro-Wilk and Kolmogorov-Smirnov tests to choose the correct analysis of variance (data not shown). Test of Levene verified the homoscedasticity (data not shown). Based on these preliminary tests, four types of analysis of variance with their corresponding post hoc tests were applied (Table 3). Table 4 shows the results of some VI and CR indices of Farbaly, since SPAD, Depth1460 and NPCI index did not show any significative differences.
* Different letters indicate significant differences (p < 0.05) among the cardinal points.
Results
A total of 23 SPAD prediction models were fitted for each variety: 8 models for wavelength ranges, full and partial (R ranges) elaborated with Partial Least Square Regression (PLSR) and 9 regarding VI and CR indices elaborated with Ordinary Least Square Regression (OLSR).
Analysis of normal distribution
Table 5 shows the basic statistics of the response variable SPAD for the two varieties in both ODs and SODs. Shapiro-Wilk and Kolmogorov-Smirnov tests were used to assess the normal distribution. When at least one of the two tests was significant at a level probability of 5% (p < 0.05), SPAD was transformed in Gaussian ranks by SAS/RANK procedure: the ranks divided by the total number of observations form values in the range 0–1, which were used in subsequent processing. Predictors (VI and CR indices and spectral data) and response variable (SPAD rank transformed) were centred and scaled to have the mean at zero and the variance at 1 and to place both on the same relative position to their variation in the process of prediction model estimation.
* Mathematic factor computed by Shapiro-Wilk test.
† Mathematic factor computed by Kolmogorov-Smirnov test.
Prediction models
Farbaly OD
Table 6 shows PLSR statistics regarding R ranges. The only value of RPD ≥ 1.4 belonged to R1R6 range with 9 factors, (highest R2 0.623 and lowest RMSEP 0.610) (Figure 2). No VI and CR indices showed a RPD value ≥ 1.4 with an explained variance that never exceeded 10% (data not shown).
* Mathematic factor computed by Shapiro-Wilk test.
† Mathematic factor computed by Kolmogorov-Smirnov test.
Farbaly subsets
Table 7 shows PLSR statistics of SODs in descending order of RPD values until the last good effective model. Although the greatest value of RPD belonged to the R1R6 range in the east SOD (10.657), that represented by the R3 range of the same SOD was chosen as the best model (2.160) (Figure 3). It, among the models with an excellent RPD value, shows the lowest number of factors. Ten models have an excellent RPD value and a normal distribution of residuals (Table 7) and eight a good one (data not shown). Considering the models with RPD ≥ 1.4, the north SOD is represented six times while the west one only once (data not shown).
In bold the models chosen for each variety.
* Mathematic factor computed by Shapiro-Wilk test.
† Mathematic factor computed by Kolmogorov-Smirnov test.
No VI and CR indices show a RPD value ≥ 1.4 both with an explained variance that never exceeds 50% (data not shown). The highest value of explained variance belongs to Depth1000 (49,90%) with a RMSEP of 0,696 and a RPD of 1.386, followed by Depth1460 (41,30%) with a RMSEP of 0,754 and a RPD of 1.280 in east SOD.
Farlis OD
Table 6 shows PLSR statistics regarding R ranges. The only value of RPD ≥ 2.0 belonged FS range, with 15 factors (highest R2 0.886 and lowest RMSEP 0.504) (Figure 2) while the other ranges never exceeded the 1.4 value. No VI and CR indices show a RPD value ≥ 1.4 both in OD with an explained variance that never exceeded 12% (data not shown).
Farlis subsets
Table 7 shows PLSR statistics of SODs in descending order of RPD values until the last good effective model. Although the greatest value of RPD belongs to the FS range in the north SOD (10.394), that represented by the R1 range of the east SOD was chosen as the best model (2.276), since it showed the lowest number of factors and a normal distribution of residuals (Figure 4). Nine models had an excellent RPD value and a normal distribution of residuals (Table 7), six a good one (data not shown). Since the west SOD model departed by the normal distribution, it is excluded by the effective models. Considering the models with RPD ≥ 1.4, the north SOD is represented six times, while the south one twice (data not shown).
No VI and CR indices show a RPD value ≥ 1.4 both with an explained variance that is around 15% (data not shown). The highest value of explained variance belongs to NDNI (15.44%) with a RMSEP of 1.470 and a RPD of 1.067, in east SOD, followed by NPCI (15.23%) with a RMSEP of 1.260 and a RPD of 1.062 in south SOD.
Discussion
The models based on the wider ranges of spectrum, FS and R1R6, were the best ones both in Farlis and Farbaly ODs (Table 6). Considering the SODs results, the cardinal point east in Farbaly and the north one in Farlis showed the highest number of excellent models (Table 7). However, in Farbaly north SOD showed the highest number of models with RPD ≥ 1.4 (data not shown). These results were consistent with those of the ANOVA for Farbaly, since the north cardinal point showed the highest number of significative differences, 14, followed by the west one and as the lowest the south one (Table 4). Farlis showed a significative difference in CI index only between north and south cardinal point at a level of 1% (data not shown). The results suggested the possibility that metabolic behaviour could vary within the plant, even according to cardinal points and the importance of row direction in orchard design as reported by Javaid et al. (Reference Javaid, Qureshi, Masoodi, Sharma, Fatima and Saleem2017).
Considering the R ranges corresponding to the absorbance peaks where Depths were calculated (Table 2), the R1 and R2 ranges included the wavelengths at which SPAD and the spectroradiometer both worked. From R3 onward, only the spectroradiometer was working. They have been associated with physiological characteristics of the plant based on previous studies. R1 (400–820 nm) corresponded to the absorbance range of chlorophyll while R2 (820–1110 nm) was associated with leaf structure (Bauer, Reference Bauer1985; Knipling, Reference Knipling1970; Peñuelas et al., Reference Peñuelas, Filella, Biel, Serrano and Save1993). The results showed excellent RPD for these two models to the north and east in both Farbaly and Farlis, and this was attributed to the fact that R1 and R2 ranges included SPAD wavelengths (650 and 930 nm). R3 range was associated with leaf water content (Clevers et al., Reference Clevers, Kooistra and Schaepman2010; González-Fernández et al., Reference González-Fernández, Rodríguez-Pérez, Marabel and Álvarez-Taboada2015). It became interesting to note that although there was no longer the overlap with the wavelengths where SPAD worked, for R3 the RPD value was still reliable in both Farbaly to the east (2.16) (Table 7) and Farlis to the north (1.88, not shown in Table 7). This could have been caused by the relationship between leaf structure, characterizing R2 where SPAD also worked, and R3 characterizing leaf water content. From the R3 range onward, the prediction models with the highest RPD no longer belonged to the same cardinal points in Farbaly and Farlis. Farbaly showed a good value of RPD for R6 model, whereas Farlis showed an excellent RPD value for R4 model. The ranges corresponding to R4 and R6, shared an association with organic compounds, involving stretching and bending deformations of O-H link as in the carbohydrates, over all in starch that represents plant energy reserve (Fourty et al., Reference Fourty, Baret, Jacquemoud, Schmuck and Verdebout1996). Therefore, these prediction models might be associated with photosynthetic activity. But whereas the R4 model would refer only to carbohydrates, the R6 one would also be associated with the presence of protein (Ecarnot et al., Reference Ecarnot, Compan and Roumet2013; Fourty et al., Reference Fourty, Baret, Jacquemoud, Schmuck and Verdebout1996). Both varieties showed that the R4 and R6 models were mutually exclusive: when one model is reliable, the other is not. RPD value for R5 model associated with cellulose (Fourty et al., Reference Fourty, Baret, Jacquemoud, Schmuck and Verdebout1996; Shenk et al., Reference Shenk, Workman, Westerhaus, Burns and Ciurczak2001) forming the structural basis of the tree (roots, stems and leaves), resulted good only for Farbaly. This could confirm that the prediction models of SPAD produced are associated with compounds concerning the metabolic activity of leaf and not of tree structure. It is noteworthy that these compounds have a chemical affinity with water, associated with the ranges showing good predictive models at the north cardinal point. So, these results should confirm that the row direction in orchard design gets maximum light interception in trees as reported by Javaid et al. (Reference Javaid, Qureshi, Masoodi, Sharma, Fatima and Saleem2017).
Conclusion
In this paper, the variation of physiological response on different cardinal points of apricot trees was studied. This study is in addition to others with similar purposes carried out in the past on other types of orchards. The choice to assess the efficiency of SPAD prediction models on cardinal points showed the best results for the north one. VI and CR indices did not produce reliable predictive models even if their analysis of variance showed a greater number of significative differences for north. Considering both the OD and the subsets of cardinal points, models referring to the widest ranges of wavelengths showed the best performance. When the wavelengths ranges where SPAD worked no longer overlapped that of the hyperspectral sensor, the cardinal points of the best predictive models were not the same in the two varieties. These preliminary results, although agreeing with those of other studies, should be confirmed using more measurements taken on each cardinal point. They suggest the possibility of identifying plant points more suitable for producing reliable prediction models of SPAD that, involving many aspects of the leaf metabolic activity thanks to hyperspectral analysis, enable to assess plant physiological state.
Author contributions
CR contributed to methodology, software, validation, formal analysis, writing – original draft preparation, visualization. CR and LD contributed to conceptualization, investigation, data curation, writing – review and editing. LD contributed to resources, supervision, project administration and funding acquisition.
Funding statement
This research was carried out in the framework of the project ‘TAGs – Technological and business innovation services to stimulate the local Agro-Food ecosystems and to support a cross-border collaboration among local action Groups’, a project co-funded by European Union, European Regional Development Funds (E.R.D.F.) and by National Funds of Greece and Italy, Interreg V-A Greece-Italy Programme 2014–2020 (MIS CODE: 5 003 507).
Competing interests
None.