Introduction

Improving the light environment within a complex fruit tree canopy has always been of great practical interest for enhancing fruit yield and quality (Jackson 1980; Wunsche and Lakso 2000; Coupel-Ledru et al. 2019). Plant architecture, i.e., the topological arrangement of branching and successive growth as well as geometrical attributes of plant components, determines the spatial light distribution and interception within the canopy (Costes et al. 2006; Barthélémy and Caraglio 2007). The resulting heterogeneous light distribution within the canopy mediates microclimate and organ activities by energy exchange with the surrounding environment, at both organ and canopy scales (Niinemets 2010; Gullo et al. 2014; Poorter et al. 2019). For instance, changes in the amount of intercepted light by organs can modify the leaf nitrogen spatial distribution, leaf transpiration (Giuliani et al. 1998; Larbi et al. 2015), carbon assimilation and allocation (van Hooijdonk et al. 2015; Afonso et al. 2017), as well as flowering, fruit production, and composition (Hampson et al. 2002; Feng et al. 2014). Moreover, light affects pathogen attacks, development, and spatial distribution within the canopy (Costes et al. 2013).

Several architectural traits have previously been confirmed to alter the light environment, notably leaf area density (LAD) (Wagenmakers and Callesen 1995), internode length (Han et al. 2012), and branching patterns (Da Silva et al. 2013; Tang et al. 2019). However, tree architectural characteristics were seldom considered in long-term breeding programs compared to disease resistance, fruit yield, and quality (Laurens et al. 2018). Orchard managers around the world conventionally manipulate apple tree canopies via training, pruning, bending, and leaf removal to optimize the light penetration into canopy and fruit attributes to obtain desired quality standards (Jackson 1980; Robinson and Lakso 1991; Li et al. 2003; Stephan et al. 2008; Xing et al. 2016). Based on those studies, canopy manipulations not only balance light interception and distribution within canopy (Corelli-Grappadelli and Lakso 2007), but also allow an adequate amount of light to reach the fruiting shoot leaves to obtain high-quality fruits (Willaume et al. 2004; Fanwoua et al. 2014). Furthermore, dwarfing rootstocks by reducing the tree size and increasing planting density have been shown to efficiently improve light utilization in orchards (Yang et al. 2016, Buler et al. 2001). In high-density orchards with dwarfing rootstocks, the tall spindle has become the predominant and most profitable training system in most areas of the world (Tustin et al. 1998; Lauri et al. 2004; Willaume et al. 2004; Robinson et al. 2007, 2013; Reig et al. 2019). Nevertheless, the inner central part of a single-axis tall spindle tree is not well illuminated, resulting in excessive annual vegetative growth on the exposed side of scaffolds and a reduced percentage of red-colored fruit (Palmer and Warrington 1998; Buler and Mika 2004; Yang et al. 2016; Reig et al. 2019). A training system called Bibaum®, based on bi-axis trees made of two primary scaffolds, was applied to apple and pear trees (Musacchi 2008a). It was reported to allow a higher leaf net photosynthesis rate and dry matter production than the tall spindle training system, presumably due to an early leaf area development that in turn increased the light interception (van Hooijdonk et al. 2015, 2016; Afonso et al. 2017). However, the impacts of bi-axis Bibaum® on the canopy structure, light interception, spatial distribution and partitioning at both tree and intra-tree scales, and on other canopy performances in terms of tree photosynthesis, transpiration, and water use efficiency have not yet been well elucidated.

Over the years, three-dimensional (3D) virtual plant modeling approaches have been successfully used to characterize canopy structures and light interception efficiency (Willaume et al. 2004; Stephan et al. 2008; Da Silva et al. 2014; Perez et al. 2019). In 3D virtual plants, the plant structure is represented in silico and its components are characterized by assigned geometric shapes with attributes, such as dimensions and spatial localization (Strauss et al. 2020). A virtual plant can be obtained by 3D digitizing (Sinoquet et al. 1997), terrestrial light detection and ranging (LiDAR) (Colaço et al. 2018), or by morphological and developmental rule-based simulations (Prusinkiewicz 1998). Three-dimensional digitizing remains time-consuming and is often limited to small plants and lower numbers of trees. Terrestrial LiDAR has emerged as a new phenotyping method to evaluate individual crown architecture (Coupel-Ledru et al. 2019) or LAD (Sanz et al. 2013), but it has a larger number of occlusions (Mack et al. 2017) and still poses some challenges. When terrestrial LiDAR was used on apple trees, the 3D reconstructions underestimated short shoots, which were the most numerous components of the canopy at the adult stage (Pallas et al. 2018). Rule-based simulation is also limited in generating complex plant structures, because it can be time-consuming to generate numerous trees in parallel at an orchard scale. Alternatively, partial 3D digitizing of a canopy, allometric relationships, and random distribution of certain organs attributes from fully digitized parts have been successively combined to reconstruct the 3D canopy structure (Sonohat et al. 2006). Thus, light interception efficiency can be precisely estimated for any organ or at any spatial scale based on the reconstructed canopy and is termed the silhouette to total area ratio (STAR) (Carter and Smith 1985). The STAR is considered the representation of light interception efficiency and can be used as a criterion to design an ideal canopy structure in fruit trees (Picheny et al. 2017; Tang et al. 2019). Furthermore, quantitative parameters have been proposed to evaluate the canopy structural properties, such as canopy volume (V), total leaf area (TLA), LAD and its variance (LADvar), as well as light interception properties, such as the projected leaf area (PLA), projected envelope area (PEA), mean optical density (MOD), porosity (Po), and leaf dispersion (μ) (Sinoquet et al. 2005; Stephan et al. 2008; Duursma et al. 2012; Yang et al. 2016). Light interception properties are usually estimated from virtual trees under isolated conditions without surrounding trees (hereafter called isolated trees) during light modeling. However, the mutual shading effect, which strongly relates to planting density, needs to be considered.

According to the equation of Monteith (1977), in addition to light interception efficiency, yield also depends on the energy conversion efficiency, which is the ratio of biomass produced over a given period through the photosynthetic process to the intercepted radiative energy of the canopy over the same period (Zhu et al. 2010). However, leaf photosynthetic capacity displays strong heterogeneity within the canopy. Therefore, whole canopy photosynthesis cannot be estimated from “an average leaf response” as it is the integrated response of leaves in different canopy positions with different physiological potentials tuned to their specific light environment (Hikosaka et al. 2016). Tackling such issue is an advantage of 3D virtual plants coupled with an eco-physiological model such as RATP that estimates the canopy net photosynthetic rate (Ac) and transpiration rate (Ec) (Sinoquet et al. 2001). RATP combines: (1) an intermediate 3D representation of a tree canopy with (2) a simulation of light interception integrated in all sky directions and (3) selected biophysical functions (energy balance and microclimate, such as temperature, wind speed, and vapor pressure deficit) with (4) stomatal conductance and photosynthesis sub-models (Farquhar et al. 1980; Harley et al. 1992). The model has been validated and applied to fruit trees to study the impact of tree structures on Ac, Ec and temperature, particularly in apple (Massonnet et al. 2008; Ngao et al. 2017; Woods et al. 2018).

To further understand bi-axis training system properties and impacts on the canopy structure and light interception, we trained ‘Fuji’ apple trees grafted onto the widely used dwarf M9-T337 rootstock in either a single- or bi-axis training system. For comparing those trees, 3D virtual plants based on digitizing were reconstructed, and canopy structure and light interception and distribution were estimated at both tree and intra-tree scale. The canopy daily Ac, Ec, and water use efficiency (WUEc) were estimated by the RATP model and compared between the two training systems.

Materials and methods

Plant materials

The experiment was carried out in a commercial apple orchard located 930 m above sea level, in Qianyang (34.62° N, 107.22° E), China. The average annual temperature, precipitation, frost-free days, and sunshine duration in the region are 11.8 °C, 532.5 mm, 197 days and 1983 h, respectively. Fully feathered nursery ‘Fujiko’ trees (2 years in the nursery) were used and planted in the spring season of the year 2015 and the experiment was conducted during the 4th and 5th year of tree growth. ‘Fujiko’ is a new apple variety that is characterized by an intense red color on the skin surface, selected from a block of the adjacent Fuji ‘NAGAFU 12’ trees (Leis and Mazzola 2008). During winter of the year 2012, scions were grafted with three buds onto the newly harvested, 1-year old M9-T337 rootstock at a height of 35 cm and planted in the orchard in spring of 2013. After bud-break in the spring of 2013, scions were de-budded to leave either single or two primary axes per tree. Trees with one primary axis (single-axis) were trained according to the tall spindle training system (Robinson et al. 2008, 2006; van Hooijdonk et al. 2015). Trees with two primary axes (bi-axis) were trained according to the method based on Bibaum® (Musacchi 2008a, b). All the trees were planted in rows spaced 3.5 m apart, with 1 m within the row for single-axis and 1.2 m for bi-axis trees, in a north–south orientation. The horizontal distance between the two primary axes in bi-axis trees was 60 cm. Horticultural practices included drip irrigation for integral control of watering, fertilization, and phytosanitary treatments. There was no pruning during the experiment and the average fruit load was 3.96 ± 0.62 fruits per trunk cross-sectional area (cm2) for the two training systems during the 2 years. The soil in orchard was a sandy loam. The contents of organic matter, total nitrogen, total phosphorus, and available potassium were 1.51 g kg−1, 1.0 g kg−1, 3.64 × 103 mg kg−1 and 159 mg kg−1, respectively.

Tree digitizing and foliage reconstruction

Before harvest (mid-October), three trees per training system for 4-year-old trees in 2016 and for 5-year-old trees in 2017 were chosen for tree digitizing. The spatial coordinates of the distal and proximal points of all of the current-year leafy shoots were measured with 3SPACE FASTRAK using a Stylus receiver (Polhemus; Cochester, VT, USA) and recorded with PiafDigit software (Donès et al. 2006). This allowed the computation of shoot length and orientation. All the shoots of the trees were classified as floral shoots or vegetative shoots. A floral shoot is composed of a bourse (B) and bourse shoot(s). Vegetative and bourse shoots were divided into short and long shoot, marked as VS and VL, and BS and BL, respectively (Costes et al. 2006). A threshold of 5 cm was used to separate the long from the short shoots.

Then, 20–27 randomly selected shoots per shoot type were digitized both at leaf- and shoot-scale. After digitizing, shoot length, individual leaf area, leaf width, leaf length, petiole length, and the number of leaves were recorded for the sampled shoots. This allowed us to estimate the allometric relationships among the shoot length, shoot leaf area, and the number of leaves using the methods of Palmer (1987). We also estimated the allometric relationships among the leaf length, leaf width, leaf area, and petiole length. The leaf Euler angle (midrib azimuth and inclination, lamina rolling around the midrib) distribution and the angle between the petiole and shoot axis were measured according to Sonohat et al. (2006). All the leaf- and shoot-scale measurements were conducted on all shoot types, for both training systems. Since the collected leaf- and shoot-scale data during the year 2016 were lost due to hard-disk damage, only data collected in the year 2017 were available.

The 3D distribution of leaves associated with each shoot type within the canopy was reconstructed based on Sonohat et al. (2006) using the same leaf- and shoot-scale parameters estimated in 2017 (Table S1). The reconstructed 3D virtual trees had a database of leaves, including leaf spatial coordinates, leaf dimensions, orientation angles, shoot type, and associated colors (Fig. 1). In total, a database of 3 (replicates) × 2 (training systems) × 2 (years) trees was created.

Fig. 1
figure 1

Three-dimensional representation of reconstructed isolated ‘Fuji’ trees grafted on M9-T337 rootstock trained either with a bi- or single-axis training systems in 2016 and 2017, respectively. Canopy volume were divided by a set of voxels. Each voxel is a 0.2 m cube. False colors were assigned to shoot type: hexagons were leaves and spheres are fruits, red bourses; blue bourse shoots; green vegetative shoot. In virtual orchard, the central color tree was the target tree which silhouette to total area ratio (STAR) was computed in virtual orchard, the surrounding trees were given in a black color. Tree size can be estimated according to the number of voxels. Virtual images were synthesized with VegeSTAR4.0 software

Full size image

Extracting canopy structure parameters of isolated trees

For each tree, canopy structure parameters were extracted from the reconstructed leaf database. Total leaf area (TLA) was estimated as the sum of individual leaf areas for the whole canopy and each shoot type. When fruit(s) settled on the bourse, the bourse and associated BS or BL were considered fruiting shoots (FS). Otherwise, the bourse and associated BS and BL without fruits were identified as VS and VL based on length. Consequently, TLA for FS, VS and VL was computed.

The canopy volume (V) was estimated using a bounding box method by dividing the canopy into cubic voxels with 20-cm-long edges based on the 3D spatial coordinates of leaves. We counted and added the number of voxels containing at least one leaf to compute V. Canopy LAD was computed as the ratio of canopy TLA to V. The relative variance of leaf area density (LADvar) was reported as a parameter indicating leaf clumping and was computed from the LAD of each voxel (LADv):

$${\text{LAD}}_{{\text{var}}} = \frac{1}{{n_{{\text{v}}} }}\sum\limits_{i = 1}^{{n_{{\text{v}}} }} {\frac{{\left( {{\text{LAD}}_{{\text{v}}} - {\text{LAD}}} \right)^{2} }}{{{\text{LAD}}}}} ,$$
(1)

where nv is the number of voxels.

Then, 2D maps of LAD were drawn, with LAD as a function of tree height and distance to the tree trunk based on the divided voxels, according to Stephan et al. (2008). Maps were drawn for the whole tree and each shoot type.

Estimating light interception of isolated trees

The light interception of isolated trees was computed from isolated 3D virtual trees for whole trees and each shoot type based on the projected leaf area (PLAi) and TLA (Sinoquet et al. 2007) using VegeSTAR software (Adam et al. 2002). VegeSTAR computes the PLAi by processing the virtual images of 3D virtual trees. In a given light direction, a photograph was taken with an orthographic degree, and PLAi values were computed by counting vegetation pixels from the image, indicating the leaf area illuminated by light in a given sun direction. STARi represents the average leaf irradiance related to incident radiation and equals the ratio of PLAi to TLA. Both PLAi and STARi depend on the direction of incident light (Ω). Light interception was computed from 46 light directions over the sky vault, according to the Turtle sky model proposed by den Dulk (1989). Finally, the sky-integrated PLAi values were computed by averaging PLAiΩ values using the weighted coefficients (WΩ) derived from the standard overcast distribution of sky radiance (Moon et al. 1942) as follows:

$${\text{PLAi}} = \sum\limits_{\Omega = 1}^{n} {\left[ {{\text{PLAi}}_{\Omega } W_{\Omega } } \right]} .$$
(2)

The projected envelope area (PEA) was computed using VegeSTAR by replacing the leaves within the canopy with associated voxels. Then, the PEA was used to calculate Po and MOD as follows:

$${\text{Po}} = 1 - {\text{PLAi}}/{\text{PEA}},$$
(3)

and

$${\text{MOD}} = {\text{TLA}}/{\text{PEA}}.$$
(4)

The MOD is similar to the LAI of an isolated tree. Leaves within the canopy were randomly distributed in the canopy volume defined by the voxels to produce a homogeneous canopy. Thus, the PLA of a homogeneous canopy (PLAiH) and the porosity of a homogeneous canopy (PoH) were computed as follows:

$${\text{PO}}_{{\text{H}}} = 1 - {\text{PLAi}}_{{\text{H}}} /{\text{PEA}},$$
(5)

Then, the leaf dispersion (μ) was computed according to Nilson (1971):

$$\mu = \ln {\text{Po}}/\ln {\text{PO}}_{{\text{H}}} ,$$
(6)

where μ represents foliage clumping, and a μ value less than 1 means a clumped canopy. Similarly, 2D maps of light interception, as a function of tree height and distance to the trunk, were drawn by combining the method used for 2D maps of LAD for the whole tree and each type of shoot (VS, VL, and FS).

Estimation of light interception in virtual orchard

To consider and quantify the effect of neighboring trees on light interception in an orchard, in the virtual orchard, trees were ‘planted’ with the same planting density (1.2 × 3.5 m) and actual row orientation than in the field. First, a series of virtual orchards with different tree numbers were created to choose a suitable orchard. For a specific treatment, an isolated tree located at the coordinates of (x = 0, y = 0) was regarded as the central tree and a ‘1-tree orchard’. A positive x value indicated an easterly direction and a positive y value indicated a northerly direction. Then, the size of the virtual orchard was increased by adding randomly selected reconstructed trees from each associated category in a clockwise direction (Fig. S1). For example, the second tree was located at (3.5, 0 m) and the third at (3.5, − 1.2 m). Second, the STAR of each central tree was computed in a series of virtual orchards. Third, the most suitable virtual orchard size was determined according to the trend of STAR versus the size of the virtual orchard. STAR dramatically decreased from the “1-tree orchard” to the “8-tree orchard,” then decreased slowly and reached a steady-state STAR varied within 5% when the number of trees in the virtual orchard reached 25 (Fig. S2). Finally, a reconstructed tree was planted in the center of its corresponding virtual orchard. Just as PLA was estimated from an isolated tree, the PLAo and STARo values for the central tree in the virtual orchard were estimated using VegeSTAR software. Accordingly, light mutual shading (Lms) results from neighboring trees in the orchard were computed, as follows:

$${\text{Lms}} = ({\text{STARi}} - {\text{STARo}})/{\text{STARi}},$$
(7)

Additionally, for single-axis trees, STARo and Lms were estimated with a density of 1.0 × 3.5 m to allow us to analyze the effect of planting density on canopy light interception (Tables S2–3). All the acronyms for canopy structural and light interception traits are listed in Table 1.

Table 1 Summary of acronyms for canopy architectural and light interception traits
Full size table

Estimating photosynthesis, transpiration and water use efficiency of isolated canopy

The daily Ac and Ec of isolated canopy were estimated with RATP (Sinoquet et al. 2001), which is available in the OpenAlea platform (Pradal et al. 2008). Canopy WUEc was computed as the ratio of Ac to Ec.

The RATP inputs included meteorological datasets, leaf spatial distribution, and leaf physiological characteristics. In the present study, only the difference in the leaf spatial distribution resulting from different training systems was considered, the other inputs between training systems were identical. In the RATP model, leaf spatial distribution was estimated from reconstructed 3D mock-ups and the canopy was divided into 20 cm cubic voxels as described in the above canopy structural section. Then, radiation transfer through the canopy was computed by considering Beer’s law (Ross 1981) for light beam penetration into a turbid medium in each voxel. The computation of Ac and Ec was based on the Jarvis (1976) and Farquhar (1980) sub-models, respectively. Meteorological datafiles and leaf physiology-related parameters were obtained from Yang (2016), where all the parameters were measured and estimated from dwarf M26 interstock ‘Fuji’ trees grafted on a vigorous Malus micromalus rootstock. The ‘Fuji’ trees were planted in an apple orchard 30 km away from the apple orchard in this study.

Data analysis

The TLA, PLA, STAR, and V data were analyzed for the whole canopy and each shoot type, and LAD, ζ, PEA, MOD, and μ data were analyzed for the whole canopy. Therefore, all the variables were analyzed combining the 2 years with two-way ANOVA including cultivar and year factors and their interaction. Before two-way ANOVAs were conducted, variables were analyzed to ensure data were normally distributed and the variances were homogeneous based on the Shapiro.test and Bartlett.test, respectively. Significant differences among treatments were distinguished by different letters using Duncan’s multiple mean comparison test at p < 0.05. Differences between training systems in the STAR distributions of voxels and shoots and LAD distributions of voxels for whole trees and each shoot type were analyzed using a non-parametric Chi-square test. All data were analyzed using the R software, with agricolae and stats packages (R Development Core Team 2019).

Results

Tree-scale canopy structure

Between years, canopy V was 51% higher and TLA was 37% higher in the year 2017 compared to 2016 (Fig. 1, Table 2). The increase in canopy V between years occurred within the canopy above a height of 1 m. The LAD in the year 2017 was significantly lower than that in year 2016. Bi-axis training had a significantly lower LAD than single-axis training. LADvar was affected by year rather than training.

Table 2 Information of canopy architectural and light interception traits at tree-scale estimated from isolated 3D virtual plants for ‘Fuji’ trees trained either with a bi-axis or single-axis training system in 2016 and 2017, respectively
Full size table

All the trees showed a LAD gradient from the outer periphery to the inner canopy in both years (Fig. 2). There were fewer zones with LAD values greater than 3 in bi-axis trees than in single-axis trees. This difference was significant in the year 2016 (30% in bi-axis and 39% in single-axis trees to the total voxel number) but was not significant in the year 2017 (both 31% relative to the total voxel number).

Fig. 2
figure 2

Leaf area density (LAD, m2 m−3) spatial distribution as a function of tree height above ground surface and distance to tree trunk based on virtual isolated trees for ‘Fuji’ trained either with a bi- or single-axis training systems in 2016 and 2017, respectively. The dashed line represents the central primary axis for single-axis trees and the ‘virtual’ middle parallel line between two primary axes for bi-axis trees

Full size image

Intra-tree scale canopy structure

In bi-axis trees, differences in leaf area among the three shoot types were not significant, irrespective of the year (Table S3). In single-axis trees, VL had the largest TLA, followed by VS and then FS in the year 2016 and then by FS and then VS in 2017 (Table S3). The TLA values of FS and VL were significantly higher in 2017 than 2016 (Table 3). Bi-axis trees had significantly higher TLAs of FS and lower TLAs of VL than single-axis trees. TLAs of VLs of bi-axis training trees were approximately 60% those of single-axis trees, irrespective of year.

Table 3 Total leaf area (TLA, m2), projected leaf area (PLA, m2) and silhouette to total area ratio (STAR) of fruiting shoot (FS), vegetative short shoot (VS) and vegetative long shoot (VL) estimated from isolated 3D virtual plants for ‘Fuji’ trees trained either with a bi-axis or single-axis training system in 2016 and 2017, respectively
Full size table

In 2016, VL was higher and more widely distributed within canopy than VS and FS; the latter two were distributed beneath a height of 2.0 m and showed a similar spatial distribution (Fig. 3). In 2017, VS and FS had a higher and wider distribution within the canopy as compared to their distribution in the previous year, irrespective of the training systems. In the year 2016, the FS of bi-axis trees had significantly more zones with LAD values > 3 than the FS of single-axis trees (24% for bi-axis and 15% for single-axis trees relative to the total voxel number), but had significantly fewer zones with LAD > 2 for VS (30% for bi-axis and 50% for single-axis trees relative to the total voxel number) and VL (28% for bi-axis and 41% for single-axis trees relative to the total voxel number). However, in 2017, there were significantly fewer zones with LAD values > 2 for VL in bi-axis trees than single-axis trees (7% for bi-axis and 18% for single-axis trees relative to the total voxel number), and LAD distribution of the VS and FS was similar between the two training systems.

Fig. 3
figure 3

Leaf area density (LAD, m2 m−3) spatial distribution as a function of tree height above ground surface and distance to tree trunk for fruiting shoot (FS), vegetative long shoot (VL) and vegetative short shoot (VS) based on virtual isolated trees for ‘Fuji’ trained either with a bi- or single-axis training systems in 2016 and 2017, respectively

Full size image

Tree-scale light interception of isolated trees

When averaged over years, year 2017 had significantly greater PLA and PEA values, than year 2016 (Table 2). Bi-axis trees had significantly lower MOD and higher Po values than single-axis trees, irrespective of training system or year. STAR was also unaffected either by year or training system.

Tree-scale STAR spatial distributions of the canopies were drawn as a function of voxel distance to tree height and to the central primary axis for single-axis trees and to the ‘virtual’ middle parallel line between two primary axes in bi-axis trees (Fig. 4). The STAR declined gradually from the top to the base and from the periphery to the inner canopy in both 2016 and 2017. Irrespective of year, there were few zones with STAR < 0.15 (0.12–0.87% relative to the total voxel number). In both years, bi-axis trees had more zones within the canopy with STAR > 0.3 compared with single-axis trees (92% and 90% for bi-axis trees in 2016 and 2017, respectively, and 83% in both years for single-axis trees relative to the total voxel number). Consistent with the STAR spatial distribution, bi-axis trees had significantly fewer shaded shoots and more shoots with STAR values > 0.30, regardless of year (67% and 62% relative to the total shoot number for bi-axis trees in 2016 and 2017, respectively, with corresponding values of 54% and 44%, respectively, for single-axis trees) (Fig. 5). Between years, the proportion of shoots with STAR < 0.15 in bi-axis trees from the year 2016 to 2017 decreased from 5 to 4%. However, the proportion of shoots with the STAR < 0.15 in single-axis trees from the year 2016 to 2017 increased from 8 to 16%.

Fig. 4
figure 4

Silhouette to total area ratio (STAR) spatial distribution as a function of tree height above ground surface and distance to tree trunk based on virtual isolated trees for ‘Fuji’ trained either with a bi- or single-axis training system in 2016 and 2017, respectively

Full size image
Fig. 5
figure 5

Silhouette to total area ratio (STAR) distribution of all shoots based on virtual isolated trees for ‘Fuji’ trained either with a bi- or single-axis training systems in 2016 and 2017, respectively

Full size image

Intra-tree light interception of isolated trees

Among shoot types, the mean STAR of the VL (range 0.46–0.55; mean 0.51) was greater than the FS (range 0.34–0.44; mean 0.39) and VS (range 0.23–0.36; mean 0.30), and the mean STAR of FS was greater than VS (Table S3).

From the year 2016 to 2017, the PLA of the FS increased and the STAR of VL decreased significantly (Table 2). When averaged over the training systems, the bi-axis trained trees had significantly lower PLA of VL as compared to single-axis trees, but their STAR was greater than that of single-axis trees.

Among shoot types, VL had more even STAR spatial distribution than the FS and VS, with few voxels having STAR values < 0.30 (Fig. 6). The STAR distributions in bi-axis trained trees were significantly improved for all three shoot types compared to single-axis trees, with fewer voxels with STAR values < 0.3 and more voxels with STAR > 0.3, regardless of year. Consequently, significantly more shoots in bi-axis trees had higher light interception compared to single-axis trees, irrespective of shoot type (Fig. 7).

Fig. 6
figure 6

Silhouette to total area ratio (STAR) spatial distribution as a function of tree height above ground surface and distance to tree trunk for fruiting shoot (FS), vegetative long shoot (VL) and vegetative short shoot (VS) based on virtual isolated trees for ‘Fuji’ trained either with a bi- or single-axis training systems in 2016 and 2017, respectively

Full size image
Fig. 7
figure 7

Silhouette to total area ratio (STAR) distribution of fruiting shoots (FS), vegetative long shoots (VL) and vegetative short shoots (VS) based on virtual isolated trees for ‘Fuji’ trained either with a bi- or single-axis training systems during 2016 and 2017, respectively

Full size image

Light interception in virtual orchards

The STARo values in virtual orchards were lower than the STAR values of isolated trees, regardless of the whole canopy or each shoot type (Tables 2, 3 and 4). For the whole canopy and all shoot types (except STARo of VS), the STARo values were significantly higher and the Lms values were significantly lower in 2016 than in 2017 (Table 4). The whole canopy and VS and VL had similar Lms values in the two training systems. However, the Lms of FS in bi-axis trees was 11% lower than that in single-axis trees. On average, the STARo for VS and VL in bi-axis trees were 25% and 10% greater, respectively, than that in single-axis trees.

Table 4 Silhouette to total area ratio (STARo) estimated from 3D plants in a virtual orchard and light mutual shading (Lms) estimated from difference of STAR between isolated trees and trees in virtual orchard for the whole canopy, fruiting shoot (FS), vegetative long shoot (VL) and vegetative short shoot (VS) in ‘Fuji’ trees trained either with a bi-axis or single-axis training system in 2016 and 2017, respectively
Full size table

Tree-scale photosynthesis, transpiration, and water use efficiency of isolated canopy

When averaged over year, no significant differences in Ac, Ec, and WUEc were found (Table 5). However, the mean values of Ac in bi-axis trained trees were 26% higher than those in single-axis trees, and a similar significant increasing trend was found in each year between the two training systems. The mean Ec for both training systems was similar, leading to a significantly higher WUEc in bi-axis than in single-axis trees.

Table 5 Canopy net photosynthetic rate (Ac), transpiration rate (Ec) and water use efficiency (WUEc) estimated from RATP model based on isolated 3D virtual plants in ‘Fuji’ trees trained either with a bi-axis or single-axis training system in 2016 and 2017, respectively
Full size table

Discussion

The present study illustrated the interest in using 3D virtual plants to study the impact of bi-axis training on canopy architecture, light interception, and photosynthetic rate (Marshall-Colon et al. 2017). The combined method of using partial 3D digitizing and random sampling enabled us to describe the detailed 3D structure of trees and extract canopy structure and light interception parameters and their spatial distributions both for whole trees and each type of shoot (Sonohat et al. 2006), considering the shading effect from the neighboring trees. The generation of 3D virtual plants from ruler grammar simulation is random and cannot fully represent the tree status in field (Da Silva et al. 2013, 2014). Laser scanning usually misses small sized shoots (Pallas et al. 2018), but with 3D digitizing, all shoots and types can be fully and precisely described.

Bi-axis training effects on canopy structure and light partitioning among shoots of isolated trees

Based on the detailed 3D virtual plant (Fig. 1), isolated tree STAR and its components (PLA, PEA, MOD, Po, and μ) (Nilson 1971; Smolander et al. 1994; Sinoquet et al. 2007) were quantified in the present study. A better illuminated canopy was characterized by higher STAR and accompanied by a higher Po and μ, and a lower MOD (Sinoquet et al. 2007; Duursma et al. 2012). The light interception components are functions of structural properties of canopy. Both the year and training systems affected the canopy properties, such as tree size, leaf area, shoot composition, and their spatial distribution.

Regarding the year effect, TLA, V, LAD, and LADvar were significantly different between the years (Table 2). Under most conditions, smaller trees (based on V) and lower LAD and LADvar benefited the light penetration into canopy. However, the light interception components (based on Po) were not improved by the decrease in the LAD and LADvar in the year 2017. This is consistent with a previous study where a large database of 1831 3D virtual plants was used and the results showed that trees with the same LAD could have totally different STAR values (Duursma et al. 2012). This may be because that trees have similar canopy structural parameters at the tree scale, but their spatial arrangement of leaves is different. Therefore, we suggest that changes in the PLA, PEA, and Po could be due to the increase in canopy TLA and V from year 2016 to 2017 and the optimized spatial distribution of leaves. Sinoquet et al. (2007) also showed that increased tree size and leaf area decreased the light interception efficiency. For each training system, the same leaf and shoot allometric relationships were used between years. As genotype, rootstock, and horticultural practices were identical between the two training systems, the climate, tree ontogeny, and increased shading within canopy in consecutive years may affect the shoot- and leaf-scale allometric relationships by affecting leaf size and growth duration and number of leaves per shoot (Chen et al. 2019). Therefore, the effect of year on tree architecture could be underestimated. If responses of shoot and leaf development to climate, tree ontogeny and shading between two training systems were not identical, the training effect could be underestimated, and this needs further study.

Accordingly, the significant improvement in the canopy MOD and Po by the application of the bi-axis training system to the apple trees might be the result of both the reduction in LAD and its spatial distribution (Table 2 and Figs. 2 and 3). Particularly, when compared to the spatial distributions of the STAR and the STAR of individual shoots between the training systems, light was more evenly distributed and more shoots had higher light interception efficiency with bi-axis training (Figs. 4, 5, 6 and 7). However, the averaged STAR value of isolated trees was not affected by the training systems. Therefore, the bi-axis training effect on canopy light interception performance evaluated by the STAR of isolated trees might be underestimated because of lack of knowledge of the heterogeneity of light within the canopy. The spatial distribution of LAD for a whole tree was decomposed into shoot types (Fig. 3). In both years, the lower proportions of VL and more uniform LAD spatial distribution of VL in bi-axis than in single-axis trained trees contributed to the improvement of light interception.

As the VLs are extension shoots, they are net carbon sinks in the early season and only a few of the fixed assimilates are exported to other shoots and fruits in the mid- and late season (Johnson and Lakso 1986a, b; Lauri and Kelner 2001). A lower proportion of VL (Fig. S3) and lower leaf area of VL in bi-axis training suggested that the fruit growth in bi-axis trees had less carbon competition from vegetative growth of VL compared to single-axis trees, which have more VL. This suggested lower tree vigor in bi-axis trees, which was also validated by a smaller annual trunk cross-sectional area compared to single-axis trees (Musacchi 2008b; Ma et al. 2020). The reduction of TLA in bi-axis trees (Table 2) may result from the higher frequency of bourse shoots (Fig. S3), which have lower shoot leaf number and shoot leaf area compared to single-axis trees (Table S1). ‘Fuji’ trees usually have strong vegetative growth due to high precipitation during late summer and autumn on the Chinese Loess Plateau, which leads to higher biennial bearing and a higher labor need for pruning (Han et al. 2008; Xing et al. 2016). Therefore, bi-axis training could help reduce the biennial bearing and required labor for ‘Fuji’ trees by reducing the tree vigor and improving the light environment within the canopy.

Among shoot types, the VL had the highest STAR, followed by FS and VS, regardless of the year, training system and the effect of neighboring trees (Tables 3 and 4 and S3 and Fig. 6). The light interception of each shoot type was in accordance with its spatial distribution and STAR values of individual shoots. The light interception partitioning among shoot types and their spatial distribution were consistent with the reported patterns in apple trees (Stephan et al. 2008; Yang et al. 2016). The STAR values of FS and VS were lower than the STAR of the whole tree, irrespective of the year and training, indicating that FS and VS were in shaded areas. The difference in the light interception among shoot types was closely related to their spatial distribution of LAD. The distribution of VL was the outermost, that of VS was the innermost, with FS intermediate between. Moreover, VLs have a longer internode length in favor of light interception (Takenaka 1994; Niinemets et al. 2005). Although the STAR of FS at the tree scale was not improved by bi-axis training as compared to single-axis training, the STAR spatial distribution of FS was significantly improved, including VL and VS. These led to a cumulative effect on the trees subjected to bi-axis training, allowing more light penetration into the canopy and reducing the inner shading compared to single-axis tall spindle trees, particularly for the FS and VS.

Mutual shading effect on light partitioning among shoots in a virtual orchard

Apart from the year and training effect, the planting density in the orchard plays an important role in the determination of the light interception efficiency (Trentacoste et al. 2015). Depending on the density, a non-negligible portion of branches may penetrate the neighboring trees in the same row. Consequently, the STAR values of trees in the virtual orchard were lower than those of the isolated trees in current study, and the difference between years or training systems were exaggerated, irrespective of tree- or intra-tree scales. The virtual orchard setting made the in silico experiment more realistic as the trees were planted as they were in a real orchard. Our virtual experiment highlighted the important shading effect due to neighboring trees which could reduce the STAR of trees in virtual orchard with a range of 12–34% compared to the STAR of isolated trees. This suggested that potential improvement of shaded leaves could be an option to enhance the canopy light use efficiency (Song et al. 2013). When the planting density of single-axis trees increased from 1.2 × 3.5 to 1.0 × 3.5 m, the differences in light mutual shading between the two training systems were significant, irrespective of the tree- or intra-tree scales (Table S2). This revealed that increased planting density will lead to more shading within the canopy, and a virtual orchard could be a way to optimize planting density. However, the single-axis trees were digitized in a planting density of 1.0 × 3.5 m. When these trees were planted in lower planting density (1.2 × 3.5 m), their LAD and crown development may be different from trees in higher planting densities due to different shading by neighboring trees. Except for FS, bi-axis trained trees had similar mutual shading to single-axis trees for both tree and vegetative shoots under the same planting density (Table 4). However, the vegetative shoots of bi-axis trees had significantly higher STARs. Thus, based on the STAR of the isolated trees or trees in the virtual orchard, the light interception efficiency of bi-axis trees improved by decreasing LAD and optimizing the spatial distribution of LAD at a tree scale rather than by limiting the mutual shading at the orchard scale. The VL had the lowest mutual shading, and VS and FS had similar mutual shading (Tables 4 and S2–3). This difference may also be related to their spatial distribution as discussed earlier.

Bi-axis training effect on photosynthetic rate of isolated canopy

A light intensity of at least 25–30% within canopy is required for continual flowering and high-quality apple fruits (Jackson 1980). The positive effects of bi-axis training on the fruit soluble solids, yield efficiency (Musacchi 2008b), dry matter production, flowering (van Hooijdonk et al. 2015; Ma et al. 2020), leaf gas exchange (Afonso et al. 2017), and net canopy photosynthetic rate could therefore be interpreted as a result of the significant improvement of the light interception and distribution at both the whole tree and intra-tree scale. A STAR of 0.15 is suggested to be equal to a light intensity of 30% (Yang et al. 2016). In the two training systems, only a few zones with the STAR less than 0.15 existed (Figs. 4 and 6). This could be because all the trees in the present study were at a younger stage than adult trees used in other studies (Stephan et al. 2008; Yang et al. 2017). Bi-axis training isolated trees had a 5% higher STAR compared to single-axis training. However, modeled photosynthetic rate of isolated canopy in bi-axis trees was 26% higher than single-axis trees. Therefore, the increase in the light interception cannot only interpret the increase in the photosynthetic rate. In the present study, the differences in microclimate and photosynthetic-related processes between two training systems were not considered during photosynthesis modeling. The improvements in the microclimate within the canopy and the photosynthetic-related processes induced by enhanced light interception may contribute to the increase in canopy net photosynthetic rate, such as leaf temperature, leaf nitrogen content, and maximal stomatal conductance (Woods et al. 2018; Kumarathunge et al. 2019). Early-season leaf area development in 1-year-old trees was identified as the main reason for a higher dry matter production in bi-axis trees compared to single-axis trees (van Hooijdonk et al. 2015). However, the present study suggested that the optimized structure in bi-axis trees could be the main reason for better performance, whereas different reasons may result from the different tree ages and environmental conditions (Foster et al. 2016).

Although the canopy volume of bi-axis trees was not significantly different compared to that of single-axis trees (Table 2), the canopy of bi-axis trees occupied more space above 2 m height than single-axis trees (Figs. 1 and 2). This suggested that the thinner canopy of bi-axis trees was potentially beneficial for light penetration into the canopy (Jackson 1970) and mechanical harvesting. The bi-axis canopy design will arouse more thin 2D planar tree canopy with more scaffolds, similar to Cordon training in grapevines.

Conclusions

In general, by taking advantages of 3D virtual plant modeling, the present study showed that the bi-axis training system improved the light interception efficiency, light spatial distribution, and canopy net photosynthetic rate and reduced the tree vegetative growth compared to the single-axis training system. The two training systems had similar light interception partitioning to fruiting shoots, however, light interception partitioning to vegetative shoots increased and light mutual shading of FS decreased in bi-axis training compared to single-axis training. The light interception increased in bi-axis trees through improvement in the canopy structure by decreasing the LAD and optimizing the spatial distribution of LAD at the tree scale, particularly for vegetative long shoots. Light mutual shadings between two training systems were similar. Further studies will be conducted to evaluate the effect of bi-axis training on the canopy structure and light interception in adult apple trees, and on the variability of leaf photosynthesis and related physiological processes within the canopy.