Effect of site, silviculture and tree social status on internal checking variation in plantation-grown Eucalyptus nitens

ABSTRACT

Internal checking is a wood-drying defect that affects many tree species, including eucalypts, and can adversely affect the recovery of high-quality timber. We studied the patterns of within- and between-tree variation in wood checking from three Eucalyptus nitens silvicultural trials and determined the predictability of wood checking from wood properties. We hypothesised that changed wind exposure and inter-tree competition due to thinning would increase wood checking. In total, 144 trees from two silvicultural treatments (thinned/unthinned) and three social status classes per site (dominant/subdominant/suppressed) were sampled. Regardless of site, social status and silvicultural treatment, longitudinal and radial variation within trees was the major source of variation in checking. Checking was concentrated longitudinally in bottom logs and radially in the middle of wedges, with a shift in checking towards the cambium in the thinned treatment. On average, dominant trees in the thinned treatment had higher levels of checking than dominant trees in unthinned treatments at all sites, but the opposite trend was found for suppressed trees. Wedge checking was poorly predicted using non-destructive traits assessed on standing trees. Our models accounted for at most 22% of the variation in wedge checking and the best predictors were the destructive measures of wedge collapse. With most checking occurring in lower logs, future studies should concentrate on reducing checking in bottom logs, which are typically pruned for clearwood products.

KEYWORDS:

• wood defects
• thinning
• near-infrared spectroscopy
• drilling resistance
• random forest modelling
• image analysis

Introduction

Internal checking is a drying defect of sawn boards that involves the separation of fibres in the interior of boards, resulting in cracks/splits (Blakemore 2011). It occurs in softwood (Putoczki et al. 2007) and hardwood (Chafe et al. 1992) species and is one of many factors that need to be considered in the use of logs from plantation-grown eucalypts for producing high-value, appearance-grade boards (Raymond 2000; McKenzie et al. 2003; Blackburn et al. 2010). Internal checks arise due to internal stresses that manifest in the early stages of board drying (Blakemore 2011). They often only become visible following the post-drying processing of boards and thus have a significant economic impact because deficient boards need to be reprocessed to achieve acceptable quality or are considered waste (Ilic 1999). According to McKenzie et al. (2003), the economic impact of checking may be due to product rejection or the need for remanufacturing to achieve acceptable quality, particularly for appearance-grade products. It may also affect product recovery if checking is more severe in bottom logs and close to the periphery, where most of the investment is made in pruning and thinning to concentrate wood of appearance-grade quality (Washusen et al. 2009; Forrester and Baker 2012).

Internal checks usually occur within individual growth rings (i.e. intra-ring), often at the earlywood to latewood transition zone (Blakemore 2011). Checking is a phenomenon closely related to wood collapse (Yang et al. 2019), which is an abnormal form of shrinkage usually appearing as a caved or corrugated board (also called ‘washboarding’) (Chafe et al. 1992). Wood collapse arises from the flattening of single cells or rows of cells during drying when the thin-walled earlywood cells are unable to withstand the hydrostatic tensions arising from the rapid removal of free water (Pang et al. 1999). Many studies on drying technologies have assisted in optimising the drying process to minimise checking and collapse (Innes 1996b; Harris et al. 2008; Ananías et al. 2020). However, some wood is inherently difficult to dry without degradation, leading to economic losses. Therefore, the reduction of drying defects such as checks and collapse is an important objective for the development of solid-wood products from eucalypts (Ilic 1999; Innes 2007), particularly for species such as Eucalyptus nitens (H.Deane & Maiden) Maiden when grown in plantations (Raymond 2000; McKenzie et al. 2003; Blackburn et al. 2010).

The incidence of internal checking in E. nitens is influenced by tree characteristics (e.g. genetics, Blackburn et al. 2010), as well as sawing methods and drying schedules (Washusen et al. 2009). At the tree level, checking may be indirectly related to physical, chemical and anatomical features of the wood. For example, Gacitúa et al. (2007) demonstrated that, in E. nitens, checked wood had higher modulus of elasticity and larger-diameter vessels than normal wood. Wood with a higher modulus of elasticity has been linked to an increased proportion of tension wood (Clair and Thibaut 2014). Wood developed under higher tensile stress could explain, in part, the observed differences in checking patterns because several relevant drying defects have been linked to this type of wood in E. globulus Labill (Washusen and Ilic 2001; Downes et al. 2014). Tension wood commonly develops at the bottom and periphery of straight, vertical and dominant E. globulus trees (Washusen et al. 2003) and increases with thinning intensity (Washusen et al. 2005). When a tree stem is impacted by mechanical stimuli (e.g. wind), tension wood is formed on the upper side of the leaning stem, which, at the cellular level, results in a cellulose-rich additional layer (called the ‘G layer’), with decreased levels of lignin and hemicellulose in the inner section of the secondary cell wall (Bowling and Vaughn 2008). In trees such as eucalypts, chemical variation during wood maturation also occurs with new annual rings and heartwood development. Heartwood is the physiologically inactive central part of a tree stem where cells accumulate extractives. It is developed as a continuum from the central part of the stem towards the periphery as trees age (Hillis 1987). Eucalypt heartwood could develop stronger water tension forces during drying because of decreased permeability due to the occlusion of pits by extractives (Kininmonth 1972), which may produce more collapse (Chafe et al. 1992) and potentially induce more checking in processed wood. To better characterise the available wood resources in solid-wood plantations, it is important to understand both the between- and within-tree variation of internal checking and the most influential external factors. This knowledge could assist plantation managers and processors in identifying and segregating logs for alternative processing strategies.

Assessing the variation in checking within and between trees involves expensive destructive sampling, as well as time-consuming sample preparation and check assessment (visual categorisation or manual measurement), but image-based assessments have shown promising results (Phonetip et al. 2017). The advantages of using digital image analysis to quantify internal checking include better accuracy, the more efficient use of labour, and the storage of digital information for future analysis, along with its relatively low cost and fast set-up (Ke et al. 2016). In addition, advanced methods based on machine learning have been used successfully for wood-defect detection in boards (Kamal et al. 2017; Rahiddin et al. 2020). However, we found no previous studies using a quantitative assessment and machine-learning approach to investigate variation in internal checking in eucalypts. The automated quantitative procedure developed by Rocha-Sepulveda et al. (2021) is combined with qualitative assessments and a machine-learning approach to identify factors affecting the within- and between-tree variation in the internal checking of plantation-grown E. nitens. Using wedges from discs extracted from felled E. nitens trees, this study aimed to determine the effect of site, thinning and tree social status on variation in checking at the within-tree (longitudinal and radial) level and the extent to which variation in internal checking can be predicted from destructive and non-destructive assessment of other wood properties. Here, we hypothesise that the effects of wind exposure on dominant trees in the thinned treatment may be accentuated due to their slender stems arising from initial inter-tree competition before thinning (Medhurst et al. 2012) because slender stems would require the development of more tension wood (more prone to checking) to achieve mechanical stability (Dassot et al. 2012). Additionally, we hypothesise that inter-tree competition in unthinned stands would produce more checking, which may be related to higher vessel frequency (Leandro-Zúñiga et al. 2008; Valenzuela et al. 2012) as a result of increased water stress (Pfautsch et al. 2016) with increased stand density.

Methodology

Trial description and sampling strategy

Three E. nitens thinning trials were examined – Gads, Urana and Florentine (Figure 1) – in Tasmania, Australia. They were established by Sustainable Timber Tasmania (formerly Forestry Tasmania) between 1997 and 1999 in locations representative of that institution’s E. nitens plantation estate. The trials were planted at 3.5 m × 2 m initial spacing and each trial comprised an unthinned control treatment, where the initial stocking was 1100 individuals ha⁻¹ (Unthinned), and a commercial thinning treatment (Thinned), where trees were thinned to 300 trees ha⁻¹ from ‘below’ (smaller trees were prioritised for removal), at age 8–9 years (Table 1) (Gendvilas et al. 2021, 2022). Commercial thinning, which is applied on better-quality sites (i.e. with a peak mean annual increment higher than 20 m³ ha⁻¹ y⁻¹), aims to obtain trees at mid-rotation thinning that are sufficiently large to be sold as pulplogs (Wood et al. 2009). The trees in these trials were pruned to 6.4 m in their earlier life (between ages 3 and 5 years) to produce high-quality knot-free sawlogs.

Figure 1. Location of the three silvicultural trials of Eucalyptus nitens (Gads, Florentine and Urana) in Tasmania, Australia. The map of Tasmania is based on the ESRI Physical map of the Quick Map Services in QGis, which uses dark greens for dense forests, lighter greens for grasslands or agricultural areas, yellows and browns for arid land and deserts, and blues for water bodies. The island state of Tasmania is shown in a black rectangle in the bottom right insert of Australia

Table 1. General information on the three silvicultural trials of Eucalyptus nitens in Tasmania, Australia. Location, elevation, mean annual precipitation and temperature and plantation age at sampling and thinning. Overall grand means for outerwood basic density (standard deviations in parentheses) and diameter at breast height (DBH; standard deviations in parentheses) at sampling age for each site are also shown

	Urana	Florentine	Gads
Location – Latitude	41°20’57.0“S 148°02’54.8“E	42°36’22.0“S 146°28’09.5“E	41°34’22.9“S 146°12’13.8“E
– Longitude
Elevation (m)	400	400	700
Mean annual precipitation (mm y^−1)	1101	1332	1556
Mean annual temperature (°C)	11.4	10.3	9.4
Plantation age (y)	22	20	20
Age at thinning (y)	8	8	9
Outerwood basic density (kg m^−3)	503 (6.1)^c	479 (6.2)^b	443 (6.1)^a
DBH (cm)	32.6 (0.9)^ab	34.8 (1.0)^b	30.0 (0.9)^a

Temperature and precipitation data were calculated over the growing period using the getAWAP function of the AUSClim package (P. A. Harrison, unpubl. R package). Plantation age is given at the time of core and disc collection in 2019. DBH was collected in 2018 when trees were drilled using an IML® PD400 tool. Different superscript letters against site means for density and DBH indicate sites that are significantly different at P ≤ 0.05 based on the Tukey’s test.

Each trial had four replicates with a plot of each treatment within each replicate. The plots were 20 m × 30 m in size, which at the initial stocking included approximately 60 individuals. Trees were 20–22 years old when felled for the present study. The trees felled within each plot were chosen using stratified random selection. Within each plot, three equal-size dominance classes, referred to as ‘tree social status’, were defined by diameter at breast height (DBH), according to Forrester and Baker (2012). These classes represented dominant, subdominant and suppressed trees, with two trees randomly selected within each class per plot. In total, 144 trees were selected (6 per treatment × 2 treatments × 4 replicates × 3 sites, Gendvilas et al. 2022).

Disc and core extraction

Prior to tree felling, a 5 mm diameter increment borer (Haglöf Sweden AB, Långsele, Sweden) was used to take four 5 cm length outerwood cores at breast height (1.3 m) from the north-oriented side of each tree for a non-destructive measure of outerwood basic density. Samples were extracted from north-oriented sections, which are expected to be less affected by tension wood because the prevailing wind direction in the trial locations is westerly. The trees were then felled, and one 5–7 cm thick wood disc was collected every 2.5 m from the bottom of the tree up to 20 m height (Figure 1). An additional 5–7 cm thick wood disc at 2.5 m was also collected from the same trees (N = 144). The additional sample at 2.5 m was used to study radial variation in checking in the middle of typical pruned 6.4 m sawlogs. In total, 1296 discs were collected and one wedge (of variable size and position within the disc) per disc from this set was extracted in the forest. The additional 144 discs collected at 2.5 m were kept intact, wrapped in polyethylene plastic to limit moisture loss and then stored in a cool room at 5°C until processing. The 144 discs were processed in the laboratory by extracting four wedges of approximately 45° each (Figure 2). For safety reasons, if the disc diameter was less than 15 cm (24% of the discs), 90° wedges were extracted. The most northerly wedge as well as the adjacent wedge were labelled and standardised to 5 cm thickness using a wood planer. Two sets of wedges were taken to study (1) longitudinal (1296 wedges) and (2) radial variation (144 wedges from 2.5 m) in checking.

Figure 2. Diagram of sample collection and processing. Disc wedges were collected at 2.5 m for radial assessment, outerwood cores and microdrill resistance (Resi) profiles from 144 Eucalyptus nitens trees in three silvicultural trials in Tasmania. The four 5 cm wood cores were taken at 1.3 m from the ground. To study longitudinal variation in checking, a single wedge was also taken in the field every 2.5 m, sampling from near ground level to 20 m up the stem. Wedge A (the most northerly) was normally selected for further processing

NIR = near-infrared

Wedge processing

The wedges collected for longitudinal assessment were air-dried for approximately 2 months and then oven-dried at 103°C to a constant weight (TAPPI Standards 2016; Gendvilas et al. 2022) and kept in tree groups throughout drying and storage. These wedges were then manually assessed using a subjective scoring system, where score 1 is less-frequent (or not visible) checks and score 6 represents the most frequent and highly visible checks on the surface of the sample (Blackburn et al. 2010). This subjective score can efficiently characterise checking area variation compared with the quantitative method (Rocha-Sepulveda et al. 2021). The wedges for radial assessment were oven-dried (starting when green) by batches using a temperature ramping schedule of 50°C, 70°C and then 103°C (Rebolledo et al. 2013) until constant mass was achieved at each set temperature. Weight stabilisation was assessed from subsamples of three wedges in each batch. Given the limited capacity of the ovens, the drying process was undertaken in six batches. Each batch (oven load) consisted of 2–4 wedges per tree from 24 (4 × 2 × 3) trees, with each tree randomly selected within each replicate (n $=$ 4), thinning treatment (n $=$ 2) and site (n $=$ 3) combination. The process was repeated three times using two identical ovens. Inside the oven, the 2–4 wedges per tree were randomly assigned to each rack (four racks per oven load). The most northerly wedge per disc was generally selected for processing (Figure 2, Wedge A). However, if this wedge was not composed of clearwood or had split prior to drying or because of the drying process, an adjacent suitable clear-wood wedge from the same disc was selected. In order to visualise internal checks more clearly, the internal face of oven-dried wedges was exposed (Lausberg et al. 1995). To do this, the selected ovendried wedges were sliced in half perpendicular to the longitudinal direction using a bandsaw (Woodfast® BS350C), and both internal cross-sectional surfaces were sanded using 180 grit sandpaper (Makita® BO3710X). The sanded internal surfaces were blown with compressed air to remove residual dust from the checks. Only one wedge surface per disc was used for analysis.

Assessment of wood drying traits

For the wedges from the 2.5 m discs used for the radial assessment, the external and internal wedge surfaces when green and oven-dried, respectively, were scanned using an optical desktop scanner (Canon Pixma MX536). The scanner had a resolution of 1200 dpi, resulting in an image size of 10 200 × 13 200 pixels. Image analysis was performed using an R-based script for checking, collapse and shrinkage assessment developed in Rocha-Sepulveda et al. (2021). For one internal face of the halved wedge, checks were allocated to rings by taking their distance from the pith and allocating them to a ring approximated using semicircular sectional areas defined by ring widths delimited along a central line from the pith to the cambium. The most outer incomplete ring was removed, and ring positions were aligned from cambium to pith. Thinning age corresponded to the ring positions 5, 7 and 8 at Urana, Florentine and Gads, respectively. Total check area was calculated at the ring level as a proportion of the ring area covered by checks. At the tree level, checking was quantified in terms of check area (Check_a) and check number (Check_n), both expressed as a proportion of the wedge area. Three measurements of collapse were considered: (1) total collapse (Collapse_T), based on the difference between the actual and reconstructed area of dried wedges; (2) depth of collapse (Collapse_D) as the average of left- and right-side maximum perpendicular length from the wedge edge to the reconstructed segment; and (3) curvature of collapse (Collapse_C) as the average of local curvature of corrugated borders along the radial edges. Cross-sectional (Shrinkage_CS), tangential (Shrinkage_T) and radial (Shrinkage_R) shrinkage were calculated from the areas calculated from the fresh and oven-dried digital images of the wedge (see Appendix A for details).

Assessment of other wood properties

Ring level

The radial profile in wood density from each 2.5 m disc used for the radial assessment was generated by resistance drilling to match the radial variation in internal checking. This was done by drilling the green discs prior to cutting along the north to south axis using an IML-RESI PowerDrill® PD 400 microdrill (referred to as ‘Resi’), which stored drill resistance values at 0.1 mm resolution (Downes et al. 2018). The Resi was set at 150 cm min⁻¹ speed of forward feed and 3500 rpm. Ring allocation on the microdrill resistance profiles was conducted using the annual ring allocation algorithm included in the Shiny web application ‘Eucalypt ResiProcessor’, written in R and developed by Forest Quality Pty Ltd for the analysis of wood drilling resistance (Downes et al. 2018), which is hosted in the website https://forestquality.shinyapps.io/FQ_ResiProcessor/. These ring allocations were then aligned with the matched wedge used for checking assessment using the annual ring delimitation determined using the high-resolution digital images of sanded wedges (see the section, ‘Assessment of wood drying traits’ above). The Resi-derived wood-density traits (see Table 2) were calculated using half the resistance profile (from cambium to pith), corrected according to Gendvilas et al. (2020) to account for the effects of the Resi needle friction. The radial profiles of wood basic density were determined from the Resi resistance values using an existing linear regression developed for the same trials (slope 6.64 and intercept 207.3 calibrated at 1.3 m, Gendvilas et al. 2021). The wood density traces from rings 4 to 9 – where most of the checking variation is thought to occur (see Results and Discussion) – were considered for assessing checking predictability. The mean (Resi_mean), maximum (Resi_max), minimum (Resi_min), standard deviation (Resi_sd), range (Resi_range; maximum minus minimum) and rate of change (Resi_rate; Savitzky-Golay first derivative calculated using the signal package of R, Ligges et al. 2015) of wood density were calculated at the ring level and then averaged across rings 4–9. These averages were used in the tree-level prediction models described below.

Table 2. Predictors used in Random Forest models (M1–M5) for predicting checking in wedges from discs taken from Eucalyptus nitens at 2.5 m above ground level (i.e. top of the first extracted log). Predictor names, units and description are shown

M1	M2	M3	M4	M5	Predictor	Units	Description
x	x				DBH	mm	DBH (1.3 m)
	x		x	x	Core_BD	kg m^−3	Outerwood core basic density
	x		x	x	KPY	(%)	NIR-predicted KPY
	x		x	x	Cellulose	(%)	NIR-predicted cellulose content
	x		x	x	Klason_lignin	(%)	NIR-predicted Klason Lignin
	x		x	x	Total_lignin	(%)	NIR-predicted total lignin content
	x		x	x	Extractives	(%)	NIR-predicted extractive content
		x	x	x	DIA_Disc	mm	Resi-derived log diameter at 2.5 m
		x	x	x	Resi_mean	kg m^−3	Average of 4–9 rings mean values
		x	x	x	Resi_max	kg m^−3	Average of 4–9 ring maximum values
		x	x	x	Resi_min	kg m^−3	Average of 4–9 ring minimum values
		x	x	x	Resi_sd	kg m^−3	Average of 4–9 ring standard deviation
		x	x	x	Resi_range	kg m^−3	Average of 4–9 ring range values
		x	x	x	Resi_rate	kg m^−3/0.1 mm	Average of 4–9 ring rate of change
				x	Collapse_T	(%)	Total collapse (reconstructed area)
				x	Collapse_C	arbitrary units	Curvature of collapse
				x	Collapse_D	unitless	Maximum depth of collapse
				x	Shrinkage_CS	(%)	Total cross-sectional shrinkage
				x	Shrinkage_R	(%)	Radial shrinkage
				x	Shrinkage_T	(%)	Tangential shrinkage

Core_BD = core wood basic density based on the water displacement method; DBH = in-field measurement of tree diameter at breast height (1.3 m) at age of assessment; KPY = kraft pulp yield; NIR = near-infrared. KPY, cellulose, Klason lignin, total lignin and extractives are NIR predictions from core sampling at 1.3 m. Resi measurements were done in the laboratory for each disc and ring delimitation was undertaken based on the collected Resi profiles. The Resi-based predictors were then calculated from a subset of the rings (4–9) where checking is more frequent. Collapse and shrinkage traits were assessed using digital images from the internal faces of halved wood wedges.

Tree level

To examine the tree-level predictors of internal checking using traits that could be obtained from non-destructive assessments, four breast-height (1.3 m) cores were extracted from trees prior to felling. Two of the four cores were milled together for near-infrared (NIR) prediction of kraft pulp yield (KPY), total lignin (Total_lignin), Klason lignin (Klason_lignin) and extractives content (Extractives). For each of these samples, NIR spectra were obtained between 1000 and 2500 nm wavelength (4000 and 10 000 wave number) on a Bruker MPA FT-NIR instrument (Downes et al. 2012). The NIR predictions were obtained from global models with validated R² ranging from 0.82 to 0.92, developed by Forest Quality Pty Ltd (Downes et al. 2011). The other two cores were oven-dried at 103°C and used to assess wood basic density with the water displacement method (TAPPI Standards 2016). The average outerwood basic density of the two cores collected at 1.3 m from each tree (Core_BD) was then used in the analysis. The NIR predictions of KPY (and associated traits) as well as the outerwood core basic density are regularly assessed in pulpwood breeding programs (Hamilton et al. 2008).

Statistical analysis

The machine-learning algorithm Random Forest (RF) (Breiman 2001) implemented in the extendedForest R package (Ellis et al. 2012) was used to investigate the factors affecting checking and its predictability from destructive and non-destructive assessment of wood properties. RF is a regression technique that combines the performance of multiple decision-tree algorithms (Breiman et al. 1984) to predict the value of a variable. The main advantages are that it does not require the assumption of normality, and it is flexible to model non-linear trends with minimal parameter specification. To explore the between- and within-tree variation in checking, models were fitted, including the main and explicit interaction terms, to analyse their importance. A permutation-based variable importance measure based on root mean square error (RMSE) loss was used to determine the most important variables/factors affecting checking, as implemented with the DALEX R package (Biecek 2018). First, the loss function (RMSE) was calculated using the original data (unshuffled), referred to as ‘reference accuracy’. Second, the loss in accuracy of predictions was calculated after randomly shuffling a single column (focal factor) of the data, leaving the response and all other columns in place. Given that the maximum number of levels of a categorical variable in the RF algorithm is 32, the Friedman’s H statistic (Frieman and Popescu 2008) was performed when the most important variable and higher interaction term produced more than the maximum number of categories/levels. The Friedman’s H statistic measures the interaction strength based on the variance decomposition of the predicted response and takes values between zero (no interaction) and one (100% of variance due to interaction). Friedman’s H statistic was calculated between the most important predictor variable and the other predictor variables in the RF model by averaging the results of 100 runs of the function Interaction using the default number of trees included in the IML R package (Molnar et al. 2018). Plots of the predicted effects of the predictor variables of interest were based on accumulated local effects (ALE), an approach to prediction that performs well when predictors are highly correlated (Apley and Zhu 2020). The ALE predictions are centred at zero, making interpretation simpler because the value at each point of the ALE curve is the difference from the mean prediction. For example, if ALE for a specific ring position Rp₈ is 5, it means the effect of Rp₈ on the response variable at that specific position increases by 5 compared with the average prediction. The model response was explored using the model_profile function of DALEX package, which allows the prediction of the response across different groups when variables are categorical (with many levels or groups). In the present case, the response variable was check score for the longitudinal-level study and check area proportion for the ring-level study. The various predictors used in the models were the main effect stem height (H), ring position (R_p), site (S), treatment (Tr), social status (So) and their interactions.

To explore the ability of destructive and non-destructive measurements in predicting checking at the tree level, five RF models (Table 2) were constructed using combinations of DBH, wood density, drying traits and NIR-predicted wood chemistry. The response variable was check area proportion (Check_a), and the goodness-of-fit of the models was evaluated using the out-of-bag (OOB) R². OOB is an n-fold cross-validation process in RF models in which approximately 30% of data is not used in building a given tree (Breiman 2001). The models were run at individual site and pooled levels and their OOB R² calculated. Variable importance was calculated for the full model (M5 in Table 2) at the pooled level.

Differences in outerwood basic density and DBH at the site level (Table 1) were tested using linear mixed models as follows: Y = Site + Treatment + Social + Site:Treatment + Site:Social + Treatment:Social + Site:Treatment:Social + Block(Site) + Plot(Block), where Y is the individual tree value of outerwood basic density or DBH. Site, Treatment and Social are the main fixed effects and Site:Treatment, Site:Social, Treatment:Social and Site:Treatment:Social are their two- and three-way interactions. Block within Site Block(Site) and Plot within Block Plot(Block) were treated as random effects. The linear mixed models were implemented using the lmerTest R package (Kuznetsova et al. 2017) and the Tukey post-hoc test was conducted to compare levels of the fixed effects using the emmeans (Lenth et al. 2020) and multcomp (Hothorn et al. 2008) R packages.

Results and discussion

Longitudinally, our study showed that checks were concentrated at lower stem heights and intermediate ring positions, but the exact pattern varied depending on tree social status, site and silvicultural treatment. At the tree level, some higher-level interactions among site, wedge height, thinning and social status were important, which makes checking variation complex to model. Dominant trees had greater checking in the thinned compared with the unthinned treatment. Within the unthinned treatment, checking increased with increasing suppression, but this effect was not evident in the thinned stands. However, opposite trends were observed between the dominant and suppressed cohorts in the thinning treatments.

The results of the present study support the hypothesis that dominant trees in the thinned treatment experience more stem mechanical instability than dominant trees in the unthinned treatment, with the associated development of tension wood and subsequent increased modulus of elasticity. This may facilitate the occurrence of internal fractures due to higher tensile stresses in the wood, leading to unfavourable drying characteristics such as increased checking (Gacitúa et al. 2007; Clair et al. 2013). In addition, we argue that more suppressed trees in the unthinned stands may increase vessel frequency because of competition and water stress, which may promote checking. These two components agree with the hypotheses of stem mechanical instability and inter-tree competition, which are discussed below. In general, thinning results in check peaks shifting more towards the stem periphery, which for dominant trees means they have more checks in the denser and stiffer outerwood. At the tree level, readily assessed non-destructive wood properties provided little power to predict variation in checking.

Longitudinal variation

In the longitudinal dataset, the height of the wedge up the stem (H) was by far the most important main factor explaining variation in the checking score, but some of its two-way interactions were of higher importance than the main effects (Figure 3a). Accordingly, the interactions involving height were studied in more detail using the Friedman’s H statistic, which indicates the strength of interaction between variables. The four-way interaction term (H:S:So:Tr) was the strongest interaction term (Figure 3b), indicating that the effect of height varied depending on multiple factors. Nevertheless, the main longitudinal trend was evident in all treatment combinations, with checking highest in the basal disc and rapidly declining over the next 10 m to generally stabilise (Gads site) or decrease slowly (Urana and Florentine) thereafter (Figure 4). The interactions mainly reflected differences in the exact height and degree of stabilisation of checking levels higher up the tree. Checking was clearly greatest in wedges from bottom logs, followed by second logs. Although there was variation in the main trend, particularly higher up the tree, the high-order interaction effect seemed to have little influence at 2.5 m height (where wedges were taken for the rest of the study) (Figure 4). There are only a few longitudinal studies of checking in eucalypts and pines, but these report similar trends. For example, in E. nitens grown in New Zealand, Shelbourne et al. (2002) reported a higher number of checks in the first bottom log from 0 to 6.4 m, with insignificant numbers at 11.4 m height. The higher development of checks at lower log positions has also been found in boards of Australian-grown E. nitens, with a significantly higher number of checks reported in boards taken from bottom logs than upper logs, regardless of sawing method (Washusen et al. 2009). For E. nitens grown in Chile, Valenzuela et al. (2012) reported a similar longitudinal pattern of check variation with higher levels of checking at the bottom stem positions (0–7 m) at two sites, results that were consistent with those of Leandro-Zúñiga et al. (2008), who found that checks were highly concentrated in the first 10 m of the tree. Similar longitudinal trends have been found in Pinus radiata D.Don, where checking severity decreased with the tree height, with the bottom of bottom log the most affected (Aguilera and Inpinza 2009).

Figure 3. Variable importance for check area (a) and higher-order interaction strength (b) estimated from the Random Forest analysis of longitudinal data from 144 Eucalyptus nitens trees. The segmented vertical dashed line (reference accuracy) represents the root mean square error (RMSE) of the full model using the original data (unshuffled). The higher the value from the dash line, the more important the term. The bars show the 95% confidence interval after 1000 permutations; when the lower interval does not overlap the reference accuracy, the model term was considered important. Higher-order interactions were calculated based on Friedman’s H statistic of 100 model runs. Main effects are wedge height and site, tree social status, and treatment (thinned/unthinned). The composite terms are the two-, three- and four-way interactions between these main effects

Figure 4. Accumulated local effect (ALE) plots showing the predicted effect of stem height on check score for each combination of trial, treatment and social class. ALE plots display how changes in a specific stem height shift the model predictions towards higher or lower checking (based on check score). For example, wedges with scores higher than two units above the grand mean are predicted to occur at the bottom of the tree stem. 95% confidence intervals of the means around the thinned and unthinned treatment responses are shown with shading

The underlying factors explaining why checks are concentrated in the lower part of the tree stem are not clear. However, it has been argued that such drying defects are likely due to tree growth stresses and the occurrence of tension wood, which are higher at the bottom of trees and mostly concentrated in the stem periphery (Washusen et al. 2003). With age, trees require higher levels of stem stability and respond by producing wood with a higher modulus of elasticity, making the wood more brittle as the capacity to absorb mechanical energy becomes lower (Lachenbruch et al. 2011). Regardless of the cause, this longitudinal variation has important economic implications for the development of high-quality products because current solid-wood silviculture regimes concentrate investment on the bottom logs (through pruning and thinning), where the levels of checking are highest.

Tree-level variation (at 2.5 m height)

At the tree level, check area (expressed as the proportion of a 2.5 m wedge area) had a high correlation with check number proportion (r $=$ 0.85, P $<$ 0.01, Appendix Figure B1). Thus, check area proportion was used in all the studies discussed below of the effects of site (S), treatment (Tr) and social status (So) on tree-level checking. The three-way interaction term (S:So:Tr) was the most important effect (Figure 5), but general trends were evident. Dominant trees in the thinned treatment had more checks than dominant trees in the unthinned treatment, but the opposite trend was found for suppressed trees (Figure 6). These trends were relatively consistent across sites. Subdominant and dominant trees tended to show a similar (Urana and Gads) or neutral (Florentine) trend in the thinning treatments (Figure 6).

Figure 5. Variable importance for the check area proportion in wedges taken at 2.5 m from 144 Eucalyptus nitens trees. The variable importance is based on the root mean square error (RMSE) loss. The segmented vertical dashed line represents the RMSE of the full model using the original data (unshuffled). The higher the value from the ‘reference accuracy’, the more important the term. The bars show the 95% confidence interval after 1000 permutations, and when the lower interval does not overlap the ‘reference accuracy’, the model term was considered important. The explained variance of the extended Forest model is R² $=$ 6.8%

Figure 6. Accumulated local effect (ALE) plots showing the predicted effect of silvicultural treatment by social status across the three sites on check area proportion. Analysis was undertaken using 1000 runs and 95% confidence intervals shown (bars). ALE plots display changes in check area proportion from the overall mean prediction (centred on zero); for example, suppressed thinned trees had lower check area proportion and these values are below the mean prediction

Few studies exist that quantify the between-tree variation of checking in plantation-grown eucalypts. Shelbourne et al. (2002) showed a significant between-tree variation in checking, expressed as the mean number of checks per ring. However, there are few previous studies and no definite conclusions in eucalypts regarding the effect of thinning on internal checking. Washusen et al. (2009) reported no thinning effect on the internal checking of E. nitens boards, but Valenzuela et al. (2012) found that checking was significantly higher in thinned than unthinned E. nitens coupes at two sites in Chile, as we found for dominant (and potentially subdominant) trees. However, the two aforementioned studies did not stratify the samples by tree social status, as we did.

The greater checking in dominant (and in some cases subdominant – Urana and Gads) trees from the thinned compared with unthinned treatments (Figure 6) may be due to their development of higher levels of tension wood in the lower stem because of the greater mechanical stresses they were subject to. Because dominant trees are more susceptible to stem mechanical failure than smaller trees (Clair et al. 2013), it is possible that their greater check development in our thinned treatment was a response to increased wind exposure after thinning, resulting in the rapid development of stiffer wood (tension wood) at the bottom of their stems, combined with their faster growth rate, which increases the proportion of thinner-walled cells (Borukanlu et al. 2021).

Checking increased with increasing tree suppression in the unthinned treatment (particularly Urana and Gads, Figure 6). This trend may be because high inter-tree competition accentuated water stress (White et al. 2009; Forrester and Baker 2012), particularly in the smaller trees, which are likely to have less-developed root systems than larger trees (Misra et al. 1998). The reduced checking observed in the suppressed trees in the thinned compared with the unthinned treatment (Figure 6) may be due to (1) removal of the extremes of a cohort of small trees with high checking during thinning (as thinning was from below); and (2) the wind-protected trees in the suppressed cohort at assessment age in the thinned treatment experiencing a phase of reduced water stress following thinning, leading to the production of wood with less checking propensity. Reduced water stress significantly reduces the density of vessels (Searson et al. 2004), which could reduce the propensity for checking. Thinning has been associated with a reduction in vessel frequency in tree species, including eucalypts (Phelps and Workman 1994; Diaconu et al. 2016; Aiso-Sanada et al. 2019), and this effect has been reported to significantly increase wood cleavage strength, which is associated with reduced checking (Soares et al. 2021). Suppressed trees in the thinned treatment are likely subject to less mechanical stress than dominant trees, and their increased DBH following thinning may be enough to ensure stem stability. We thus suggest that the opposite social-class trends in checking in the thinned compared with unthinned treatments is driven by (1) the greater susceptibility of dominant trees in the thinned treatment to mechanical stresses (i.e. tension wood formation); and (2) the high susceptibility of the suppressed trees in the unthinned treatment to water-stress-induced checking and the removal of the extremes of this cohort or alleviation of the water stress following thinning.

Radial variation (at 2.5 m height)

When the ring-level data from the 2.5 m wedges were considered, ring position (Rp, representing years from core at 2.5 m) was the most important variable in the RF model (Figure 7a). The main effect response was non-linear, with check area proportion increasing from rings nearby the pith towards the cambium and then declining in the outer rings (Figure 8). However, interaction terms were important, of which the four-way interaction (Rp:S:So:Tr) was the strongest (Figure 7b).

Figure 7. Variable importance for ring-level check area proportion (a) and higher-order interaction strength (b) of 144 Eucalyptus nitens studied trees. The segmented vertical dashed line represents the root mean square error (RMSE) loss from the full model using the original data and the bars show the 95% confidence interval after 1000 permutations. Higher-order interaction strength was calculated based on Friedman’s H statistic of 100 model runs. The main factors modelled were site, treatment (thinned/unthinned), ring position and social status

Figure 8. Accumulated local effect (ALE) plots showing the predicted effect of ring position on check area proportion within each Eucalyptus nitens tree ring. Analysis was undertaken using 1000 runs and 95% confidence intervals shown (shaded area). ALE plots display how changes in a specific ring position shift model predictions towards higher or lower check area proportion (overall mean centred to zero). The vertical dash lines indicate the rings of peak checking in the unthinned and thinned treatments. Trials were thinned at 8–9 years of age and radial variation in checking was recorded 2.5 m up the stem. Age at thinning corresponds to the ring positions 5, 6 and 7 at Urana, Florentine and Gads, respectively

The interaction strength indicated that the exact radial response depended on the various combinations of site, social status and treatment, although general trends were evident. For example, the ring position(s) of peak checking was in rings closer to the cambium in the thinned compared with the unthinned treatment (Figure 8), meaning that there is more checking in the outerwood of the thinned treatment (Figure 9). In general, the delay in check peaks was consistent across sites and tree social status, but this was less evident at Urana. The differences in the radial variation of checking between the thinned and unthinned treatments prior to when the thinning occurred (Figure 8) may be explained in part by the implicit effect of greater ring width in the thinned stand (Appendix Figure C1), which could potentially increase the effect of the earlywood–latewood variation. For example, a higher intra-ring variance in wood-density profiles was observed in checked wood of E. regnans F. Muell (Ilic 1999). This radial pattern of variation in checking was relatively consistent across our sites, but the exact pattern and peak checking area proportion was variable, depending on site and social class. In all cases, peak checking generally occurred in the band of rings where the transition occurred from low-density corewood (between rings 6 and 8 from the pith at 2.5 m) to higher-density outerwood (between rings 11 and 13 at 2.5 m) in E. nitens trees and where the radial variation in microfibril angle had become relatively stable (commencing between rings 6 and 10, depending on site) (Vega et al. 2020). However, this coincidence does not explain the shift in checking peak towards ring positions closer to the cambium in dominant trees.

Figure 9. Spatial representation of check distribution on wood discs (sampled at 2.5 m) from wedge-derived radial prediction of checking. Predictions at ring level were obtained from Random Forest models. Dark blue regions represent areas with a higher proportion of checks. Ring widths were calculated based on the average ring width at the wedge level for each site, social status and treatment combination shown. The thinning treatment slightly shifts check concentrations towards the stem periphery, with the extent of the shift depending on site and social status

The radial variation of checks observed in the present study was consistent with the study of 14-year-old E. nitens reported by Leandro-Zúñiga et al. (2008). They found checking to be highly dependent on radial position, with the highest (peak) number of checks located between rings 4 and 9 from pith, and this between-ring variation in checking was negatively related to wood density and positively related to frequency and diameter of vessels measured at the base of the tree. Similarly, Valenzuela et al. (2012) found that the highest (peak) number of checks occurred between rings 5 and 7, and its variation may be related to an increased vessel size at those ring positions in 12-year-old E. nitens discs collected at 3 m. Ball et al. (2005) found that decreasing wood density, increasing tracheid radial diameter and decreasing cell wall thickness were associated with increased incidence of intra-ring checking in the softwood, P. radiata. In P. radiata, tracheid collapse has been observed adjacent to checks and between checks, with cell walls appearing to have altered lignin distribution, particularly in the xylem S₁ cell wall layer (Putoczki et al. 2007). Hardwood species such as E. nitens are characterised by a more diverse wood microstructure, including vessels, so comparisons between softwood and hardwood species in terms of checking and its relationship with anatomical or physical properties should be made with caution. However, regardless of species, collapsed wood cells seem to promote the development of checking. There is general agreement that checking occurs predominantly in middle sections of wedges/discs, where wood traits such as wood density and stiffness rapidly transit from lower to higher values as trees age (Vega et al. 2020). Radially, it may suggest a possible association at multiple scales (i.e. within and across rings) in which fibre cell walls may be more prone to collapse, or a sudden change in physical wood features may produce uneven water flow during drying and thus potentially increase checking.

In E. nitens plantations, Shelbourne et al. (2002) found that heartwood proportion was positively correlated with tree growth, with larger-diameter stems having a proportionally larger heartwood area. In the same study, most of the checking occurred in the transition zone between heartwood and sapwood. Thus, it is possible that the delayed peak in checks on the radial profile in the thinned treatment in the present study may be a consequence of a greater heartwood area stimulated by thinning, which would shift the transition zone towards the cambium.

It is common in E. nitens that many trees produce high longitudinal tensile stress around their circumference (Clair et al. 2013). Higher levels of checks closer to the periphery of the stem would lead to issues in wood processing of timber of higher quality. Such is the case for pruned logs, where clearwood products with higher appearance-grade quality are expected to be obtained from the free-of-defects external parts of logs (McKenzie et al. 2003). From a processing perspective, sawing techniques such as quarter-sawn cutting patterns have been shown to reduce the impact of checking in final products and could potentially increase the recovery of selected products (Washusen et al. 2009). Nevertheless, as pruning is undertaken in conjunction with thinning in solid-wood silvicultural regimes, the present study suggests that increased clearwood recovery may be countered by an increase in checking in the wood recovered from the larger dominant trees.

Prediction of internal checking at the tree level

In general, checking area proportion in the wedge sampled at 2.5 m (top of bottom logs) was poorly predicted by non-destructive measurements on standing trees taken at 1.3 m above ground level. Virtually no inter-tree variance in checking was explained by DBH alone (model M1) or when combined with wood properties assessed from the 1.3 m cores (model M2) (models are detailed in Table 2). The absence of an association of checking with core basic density at the tree level is consistent with the non-significant relationship between wedge basic density and wedge checking score reported for E. nitens by Blackburn et al. (2010). Using just the wood-density parameters derived from the resistance profile (Resi) of green 2.5 m discs (ring mean, minimum, maximum, standard deviation and range) and disc diameter slightly improved the predictive power, but the R² was never greater than 10% (model M3). However, when including destructive predictors assessed on the same wedge associated with wedge collapse and shrinkage (model M5, Table 2), a greater proportion of the variance in wedge checking was explained across all sites (M5 was R² $=$ 22%) and within the Gads (R² $=$ 22%) and Florentine (R² $=$ 14%) sites.

The greater proportion of the variance explained by the full model, M5, is likely due to the significant positive correlation between collapse and checking (Appendix Figure B1). Among the collapse measurements, collapse curvature or ‘washboarding’ (Collapse_C) was the most important variable in the RF model, followed by the maximum depth of collapse (Collapse_D) (Figure 10) (ALE plots are in Appendix Figure D1). Collapse is a drying trait that appears closely related to check development in eucalypts (Blakemore 2011). Internal checking generally increased with increasing wood collapse (Ilic and Hillis 1986; Innes 1996a; Yuniarti et al. 2015), and both drying defects are highly dependent on radial position within the stem (Rebolledo et al. 2013; Ananías et al. 2014). In addition to the three measures of collapse (top-three most important variables), the next most important predictor among the other drying traits was radial shrinkage (Shinkage_R) (Figure 10), although, at the univariate level, this was only significantly positively correlated with the check number per unit wedge area (Appendix Figure B1). Shrinkage and collapse have usually been considered distinct processes occurring during wood drying. However, they act in combination, so it is possible that both variables influence check development in eucalypts, even at the initial drying stages (Yang et al. 2014).

Figure 10. Variable importance for check area proportion at the wedge level for 144 Eucalyptus nitens trees based on the model M5 (all traits included; see details in Table 2). The segmented vertical line represents the RMSE loss of the full model using the original data and the bars show the 95% confidence interval after 1000 permutations

Of the destructive measurements of wood characteristics taken at the disc level, within-ring standard deviation of microdrilling resistance was the fourth most important variable (Figure 10). Eucalyptus nitens has a high intra-annual range in wood density, with low-density wood formed early in the growing season and high-density wood later (Wimmer et al. 2002). This fourth variable is considered to mainly reflect the within-ring seasonal variation in density, although other factors such as release from water stress can also contribute to low-density wood production within a ring (Wimmer et al. 2002). The general trend was for checking area to increase with increasing within-ring variation in wood density (Appendix Figure D1). Blackburn et al. (2010) noted that the concentration of internal checks within the growth rings of E. nitens suggests that internal checking within a ring may be driven by variation in collapse between latewood and earlywood. Ilic (1999) observed that radial density profiles in checked E. regnans wood showed a markedly lower earlywood density than in non-checked samples. Although not significant at the univariate level (Appendix Figure B1), the present study suggests that variation in the rate of change in intra-ring density profile may be more important than the mean or minimum values.

Among the non-destructive measurements, NIR-predicted extractives content was the most important variable (fifth most important variable), and although it was not significantly associated with checking at the univariate level (Appendix Figure B1), it was well above other NIR-predicted wood properties in importance (Figure 10). Wood extractives may be positively associated with checking because extractives could block water and air flow during drying, producing uneven moisture content gradients that may contribute to drying defects such as excessive shrinkage and collapse (Kininmonth 1972; Dawson et al. 2020). Drying can potentially relocate extractives, which can decrease porosity in cell walls and thereby increase the potential for excessive shrinkage and collapse. For example, Meyer and Barton (1971) found that collapse was related to high extractive content in western red cedar (Thuja plicata Donn ex. D.Don) wood (a softwood) because pit chambers and membranes of wood cells were heavily deposited with extractives. To avoid this issue, methods such as ‘partial removal of extractives’ by steaming was found to alleviate the blocking effect of extractives, allowing greater permeability and resulting in more even radial and tangential diffusion during drying in eucalypts (Alexiou et al. 1990) and potentially resulting in less checking. However, in the present case, increased checking was associated with decreasing levels of extractives (Appendix Figure D1).

Although the present study did not enable the derivation of a usable model for predicting checking at the tree level, it has revealed some of the most important variables that warrant further exploration. Future research using the direct relationship between NIR spectra signatures and checking, in combination with anatomical wood features, may assist in developing better checking predictive models, not only at the tree level but also along the radial continuum of variability. The RF models including these traits (M1 and M2) had no power to predict wedge checking, and these traits or their equivalent were of low importance in the full model (M5). Combined with the absence of a genetic correlation between internal checking and basic density (Blackburn et al. 2010), these results are consistent with the argument that the current genetic improvement of E. nitens for a pulpwood breeding objective would not degrade the plantation resource for solid-wood production.

Conclusion

The importance of higher-level interactions and low predictability of internal checking in E. nitens argue it is a complex trait influenced by many factors. However, general trends are evident at between- and within-tree levels (Figure 11). Our study suggests that thinning may have adverse effects on checking in dominant trees. This trend, combined with a shift in checking towards the stem periphery, could potentially affect the recovery of higher-quality timber products from pruned bottom logs of E. nitens. However, the significant reduction of checking in upper logs represents an opportunity to develop products with reduced levels of checks and collapse that do not require clearwood. There is a need for future research to focus on reducing checking of the most valuable bottom logs where investment in pruning is concentrated to produce clearwood.

Figure 11. Schematic diagram of the two main findings of the present study. The general trend at the ring level depicts the delay in checking in trees from thinned stands (top panel). The bar plots (bottom panel) depict the main tree-level differences between dominant and suppressed cohorts in the unthinned and thinned treatments. The wedges represent examples of each situation

Acknowledgements

The authors acknowledge Sustainable Timber Tasmania for providing access to the trials and allowing sample collection from selected trees. The authors thank Dr Thomas Baker and Hugh Fitzgerald (School of Natural Sciences and ARC Industrial Transformation Training Centre, University of Tasmania) for assisting in the fieldwork. Thanks to Dr Phillip Blacklow (School of Creative Arts and Media, University of Tasmania) for his technical support with the wood sample processing. Thanks to Dr. Geoff Downes for conducting the near-infrared readings. Thanks also to the anonymous reviewers, whose comments helped improve the manuscript.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/00049158.2024.2341534

Additional information

Funding

This research was funded by Australian Research Council Industrial Transformation Training Centre for Forest Value (project number IC150100004).

Appendices

Appendix A.

Assessment of wood drying traits

Normal cross-sectional shrinkage (Shrinkage_CS) was calculated as:

(1) $\mathrm{Shrinkage}\mathrm{\_}\mathrm{CS}=\frac{R_g-R_d}{R_g}\times 100\left(\mathrm{\%}\right)$(1)

where R_g and R_d are the reconstructed cross-sectional distance of green and oven-dried wedges (see Appendix Figure A1a), respectively. Tangential shrinkage (Shrinkage_T) was calculated as:

(2) $\mathrm{Shrinkage}\mathrm{\_}\mathrm{T}=\frac{{\mathit{D}}_{\left(1\mathit{g},3\mathit{g}\right)}-D_{\left(1\mathit{d},3\mathit{d}\right)}}{D_{\left(1\mathit{g},3\mathit{g}\right)}}\times 100\left(\mathrm{\%}\right)$(2)

where $D_{\left(1\mathit{g},3\mathit{g}\right)}$ is the distance between the left and right outer corners of a green wedge and $D_{\left(1\mathit{g},3\mathit{g}\right)}$ is the same distance measured from the wedge when oven-dried, as depicted in Appendix Figure A1a. Radial shrinkage (Shrinkage_R) was determined as:

(3) $\mathrm{Shrinkage}\mathrm{\_}\mathrm{R}=\frac{{\mathit{D}}_{\left(2\mathit{g},4\mathit{g}\right)}-D_{\left(2\mathit{d},4\mathit{d}\right)}}{D_{\left(2\mathit{g},4\mathit{g}\right)}}\times 100\left(\mathrm{\%}\right)$(3)

where $D_{\left(2\mathit{g},4\mathit{g}\right)}$ is the distance between the pith and mid-point of the tangential segment of a green wedge and $D_{\left(2\mathit{g},4\mathit{g}\right)}$ is the same distance measured after the wedge was oven-dried, as depicted in Appendix Figure A1a. Collapse (Collapse_T) was estimated as:

(4) $\mathrm{Collapse}\mathrm{\_}\mathrm{T}=\frac{R_d-S_d}{R_d}\times 100\left(\mathrm{\%}\right)$(4)

where Collapse_T is the total non-reconditioned collapse, $R_d$ is the area of the reconstructed section of the oven-dried wedge and $S_d$ is the actual cross-section area. The difference between $R_d$ and $S_d$ is coloured red in Appendix Figure A1b.

The depth of collapse (Collapse_D) for a given wedge was calculated as:

(5) $\mathrm{Collapse}\mathrm{\_}\mathrm{D}=\frac{D\mathrm{ept}{h}_L}{L_L}+\frac{D\mathrm{ept}{h}_R}{L_R}$(5)

where$\mathrm{Collapse}\mathrm{\_}\mathrm{D}$ is the total depth of collapse, $\mathit{Dept}{h}_L$ is the maximum depth of collapse for the left side of the wedge, and $\mathit{Dept}{h}_R$ is the maximum depth of collapse for right side of the wedge. $L_L$ and $L_R$ are the left- and right-side length, respectively, from reconstructed sections (Appendix Figure A1c). The curvature of collapse was calculated as the total area under the absolute local curvature profile representing concave deformations (red area in Appendix Figure A1b). The details of these calculations are explained in Rocha-Sepulveda et al. (2021).

Appendix B.

Correlation matrix

Appendix C.

Predicted ring width variation

Appendix D.

Predicted effect of wood properties on checking area proportion

Appendix Figure A1. Image segmentation for a green (a, top) and its corresponding oven-dried wedge (a, bottom). The overlapped images (b) and the collapse area (in red) are also shown. The absolute local curvature along the radial face of an oven-dried wedge and an expanded area of the right-side contour is shown (c). The maximum depth of collapse (Depth_R) along the reconstructed straight edges of a wedge is also depicted in (c). The coding is explained in the text. Figure modified from Rocha-Sepulveda et al. (2021)

Appendix Figure B1. Correlation matrix showing check area proportion and number of checks in wedges taken at 2.5 m (tree-level analysis; n $=$ 144) and their correlation with the most important variables in the Random Forest model (M5; Figure 11). Correlations significantly different from zero are indicated: *0.01 ≤ P < 0.05; **0.001 ≤ P < 0.01; ***P < 0.001. Check_a: check area proportion; Collapse_C: collapse curvature; Collapse_D: depth of collapse (%); Check_n: check number proportion; Collapse_T: collapse total area (%); Extractives: Near-infrared (NIR)-predicted extractives content (%); Resi_sd: standard deviation of microdrill resistance-derived wood density (ring position 4–9); Shrinkage_R: radial shrinkage (%)

Appendix Figure C1. Accumulated local effect (ALE) plots showing the predicted effect of ring position on ring width within each Eucalyptus nitens tree ring. Analysis was undertaken using 1000 runs, and 95% confidence intervals are shown (shaded area). ALE plots display how changes in a specific ring position shift model predictions toward higher or lower check area proportion (overall mean centred to zero). Trials were thinned at 8–9 years of age and ring data were recorded 2.5 m up the stem. Age at thinning correspond to the ring positions 5, 6 and 7 at Urana, Florentine and Gads, respectively

Appendix Figure D1. Accumulated local effect (ALE) plots showing the predicted effect of Collapse_C, Collapse_D, Resi_sd, Resi_rate, Extractives and Shrinkage_R on check area proportion. Analysis was undertaken using 1000 runs and 95% confidence intervals are shown (shaded area). Rug plots along the bottom indicate the distribution of the data (individual trees)