Data collection

Rowan fruit production was monitored, by counting infructescences (using binoculars) in the period 2000-2022. Canopy openness was measured using hemispherical photographs, taken with a Kodak PIXPRO SP360 camera, directly above the crown of each rowan tree, in July of 2021. The diameter at breast height (DBH) and tree height (H) were measured in 2018. The architectural data collection (in September 2021) employed the single-image-photogrammetry (SIP) method, following \cite{Gazda_2017}, using  a Nikon D5600 camera (6000 × 4000 pixels images, oriented horizontally). 

Semi-automated crown delineation

The digital photographs were corrected for tilt (vertical inclination of camera line of sight) and scaled following \cite{K_dra_2022}.  The R  \cite{team2020} algorithm "SIP_SLIC0" (see Supplementary material) runs the image segmentation method SLIC0 – the Simple Linear Iterative Clustering with 0 parameter \cite{Achanta_2012,pu50y} over each image. The output is a raster layer where individual pixels are grouped into "superpixels" (representing homogeneous portions of the input image). One advantage of the SLIC0 method is that all superpixels are similar in size,  the requested superpixel area was set to 400 cm2 (equivalent to a 20 cm × 20 cm board), as a basic unit of foliage clumps automated delineation. The superpixel raster layer was digitized and all polygons representing foliage of the target individual tree were manually selected. The aggregated foliage superpixels was the final vertical foliage area shapefile layer, and the minimal convex hull of the former was the final vertical crown area shapefile layer. Finally, to account for a possible tree asymmetry \cite{Szwagrzyk_2015}, a single-point layer was created representing stem position below the live crown. The QGIS software was used as a user interface and to display the images. All measurements (extraction of the crown variables) were subsequently performed in R (in Cartesian coordinates), from the three shapefiles per tree.  

Data validation

To check the robustness of the manual part of the crown delineation method (selection of superpixels representing clumps of foliage) we selected 20 trees (10 difficult trees, and 10 random trees) to be evaluated by a second observer (see Supplementary material). We compared three output variables: Foliage Area (sum of the selected superpixel areas), Crown Area (convex hull of Foliage superpixels), and Transparency (1 - Foliage/Crown Area). We used the Standard Major Axis (SMA) regression \cite{pierre2014}, which deals with errors in both X and Y observations (measurements of the independent observers). This analysis confirmed very high robustness of the convex hull method (r=0.99, SMA_slope=43.1°), while Foliage was consistently estimated higher by the observer #1 than by the observer #2 (r=0.92, SMA_slope=29.4°). Still, for Crown Area and for Foliage the inter-observer agreement was satisfactory. Crown Transparency (based on the ratio between the former) exhibited lower inter-observer agreement (r=0.54), but less bias than for Foliage (SMA_slope=37.7°), and we decided to keep this variable as well. 

Tree and crown variables  

Two sets of structural traits were considered: seven absolute traits and eight relative traits. The seven size traits included: tree height (H) and diameter at breast height (DBH) - both measured in the field, and five image-derived traits: vertical foliage area (Foliage; sum of delineated foliage clumps), crown area (CA; convex hull area of delineated foliage clumps), crown width (CW), crown length (CL), and tree asymmetry (Asymm; the horizontal distance between crown center and stem center below live crown). 
The relative traits were expected to provide sharper structure-function linkages \cite{Iida_2011,MacFarlane_2017}, while controlled for similar components in the absolute trait pairs. Dividing one linear trait (vertical or horizontal extent) by another should account for, and largely cancel out, the effects of size and metabolic costs. Four such traits were included: tree slenderness (HD, tree height per DBH), the relative crown length (CL_H), the relative crown width (CW_H), and crown aspect ratio (CW_CL = CW/CL). A ratio between the quadratic traits provided a measure of crown transparency (Transp = 1 - Foliage/CA, the complement of foliage area per crown convex hull area).
The ratios between different power traits (linear and quadratic) were included to inform about the relative costs of the lower-order trait in relation to the higher-order trait. Two such traits were: crown width construction cost (CW_cost, units of crown area per unit of crown width) and crown length construction cost (CL_cost, units of crown area per unit of crown length). For instance, large values of the both ratios would indicate full and costly crowns, here expected to be rare. The third similar measure was the crown density index (DBH_CA, stem diameter fraction per unit crown area), considering stem size as a proxy of total branch length \cite{K_dra_2022,Gazda_2017} and foliar area \cite{hinckley1979,2002}.