Leaf-Centric Paradigm in Plant Analysis

• Leaf-Centric Paradigm is a framework that treats individual leaves as the primary analytical unit, capturing their vascular, hydraulic, mechanical, and morphogenetic details.
• It employs methodologies like topological analysis and conformal mapping to reveal detailed growth dynamics and network intricacies at the leaf level.
• This approach enhances phenotyping, digital reconstruction, and understanding of post-abscission function, leading to improved assessments of plant structure and resilience.

The Leaf-Centric Paradigm denotes a family of research frameworks in which the leaf is treated as the primary analytical unit, rather than as a passive appendage summarized by whole-plant averages or by a small set of coarse geometric descriptors. In this organ-centered view, a leaf may be analyzed as a reticulate vascular network with intrinsic topology, a distributed hydraulic system with spatially varying water status, a mechanically regulated thin lamina, an autonomous aerodynamic body after abscission, and a discrete phenotyping instance in computer vision and 3D reconstruction. Across these domains, the common claim is that biologically relevant structure and function emerge at the level of the individual leaf and are often lost when analysis is restricted to global plant traits or purely local measurements (Ronellenfitsch et al., 2015, Luo et al., 2021, Biviano et al., 2024).

1. Conceptual scope and defining claims

A leaf-centric framework is defined by the choice of the leaf itself as the locus of explanation. In venation studies, this means that a leaf is not adequately characterized by geometry alone, because the topology of nested loops adds a new, approximately independent axis of phenotypic variation (Ronellenfitsch et al., 2015). In hydraulics, it means that the leaf is not a passive hydraulic terminus but a distributed organ whose xylem, stomata, and storage tissues generate internal spatial gradients in water potential and flow (Luo et al., 2021). In morphogenesis and mechanics, it means that leaf form must be explained from whole-lamina growth fields and the regulation of flatness, rather than from a single mean growth rate (Armon et al., 2020, al-Mosleh et al., 2022). In post-abscission physics, it means that leaf shape remains functionally relevant after detachment because settling speed affects nutrient return near the parent tree (Biviano et al., 2024).

This perspective is also methodological. The 2016 conformal-growth study argued that, for nearly planar leaves with locally isotropic growth, the changing leaf contour is sufficient to recover most of the internal displacement field (Alim et al., 2016). The 2020 growth-field study argued that the biologically relevant object is the full spatiotemporal statistics of a tensorial growth field over the lamina, not just its mean (Armon et al., 2020). The 2022 flat-leaf mechanics study went further by formulating flatness as a control problem: long-wavelength bending modes are cheap in thin sheets, so a growing leaf must regulate growth through feedback to remain flat (al-Mosleh et al., 2022). Taken together, these works define the paradigm not as a single theory but as a consistent re-centering of analysis on leaf-scale structure, dynamics, and function.

2. Venation as a multiscale topological and hydrodynamic system

One of the clearest statements of the leaf-centric paradigm appears in work on reticulate venation. Conventional phenotyping had emphasized vein density, areole area, intervein distance, diameter distributions, branching angles, and segment lengths. The 2015 venation-topology paper argued that these descriptors miss the hierarchy of recursively nested loops in angiosperm leaves and introduced a topological phenotype derived from a nesting tree built by hierarchical decomposition of planar areoles (Ronellenfitsch et al., 2015). For each internal tree node $j$, with subtree leaf counts $r_j \ge s_j$, the nesting ratio is

$q_j=\frac{s_j}{r_j},$

and the overall nesting number is

$i=\sum_j w_j q_j,\qquad \sum_j w_j=1.$

The paper used both an unweighted nesting number $i_u$ and a degree-weighted version $i_w$, and paired them with a tapering descriptor, the mean topological length $L_{\mathrm{top}}$. Combined with five geometric variables—vein density $\sigma$, mean distance between veins $a$, mean areole area $A$, areole density $r_j \ge s_j$0, and average vein diameter $r_j \ge s_j$1—these form an 8-dimensional “leaf venation fingerprint” (Ronellenfitsch et al., 2015).

The evidence for a distinct topological axis was multivariate and practical. On a database of 186 leaves and leaflets representing 137 species, predominantly Burseraceae, principal component analysis showed that the first two principal components explained $r_j \ge s_j$2 of total variance, with $r_j \ge s_j$3 on component 1 and $r_j \ge s_j$4 on component 2; component 1 was dominated by geometric variables, whereas component 2 was dominated by topological variables $r_j \ge s_j$5, with $r_j \ge s_j$6 contributing to both (Ronellenfitsch et al., 2015). The same study showed that fragment identification improved substantially when topology was added: using only geometric descriptors, 10-fold cross-validated Linear Discriminant Analysis accuracy was $r_j \ge s_j$7 with $r_j \ge s_j$8 confidence interval $r_j \ge s_j$9; adding topology raised accuracy to $q_j=\frac{s_j}{r_j},$0 with $q_j=\frac{s_j}{r_j},$1 confidence interval $q_j=\frac{s_j}{r_j},$2, with Welch’s $q_j=\frac{s_j}{r_j},$3, $q_j=\frac{s_j}{r_j},$4 (Ronellenfitsch et al., 2015). The proposed developmental interpretation was that nestedness records the stochastic history of loop subdivision, with low noise preserving a legible hierarchy and high noise obscuring it.

A complementary leaf-centric development appeared in full-scale hydrodynamic modeling of venation. The 2024 full-network study moved from idealized networks to complete leaf graphs extracted from images, preserving node-edge topology, lengths, and widths, and then ran a transport optimization model directly on those graphs (Skjegstad et al., 2024). Flow obeyed

$q_j=\frac{s_j}{r_j},$5

with conductivity-width scaling $q_j=\frac{s_j}{r_j},$6. Under a material-cost constraint $q_j=\frac{s_j}{r_j},$7 with $q_j=\frac{s_j}{r_j},$8, optimal conductivities satisfy

$q_j=\frac{s_j}{r_j},$9

To produce loops, the model averaged squared flows over fluctuating sink configurations and fitted a sink-fluctuation amplitude $i=\sum_j w_j q_j,\qquad \sum_j w_j=1.$0. The fitted species means were approximately $i=\sum_j w_j q_j,\qquad \sum_j w_j=1.$1 for Symphoricarpos albus, $i=\sum_j w_j q_j,\qquad \sum_j w_j=1.$2 for Lonicera xylosteum, and $i=\sum_j w_j q_j,\qquad \sum_j w_j=1.$3 for Crataegus monogyna, with consistency within species but substantial edgewise residual error (Skjegstad et al., 2024). The same framework defined a Murray exponent for reticulate networks by averaging incoming and outgoing radius sums over moving-sink states; it recovered $i=\sum_j w_j q_j,\qquad \sum_j w_j=1.$4 in the tree limit and near-3 values with slight upward shifts when sink fluctuations preserved loop redundancy (Skjegstad et al., 2024).

3. Morphogenesis, growth fields, and the mechanics of flatness

A leaf-centric view of morphogenesis was formulated explicitly in the conformal-growth study of 2016. There, the contour of a leaf at one time was mapped to the contour of the same leaf at a later time using a conformal map $i=\sum_j w_j q_j,\qquad \sum_j w_j=1.$5, with $i=\sum_j w_j q_j,\qquad \sum_j w_j=1.$6, and the predicted displacement field was taken from

$i=\sum_j w_j q_j,\qquad \sum_j w_j=1.$7

Under the assumption of locally isotropic growth, conformality is equivalent to the Cauchy–Riemann-type conditions

$i=\sum_j w_j q_j,\qquad \sum_j w_j=1.$8

For relatively planar Petunia and Tobacco leaves imaged every hour over intervals from 3 hours to 3 days, with overall size increases from 10\% to 42\%, the measured and predicted displacement fields agreed with more than 92\% correlation, reaching 97\% in the best specimen (Alim et al., 2016). The large-scale growth dynamics of the leaves studied were reproduced by the first two terms of the expansion

$i=\sum_j w_j q_j,\qquad \sum_j w_j=1.$9

indicating a low-dimensional organ-scale structure in which the boundary evolution encodes most of the internal deformation (Alim et al., 2016).

This smooth organ-scale description was later complicated, not replaced, by measurements of growth intermittency. The 2020 study of Tobacco leaf #6 tracked a single wild-type leaf every 15 minutes for 2 days, during which its area increased from 28 to 89 mm$i_u$0 at about 4\% per hour on average (Armon et al., 2020). Using 3D profilometry and PIV in Lagrangian coordinates, the authors constructed a local growth tensor with eigenvalues $i_u$1, local areal growth

$i_u$2

anisotropy

$i_u$3

and principal-direction angle $i_u$4 relative to the main vein. At 15-minute resolution, the $i_u$5 field was broad and non-Gaussian, with abundant local shrinkage even though the leaf as a whole was expanding. The normalized histograms retained approximately constant higher moments, with skewness $i_u$6 and kurtosis $i_u$7, across temporal coarse-graining (Armon et al., 2020). A temporal spectral peak at $i_u$8 implied a characteristic timescale of roughly 45 minutes, and spatial decorrelation lengths were about $i_u$9 mm during the day and $i_w$0 mm at night, with night growth more intermittent and lacking global directionality (Armon et al., 2020). The central claim was that a leaf remains flat not because local growth is smooth, but because fluctuations are regulated and correlated.

The 2022 mechanics paper placed that claim in a control-theoretic setting. Modeling a leaf as a thin elastic plate near the flat state, it wrote the linearized quasi-static equations as

$i_w$1

where $i_w$2 is a scaled Airy stress function, $i_w$3 the out-of-plane deflection, $i_w$4 the in-plane growth incompatibility, and $i_w$5 the growth-induced transverse forcing (al-Mosleh et al., 2022). Purely local instantaneous feedback,

$i_w$6

can stabilize deterministic perturbations only if $i_w$7 and $i_w$8, but it fails to suppress long-wavelength stochastic fluctuations because the angular roughness diverges like $i_w$9 (al-Mosleh et al., 2022). The paper therefore introduced spatially nonlocal and temporally delayed feedback kernels,

$L_{\mathrm{top}}$0

and showed that such feedback suppresses the long modes that would otherwise make a flat lamina mechanically implausible (al-Mosleh et al., 2022). The broader implication is that flatness is an actively maintained state of the individual leaf.

4. The leaf as a distributed hydraulic organ

The hydraulic version of the leaf-centric paradigm treats the leaf interior as a spatially extended transport-and-storage network, rather than as a single bulk water-status variable. In the 2021 capacitive model, nodes $L_{\mathrm{top}}$1 along a grass-leaf xylem conduit are linked by axial resistances $L_{\mathrm{top}}$2, each node leaks to the atmosphere through stomatal resistance $L_{\mathrm{top}}$3, and each node exchanges with local storage through a capacitor $L_{\mathrm{top}}$4 and storage-pathway resistance $L_{\mathrm{top}}$5 (Luo et al., 2021). The local conservation law is

$L_{\mathrm{top}}$6

with Ohmic relations for axial flow and transpiration and a capacitive storage law

$L_{\mathrm{top}}$7

Under the uniform 1D approximation, this yields a continuum PDE coupling space and time, and for an excised leaf the mean xylem potential and total transpiration decay exponentially with characteristic time

$L_{\mathrm{top}}$8

This result makes leaf capacitance and xylem-to-storage accessibility central hydraulic traits, rather than secondary corrections (Luo et al., 2021).

The biological significance of the spatial formulation is that leaf water status is strongly heterogeneous. In the steady-state example given in the paper, with $L_{\mathrm{top}}$9, $\sigma$0 MPa, $\sigma$1 MPa m$\sigma$2 s mmol$\sigma$3, and $\sigma$4 MPa m$\sigma$5 s mmol$\sigma$6, water potential declines monotonically from base to tip, reaching about $\sigma$7 MPa at the tip while the mean is only $\sigma$8 MPa, and total transpiration is $\sigma$9 mmol m$a$0 s$a$1 (Luo et al., 2021). Under a humidity perturbation from $a$2 to $a$3 MPa, the mean potential falls gradually to $a$4 MPa, but the distal region can cross a severe-stress threshold of $a$5 MPa after about 20 min even while the organ-average state remains apparently safer (Luo et al., 2021). The paper concludes that large capacitance $a$6 and, to a lesser extent, larger $a$7 and $a$8, delay dehydration and increase robustness to intermittent drought. In leaf-centric terms, this means that the biologically relevant state variable is not a single leaf-average potential, but a spatial field whose extremes occur within the blade.

5. Leaf-centric phenotyping, segmentation, and digital reconstruction

In plant vision and phenotyping, the leaf-centric paradigm appears as a shift from whole-plant masks to leaf-level semantic or instance representations. One branch emphasizes category-level leaf segmentation under practical greenhouse constraints. The 2022 self-supervised framework for complex lighting combined a self-supervised semantic segmentation model, a color-based leaf segmentation algorithm, and a self-supervised color correction model (Lin et al., 2022). On natural-light images it achieved Foreground-Background Dice scores of 94.8 on cannabis, 94.7 on A1, 92.0 on A2, 95.2 on A3, and 96.1 on A4; under yellow lighting after correction, scores were 87.1, 88.7, 92.5, 93.9, and 92.3; under purple lighting after correction, 83.9, 94.7, 92.6, 94.8, and 83.8 (Lin et al., 2022). The paper treated leaf pixels as the practical entry point to canopy-related phenotyping when instance-level separation is difficult.

A second branch addresses explicit individual leaf instances. The 2019 point-cloud method for overlapping canopies used 3D joint filtering plus facet over-segmentation and facet-based region growing to recover separate leaves from crowded plant point clouds (Li et al., 2019). Across Epipremnum aureum, Monstera deliciosa, Calathea makoyana, and Hedera nepalensis, it reported average leaf-level Recall $a$9, Precision $A$0, and F-measure $A$1, with point-level cover rates of 97\%, 99\%, 99\%, and 87\%, respectively (Li et al., 2019). In 2D, the 2021 LeafMask system combined an anchor-free detector, a mask assembly module, a mask refining module, and a dual attention-guided mask branch, reaching 90.09\% BestDice on the Leaf Segmentation Challenge dataset (Guo et al., 2021). The 2026 ReLeaf benchmark then broadened the question from accuracy to generalization: a YOLO26 Medium model at $A$2 provided the best accuracy-latency trade-off, and a model trained on all four selected public datasets achieved mean mAP$A$3 of $A$4 across their test sets but only $A$5 on the new 23-species CropAndWeedAndLeaf benchmark, exposing strong cross-domain and cross-species degradation (Martinko et al., 5 May 2026).

The same paradigm extends to 3D generative modeling. NeuraLeaf, introduced in 2025, treats the individual leaf as the unit of a neural parametric model by disentangling geometry into a 2D base shape, a 3D deformation, and a texture field (Yang et al., 17 Jul 2025). The base shape is the zero-level set of a 2D neural signed distance function,

$A$6

while deformation is applied by a skeleton-free skinning model with $A$7 control points and vertex updates of the form

$A$8

This representation was trained using large 2D leaf datasets for base shapes and a new DeformLeaf dataset of about 300 base–deformation pairs for 3D deformation, enabling fitting to point clouds and depth maps while transferring deformation patterns across shape classes (Yang et al., 17 Jul 2025). A plausible implication is that leaf-centric computer vision and leaf-centric geometry are beginning to converge on the same object: a single leaf as a disentangled, measurable, and reconstructible entity.

6. Post-abscission function and terminological divergence

The leaf-centric paradigm also includes research on the detached leaf. The 2024 settling-aerodynamics study argued that deciduous leaves should be understood partly as anti-dispersal organs whose shapes affect sedimentation and thus nutrient return (Biviano et al., 2024). Using an Automated Sedimentation Apparatus capable of roughly $A$9 free-fall experiments per day on biomimetic paper leaves, the study found that most of 25 representative leaves settled at rates within about $r_j \ge s_j$00 of a circular disk control, whereas the Arabidopsis asymmetric leaves1 mutant fell about 15\% slower than wild type (Biviano et al., 2024). Digitally imposing as1-like asymmetry on deciduous tree leaves produced a similar $r_j \ge s_j$01 reduction, and a mutated Amelanchier arborea mimic with $r_j \ge s_j$02 compared with a wild-type mimic at $r_j \ge s_j$03 was nearly 25\% slower when normalized by the circular control (Biviano et al., 2024). The paper’s “fast-leaf hypothesis” states that deciduous leaves are symmetric and relatively unlobed in part because those traits maximize settling speed and nutrient retention. In this form, leaf-centricity means following the leaf beyond the canopy into free fall, deposition, and ecosystem recycling.

A separate issue is terminological. Outside botany, “leaf-centric” and the acronym LEAF are used in unrelated technical senses. In probability on random trees, a “leaf-growth measure” governs growth on a fractal subset of leaves, with typical support size $r_j \ge s_j$04 and a Brownian Continuum Random Tree limit of Hausdorff dimension $r_j \ge s_j$05 (Caraceni et al., 2024). In random binary-tree sources, “leaf-centric” means that the split law is defined by how leaf mass is divided between subtrees (Benkner et al., 2018). In reverse mathematics, “leaf management” refers to a transformation

$r_j \ge s_j$06

that equips arbitrary trees with explicit leaf sets without changing the strength of several tree principles (Hirst, 2018). In machine learning, LEAF is a benchmark for federated settings organized around client or task “leaves” (Caldas et al., 2018), LEAFAGE is a local example-based explanation method (Adhikari et al., 2018), and LEAF for few-shot continual event detection is a LoRA-expert architecture with semantic routing (Dao et al., 29 Sep 2025). These usages are conceptually separate from the botanical leaf-centric paradigm. The shared term reflects a general emphasis on terminal units or local entities, but only the botanical literature uses “leaf-centric” to mean that the biological leaf itself is the primary explanatory object.