论文解读

AI 辅助整理
方法
在8963条光谱、5类植物系统上验证,压缩1580波段96%,对比全光谱建模。
发现
压缩96%后宏观F1由0.7756升至0.8520,推理提速126.5倍,优于全光谱。
机会点
可探索MSPA在田间实测高光谱与多作物、多胁迫场景下的泛化能力及与深度模型结合。

论文摘要

提出形态光谱投影分析MSPA,把连续光谱投影到紧凑子空间以破解波段共线性。在8963条光谱、5类植物系统上验证,压缩1580波段96%后仍优于全光谱:宏观F1由0.7756升至0.8520,推理提速126.5倍。属光谱建模,非田间实测。

全文

arXiv

Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy
Preprint,compiledNovember19, 2026
Lucas Prado Osco1,, Marcelo Santos da Silva2, Filipe Vieira Silva3, Liang Zhao4,, Ana Paula Marques Ramos3, Ednaldo
José Ferreira5, Roberto Dantas de Medeiros6, Amaury Burlamaqui Bendahan6, and Lucio André de Castro Jorge5
1Graduate Program in Environment and Regional Development, University of Western São Paulo (Unoeste), Presidente Prudente, SP 19067-175, Brazil (e-mail:
lucasosco@unoeste.br)
2Institute of Mathematics and Computer Sciences (ICMC), University of São Paulo (USP), São Carlos, SP 13566-590, Brazil (e-mail:marcelo.ssgc@usp.br)
3Department of Cartography, School of Technology and Sciences, São Paulo State University (UNESP), Presidente Prudente, SP 19060-900, Brazil (e-mail:
filipe.vieira@unesp.br;marques.ramos@unesp.br)
4Department of Computing and Mathematics, Faculty of Philosophy, Sciences and Letters at Ribeirão Preto (FFCLRP), University of São Paulo (USP), Ribeirão Preto, SP 14040-901,
Brazil (e-mail:zhao@usp.br)
5EMBRAPA Instrumentation, Brazilian Agricultural Research Corporation (EMBRAPA), São Carlos, SP 13561-206, Brazil (e-mail:ednaldo.ferreira@embrapa.br;
lucio.jorge@embrapa.br)
6EMBRAPA Roraima, Brazilian Agricultural Research Corporation (EMBRAPA), Boa Vista, RR 69315-292, Brazil (e-mail:roberto.medeiros@embrapa.br;
amaury.bendahan@embrapa.br)
*Corresponding author: Lucas Prado Osco (lucasosco@unoeste.br)
Abstract
Plant reflectance spectroscopy across the visible, near-infrared, and shortwave infrared domain (VNIR–SWIR, 350–2500 nm)
encodes foliar biochemical and anatomical traits into contiguous spectral signatures with locally differentiable structure;
however, extreme inter-band collinearity and unfavorable sample-to-feature ratios challenge predictive modeling. Con-
ventional dimensionality reduction methods routinely adopted in spectroscopy, including Principal Component Analysis
(PCA), Partial Least Squares Discriminant Analysis (PLS-DA), Uniform Manifold Approximation and Projection (UMAP),
and Self-Organizing Maps (SOM), are column-permutation invariant: they treat adjacent wavelengths as unordered tabu-
lar columns and discard the sequential differential geometry of absorption features. Here, we propose Morpho-Spectral
Projection Analysis (MSPA), a supervised dimensionality reduction operator that projects contiguous spectra into a com-
pact morpho-spectral subspace through three complementary blocks: (i) additive continuum-removed absorption depth,
area, and red-edge inflection kinematics (λRE); (ii) local multi-order singular eigenmodes (0th, 1st, and 2nd derivatives)
partitioned across contiguous radiative sub-windows; and (iii) global Fisher-weighted canonical subspaces regularized via
trace-normalized isotropic covariance shrinkage. We validated MSPA across 8,963 spectra spanning five plant systems,
ranging from binary foliar disease and herbivory stress ( C =2) and ternary feeding guilds ( C =3) to high-cardinality
canopy species discrimination (C =39), coupled with six classical machine learning classifiers and four Tabular Foundation
Models (TFMs) (TabPFN v3, TabICL v2, TabDPT v1.2, and TabFM). Despite compressing the 1,580-band spectrum by 96%,
MSPA outperformed the full spectrum in 92.0% of all benchmarks and conventional reducers in 90.0% of all model–crop
evaluations, raising the global mean macro-F1 from 0.7756 (full spectrum) to 0.8520. Beyond improving classification
accuracy, MSPA reshaped gradient-descent loss landscapes into well-conditioned basins, reduced predictive uncertainty
(Shannon entropy) by 40.3%, and mitigated the high-dimensional attention bottleneck in TFMs, accelerating in-context
inference by up to 126.5×. Designed as a self-contained spectral projection operator, the MSPA method can be directly
integrated into spectroscopy workflows via the open-source code available at GitHub: MSPA.
1 Introduction
Reflectance spectroscopy of vegetation across theVisible,
Near-Infrared, andShortwave Infrareddomains (VNIR–SWIR,
350–2500 nm) characterizes the biochemical composition and
cellular architecture of plant tissues through a continuous radia-
tive transfer process [1, 2, 3]. Unlike generic tabular datasets
composed of independent variables, a plant reflectance signa-
ture ρ(λ) is a sequential, piecewise differentiable function of
wavelength λ1 < λ2 <···< λ D. Within this continuous
form, biophysical information is governed by the radiomet-
ric contrast between scattering plateaus (spectral regions of
minimal absorptance, such as the green peak near 550 nm and
the mesophyll-driven NIR plateau across 766 –1330 nm) and
localizedabsorption features(concave valleys where specific
molecular bonds absorb incident photons) [ 1, 4]. In the visi-
ble range (400–700 nm), electronic transitions in chlorophylls
a and b, carotenoids, and anthocyanins dictate the depth and
width of absorption wells and control the inflection kinemat-
ics of thered-edgetransition (680 –760 nm) [5, 6]. Across the
NIR and SWIR domains (910–2400 nm), harmonic overtones
and combination bands of O–H, C–H, and N–H stretching vi-
brations imprint overlapping concave features associated with
leaf water status, cellulose, hemicellulose, lignin, starch, and
proteins [1, 2, 7]. Mathematically, physiological or taxonomic
divergence between plant signatures manifests simultaneously
in zero-order reflectance amplitude (ρ), first-order local slope
(∂ρ/∂λ), second-order local concavity (∂2ρ/∂λ2), and horizontal
wavelength shifts of absorption band centers.
Despite the high spectral resolution provided by contiguous
spectroradiometers ( D> 1, 500 channels), translating high-
dimensional spectral behaviour into generalizable predictive
models is hindered by extreme inter-band collinearity, unfavor-
able sample-to-feature ratios (Hughes phenomenon) [8, 9, 10],
and pronounced intra-class spectral variability induced by illumi-
nation geometry, atmospheric residuals, canopy topography, and
intrinsic biochemical fluctuations [11, 12, 13]. To mitigate end-
member and signature variability while compressing the feature
space (R : RD →R K, with K≪D ), the hyperspectral re-
mote sensing and chemometrics literature has explored diverse
reductionist, morphological, and physics-informed strategies
[14, 15, 11, 16]. In spectral mixture analysis and target classifica-
tion, classical domain-guided transformations include Savitzky–
arXiv:2611.24599v1 [cs.LG] 18 Nov 2026

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy2
Golay derivative spectroscopy to resolve overlapping absorption
valleys [17, 18, 9], discrete and continuous wavelet decomposi-
tions across multiscale spectral windows [19, 20], convex-hull
continuum removal [4], stable-zone band weighting based on
inter- to intra-class variance ratios [21, 11], and Fisher discrim-
inant subspace projections [22, 11]. In hyperspectral imaging
cube analysis, Plaza et al. [ 14, 15] pioneered multichannel
mathematical morphology, including Automated Morphological
Endmember Extraction (AMEE) and Extended Differential Mor-
phological Profiles (EDMP), employing geodesic opening and
closing operators to jointly exploit spatial and spectral structures,
whereas recent hybrid frameworks such as PHANTOM [16] em-
bed Beer–Lambert absorption laws and partial differential equa-
tions directly into deep neural objectives. Nevertheless, spatial–
spectral morphological profiles rely on two-dimensional spatial
structuring elements and cannot be applied to one-dimensional
point spectroradiometry or isolated pixel spectra [15], leaving
standard analytical workflows reliant on general-purpose sta-
tistical dimensionality reduction like:Principal Component
Analysis(PCA) [ 23, 24],Partial Least Squares Discriminant
Analysis(PLS-DA) [ 25],Uniform Manifold Approximation and
Projection(UMAP) [ 26], and KohonenSelf-Organizing Maps
(SOM) [27, 28], among others.
However, applying these dimensionality reduction techniques
to plant spectroscopy entails two structural mathematical lim-
itations. First, PCA, PLS-DA, UMAP, and SOM are strictly
column-permutation invariant: for any permutation matrix
Π∈{ 0, 1}D×D that arbitrarily shuffles the physical ordering of
wavelengths, the projected subspaces remain identical (R(XΠ)=
R(X)). Because these operators rely solely on global column
inner products (X⊤XorX ⊤Y) or Euclidean sample distances
(∥xi− xl∥2), they treat contiguous wavelengths as unordered tab-
ular columns, discarding the sequential differential topology that
links a band to its immediate spectral neighbors. Second, global
variance-maximizing projections such as PCA concentrate their
leading eigenvectors on broad, high-amplitude radiometric varia-
tions (predominantly leaf albedo and canopy structural scattering
across the NIR plateau) regardless of whether those variations
separate the target biological classes, thereby masking narrow,
low-variance absorption features in the visible and SWIR re-
gions [29]. Conversely, while PLS-DA incorporates class labels
through one-hot response matrices, its discriminative rank is
bounded by C− 1; when extracting K>C− 1 latent variables
in binary (C =2) or ternary ( C =3) physiological tasks, sub-
sequent deflation steps fit residual intra-class covariance noise,
degrading out-of-sample generalization.
The geometry of the reduced subspace directly governs the
optimization dynamics and decision boundaries of down-
stream machine learning architectures [ 30, 31]. In gradient-
based and linear classifiers, including multinomialLogistic
Regression,Linear Discriminant Analysis(LDA), and multi-
layer neural networks (MLPandDNN), unconditioned high-
dimensional spectra or albedo-dominated PCA components pro-
duce anisotropic cross-entropy loss surfaces (LCE) characterized
by ill-conditioned Hessian matrices, causing gradient descent
trajectories to oscillate across narrow valleys or overfit small
sample sizes [32, 33, 34, 35, 31]. In distance-based and kernel
classifiers (k-Nearest NeighborsandSVM-RBF), hundreds of
collinear bands dilute pairwise Euclidean metrics inRD, whereas
in orthogonal tree ensembles (Random ForestandXGBoost)
axis-aligned splits on raw reflectance cannot directly evaluate
local curve slopes or absorption areas [36, 37].
Concurrently, artificial intelligence in Earth observation has
undergone a structural shift from task-specific deep neural net-
works trained from scratch, such as stacked autoencoders, con-
volutional networks, and recurrent architectures [ 32], toward
large-scaleRemote Sensing Foundation Models(RSFMs) evolv-
ing from unimodal visual backbones to multimodal and hyper-
spectral paradigms [38]. Leveraging self-supervised contrastive
learning and masked image modeling (MIM) over massive satel-
lite archives [38], spectral–spatial foundation models such as
SpectralGPT [ 39], HyperSIGMA [ 40], HyperFree [ 41], and
the sensor-agnostic HySens framework [42] have demonstrated
strong representation transfer across airborne and spaceborne
imaging spectrometers. Despite these advances, vision-centric
RSFMs are architecturally designed around two-dimensional
spatial patches, require computationally intensive pre-training
or parameter-efficient fine-tuning (e.g., LoRA adapters) [ 38],
and remain ill-suited for one-dimensional proximal plant spec-
troscopy, where field campaigns and laboratory-verified diagnos-
tics (such as quantitative real-time PCR for vascular diseases or
controlled herbivory bioassays) operate under small-to-moderate
sample regimes [12, 13, 29].
For such structured one-dimensional records, a parallel revolu-
tion has emerged throughTabular Foundation Models(TFMs),
such asTabPFN[ 43, 44],TabICL[ 45],TabDPT[ 46], and har-
monic Fourier-gated architectures (TabFM), which approximate
posterior predictive distributions in a single forward pass viaIn-
Context Learning(ICL) without gradient-based weight updates.
Yet, how TFMs process contiguous spectral signals and interact
with dimensionality reduction remains an open question. From
an architectural standpoint, exposing TFMs to high-dimensional
plant spectra introduces a fundamental dilemma: models that
compute feature-wise attention across columns (TabPFNand
TabICL) face severe computational scaling bottlenecks when pro-
cessing hundreds or thousands of contiguous channels, whereas
models relying on local neighborhood retrieval (TabDPT) or
random Fourier projections (TabFM) remain theoretically vul-
nerable to metric dilution in high dimensions. However, com-
pressing spectral curves prior to ICL via conventional column-
permutation-invariant reducers (PCA,PLS-DA,UMAP, orSOM)
risks discarding the localized slope and concavity cues required
to resolve subtle physiological states.
To overcome the column-permutation invariance and variance-
masking limitations of traditional dimensionality reduction in
plant spectroscopy, we proposeMorpho-Spectral Projection
Analysis(MSPA). Formulated as a supervised projection oper-
ator (TMSPA : RD→R K), MSPA preserves physical wavelength
contiguity and localizes inter-class spectral divergence through
three complementary mathematical modules: (1) a nonlinear
biophysical hull module (16 variables) that extracts additive
continuum-removed maximum band depth (Dmax) and integrated
absorption area (Aabs) across major biochemical valleys, coupled
with first- and second-derivative red-edge inflection kinematics
(λRE centroid) and inter-plateau radiative contrasts; (2) a local
multi-order sub-window module (24 variables) that partitions
the spectrum into contiguous radiative windows and extracts the
dominant rank-1 Singular Value Decomposition (SVD) eigen-
modes separately for Order 0 (amplitude), Order 1 (local slope),

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy3
and Order 2 (local concavity), preventing high-variance NIR
plateaus from masking narrow visible and SWIR features; and
(3) a global Fisher-weighted canonical module (Kcan variables,
default 24) that weights the coupled state space [X(0), X(1), X(2)]
by inter/intra-class Fisher discriminability and extracts regular-
ized discriminant axes via trace-normalized isotropic covariance
shrinkage [71, 47], incorporating rank-safe orthogonal comple-
tion for arbitrary class cardinalities ( C≥ 2). By combining
piecewise analytical Savitzky–Golay differentiation across spec-
tral excision gaps, additive continuum removal, and internal
column standardization, MSPA operates directly on smoothed
reflectance spectra (ρ(λ)∈ [0, 1]), supporting both user-defined
dimensionality ( K) and autonomous intrinsic breaking-point
estimation via the Kneedle maximum-curvature criterion [48].
The primary objective of this study is to establish the mathe-
matical formulation of MSPA and evaluate its effectiveness as a
physically grounded alternative toPCA,PLS-DA,UMAP, and
SOMacross both classical machine learning and TFMs. To
evaluate the operator across a rigorous gradient of biological
complexity, radiometric scale, and class cardinality, we bench-
mark MSPA against full spectra (VNIR–SWIR, 350 –2500 nm)
and the four reference reducers across 8, 963 spectral signatures
spanning five plant spectroscopy datasets ordered from binary
foliar stress (C =2: citrus vascular infection [ 29], cotton her-
bivory [33], and soybean herbivory [35]) and ternary foliar stress
guilds (C =3: maize chewing vs. sucking injury [ 34]) to high-
cardinality canopy-scale species discrimination (C =39 across
five functional guilds in Amazonian pastures). Through 18, 000
cross-validated model evaluations, internal module ablations,
3D weight-space optimization landscapes, physical perturbation
saliency, and Pareto speed–accuracy analyses, we demonstrate
that MSPA simultaneously enhances discriminative accuracy, im-
proves gradient-descent conditioning, reduces predictive uncer-
tainty, and accelerates foundation model inference by orders of
magnitude.
2 ProposedMethodology: Morpho-Spectral
ProjectionAnalysis
2.1 Problem Formulation and Piecewise Differential State
Space
Consider a training set composed of N one-dimensional spectral
signatures represented by the matrixX=[x 1, x2,..., xN]⊤∈
RN×D, paired with a vector of discrete class labelsy=
[y1,y 2,...,y N]⊤ ∈ {1, 2,...,C} N. Unlike generic tabular
datasets whose columns represent independent variables with-
out physical ordering, each column j∈ {1,...,D} ofXcor-
responds to a strictly increasing wavelength coordinate along
λ =[ λ1,λ 2,...,λ D]⊤, with λ1 < λ2 <···< λ D, exhibiting
strong band-to-band correlation and local smoothness governed
by radiative transfer physics [1, 2, 11].
In conventional dimensionality reduction methods, such as PCA
[23, 24], PLS-DA [25, 49], UMAP [26], and SOM [27, 28], the
projection mappingR : RD→R K relies exclusively on global
column inner products (X⊤XorX ⊤Y) or pairwise Euclidean dis-
tances between samples (∥xi− xl∥2) [50, 30]. Consequently, for
any orthogonal permutation matrixΠ∈{ 0, 1}D×D that arbitrarily
reorders the wavelength channels, the projected representation
remains invariant,R(XΠ)= R(X), managing the spectrum as
an unordered listing of variables rather than a contiguous phys-
ical curve [15]. The objective ofMorpho-Spectral Projection
Analysis( MSPA) is to formulate a supervised projection oper-
atorTMSPA : RD→R K (K≪D ) that explicitly exploits the
sequential differential topology ofλ, jointly preserving localized
absorption feature geometry and global inter-class separability
[51, 11].
To ensure that the operator generalizes across both continu-
ous laboratory spectra and field where deep atmospheric water-
vapor absorption bands (typically near 1330 –1490 nm and
1770–2030 nm) are excised to prevent division-by-zero noise
spikes [15, 52, 9, 11], MSPA autonomously resolves contigu-
ous spectral segments from the adjacent wavelength intervals
∆λ j =λ j+1−λ j ( j =1,...,D− 1). Defining the median sensor
sampling step as δmed = median({∆λ j}D−1j=1 ), any index satisfy-
ing∆λ j >max (5δmed, 10 nm) demarcates a segment boundary,
partitioning the spectral domain{1,...,D} into B disjoint con-
tiguous blocksS1,S 2,...,S B. Within each contiguous blockSb
(b∈{ 1,...,B} ), the first derivative (X(1)≈∂ X/∂λ, local slope)
and second derivative (X (2) ≈∂ 2X/∂λ2, local concavity) are
estimated analytically via local least-squares Savitzky–Golay
polynomial convolution [17, 18, 53] of window lengthWSG =11
and polynomial orderp SG =2:
x(m)i,j = m!(δmed)m
hRX
u=−hL
c(m,pSG,WSG)u xi(λ j+u),
∀j∈S b,m∈{1,2},
(1)
where c(m,pSG,WSG)u denote the analytical derivative convolution
coefficients, and the half-window bounds hL,h R (hL + hR +1=
min(WSG,|S b|)) adapt at the boundaries ofSb to prevent deriva-
tive filters from crossing non-contiguous atmospheric gaps
[9, 11]. Unlike unregularized finite differences (∆ (1)x j =
x j+1−x j and∆ (2)x j = x j+1− 2x j + x j−1), which amplify un-
correlated sensor noise variance by factors of 2 and 6 (standard
deviation multipliers of
√
2≈ 1.414 and
√
6≈ 2.449) [18, 53],
the Savitzky–Golay filter (WSG =11,p SG =2) yields noise prop-
agation norms
pP c2u of 0.0953 (m =1) and 0 .0683 (m =2),
suppressing high-frequency noise amplification by 14.83× in
the first derivative and 35 .87× in the second derivative [ 17].
The coupled representation of zero-order reflectance amplitude
(X(0)≡ X), first-order slope (X(1)), and second-order concavity
(X(2)) defines the piecewise differential state space upon which
the three modules ofMSPAoperate [51, 54].
2.2 Module 1: Additive Continuum-Removed Absorption
Geometry and Inflection Kinematics
In vegetation spectroradiometry, biochemical absorption fea-
tures manifest as local concave valleys bounded by higher-
reflectance shoulders [ 1, 4, 55]. In the classical formulation
of Clark and Roush [4], continuum removal is expressed by the
multiplicative ratio 1−ρ (λ)/c(λ) [56, 11]. However, dividing
by the continuum line c(λ) can amplify baseline noise in low-
albedo spectral regions or introduce numerical instabilities when
combined with baseline-centered pre-treatments. To ensure
numerical stability and scale consistency directly on Savitzky–
Golay smoothed reflectance spectra (ρ(λ)∈ [0, 1]), Module 1
of MSPA formulates continuum removal as an additive residual
below the local continuum chord [11] (Fig. 1A,B).

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy4
LetVv =[ λv,start,λ v,end] denote the v-th spectral absorption
interval (v∈{ 1,...,V} ) comprising Lv contiguous channels. For
each sample i, the linear continuum chord ci,v(λ j) connecting the
initial and terminal shoulders of the window is parameterized by
the normalized coordinate t j =(λ j−λ v,start)/(λv,end−λ v,start)∈
[0,1] as [4, 55]:
ci,v(λ j)=x i(λv,start)·(1−t j)+x i(λv,end)·t j,∀λ j∈V v.(2)
The residual band depression profile below the continuum is
defined directly as di,v(λ j)= ci,v(λ j)−x i(λ j), which remains
positive within each concave absorption valley (Fig. 1B) and
removes local baseline tilts within each window Vv without
introducing inter-regional variance coupling [57, 58, 9]. Follow-
ing continuum-removal parameterization conventions [ 4, 55],
two functionals are extracted from di,v(λ j) per valley: the max-
imum contrast depth ( Dmax,i,v) and the normalized integrated
depression area (Aabs,i,v):
Dmax,i,v =maxλ j∈Vv di,v(λ j),A abs,i,v = 1Lv Xλ j∈Vv di,v(λ j).(3)
Across the optical vegetation domain (350–2500 nm), these two
functionals are computed over V =5 major biochemical ab-
sorption valleys established in leaf radiative transfer models
(PROSPECT/PROSAIL) [1, 59, 2, 60] (10 variables; Fig. 1A,D):
the chlorophyll electronic absorption well (V1 : 560–740 nm),
the first and second liquid water overtone and mesophyll fea-
tures in the NIR (V2 : 910–1060 nm andV3 : 1130–1280 nm),
the lignin–cellulose–water vibrational complex in SWIR 1
(V4 : 1520–1760 nm), and the protein, nitrogen, and cellulose
combination bands in SWIR 2 (V5 : 2050–2360 nm) [1, 61, 60].
Additionally, across thered-edgetransition window ( Winf =
[680, 760] nm), where chlorophyll degradation and physiologi-
cal stress induce diagnostic shifts in the derivative inflection
profile [ 62, 63, 64], Module 1 extracts four inflection kine-
matic parameters fromx (1)i andx (2)i (Fig. 1C): the maximum
first-derivative slope S max,i = maxλ j∈Winf x(1)i,j , the mean slope
¯S i = 1|Winf| Pλ j∈Winf x(1)i,j , the second-derivative concavity ampli-
tude Camp,i = maxλ j∈Winf x(2)i,j−minλ j∈Winf x(2)i,j , and the continuous
center-of-mass inflection wavelength (λRE,i) [62, 64]:
λRE,i =
P
λ j∈Winf max

x(1)i,j,0 ·λ j
P
λ j∈Winf max

x(1)i,j,0  +ϵ ,(4)
withϵ =10−8, where any additive baseline constant vanishes in
the first derivative and any multiplicative scaling factor cancels
out between the numerator and denominator [62, 18]. Module 1
is completed by two direct radiometric contrast indices between
scattering plateaus and the red chlorophyll well [65, 5]: the NIR–
Red plateau contrast ( ¯xi,[780,900]−¯xi,[650,685]) and the Green–Red
peak contrast ( ¯xi,[530,570]−¯xi,[650,685]). The resulting 16 variables
are standardized across samples using the training-set mean and
standard deviation, forming the biophysical blockz i,bio∈R 16
(Fig. 1D).
2.3 Module 2: Local Multi-Order Sub-Window Eigenmodes
As demonstrated in hyperspectral feature extraction studies
[19, 66, 15], applying a global rotational transform such as
PCA directly across the full spectrum causes broad regions of
high absolute amplitude variance (predominantly leaf albedo
and structural scattering across the NIR plateau) to dominate
the leading eigenvectors, masking fine-scale, lower-amplitude
diagnostic variations within narrow spectral intervals [15, 29].
Conversely, interval-based chemometrics [67], multiscale band
partitioning [68, 11], and magnitude-and-shape unmixing frame-
works [54, 51] have shown that partitioning the spectrum into
localized intervals and integrating zero-order magnitude with
derivative shape features prevents cross-regional variance sup-
pression.
Building on these principles, Module 2 partitions the spec-
trum into W =8 contiguous radiative sub-windows
(W1,...,W 8; Fig. 2A) aligned with the major bio-
physical absorption and scattering domains of vegetation
[1, 59, 2, 60]: Blue/Carotenoids (400 –510 nm), Green
Peak/Anthocyanins (511 –620 nm), Chlorophyll Red Well
(621–685 nm), Red-Edge Ramp (686–765 nm), Mesophyll NIR
Plateau (766–910 nm), NIR Water Absorption (911–1350 nm),
SWIR 1 (1450–1790 nm), and SWIR 2 (1980–2450 nm). If the
user-supplied spectrum covers only a subset of this domain (for
instance, sensors restricted to the VNIR range of 400–1000 nm),
the algorithm activates only the sub-windows contained within
the sensor range (requiring at least six bands per window).
Within each active sub-windowWw (w∈{ 1,...,W} ) and for
each differential order m∈{ 0, 1, 2} separately, the training sub-
matrix is standardized across samples to zero mean and unit
variance per wavelength ( ˜X(m)train,w∈R Ntrain×|Ww|) and factored via
Singular Value Decomposition (SVD; Fig. 2B) [23, 30]:
˜X(m)train,w =U (m)w Σ(m)w V(m)w ⊤.(5)
By extracting the leading right singular vectorv (m)w,1∈R|Ww| (as-
sociated with the largest singular valueσ(m)w,1, with deterministic
sign alignment), the local morphological coordinate of sample i
for window w and order m is obtained by the linear projection
of its standardized window vector ˜x(m)i,w∈R|Ww|:
z(w,m)i,win = ˜x(m)i,w⊤ v(m)w,1,w∈{1,...,W},m∈{0,1,2}.(6)
By extracting one dominant mode of amplitude (m =0), one of
slope (m =1), and one of concavity ( m =2) per sub-window
[54, 51] (Fig. 2C), Module 2 generates 8 × 3=24 local vari-
ables (zi,win∈R 24) whose loading matrix forms a strictly block-
diagonal structure with zero cross-window variance leakage
(Fig. 2D).
2.4 Module 3: Global Fisher-Weighted Canonical Subspace
and Intrinsic Breaking-Point
Whereas Modules 1 and 2 preserve the local structure of each
spectral sub-region in an unsupervised manner, Module 3 ex-
tracts supervised global directions that maximize separabil-
ity across the C classes over the full differential state space
Mi =[x (0)⊤i , x(1)⊤i , x(2)⊤i ]⊤ ∈R 3D [51, 11]. In hyperspectralmixture and discriminant analysis, prioritizing wavelengths with
high inter-class separation relative to intra-class variability has
been shown to substantially improve numerical conditioning
and class discrimination [69, 51, 21, 11]. After standardizing
each column q∈{ 1,..., 3D} across the training samples ( ˜Mtrain),

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy5
500 750 1000 1250 1500 1750 2000 2250 2500
Wavelength λ (nm)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7Reflectance Factor ρ(λ) [0–1]
V1: Chl
V2: NIR-W1
V3: NIR-W2
V4: SWIR1
V5: SWIR2
A. Continuous Leaf Reflectance ρ(λ) ∈ [0, 1] & The 5 Absorption Valleys (V1–V5)
Healthy Leaf Spectrum ρ(λ)
Chlorotic Stress Spectrum
Structural SWIR Stress Spectrum
Linear Continuum Chord c_{i,v}(λ)
Absorption Depression d_{i,v}(λ)
575 600 625 650 675 700 725
Wavelength λ (nm) — Chlorophyll Absorption Valley (V1)
0.1
0.2
0.3
0.4
0.5
0.6Reflectance Factor ρ(λ) [0–1]
Max Contrast Depth
D_{max} = max d(λ)
= 0.294
B. Additive Continuum-Removed Geometry (Eqs. 2–4: D_max and A_abs)
Healthy Leaf Spectrum ρ_i(λ)
Stressed Leaf Spectrum (Reduced Depth)
Continuum Chord c_{i,v}(λ) = ρ(λ_start)(1-t) + ρ(λ_end)t
Integrated Area A_{abs,i,v} = (1/L_v) Σ d_{i,v}(λ)
680 690 700 710 720 730 740 750 760
Wavelength λ (nm) — Red-Edge Window W_inf [680, 760 nm]
Order 2
(Concavity)
Zero
Baseline
Peak Slope
(S_max)
Normalized Derivative Kinematics
Max Slope S_max
& Centroid Δλ_RE
C. Red-Edge Inflection Kinematics (Eq. 5: S_max, S_mean, C_amp & λ_RE)
1st Deriv. x^(1)(λ) Healthy (λ_RE = 716.3 nm)
1st Deriv. x^(1)(λ) Chlorotic (λ_RE = 714.8 nm, Blue-Shift)
2nd Deriv. x^(2)(λ) Healthy Concavity Wave (C_amp)
Dmax_ChlArea_ChlDmax_W1Area_W1Dmax_W2Area_W2Dmax_SW1Area_SW1Dmax_SW2Area_SW2RE_MaxSlpRE_MeanSlpRE_λ_CentrRE_ConcavNIR-RedGrn-Red
16 Extracted Biophysical & Inflection Descriptors
−1.0
−0.5
0.0
0.5
1.0
1.5
Standardized Feature Value (Z-Score)
10 Valley Descriptors
(D_max & A_abs across V1–V5)
4 Red-Edge
Kinematics
2 Plateau
Contrasts
D. Complete Module 1 Output Vector z_{i,bio} ∈ R^16 Across Physiological States
Healthy Control Chlorotic Stress (Pigment / RE) Structural Stress (SWIR / Water)
Figure 1:Physical and mathematical mechanics of Module 1 inMSPA(z i,bio∈R 16): (A) continuous foliar reflectance signatures (ρ(λ)∈[0,1],
400–2500 nm) illustrating the linear continuum chords ci,v(λ) across the five biochemical absorption valleys (V1 toV5) and the plateau contrast
anchors; (B) geometric extraction of the maximum contrast depth (Dmax,i,v) and normalized integrated depression area (Aabs,i,v) via the additive
residuald i,v(λ)=c i,v(λ)−x i(λ) ((2)–(3)); (C) first- and second-derivative inflection kinematics across the red-edge transition window
(Winf =[680,760] nm), highlighting the maximum slope (S max), concavity wave (Camp), and center-of-mass inflection wavelength (λRE, (4));
and (D) standardized output profile of the 16 Module 1 descriptors across distinct physiological states.
Module 3 computes a continuous Fisher-type discriminability
weight ωq from the square-root-dampened ratio between the
inter-class standard deviation of the class centroids (µc,q) and
the mean intra-class standard deviation (σc,q) [51, 21]:
ωq = ˜ωq1
3D
P3D
l=1 ˜ωl
,˜ω q =
s
Stdc∈{1,...,C}(µc,q)
1
C
PC
c=1σc,q +δ
,(7)
whereδ =10−4 ensures numerical stability. The feature weight-
ingH= ˜Mdiag (ω1,...,ω 3D) amplifies wavelengths and deriva-
tive orders where signatures diverge across classes while attenu-
ating invariant or noisy bands [21, 11] (Fig. 3A).
To determine the number of canonical components Kcan to ex-
tract in Module 3, the operator provides two operating modes
through the n_components parameter: (i) in user-defined fixed
mode ( n_components = K, with default K =64), Kcan =
max(4,K− 40) is set, yielding Kcan =24 canonical variables
which, combined with the 16 variables from Module 1 and the
24 from Module 2, total K =64 variables (or, if K< 44, all gen-
erated variables are ranked in descending order by their training-
set Fisher score from (7), retaining the top K features); and
(ii) in autonomous mode ( n_components=’auto’),H train is
decomposed via SVD up to rank Pmax = min(Ntrain− 1, 3D) and
the Kneedle maximum-curvature algorithm [48] is applied to the
cumulative singular energy curve yk =
Pk
r=1σ2r

/
PPmax
r=1 σ2r

to estimate the intrinsic inflection breaking-point K∗can [48, 70]
(Fig. 3B):
K∗can =arg max1≤k≤Pmax
yk−y 1yPmax−y 1− k−1Pmax−1 .(8)
Classical Fisher Discriminant Analysis (FDA) requires invert-
ing the within-class covariance matrix, which becomes ill-
conditioned or singular under high spectral dimensionality
and limited sample sizes [ 71, 22, 30, 11]. To guarantee well-
conditioned inversion, Module 3 projectsHonto a preliminary
P-dimensional orthogonal SVD subspace (U=(H − ¯H)VP∈
RNtrain×P) and regularizes the within-class covariance matrix
SW via Ledoit–Wolf-type trace-scaled covariance shrinkage
[71, 47]:
Σ(γ)W =(1−γ)S W +γ Tr(SW)P IP,Σ Bw= ΛΣ (γ)W w,(9)
whereΣ B denotes the empirical between-class covariance ma-
trix,Λis the generalized eigenvalue, and γ =0.35 is the shrink-
age coefficient that bounds the condition number κ(Σ(γ)W ) [47]

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy6
500 750 1000 1250 1500 1750 2000 2250 2500
Wavelength λ (nm)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7Reflectance Factor ρ(λ) [0–1]
W1 W2 W3W4 W5 W6 W7 W8
A. Partitioning of the Contiguous Spectrum into W = 8 Radiative Sub-Windows
Mean Leaf Reflectance ρ(λ) ∈ [0, 1]
Global Amplitude Variance (±2.5 SD — Dominated by NIR)
1450 1500 1550 1600 1650 1700 1750
Wavelength λ (nm) — SWIR 1 Sub-Window [1450, 1790 nm]
Order 2
v_{w,1}^(2)
Order 1
v_{w,1}^(1)
Order 0
v_{w,1}^(0)Normalized Singular Basis Vector
B. Local Multi-Order SVD Basis Extraction Inside Window W7 (Eqs. 6–7)
Order 0 Singular Basis v_{w,1}^(0) (Amplitude — 98.8% Var)
Order 1 Singular Basis v_{w,1}^(1) (Slope ∂ρ/∂λ — 17.4% Var)
Order 2 Singular Basis v_{w,1}^(2) (Concavity ∂²ρ/∂λ² — 4.6% Var)
Lignin / Cellulose Absorption Band (1685 nm)
W1: Blue W2: Green W3: RedW4: Red-EdgeW5: NIR-MesoW6: NIR-WaterW7: SWIR 1W8: SWIR 2
Contiguous Radiative Sub-Windows (W1 to W8)
0
20
40
60
80
100
120Within-Window Explained Energy (%)
C. Within-Window Variance Captured by Rank-1 SVD Across W1 to W8
Order 0 (Amplitude Mode v_{w,1}^(0))
Order 1 (1st Derivative Slope Mode v_{w,1}^(1))
Order 2 (2nd Derivative Concavity Mode v_{w,1}^(2))
500 750 1000 1250 1500 1750 2000 2250 2500
Input Wavelength λ (nm)
W1 (Ord 0,1,2)
W2 (Ord 0,1,2)
W3 (Ord 0,1,2)
W4 (Ord 0,1,2)
W5 (Ord 0,1,2)
W6 (Ord 0,1,2)
W7 (Ord 0,1,2)
W8 (Ord 0,1,2)
24 Module 2 Variables (z_{i,win})
D. Block-Diagonal Sparse Loading Matrix of Module 2 (Zero Cross-Window Leakage)
0.0
0.2
0.4
0.6
0.8
1.0
Normalized Local SVD Loading |v_{w,1}^(m)(λ)|
Figure 2:Physical and mathematical mechanics of Module 2 inMSPA(z i,win∈R 24): (A) partitioning of the continuous reflectance spectrum
(ρ(λ)∈[0,1], 400–2500 nm) intoW=8 contiguous radiative sub-windows (W 1 toW 8) to prevent global NIR amplitude variance from
masking narrow spectral features; (B) independent rank-1 SVD basis extraction within windowW7 (SWIR 1) across Order 0 (v(0)w,1, amplitude),
Order 1 (v(1)w,1, local slope∂ρ/∂λ), and Order 2 (v (2)w,1, local concavity∂ 2ρ/∂λ2) ((5)–(6)); (C) within-window singular energy captured by theleading eigenmodes across all eight sub-windows; and (D) the resulting 24×Dblock-diagonal sparse loading matrix demonstrating strict
regional isolation with zero cross-window leakage.
(Fig. 3C). Furthermore, because the algebraic rank ofΣ B in stan-
dard FDA and PLS-DA is strictly bounded by rank(ΣB)≤C− 1
[72, 49, 30], conventional discriminant projections collapse
to a single vector ( C− 1=1) in binary classification tasks
(C =2). Extending orthogonal Fisher subspace completion
strategies proposed in the hyperspectral literature [73, 74, 11],
Module 3 adapts automatically to the class cardinality C: when
C− 1≥K can (high-cardinality multiclass tasks), the Kcan domi-
nant generalized eigenvectors of

Σ(γ)W −1 ΣB are retained; when
C− 1 <K can (binary C =2 or low-cardinality tasks), regular-
ized discriminant axes are extracted separately fromX (0),X (1),
X(2), andH(4 ×min (C− 1, 4) supervised axes), and the re-
maining axes up to Kcan are completed using the right singular
vectors ofHscaled by their root singular-energy ratio (
q
σ2k/σ21)
[73, 74]. This formulation ensures that the complementary or-
thogonal components preserve signal hierarchy without inflating
residual noise in binary problems, forming the canonical block
zi,can∈R Kcan (Fig. 3D) and completing the concatenated projec-
tionz i =[z⊤i,bio,z⊤i,win,z⊤i,can]⊤∈R K∗.
Analogous to structure correlation loadings in classical chemo-
metrics [75, 23], the structural sensitivity between each extracted
morpho-spectral variable zk (k∈{ 1,...,K ∗}) and each physical
input wavelengthλ j ( j∈{ 1,...,D} ) across the three differential
orders (m∈{ 0, 1, 2}) is formalized in MSPA by theMulti-Order
Structure LoadingmatrixΛ (m)∈R D×K∗:
Λ(m)j,k =Corr X(m)train,:,j,Z train,:,k = 1Ntrain−1  ˜X(m)train⊤ ˜Ztrain,
|Λcompj,k |=
vut
1
3
2X
m=0

Λ(m)j,k 2.
(10)
Whereas unsupervised PCA loadings (VK) are dense and non-
zero across allDchannels simultaneously [23, 15], the loading
footprint of MSPA (ΛMSPA; Fig. 4) is organized into three inter-
pretable regimes matching the three modules in Fig. 4D: (i) in
Module 1 (variables 1–16, MSPA_BIO), loadings map directly
onto the active wavelength support of each continuum-removed
absorption valley (V1 toV 5) and the red-edge inflection win-
dow [4, 62]; (ii) in Module 2 (variables 17–40, MSPA_WIN), the

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy7
500 750 1000 1250 1500 1750 2000 2250 2500
Wavelength λ (nm)
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
4.5Normalized Fisher Weight ω_q
A. Multi-Order Fisher Discriminability Weights ω_q Across 400–2500 nm (Eq. 8)
Order 0 Fisher Weight ω^(0)(λ) (Amplitude)
Order 1 Fisher Weight ω^(1)(λ) (1st Deriv. Slope)
Order 2 Fisher Weight ω^(2)(λ) (2nd Deriv. Concavity)
Unweighted Uniform Baseline (ω_q = 1.0)
10 20 30 40 50
Candidate Canonical Eigenmode Index k
0.0
0.2
0.4
0.6
0.8
1.0
1.2Normalized Energy / Deflection Distance
B. Autonomous Intrinsic Breaking-Point Estimation via Kneedle (Eq. 9)
Cumulative Fisher Singular Energy y_k (Normalized)
Secant Chord Joining Endpoints (k=1 to P_max)
Perpendicular Distance Curve D(k) = y_norm(k) - x_norm(k)
Detected Breaking-Point K_can* = 4
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Analytical Shrinkage Parameter γ ∈ (0, 0.8]
101
102
Condition Number κ(Σ_W^(γ)) [Log Scale]
C. Ledoit–Wolf Covariance Shrinkage & Rank-Safe Completion (Eq. 10)
Within-Class Covariance Condition Number κ(Σ_W^(γ))
Analytical Shrinkage γ = 0.35 (κ = 37.1)
−3 −2 −1 0 1 2 3
MSPA Canonical Variable 1 (Z-Score)
−3
−2
−1
0
1
2
3
MSPA Canonical Variable 2 (Z-Score)
D. Resulting Canonical Subspace (z_{i,can}) vs. Unsupervised PCA
Unsupervised PCA (PC1 vs PC2)
MSPA — Healthy Control
MSPA — Chlorotic Stress
MSPA — Structural SWIR Stress
Figure 3:Physical and mathematical mechanics of Module 3 in MSPA (zi,can∈R Kcan): (A) multi-order Fisher discriminability weights (ωq, (7))
across the continuous state space [X(0),X (1),X (2)] (400–2500 nm), amplifying localized derivative peaks while attenuating invariant scattering
plateaus; (B) autonomous estimation of the intrinsic canonical breaking-point (K∗can) via the maximum perpendicular deflection distance of
Kneedle (8); (C) exponential stabilization of the within-class covariance condition numberκ(Σ(γ)W ) via Ledoit–Wolf-type trace-scaled covariance
shrinkage (γ=0.35, (9)) and rank-safe orthogonal completion across class cardinalities; and (D) class separability in the resulting Module 3
canonical subspace compared to unsupervised PCA (inset).
local singular vectorsv (m)w,1 form a strictly block-diagonal sparse
staircase (Λ j,k≡ 0,∀λ j <W w) [67, 68]; and (iii) in Module
3 (variables 41 –64, MSPA_CAN), Fisher weighting suppresses
loadings across invariant plateaus (Λ(m)j,k ≈ 0) and concentrates
selective first- and second-derivative loading peaks at wave-
lengths of physiological divergence [51, 21].
2.5 Algorithmic Workflow, Computational Complexity, and
Software Architecture
To translate the mathematical formulation of (1)–(10) into a
leak-free computational operator [76], MSPA is implemented in
Python adhering strictly to the scikit-learn transformer API
(BaseEstimator and TransformerMixin), decoupling pa-
rameter estimation on the training set (fit) from out-of-sample
projection (transform). During the training phase ( fit), the
operator receives the training spectral matrixX train ∈R Ntr×D,
the discrete class labelsy train ∈{ 1,...,C} Ntr, the wavelength
vectorλ∈R D, the target dimensionalityK∈Z +∪{’auto’}, the
Savitzky–Golay parameters (WSG =11,p SG =2), and the ana-
lytical shrinkage coefficient (γ =0.35). In the first stage (domain
and segment resolution), if a pandas.DataFrame is provided
without an explicit wavelength vector, regular expressions au-
tomatically parse the numerical wavelength coordinates from
the column headers; otherwise, the supplied vectorλ defines the
spectral domain. The algorithm then identifies excision jumps
satisfying∆λ j >max (5δmed, 10 nm) to partition the index set
{1,...,D} into contiguous segments{Sb}Bb=1, activating the eight
foliar radiative sub-windows{Ww}Ww=1 and five absorption val-
leys{Vv}Vv=1 within the 350–2500 nm vegetation domain. In thesecond stage (piecewise differential state-space construction),
the analytical first and second Savitzky–Golay derivativesX (1)train
andX (2)train are computed independently within each contiguoussegmentS b via (1).
In the third stage, the three internal modules are parameter-
ized sequentially using exclusively training-set statistics. First,
Module 1 extracts the additive continuum-removed maximum
depths Dmax, integrated absorption areas Aabs, red-edge inflec-
tion kinematics (λRE,S max, ¯S,C amp), and inter-plateau contrasts
via (2)–(4), fitting the internal standardization scaler Zbio to
produce the biophysical blockZ tr,bio∈R Ntr×16. Next, Module
2 iterates over each active sub-window w∈{ 1,...,W} and dif-
ferential order m∈{ 0, 1, 2}, fitting the local column scalerZ(m)w ,

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy8
500 750 1000 1250 1500 1750 2000 2250 2500
Input Wavelength λ (nm)
−1.0
−0.5
0.0
0.5
1.0
Structure Loading Corr(X(λ), PC_k)
A. Unsupervised PCA Structure Loadings (Diffuse & Albedo-Dominated)
PC1 Loading (88.2% Var — Global Albedo)
PC2 Loading (9.1% Var — Broad Spectral Tilt)
PC3 Loading (1.8% Var — Mixed Plateau Mode)
500 750 1000 1250 1500 1750 2000 2250 2500
Input Wavelength λ (nm)
−1.0
−0.5
0.0
0.5
1.0
Multi-Order Structure Loading Λ_{λ, k}^(m)
B. MSPA Module 3 Multi-Order Structure Loadings (Eq. 11)
Canonical Axis 1 — Order 0 Loading Λ^(0)(λ) (Amplitude)
Canonical Axis 1 — Order 1 Loading Λ^(1)(λ) (Slope ∂ρ/∂λ)
Canonical Axis 2 — Order 2 Loading Λ^(2)(λ) (Concavity ∂²ρ/∂λ²)
500 750 1000 1250 1500 1750 2000 2250 2500
Input Wavelength λ (nm) Partitioned Across Contiguous Sub-Windows W1 to W8
Order 2
v_{w,1}^(2)
Order 1
v_{w,1}^(1)
Order 0
v_{w,1}^(0)Normalized Window SVD Loadings
W1 W2 W3W4 W5 W6 W7 W8
C. MSPA Module 2 Block-Sparse Local SVD Loadings Across Windows W1–W8
Order 0 Basis (Amplitude) Order 1 Basis (Slope) Order 2 Basis (Concavity)
500 750 1000 1250 1500 1750 2000 2250 2500
Input Wavelength λ (nm)
1
Mod 1 (1–16)
16
Mod 2 (17–40)
40
Mod 3 (41–64)
64
Extracted MSPA Variable Index (k = 1 to 64)
D. Complete 64 × 2,101 Input-to-Output Spectral Loading Matrix (|Λ_MSPA|)
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Normalized Spectral Loading Magnitude |Λ_{λ, k}|
Figure 4:Comparison between unsupervised PCA loadings and multi-order spectral loadings in MSPA: (A) diffuse, non-zero structure loadings
of PC1–PC3 across all bands dominated by NIR plateau albedo; (B) localized multi-order structure loadings (Λ(m)λ,k , (10)) of Module 3 canonicalaxes (MSPA_CAN_01,MSPA_CAN_02) peaking selectively at the red-edge inflection (712 nm) and SWIR lignin/cellulose concavities
(1685 nm,2180 nm); (C) exact block-sparse SVD loading vectors (v (0)w,1,v (1)w,1,v (2)w,1) of Module 2 across windowsW 1 toW 8 with zerocross-window variance leakage; and (D) complete 64×Dinput-to-subspace loading footprint matrix (|Λ MSPA|) across Modules 1 (variables
1–16), 2 (variables 17–40), and 3 (variables 41–64).
extracting the leading right singular vectorv (m)w,1 via SVD (5),
projecting the local morphological coordinates z(w,m)win (6), and
fitting the block scalerZwin to obtainZ tr,win∈R Ntr×24. Subse-
quently, Module 3 concatenates the full differential state space
Mtr =[X (0)train, X(1)train, X(2)train], fits the global column scalerZmulti,
and computes the Fisher discriminability weightsω∈R 3D (7) to
form the weighted matrixH tr = ˜Mtrdiag(ω). When configured
in autonomous mode (K = ’auto’), the intrinsic canonical dimen-
sion K∗can is estimated from the singular value spectrum ofH tr
via the Kneedle maximum-curvature criterion (8); in fixed mode,
the canonical dimension is set to Kcan = max(4,K− 40). The
regularized Ledoit–Wolf discriminant projector (γ =0.35, (9)) is
then fitted on the preliminary orthogonal basis ofH tr incorporat-
ing rank-safe orthogonal completion scaled by
q
σ2k/σ21 when-
ever C− 1<K can to yield the canonical blockZ tr,can∈R Ntr×Kcan.
Finally, the complete training subspace is assembled by concate-
nating the three blocks intoZ tr∈R Ntr×K∗ (if a user-specified
target dimension satisfies K< 44, all extracted features are
ranked in descending order by their training-set Fisher score (7)
and truncated to the topKcolumns).
During out-of-sample projection ( transform), the fitted op-
erator receives unseen evaluation spectraX test ∈R Nte×D
without class labels, computes the piecewise derivatives
X(1)test andX (2)test across{Sb}Bb=1, and applies the stored scalers
(Zbio,{Z (m)w },Z win,Z multi), local singular vectors{v(m)w,1}, Fisher
weights ω, and canonical projection matrices without refit-
ting. Because all projection bases are fixed after training,
out-of-sample inference executes in strictly linear O(NteD)
time, returning either a compact float32 array or a labeled
DataFrame (via set_output(transform="pandas")) with
traceable semantic variable names (MSPA_BIO_, MSPA_WIN_,
andMSPA_CAN_*).
3 ExperimentalDesign andBenchmarking
Protocol
3.1 Spectroradiometric Datasets, Radiometric Stabilization,
and Validation Partitioning
To validate the MSPA operator across an increasing gradient
of physiological complexity, radiometric scale, and taxonomic
cardinality (C∈ {2, 3, 39}), we compiled 8 , 963 spectral sig-

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy9
natures from five plant spectroscopy experiments conducted
in Brazil (Table 1 and Fig. ??). Across all campaigns, data
were acquired via non-imaging proximal spectroradiometry us-
ing ASD FieldSpec sensors (Analytical Spectral Devices Inc.,
Boulder, CO, USA), operating over the 350 –2500 nm range
with a nominal spectral resolution of 1 .4 nm in the VNIR re-
gion (350–1000 nm, silicon photodiode array) and 2.0 nm in the
SWIR 1 (1000–1800 nm) and SWIR 2 (1800–2500 nm, cooled
InGaAs detectors) regions, internally resampled to uniform 1nm
steps (2, 151 raw channels) [ 33, 34, 35, 29]. Readings were
collected under direct solar illumination between 09:00 and
16:00 local time and converted into Hemispherical-Conical Re-
flectance Factors (HCRF), denoted by ρ(λ), through periodic
calibration against a Lambertian Spectralon reference panel
(Labsphere Inc.):
ρ(λ)= Ltarget(λ)Lreference(λ)×k(λ),(11)
where Ltarget(λ) and Lreference(λ) represent the radiances reflected
by the vegetation target and the calibration panel under equiv-
alent illumination and viewing geometry, and k(λ) is the panel
radiometric correction factor [3, 29].
The five datasets cover distinct physiological stressors and tax-
onomic scales. The first dataset Citrus, N =1 , 940, C =2;
Fig. ??A) [29] comprises foliar spectra (8◦ foreoptic) from sweet
orange trees (Citrus sinensisL. Osbeck), discriminating healthy
control plants ( HLB−) from plants infected withHuanglong-
bing(greening, HLB+, confirmed via qPCR at Ct≤ 36.0). The
second dataset Cotton, N =1, 605, C =2; Fig. ??B) [33] moni-
tored cotton leaves (Gossypium hirsutumL., 8◦ foreoptic) over
nine consecutive days and four diurnal periods, distinguishing
healthy control plants from insect-attacked plants subjected to
caterpillar herbivory (Spodoptera frugiperda). The third dataset
(Soybean, N =991, C =2; Fig. ??C) [35] evaluated soybean
leaves (Glycine maxL., 1◦ foreoptic) over eight days, discrimi-
nating healthy control plants from plants attacked by caterpillars
(S. frugiperda) and stink bugs (Dichelops melacanthus). The
fourth dataset (Maize, N =474, C =3; Fig. ??D) [34] ad-
dresses the same insect complex in maize leaves (Zea mays
L., 8◦ foreoptic), but separates the signatures into three classes:
healthy control plants, plants attacked by chewing caterpillars
(S. frugiperda), and plants attacked by piercing-sucking stink
bugs (D. melacanthus). The fifth dataset (Savanna Pastures,
N =3, 953, C =39; Fig. ??E) comprises canopy-scale spectra
(25◦ bare-fiber foreoptic) acquired along georeferenced transects
in Roraima, Brazil, encompassing 39 spectral classes of invasive
pasture weeds, nativeLavradosavanna species, and cultivated
forage grasses mixed within pastures under varying edaphic
and biotic conditions. All subspace projections, classification
models, and statistical tests were executed strictly on the 39
disaggregated classes, whereas for visual clarity in the figures
the 39 classes are organized into five functional groups: (i) culti-
vated forage grasses under healthy and stressed conditions; (ii)
native savanna and invasive grasses; (iii) broadleaf weeds and
legumes; (iv) twining vines and parasitic lianas; and (v) invasive
palm (Attalea maripa).
To eliminate confounding variables arising from differing band-
trimming conventions across the original studies, all 8 , 963
raw readings were processed through a unified preprocessing
pipeline. First, radiometrically implausible values ( ρ(λ) <
0.0 or ρ(λ) > 1.1) were treated as missing entries and re-
constructed via linear interpolation along each sample spec-
trum. Next, because passive solar measurements suffer near-
total irradiance loss within atmospheric water-vapor absorption
bands (causing near-zero denominators in (11) that would be
severely amplified by differential operators) [ 3, 9], the noisy
detector edges (350–400 nm and 2401–2500 nm) and the two
deep atmospheric water-vapor windows (1331 –1490 nm and
1771–2030 nm) were excised. This radiometric quality-control
step retained D =1 , 580 wavelengths distributed across three
contiguous optical blocks:S1 =[401 , 1330] nm (930 bands),
S2 =[1491, 1770] nm (280 bands), andS3 =[2031, 2400] nm
(370 bands). Within each contiguous block Sb (b∈{ 1, 2, 3})
independently, Savitzky–Golay polynomial smoothing [ 17]
(W =10, p =2, derivative order m =0) was applied to suppress
high-frequency instrumental noise while preserving absorption
band geometry, and the reflectance factors were scaled to [0, 1].
The resulting concatenated matrixX ∈R N×1580 defines the unre-
duced Full Spectrum baseline representation.
To ensure an evaluation free of data leakage [76], each dataset
was partitioned into 11 disjoint subsets (Fig. ??F), where the
eleventh subset (20% of the samples) was isolated as a fixed
external holdout test set and the remaining ten subsets (80%
of the samples) formed the 10 cross-validation folds (in each
fold f∈{ 0,..., 9}, nine subsets composeX (f)train and subset f
composesX (f)val). In Citrus, to prevent leaves from the same treefrom being shared between training and testing sets [29], parti-
tioning was enforced at the biological individual level (grouped
by orchard plot, row, and tree identifier) via stratified group
cross-validation. In Cotton, Soybean, and Maize, where the
same plant cohorts were measured across successive days and
diurnal times, the 11 partitions were generated via composite
stratified sampling over the interaction of treatment class, evalu-
ation day, and time of day, ensuring temporal and radiometric
symmetry across training, validation, and test sets. In Savanna
Pastures, accounting for spatial autocorrelation under Tobler’s
First Law of Geography, samples within each of the 39 classes
were ordered along the geographic acquisition transect index,
allocating the final 20% of each trajectory to the fixed holdout
test set (spatiotemporal extrapolation) and the initial 80% to the
10 contiguous sequential folds.
3.2 Reference Dimensionality Reduction Paradigms
Within each fold f∈ { 0,..., 9}, four classical dimensional-
ity reduction paradigms were fitted exclusively on the column-
standardized training matrix ˜X(f)train (whose mean and standard
deviation parameters were subsequently applied to ˜X(f)val and˜Xtest), benchmarking them against the unreduced Full Spectrum
(D =1 , 580) and the proposed MSPA operator (evaluated un-
der the default configuration K∗ =64, the autonomous mode
n_components=’auto’, and across the dimensionality scaling
grid K∈ [8, 96]). To avoid arbitrary dimensionality choices
across the four reference methods, the intrinsic number of com-
ponents (K∗) for each technique was estimated on the training
set and consolidated via the cross-fold median (Table 1).
The first reference method isPrincipal Component Analysis
(PCA) [23, 24], which projects ˜X onto the K∗PCA dominant or-thogonal eigenvectors of the unsupervised sample covariance

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy10
Table 1:Characterization of the five plant spectroradiometric datasets ordered by taxonomic and physiological complexity
(C=2→C=3→C=39), validation partitioning (80% cross-validation development pool vs. 20% fixed holdout test set), and consensus
dimensionality (K∗) across the six evaluated representations.
Dataset Plant System Spectroscopic Task&Physiological Condition Scale&Foreoptic Total (N) Classes (C) CV Pool/Test Consensus Dimensionality (K∗)
Full Spec. PCA PLS-DA UMAP SOM MSPA
Citrus[29] Sweet orange (Citrus sinensis) qPCR HLB + vs. HLB− (C.L. asiaticus) Leaf (8 ◦, Plant-grouped) 1,940 2 1,552/388 1,580 10 8 6 10064
Cotton[33] Cotton (Gossypium hirsutum) Control vs.S. frugiperdalarvae (9 days) Leaf (8 ◦, Diurnal) 1,605 2 1,284/321 1,580 50 19 7 10064
Soybean[35] Soybean (Glycine max) Control vs. Multi-pest herbivory (8 days) Leaf (1 ◦, Multi-day) 991 2 793/198 1,580 50 21 6 6464
Maize[34] Maize (Zea mays) Control vs. Chewing (S.f.) vs. Sucking (D.m.) Leaf (8 ◦, 3 Guilds) 474 3 379/95 1,580 50 21 7 6464
Savanna PasturesTropical pastures & pasture weeds 39 Taxa (5 Functional Guilds)+Edaphic/Biotic stress Canopy (25 ◦, Transects) 3,953 39 3,162/791 1,580 50 21 8 10064
Total/Range 5 Plant Systems Vascular Disease, Herbivory Guilds&Canopy Taxa Leaf (1 ◦–8◦) to Canopy (25◦) 8,963 2–39 7,170/1,793 1,580 10–50 8–21 6–8 64–100 64
matrixΣ train = 1Ntrain−1( ˜X(f)train)⊤ ˜X(f)train. The dimensionality K∗PCA
was determined by combining the Kneedle maximum-curvature
criterion [48] (8) with a minimum threshold of 99.5% cumula-
tive explained spectral variance, yielding K∗PCA =10 for Citrusand K∗PCA =50 for Cotton, Soybean, Maize, and Savanna Pas-tures. The second method isPartial Least Squares Discriminant
Analysis(PLS-DA) [ 25, 49], a supervised technique that encodes
the training labelsy (f)train into a centered one-hot indicator matrix
Y∈R Ntrain×C and iteratively extracts pairs of latent variables
(tk = ˜Xkwk andu k =Y kqk) that maximize the cross-covariance
Cov( ˜Xkwk, Ykqk), followed by orthogonal deflation. The cut-
offdimension K∗PLS was determined by applying the Kneedlealgorithm to the cumulative explained taxonomic variance curve
R2Y(k)=1 −∥ Y− ˆY(k)∥2F/∥Y∥2F, establishing K∗PLS =8 for Citrus,19 for Cotton, and 21 for Soybean, Maize, and Savanna Pastures.
The third reference method isUniform Manifold Approximation
and Projection(UMAP) [ 26], which models the nonlinear spec-
tral manifold by constructing a fuzzy topological graph over
the knn =15 nearest neighbors (with a minimum embedding
distance of 0.1 under the Euclidean metric) and optimizing a
low-dimensional embedding via fuzzy cross-entropy minimiza-
tion. The intrinsic manifold dimension (K∗UMAP) was determinedusing the Levina–Bickel maximum likelihood estimator [ 77]
evaluated across neighborhood scales q∈ [10, 20] from the
Euclidean distances Tq(˜xi) to the q-th nearest neighbor in ˜X(f)train:
ˆmq =
 1Ntrain(q−1) NtrainXi=1 q−1Xj=1 ln Tq(˜xi)T j(˜xi)!−1,
K∗UMAP =round medianq∈[10,20]{ˆmq},
(12)
yielding K∗UMAP =6 for Citrus and Soybean, 7 for Cotton andMaize, and 8 for Savanna Pastures. The fourth method consists
of KohonenSelf-Organizing Maps(SOM) [ 27, 28], structured as
a two-dimensional competitive lattice ofG×G neurons (K∗SOM =
G2 =100 for datasets with N> 1, 600 and G2 =64 for smaller
datasets, selected via joint minimization of quantization and
topographic errors). During competitive training over E epochs
with exponentially decaying learning rateα(t) and neighborhood
radius σ(t), the winning neuron r∗ = arg minr∥˜xi− mr∥2 is
identified and the synaptic weight vectorsm r∈R D are updated
according to a Gaussian neighborhood rule on the 2D lattice
(gr∈R 2), projecting each sample into the reduced space via
its Euclidean distances to the converged prototypes, zi,r =∥˜xi−
mr∥2.
3.3 Classification Architectures, Symmetric Class Balancing,
and Evaluation Metrics
To investigate the interaction between the six representation
spaces (Full Spectrum, PCA, PLS-DA, UMAP, SOM, and
MSPA) and distinct mathematical decision mechanisms, ten
supervised architectures were evaluated across the identical data
partitions and three initialization seeds ({42, 2026, 777}), total-
ing 9, 000 fitted models and 18, 000 full validation and hold-
out test evaluations (Table 2). The classical machine learning
group comprises six algorithms representing the major statistical
learning families [30]: (1)Linear Discriminant Analysis(LDA),
with linear decision boundaries derived from multivariate Gaus-
sian densities and automatic Ledoit–Wolf regularization when
K> 200; (2)Logistic Regression(LR), optimized via L-BFGS
(400 maximum iterations) with anL2 Ridge penalty (Creg =1.0);
(3)Support Vector Machinewith aRadial Basis Functionkernel
(SVM-RBF, Creg =1.0,γrbf =[ K·Var (Z)]−1) and Platt proba-
bilistic calibration; (4)k-Nearest Neighbors( k-NN), weighted
by the inverse Euclidean distance to the k =5 nearest neighbors;
(5)Random Forest(RF) [ 36], composed of 500 decision trees
with bootstrap sampling and⌊
√
K⌋ features per split; and (6)
Extreme Gradient Boosting(XGBoost) [ 37], configured with up
to 500 additive trees, learning rateη =0.05, maximum depth 6,
row subsampling ratio of 0.8, column subsampling ratio per tree
of 0.8, and early stopping after 30 rounds on the validation split.
The second group comprises fourTabular Foundation Model
(TFM) architectures executed on GPU (NVIDIA GeForce RTX
4070, CUDA FP16/FP32) viaIn-Context Learning(ICL): (7)
TabPFN v3 [ 43, 44], based on networks pre-trained on syn-
thetic data from structural causal models with two-dimensional
induced attention across samples and features (8 ensemble es-
timators with column and class permutations); (8) TabICL v2
[45], a two-stage architecture that compresses inter-column in-
teractions via aSet-Transformerprior to row-wise in-context
attention, augmented with an iterative Latin-square generator
free of recursive stack overflow to support D =1, 580-channel
matrices (8 estimators); (9) TabDPT v1.2 [ 46], aRetrieval-
Augmented ICLparadigm that retrieves the kret =96 nearest
training neighbors via cosine similarity to weight local sub-
contexts; and (10) TabFM, which projects continuous vari-
ables onto 32 geometric scales of harmonic Fourier frequencies
(
h
sin(2π ˜ZW f ),cos(2π ˜ZW f )i∈R 256) modulated by nonlinear
SwiGLUgates ( ϕ(u)= u· [1+ exp(−u)]−1) coupled to an in-
context attention head.
To account for natural class imbalance without introducing in-
terpolated synthetic samples while maintaining symmetric treat-
ment across all ten architectures, an inverse-frequency weight
was computed in each fold f for every class c in the training set,

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy11
Table 2:Mathematical taxonomy, decision mechanisms, hyperparameter configurations, and symmetric class-balancing implementations across
the ten evaluated classification architectures.
Architecture Model Family Decision Paradigm Key Configuration&Regularization Class Balancing Mechanism Execution Device
LDAClassical Linear Mahalanobis covariance hyperplanes SVD/Eigen solver with Ledoit–Wolf shrinkage Post-hoc Bayesian prior reweighting CPU (64-bit float)
Logistic RegressionClassical Linear Multinomial log-odds linear modelL 2 Ridge penalty (C=1.0), L-BFGS (400 iter.) Native loss sample weights (w (f)c ) CPU (Multi-core)
SVM-RBFClassical Kernel Maximum-margin RKHS hyperplane RBF kernel (C=1.0,γ=scale), Platt scaling Native margin sample weights (w (f)c ) CPU (LibSVM)
k-NNClassical Distance Local Euclidean manifold votingk=5 neighbors, inverse-distance weighting Post-hoc Bayesian prior reweighting CPU (Multi-core)
Random ForestClassical Tree Bagging Orthogonal recursive partitioning 500 trees, max features=
√
K, bootstrap Node impurity sample weights (w (f)c ) CPU (Multi-core)
XGBoostClassical Tree Boosting 2nd-order gradient boosted trees Max 500 trees,η=0.05, depth 6, early stop 30 Hessian/gradient sample weights (w (f)c ) GPU (CUDA Hist)
TabPFN v3Tabular Foundation Model 2D row–column induced attention ICL 8 ensemble permutations, memory-saving mode Post-hoc Bayesian prior reweighting GPU (CUDA FP16/32)
TabICL v2Tabular Foundation Model Set-Transformer column-to-row ICL 8 estimators, batch size 2, iterative shuffler Post-hoc Bayesian prior reweighting GPU (CUDA FP16/32)
TabDPT v1.2Tabular Foundation Model Retrieval-augmented localk-NN ICLk ret =96 cosine neighborhood context fusion Post-hoc Bayesian prior reweighting GPU (CUDA FP32)
TabFMTabular Foundation Model Fourier harmonic+SwiGLU gated ICL 32 geometric Fourier frequencies, 256-dim gate Post-hoc Bayesian prior reweighting GPU (CUDA FP32)
w(f)c = N(f)train/(C·N (f)train,c). In algorithms based on loss optimiza-
tion or node impurity (Logistic Regression, SVM-RBF, Random
Forest, and XGBoost), the weights w(f)yi were applied directly
during model fitting; in generative, neighborhood-based, and
in-context foundation models (LDA, k-NN, and the four TFMs),
post-hoc Bayesian posterior recalibration (prior reweighting)
was applied to the raw predicted probabilitiesp i,c:
˜pi,c = pi,c·w (f)cPC
l=1 pi,l·w (f)l .(13)
Quantitative evaluation on the fixed holdout test set spanned
four complementary dimensions. First, multiclass and binary
discrimination capability was quantified via theMacro F1-
score( F1Macro),Balanced Accuracy( BalAcc) [78, 79], abso-
lute difference relative to the unreduced spectrum (∆ F1 =
F1Macro(d)−F1 Macro(Full Spectrum)), Performance Retention
and Gain Ratio ( PRR = F1Macro(d)/F1Macro(Full Spectrum)),
and the Fisher cluster separability ratio in the projected plane
(JF = Tr(SB)/Tr(SW)). Second, predictive uncertainty, prob-
abilistic quality, and inter-model consensus of the posterior
distributions ˜pi were assessed through Normalized Shannon
Entropy (Hnorm(˜pi)= − 1lnC PCc=1 ˜pi,c ln ˜pi,c∈ [0, 1]), Effective
Class Perplexity ( Perp(˜pi)= exp[−PCc=1 ˜pi,c ln ˜pi,c]∈ [1,C ]),
15-binExpected Calibration Error( ECE), area under the ROC
curve for selective prediction-error detection (AUROCErr), and
pairwise Jensen–Shannon divergence ( DJS) across architec-
tures. Third, weight-space optimization conditioning was au-
dited following the filter-normalized perturbation protocol of
Liet al.[ 31], evaluating the three-dimensional cross-entropy
loss surface LCE(Θ∗ + αu1 + βu2), basal Shannon entropy,
and iterative Gradient Descent trajectory (Θ (t)) on the test set
along orthogonal directions (u 1, u2) normalized by logit vari-
ance (α,β∈ [−1.5, +1.5]), complemented by 3D response
surfaces of in-context sample efficiency across geographic
extrapolation horizons and softmax temperature calibration
(τ∈ [0.25, 2.50]). Fourth, fitting time and per-sample infer-
ence latency ( ms/sample) were recorded, applying the non-
parametric Friedman test (χ2F), Demšar Critical Difference (CD)diagrams [80], and pairwise Wilcoxon signed-rank tests with
Holm–Bonferroni correction (α=0.05).
4 Results
4.1 Global Performance, Subspace Class Separability, and
Inter-Model Consensus
The consolidated evaluation across the five plant systems and
ten classification architectures (50 paired system–model combi-
nations, totaling 1, 500 holdout test evaluations) revealed highly
significant differences among the six dimensionality represen-
tations (χ2F =173 .82, p =1 .12× 10−35; Table 3). Relativeto the unreduced Full Spectrum baseline ( D =1 , 580, mean
F1Macro =0 .7756), all three unsupervised dimensionality re-
duction methods, PCA, SOM, and UMAP, degraded global pre-
dictive performance, with UMAP incurring the sharpest drop
(∆F1 =−0.1449). Among the reference methods, only super-
vised PLS-DA achieved a marginal positive mean gain over the
Full Spectrum (∆ F1 = +0.0110), yet its median differential
remained negative (−0.0011).
In contrast, compressing the 1, 580 bands into K∗ =64 morpho-
spectral variables via MSPA (24 .7× compression ratio, or
95.95% dimensionality reduction) yielded the highest overall
performance in the benchmark ( F1Macro =0 .8520± 0.0999),
representing an average gain of+7.64 percentage points (PRR =
1.1280) over the Full Spectrum while reducing inter-model
dispersion by 36.8% (Table 3). Pairwise Holm–Bonferroni-
adjusted Wilcoxon signed-rank tests confirm that MSPA sig-
nificantly outperformed the Full Spectrum in 92 .0% (46/50,
p =5 .32× 10−10), PCA in 98 .0% (49/50), PLS-DA in
92.0% (46/50), and both UMAP and SOM in 100.0% (50/50)
of all evaluated pairs. Compared directly against the best-
performing reference reducer for each individual system–model
pair (max{PCA, PLS-DA, UMAP, SOM}), MSPA prevailed in
90.0% (45/50) of the scenarios.
Stratifying the results across the five plant systems (bottom
block of Table 3 and Fig. 6) shows that the discriminative benefit
of MSPA scales directly with spectral overlap and taxonomic
cardinality. In binary Citrus diagnostics (C =2), where baseline
separability on the Full Spectrum is already high (0.9622), PCA
degraded average performance, whereas PLS-DA and MSPA
preserved inter-class separation (+1.20 pp). In multi-day foliar
herbivory monitoring across Soybean (C =2) and Cotton ( C =
2), where diurnal and developmental variations caused all four
conventional reducers to underperform the unreduced spectrum,
MSPA improved all ten classifiers (10/10 win rate), raising mean
F1Macro by+6 .05 pp and+7 .89 pp, respectively. The largest
gains occurred in the multiclass regimes:+9 .08 pp in ternary
Maize feeding groups (C =3) and+13 .98 pp (PRR =1.2787,
10/10 win rate) in 39-class Savanna Pastures (C=39).
Demšar Critical Difference ( CD) analysis ( CD =1 .07 at
α =0 .05) confirms that MSPA occupies the top rank in iso-
lation (Mean Rank =1.14), statistically separated from the tied
middle tier formed by PLS-DA (2.68), Full Spectrum (3.04), and
PCA (3.52), as well as from SOM (4.66) and UMAP (5.96). Fur-
thermore, evaluating the predicted posterior distributions across

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy12
Table 3:Global, paradigm-stratified, and plant-system-stratified performance summary across the six dimensionality representations (N=50
paired system–model combinations across 10 cross-validation folds). Values report mean F1Macro, standard deviation (SD), median, Performance
Retention and Gain Ratio (PRR vs. Full Spectrum), absolute difference (∆F1), and pairwise Wilcoxon signed-rank test statistics comparing
MSPA against each reference representation.
Aggregation Scope Representation Dim. (K ∗) Mean F1 Macro±SD Median F1 Macro Mean PRR±SD Mean∆F1 vs. Full MSPA Win Rate Wilcoxonp-value
Global (5Systems×10Models)
Full Spectrum(Unreduced) 1,580 0.7756±0.1581 0.7921 1.0000±0.0000 0.0000 46/50 (92.0%) 5.32×10 −10
PCA(Unsupervised Orthogonal) 10–50 0.7629±0.1544 0.7440 0.9871±0.0906−0.0126 49/50 (98.0%) 5.33×10 −15
PLS-DA(Supervised Latent) 8–21 0.7866±0.1297 0.7628 1.0325±0.1450+0.0110 46/50 (92.0%) 1.61×10 −12
UMAP(Nonlinear Manifold) 6–8 0.6307±0.2007 0.6918 0.8060±0.1562−0.1449 50/50 (100.0%) 1.78×10 −15
SOM(Kohonen Topological) 64–100 0.7237±0.1724 0.7335 0.9334±0.1172−0.0519 50/50 (100.0%) 1.78×10 −15
MSPA(Proposed Operator)64 0.8520±0.0999 0.8272 1.1280±0.1781+0.0764 Reference (Rank 1)—
By Classifier Paradigm Classical ML (6 Models×5 Systems) Full→MSPA0.7467→0.83580.7517→0.82381.0000→1.1481+0.089129/30 (96.7%) 1.16×10 −8
Tabular Foundation (4 TFMs×5 Systems) Full→MSPA0.8189→0.87620.8407→0.86301.0000→1.0978+0.057317/20 (85.0%) 6.29×10 −3
By Plant System (10Models)
Citrus(C=2) Full→MSPA0.9622→0.97430.9705→0.98051.0000→1.0128+0.01208/10 (80.0%) 1.37×10 −2
Cotton(C=2) Full→MSPA0.7937→0.87250.7779→0.87661.0000→1.1032+0.078910/10 (100.0%) 1.95×10 −3
Soybean(C=2) Full→MSPA0.8393→0.89970.8273→0.89671.0000→1.0748+0.060510/10 (100.0%) 1.95×10 −3
Maize(C=3) Full→MSPA0.6520→0.74280.6282→0.73011.0000→1.1705+0.09088/10 (80.0%) 1.95×10 −2
Savanna Pastures(C=39) Full→MSPA0.6307→0.77050.6537→0.78061.0000→1.2787+0.139810/10 (100.0%) 1.95×10 −3
distinct classifier families reveals that projecting spectra through
MSPA reduces the mean pairwise Jensen–Shannon divergence
(DJS) among architectures by 39.7% (0.300→ 0.181 bits), ele-
vating Top-1 prediction consensus among the leading classifiers
(TabPFN v3, TabICL v2, Random Forest, and XGBoost) from
79.5% on the Full Spectrum to 92.8% on MSPA.
4.2 Architectural Interaction Between Classical Classifiers
and Tabular Foundation Models
The disaggregated performance matrix (Table 4 and Fig. 7)
reveals distinct interactions between classifier decision mecha-
nisms and spectral representations. On the unreduced Full Spec-
trum (D =1, 580), the two column-attention Tabular Foundation
Models, TabPFN v3 (cross-system mean of 0.9242) and TabICL
v2 (0.9051), outperformed all other architectures. Conversely,
TFMs relying on cosine neighborhood retrieval (TabDPT v1.2)
or random Fourier projections (TabFM), as well as distance-
based k-NN, suffered severe accuracy degradation on the 1, 580
raw bands in multiclass tasks (Maize and Savanna Pastures),
where collinear wavelengths dilute pairwise distance metrics.
When preceded by the four reference reducers (PCA, PLS-DA,
UMAP, and SOM), the ten classifiers bifurcated into two re-
sponse regimes (Table 4 and Fig. 7B). For column-attention
TFMs (TabPFN v3 and TabICL v2) and regularized linear dis-
criminants (LDA and Logistic Regression), all four conventional
reducers caused systematic accuracy losses relative to the Full
Spectrum, as discarding low-variance components and deriva-
tive shape cues deprived their attention and linear projection
heads of diagnostic information. Conversely, distance-based,
kernel, and tree-based models (k-NN, TabDPT v1.2, SVM-RBF,
Random Forest, and TabFM) gained accuracy when moving
from the Full Spectrum to supervised PLS-DA, benefiting from
collinearity reduction.
Coupling the classifiers with MSPA ( K∗ =64) resolved this
dichotomy, substantially improving both classical machine learn-
ing models (+8.91 pp average gain, 96.7% win rate) and Tabular
Foundation Models (+5.73 pp average gain, 85.0% win rate;
Table 3). Across Cotton, Soybean, and Savanna Pastures, 100%
of the evaluated architectures (30 /30) achieved their highest
F1Macro on the MSPA subspace (Table 4). In Citrus, SVM-
RBF coupled with MSPA achieved the highest overall score
(0.9860), while in Cotton and Soybean the top positions were
led by TabPFN v3 (0.9821 and 0.9840), TabICL v2 (0.9712 and
0.9800), and TabDPT v1.2 (0.9547 and 0.9730). In 39-class Sa-
vanna Pastures, MSPA enabled TabICL v2 (0.8072), TabPFN v3
(0.8070), and Random Forest (0.8021) to surpass the 0.80 thresh-
old while promoting large recoveries for distance- and retrieval-
based architectures (+36.99 pp for k-NN and+27.59 pp for Tab-
DPT v1.2), with consistent gains of+10.7 to +18.6 pp across all
five morpho-ecological functional groups. The only exception
occurred in Maize (C =3,N =474): while eight of the ten archi-
tectures achieved substantial gains of+6.68 to +22.27 pp under
MSPA compared to the Full Spectrum, TabPFN v3 and TabICL
v2 scored higher directly on the unreduced bands (0.9111 and
0.8773 vs. 0 .8273 and 0.7972), reflecting the higher sample
variance of empirical covariance estimation in small training
cohorts (Ntrain≈341).
4.3 Internal Module Ablation, Physical Explainability, and
Intrinsic Dimensionality Scaling
An internal ablation study across Random Forest, SVM-RBF,
and LDA over the 10 cross-validation folds confirmed that all
three mathematical blocks of MSPA contribute additively to dis-
crimination accuracy (Table 5). Evaluated in isolation, the bio-
physical hull block (Module 1, 16 variables) achieved a global
mean F1Macro of 0.7427, the local multi-order window SVD
block (Module 2, 24 variables) reached 0.7816, and the global
Fisher-weighted canonical block (Module 3, 24 variables) at-
tained 0.8304. Combining the local window eigenmodes with
the global canonical axes (Modules 2+3, 48 linear variables) im-
proved performance over Module 3 alone across four of the five
plant systems (0.8337 global mean), while integrating the non-
linear continuum-removed and red-edge inflection descriptors
of Module 1 (forming the complete 64-variable MSPA operator)
yielded the highest overall accuracy (0.8381).
Perturbation saliency (∆LCE) and feature importance auditing
across the 64 MSPA variables (Fig. 8) demonstrate that the
top-ranked predictors draw from all three modules, led by the
leading global canonical axes (Module 3), the first-derivative
red-edge slope (686–765 nm, Module 2), the red-edge inflec-
tion wavelength centroid (λRE, Module 1), the SWIR 1 lignin
concavity mode (Module 2), and the maximum chlorophyll ab-
sorption depth ( Dmax, Module 1). Partitioned by differential
order (Fig. 8C), 73.4% of the cumulative discriminative weight
across the five systems originates from first-derivative slope
(42.9%) and second-derivative concavity (30.5%) rather than
zero-order amplitude (26.6%), with high rank concordance be-

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy13
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
Citrus (C = 2 | Leaf HLB)
Component 2 (Z-Score)
PCA (PC1 vs. PC2)
Fisher JF = 0.59
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
PLS-DA (LV1 vs. LV2)
Fisher JF = 0.58
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
UMAP (Dim 1 vs. Dim 2)
Fisher JF = 0.10
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
SOM (BMU Dist 1 vs. 2)
Fisher JF = 0.79
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
MSPA (CAN 1 vs. CAN 2)
Healthy Control (HLB-)
Greening Infected (HLB+)
Fisher JF = 3.53
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
Cotton (C = 2 | Leaf Herbivory)
Component 2 (Z-Score)
Fisher JF = 0.00
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
Fisher JF = 0.06
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
Fisher JF = 0.01
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
Fisher JF = 0.02
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
Healthy Control
Herbivory Infested
Fisher JF = 0.25
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
Soybean (C = 2 | Leaf Multi-Pest)
Component 2 (Z-Score)
Fisher JF = 0.06
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
Fisher JF = 0.10
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
Fisher JF = 0.09
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
Fisher JF = 0.06
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
Healthy Control
Multi-Pest Infested
Fisher JF = 0.41
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
Maize (C = 3 | Feeding Guilds)
Component 2 (Z-Score)
Fisher JF = 0.03
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
Fisher JF = 0.05
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
Fisher JF = 0.02
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
Fisher JF = 0.02
−3 −2 −1 0 1 2 3
−3
−2
−1
0
1
2
3
Healthy Control
Chewing Guild
Sucking Guild
Fisher JF = 0.33
−3 −2 −1 0 1 2 3
Component 1 (Z-Score)
−3
−2
−1
0
1
2
3
Savanna Pastures (C = 39 | 5 Guilds)
Component 2 (Z-Score)
Fisher JF = 1.99
−3 −2 −1 0 1 2 3
Component 1 (Z-Score)
−3
−2
−1
0
1
2
3
Fisher JF = 2.61
−3 −2 −1 0 1 2 3
Component 1 (Z-Score)
−3
−2
−1
0
1
2
3
Fisher JF = 2.37
−3 −2 −1 0 1 2 3
Component 1 (Z-Score)
−3
−2
−1
0
1
2
3
Fisher JF = 2.59
−3 −2 −1 0 1 2 3
Component 1 (Z-Score)
−3
−2
−1
0
1
2
3
I. Cultivated Forage Grasses
II. Native & Invasive Grasses
III. Broadleaf Weeds & Legumes
IV. Vines & Parasitic Lianas
V. Invasive Palm
Fisher JF = 5.85
Figure 6:Two-dimensional subspace projection matrix (5 Plant Systems× 5 Dimensionality Reduction Methods) on out-of-training evaluation
samples (90% confidence covariance ellipses and 2D Fisher cluster separability indexJ F =Tr(S B)/Tr(SW)). Rows follow the taxonomic
complexity gradient from binary foliar stress inCitrus,Cotton, andSoybean(C=2) and ternary feeding groups inMaize(C=3) to
canopy-scale species discrimination inSavanna Pastures(C=39, colored across the five morpho-ecological functional groups).
tween tree-ensemble importance and TFM perturbation saliency
(Spearmanρ=0.91,p<10 −15; Fig. 8D).
Evaluating the autonomous Kneedle breaking-point estimator
(n_components=’auto’) and dimensionality scaling across
K∈ [8, 96] (Fig. 9) shows that the detected canonical inflection
point K∗can scales monotonically with task complexity, shifting
from K∗can =12 –14 (K∗auto =52 –54) in binary Citrus to K∗can =
20–22 (K∗auto =60 –62) in 39-class Savanna Pastures. Across
the scaling grid K∈ [8, 96], PLS-DA peaks near K =16 and
degrades progressively for K≥ 24 due to rank-deficient noise
fitting when K>C− 1, whereas MSPA rises steeply between
K =8 and 48 and plateaus across the Kneedle equilibrium
zone (K∗∈ [52, 66]), achieving statistical equivalence between
autonomous ‘’auto’‘ mode and the fixed K∗ =64 configuration
(|∆F1|≤0.0015 across all five systems; Fig. 9D).
4.4 Optimization Landscapes, Probabilistic Uncertainty, and
Computational Speedup
Three-dimensional weight-space cross-entropy loss landscapes
LCE(Θ∗ +αu1 +βu2) and iterative Gradient Descent trajectories
(Θ(t); Fig. 10) illustrate the optimization conditioning induced
by each subspace. PCA produces an elongated, anisotropic

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy14
Full Spectrum (1,580)
PCA (K*) PLS-DA (K*) UMAP (K*) SOM (K*) MSPA (64)
Spectral Dimensionality Representation
TabPFN v3
TabICL v2
TabDPT v1.2
Random Forest
XGBoost
SVM-RBF
k-NN
LDA
Logistic Regression
TabFM
Classification Architecture (4 TFMs & 6 Classical ML)
0.9242 0.8851 0.8756 0.6649 0.8531 0.9145
0.9051 0.8645 0.8414 0.6567 0.8332 0.9075
0.7394 0.8123 0.8234 0.6427 0.7854 0.8871
0.7724 0.7715 0.7919 0.6352 0.6749 0.8597
0.7871 0.7807 0.7800 0.6311 0.6800 0.8503
0.7348 0.7669 0.7739 0.6183 0.6346 0.8402
0.6678 0.6342 0.7756 0.6242 0.6462 0.8400
0.7647 0.7297 0.7297 0.5866 0.7170 0.8130
0.7534 0.7348 0.7309 0.5956 0.6813 0.8118
0.7069 0.6498 0.7437 0.6513 0.7311 0.7956
A. Cross-System Mean Macro-F1 Across 10 Models & 6 Representations
−0.05 0.00 0.05 0.10 0.15
Mean Δ Macro-F1 Relative to Full Spectrum (1,580 Bands)
TabFM
Logistic Regression
LDA
k-NN
SVM-RBF
XGBoost
Random Forest
TabDPT v1.2
TabICL v2
TabPFN v3
B. Mean Macro-F1 Gain/Loss Relative to Full Spectrum
Representation
MSPA (Proposed)
PCA (Unsupervised)
PLS-DA (Supervised)
Full Spectrum Baseline (Δ = 0)
0.60
0.65
0.70
0.75
0.80
0.85
0.90
Mean Macro-F1 Across the 5 Plant Systems
Figure 7:Cross-system mean F1Macro heatmap across the ten classification architectures and six dimensionality representations (A), and mean
absolute∆F1 gain/loss relative to the Full Spectrum comparing PCA, PLS-DA, and the proposed MSPA operator (B).
0.00 0.02 0.04 0.06 0.08 0.10
Mean Normalized Feature Importance (Gain / Saliency)
W2 Slope (Green)
NIR–Red Contrast
W8 Concav. (SW2)
V4 Depth (SW1)
CAN 4 (Global)
W3 Concav. (Red)
S_max Slope (RE)
W6 Slope (NIR)
V1 Depth (Chl)
W7 Concav. (SW1)
CAN 3 (Global)
λ_RE Centroid
W4 Slope (RE)
CAN 2 (Global)
CAN 1 (Global)
Extracted MSPA Feature
0.027
0.029
0.031
0.034
0.036
0.038
0.041
0.044
0.048
0.051
0.055
0.061
0.066
0.074
0.088
A. Top 15 Most Discriminative Variables in MSPA
MSPA Architectural Block
Module 1 (Biophysical Hull)
Module 2 (Local Window SVD)
Module 3 (Global Canonical)
W1: Blue
(401–510) W2: Green(511–620) W3: Red(621–685)W4: Red-Edge(686–765)W5: NIR-Meso(766–910)W6: NIR-Water(911–1330)W7: SWIR 1(1491–1770)W8: SWIR 2(2031–2400)
Biophysical Radiative Sub-Window (W1 to W8)
Citrus
(C = 2)
Cotton
(C = 2)
Soybean
(C = 2)
Maize
(C = 3)
Savanna
Pastures (C = 39)Plant System
0.08 0.24 0.29 0.42 0.14 0.18 0.36 0.31
0.11 0.19 0.38 0.46 0.15 0.21 0.33 0.28
0.09 0.22 0.31 0.44 0.19 0.39 0.35 0.22
0.12 0.25 0.28 0.39 0.34 0.37 0.29 0.26
0.16 0.23 0.32 0.48 0.29 0.31 0.43 0.38
B. Spectral Window Perturbation Saliency (ΔL_CE) Across Systems
0.05
0.10
0.15
0.20
0.25
0.30
0.35
0.40
0.45
0.50
Perturbation Saliency ΔL_CE (nats ↑)
Citrus
(C = 2)
Cotton
(C = 2)
Soybean
(C = 2)
Maize
(C = 3)
Savanna
Pastures
Plant System
0
20
40
60
80
100
Share of Total Importance (%)
28.5%
42.0%
29.5%
25.0%
44.5%
30.5%
26.5%
43.0%
30.5%
29.0%
41.5%
29.5%
24.0%
43.5%
32.5%
C. Cumulative Decision Weight by Differential Order
Order 0: Amplitude ρ (26.6%) Order 1: Slope ∂ρ/∂λ (42.9%) Order 2: Concavity ∂²ρ/∂λ² (30.5%)
0.2 0.4 0.6 0.8 1.0
Tree Ensemble Normalized Importance (Random Forest + XGBoost Gain)
0.2
0.4
0.6
0.8
1.0TFM Perturbation Saliency (TabPFN v3 + TabICL v2)
D. Cross-Paradigm Explainability Concordance (64 Variables)
Module 1 (Biophysical)
Module 2 (Window SVD)
Module 3 (Canonical)
Identity Line (Spearman ρ = 0.91, p < 10⁻¹⁵)
Figure 8:Physical explainability and spectral perturbation saliency across the 64 semantic variables of MSPA: (A) ranking of the top 15 most
discriminative variables color-coded by internal module; (B) perturbation saliency heatmap (∆LCE in nats) across the eight radiative
sub-windows (W1 toW8) and five plant systems; (C) cumulative share of decision weight partitioned by differential order (showing that 73.4%
of total importance originates from 1st and 2nd derivative features); and (D) cross-paradigm explainability concordance between tree ensemble
gain (Random Forest+XGBoost) and TFM perturbation saliency (TabPFN v3+TabICL v2).

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy15
Table 4:Complete cross-validated holdout test performance matrix (mean F1Macro across 10 folds) over the five plant spectroscopy datasets, ten
classification models, and six dimensionality representations, along with the MSPA standard deviation (±SD), Performance Retention and Gain
Ratio (PRR), and absolute difference (∆F1) relative to the Full Spectrum. Bold values denote the best representation per row. Superscripts
indicate statistically significant MSPA gains (p< 0.05, paired Wilcoxon test across folds) over the Full Spectrum (∗) and over the best reference
reducer (†).
Plant System (C) Model Architecture Full Spectrum (1580) PCA PLS-DA UMAP SOM MSPA (64, Mean±SD) PRR MSPA ∆F1MSPA
Citrus(C=2)
SVM-RBF(Kernel Margin) 0.9742 0.9501 0.9782 0.9417 0.97920.9860±0.00411.0121+0.0118 ∗†
Random Forest(Tree Bagging) 0.9512 0.96390.98660.9310 0.9313 0.9821±0.0052 1.0326+0.0310 ∗
TabICL v2(TFM) 0.9766 0.9397 0.9779 0.9218 0.95030.9819±0.00451.0054+0.0052 ∗
XGBoost(Tree Boosting) 0.9661 0.95800.98230.9322 0.9332 0.9814±0.0058 1.0159+0.0153 ∗
TabDPT v1.2(TFM) 0.9750 0.9389 0.9774 0.9177 0.92700.9809±0.00491.0061+0.0059 ∗
LDA(Linear Covariance) 0.9792 0.9591 0.9766 0.9387 0.94880.9801±0.00441.0009+0.0009
TabPFN v3(TFM)0.97740.9335 0.9709 0.9183 0.9530 0.9721±0.0051 0.9946−0.0053
TabFM(TFM) 0.9405 0.94610.97830.9291 0.9192 0.9632±0.0074 1.0243+0.0227 ∗
Logistic Regression(Linear Prob.)0.96680.9575 0.9565 0.9243 0.9502 0.9610±0.0063 0.9940−0.0058
k-NN(Distance Manifold) 0.9154 0.93700.96030.9089 0.9064 0.9540±0.0068 1.0422+0.0386 ∗
Cotton(C=2)
TabPFN v3(TFM) 0.9598 0.9483 0.9325 0.7327 0.88980.9821±0.00541.0233+0.0223 ∗†
TabICL v2(TFM) 0.9370 0.9112 0.8752 0.7275 0.86930.9712±0.00621.0366+0.0343 ∗†
TabDPT v1.2(TFM) 0.7651 0.8273 0.8464 0.7234 0.84250.9547±0.00851.2479+0.1896 ∗†
XGBoost(Tree Boosting) 0.7936 0.7950 0.7795 0.7014 0.73320.8778±0.01121.1062+0.0842 ∗†
k-NN(Distance Manifold) 0.7495 0.6971 0.8047 0.7032 0.70860.8769±0.01081.1700+0.1274 ∗†
Random Forest(Tree Bagging) 0.8045 0.7391 0.7815 0.7104 0.73460.8762±0.01041.0892+0.0717 ∗†
TabFM(TFM) 0.7906 0.6713 0.7537 0.7214 0.79890.8501±0.01261.0753+0.0595 ∗†
SVM-RBF(Kernel Margin) 0.7539 0.7238 0.7543 0.6852 0.66170.8046±0.01191.0673+0.0507 ∗†
LDA(Linear Covariance) 0.6794 0.6449 0.6965 0.6093 0.72530.7667±0.01341.1285+0.0873 ∗†
Logistic Regression(Linear Prob.) 0.7032 0.6471 0.6838 0.6097 0.67850.7650±0.01311.0879+0.0618 ∗†
Soybean(C=2)
TabPFN v3(TFM) 0.9698 0.9724 0.9476 0.7236 0.91970.9840±0.00581.0148+0.0143 ∗†
TabICL v2(TFM) 0.9457 0.9379 0.8907 0.7168 0.90020.9800±0.00641.0364+0.0343 ∗†
TabDPT v1.2(TFM) 0.8336 0.9057 0.9079 0.7208 0.88930.9730±0.00791.1675+0.1394 ∗†
Random Forest(Tree Bagging) 0.8384 0.8363 0.8353 0.6942 0.76240.8997±0.01151.0732+0.0613 ∗†
k-NN(Distance Manifold) 0.7971 0.6953 0.8136 0.7006 0.73380.8992±0.01211.1281+0.1021 ∗†
XGBoost(Tree Boosting) 0.8583 0.8334 0.8208 0.6914 0.76520.8942±0.01181.0421+0.0360 ∗†
SVM-RBF(Kernel Margin) 0.7489 0.8228 0.8176 0.6987 0.70130.8928±0.01091.1921+0.1439 ∗†
TabFM(TFM) 0.8002 0.6768 0.7670 0.7123 0.80360.8270±0.01421.0335+0.0267 ∗†
Logistic Regression(Linear Prob.) 0.7798 0.8042 0.7572 0.6921 0.74310.8265±0.01281.0599+0.0467 ∗†
LDA(Linear Covariance) 0.8210 0.8041 0.7587 0.6901 0.77120.8211±0.01241.0001+0.0001 †
Maize(C=3)
TabPFN v3(TFM)0.91110.8301 0.7699 0.5117 0.7731 0.8273±0.0165 0.9081−0.0838
TabICL v2(TFM)0.87730.7963 0.7133 0.4895 0.7273 0.7972±0.0172 0.9090−0.0800
TabDPT v1.2(TFM) 0.6364 0.7469 0.6910 0.4931 0.68030.7640±0.01841.2024+0.1275 ∗†
Random Forest(Tree Bagging) 0.6338 0.6531 0.6459 0.4394 0.46240.7385±0.01911.1661+0.1047 ∗†
SVM-RBF(Kernel Margin) 0.5543 0.6505 0.5898 0.4075 0.41290.7301±0.01781.3172+0.1758 ∗†
LDA(Linear Covariance) 0.6225 0.6095 0.5569 0.4167 0.55070.7300±0.01821.1727+0.1075 ∗†
Logistic Regression(Linear Prob.) 0.6068 0.5954 0.5564 0.4168 0.47200.7236±0.01861.1925+0.1168 ∗†
XGBoost(Tree Boosting) 0.6528 0.6481 0.6224 0.4239 0.47240.7196±0.01951.1043+0.0668 ∗†
k-NN(Distance Manifold) 0.4964 0.5505 0.6496 0.4252 0.45480.7191±0.02041.4486+0.2227 ∗†
TabFM(TFM) 0.5285 0.5156 0.6116 0.4732 0.55630.6782±0.02181.2839+0.1497 ∗†
Savanna Pastures(C=39)
TabICL v2(TFM) 0.7892 0.7376 0.7497 0.4281 0.71920.8072±0.00881.0230+0.0181 ∗†
TabPFN v3(TFM) 0.8027 0.7410 0.7572 0.4382 0.73000.8070±0.00841.0053+0.0042 †
Random Forest(Tree Bagging) 0.6342 0.6651 0.7104 0.4012 0.48370.8021±0.00961.2652+0.1679 ∗†
SVM-RBF(Kernel Margin) 0.6426 0.6870 0.7296 0.3587 0.41820.7878±0.01021.2263+0.1452 ∗†
Logistic Regression(Linear Prob.) 0.7102 0.6699 0.7008 0.3353 0.56290.7826±0.00991.1022+0.0724 ∗†
XGBoost(Tree Boosting) 0.6647 0.6691 0.6950 0.4067 0.49620.7786±0.01081.1725+0.1139 ∗†
LDA(Linear Covariance) 0.7212 0.6308 0.6599 0.2780 0.58880.7671±0.00951.0636+0.0459 ∗†
TabDPT v1.2(TFM) 0.4868 0.6428 0.6943 0.3587 0.58780.7627±0.01141.5673+0.2759 ∗†
k-NN(Distance Manifold) 0.3808 0.2908 0.6499 0.3831 0.42760.7507±0.01221.9738+0.3699 ∗†
TabFM(TFM) 0.4746 0.4391 0.6077 0.4207 0.57760.6597±0.01451.3904+0.1851 ∗†
Table 5:Internal architectural ablation of the MSPA operator across the five plant spectroscopy datasets ordered by class cardinality (mean
F1Macro across Random Forest, SVM-RBF, and LDA over 10 cross-validation folds). Columns compare isolated modules against the linear
multi-order combination (Modules 2+3) and the complete three-module operator.
Plant System (C) Mod. 1 Only (16vars) Mod. 2 Only (24vars) Mod. 3 Only (24vars) Mod. 2+Mod. 3 (48vars) Full MSPA (64vars)∆(Mod. 2+3−Mod. 3)∆(Full−Mod. 2+3)
(Continuum&Inflection) (Local Window SVD 0,1,2) (Global Fisher Canonical) (Pure Linear Multi-Order) (Complete Architecture) (Local Window Effect) (Nonlinear Hull Effect)
Citrus(C=2) 0.9639 0.9785 0.9777 0.98090.9842+0.0032+0.0033
Cotton(C=2) 0.7441 0.7547 0.8053 0.80990.8182+0.0046+0.0083
Soybean(C=2) 0.7840 0.8253 0.86450.87300.8729+0.0085−0.0001
Maize(C=3) 0.6272 0.6660 0.7175 0.72560.7310+0.0081+0.0054
Savanna Pastures(C=39) 0.5944 0.68350.78690.7792 0.7844−0.0077+0.0052
Global Mean (5Systems) 0.7427 0.7816 0.8304 0.8337 0.8381+0.0033+0.0044
valley with an elevated test loss floor ( minL∗CE =0 .30) thatforces the gradient descent trajectory to zig-zag across the valley
walls (Fig. 10A), while PLS-DA forms a steep, narrow funnel
(minL∗CE =0.34) sensitive to weight perturbations (Fig. 10B).Conversely, MSPA conditions the Hessian into a wide, isotropic
basin with the lowest test loss floor (minL∗CE =0.07) and basal
Shannon entropy (0.12), guiding Gradient Descent along a di-
rect, monotonic path to the global minimum in fewer than 15
iterations (Fig. 10C).
In terms of posterior probability calibration and uncertainty
(Table 6 and Fig. 11), unsupervised reducers inflated mean Nor-
malized Shannon Entropy (Hnorm) relative to the Full Spectrum

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy16
10 20 30 40 50 60
Canonical Eigenmode Index k in Module 3 (1 to 60)
0.0
0.2
0.4
0.6
0.8
1.0Normalized Cumulative Fisher Singular Energy y_k
A. MSPA('auto') Breaking-Point Mechanics Across the 5 Plant Systems
Citrus (C = 2) → K_can*=12 (T otal K*=52)
Cotton (C = 2) → K_can*=15 (T otal K*=55)
Soybean (C = 2) → K_can*=16 (T otal K*=56)
Maize (C = 3) → K_can*=18 (T otal K*=58)
Savanna Pastures (C = 39) → K_can*=22 (T otal K*=62)
Secant Line (Uniform Energy Growth)
Consensus K_can* Zone [12–26] → T otal K* ∈ [52, 66] 0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
Kneedle Deflection Distance D(k) = y_norm(k) - x_norm(k)
20 40 60 80
Number of Retained Subspace Dimensions (K)
0.70
0.72
0.74
0.76
0.78
0.80
0.82
0.84Cross-System Mean Macro F1-Score
B. Subspace Dimensionality Scaling Law (K ∈ [8, 96]) Across Reducers
MSPA (Fisher-Ranked & Multi-Order Basis)
PLS-DA (Supervised Latent Variables)
PCA (Unsupervised Orthogonal Components)
MSPA('auto') Equilibrium Zone (K* ∈ [52, 64])
Default Benchmark K* = 64 (F1 = 0.8385)
10
20
30
40
50
60
70
80
Subspace Dimension K20
30
40
50
60
70
80
90
100
Training Context (% N_train)
0.60
0.65
0.70
0.75
0.80
0.85
Test Macro F1-Score
C. 3D Scaling Topology: Subspace K × Context Size (% N) × Macro-F1
PCA Reference Surface
MSPA('auto') K* Trajectory
Converged K* = 64 Optimum
Citrus
(C = 2)
Cotton
(C = 2)
Soybean
(C = 2)
Maize
(C = 3)
Savanna Pastures
(C = 39)
Plant System (Ordered by Taxonomic Complexity C = 2 → 39)
0.6
0.7
0.8
0.9
1.0
1.1Mean Macro F1-Score (Random Forest, SVM-RBF & LDA)
D. Equivalence of Automatic Kneedle K* ('auto') and Fixed K* = 64
PCA (K*)
PLS-DA (K*)
MSPA (n_components='auto')
MSPA (Fixed Benchmark K* = 64)
K*=52
K*=55
K*=56
K*=58
K*=62
Figure 9:Intrinsic breaking-point mechanics (MSPA(n_components=’auto’)) and subspace dimensionality scaling laws: (A) normalized
cumulative Fisher-weighted singular energy (yk, left axis) and Kneedle perpendicular deflection distance (D(k), right axis) across the five plant
systems; (B) cross-system mean F1Macro scaling curves across subspace dimensions K∈ [8, 96] for MSPA, PLS-DA, and PCA; (C) 3D response
surface of test F1Macro as a function of subspace dimension (K) and training context size (%N train) tracing theMSPA(’auto’)ridge trajectory;
and (D) empirical equivalence between autonomous Kneedle estimation (’auto’) and the fixed benchmark configuration (K∗ =64).
−1.5
−1.0
−0.5
0.0
0.5
1.0
1.5
Orthogonal Dir.
α (SD)
−1.5
−1.0
−0.5
0.0
0.5
1.0
1.5
Orthogonal Dir.
β (SD)
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
Cross-Entropy Loss
A. PCA Subspace (K* = 50)
−1.5
−1.0
−0.5
0.0
0.5
1.0
1.5
Orthogonal Dir.
α (SD)
−1.5
−1.0
−0.5
0.0
0.5
1.0
1.5
Orthogonal Dir.
β (SD)
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
Cross-Entropy Loss
B. PLS-DA Subspace (K* = 21)
−1.5
−1.0
−0.5
0.0
0.5
1.0
1.5
Orthogonal Dir.
α (SD)
−1.5
−1.0
−0.5
0.0
0.5
1.0
1.5
Orthogonal Dir.
β (SD)
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
Cross-Entropy Loss
C. Proposed MSPA Subspace (K* = 64)
Figure 10:Three-dimensional weight-space cross-entropy loss landscapes LCE(Θ∗ +αu1 +βu2), basal Shannon entropy contours (Hnorm), and
projected Gradient Descent trajectories (Θ(t)) on the holdout test set following filter-normalized perturbation [31] for (A) PCA, (B) PLS-DA, and
(C) the proposed MSPA subspace.
(0.4305), whereas MSPA reduced global predictive entropy by
40.3% (to 0.2571), with reductions of 57 .2%–72.0% across
the leading TFMs (compressing effective perplexity in 39-class
Savanna Pastures from 8.57 to 1.32–1.76 species) alongside

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy17
consistent gains in selective error detection ( AUROCErr up to
0.9240). Furthermore, 3D in-context learning and temperature
calibration surfaces (Figs. 12 and 13) reveal that TabICL v2 and
TabPFN v3 supplied with only 30% of the training data (854
spectra) under MSPA outperform 100% training data (2 , 846
spectra) under PCA or PLS-DA, maintaining superior gener-
alization across distal geographic transect horizons and lower
Expected Calibration Error (ECE) acrossτ∈[0.25,2.50].
Finally, the Pareto speed–accuracy trade-offanalysis (Table 6
and Fig. 14) demonstrates that MSPA shifts both modeling
paradigms toward the ideal efficiency frontier. For classical
machine learning classifiers (Fig. 14A), compressing the 1, 580
channels into 64 variables accelerated fold training and evalua-
tion by 3.8× (Random Forest) to 41.2× (LDA) while simultane-
ously raising mean F1Macro by+4.8 to +17.2 pp. For Tabular
Foundation Models (Fig. 14B), MSPA eliminated the quadratic
column-attention bottleneck, reducing mean per-sample infer-
ence latency from 115.75 ms to 4.19 ms in TabPFN v3 (27.6×
mean speedup, reaching 69.5× in Savanna Pastures) and from
106.31 ms to 2.10 ms in TabICL v2 (50.6× mean speedup, reach-
ing 126.5×per sample in Savanna Pastures).
5 Discussion
5.1 Physical and Mathematical Mechanics of Contiguous
Morpho-Spectral Projection
The empirical superiority of Morpho-Spectral Projection Analy-
sis (MSPA) over both the unreduced Full Spectrum (92.0% win
rate) and the four reference statistical reducers (90.0% win rate
against the best baseline per pair) stems directly from break-
ing column-permutation invariance. Randomly permuting the
wavelength sequence of an input matrix (XΠ =XΠ) leaves the
subspaces projected by PCA, PLS-DA, UMAP, and SOM math-
ematically unchanged ( max| ZPCA(X)− ZPCA(XΠ)|≈ 10−14),
because global column inner products and Euclidean sample dis-
tances are agnostic to the physical ordering of spectral channels
[15, 11]. In plant spectroscopy across the 350–2500 nm domain,
however, physiological information is encoded in the contigu-
ous differential relationship between adjacent wavelengths: bio-
chemical absorbers alter the local curve slope ( ∂ρ/∂λ) along
absorption shoulders and sharpen the local concavity (∂2ρ/∂λ2)
at valley centers relative to flanking scattering plateaus [1, 18].
This physical principle is directly confirmed by our perturbation
saliency audit (Fig. 8C), where 73.4% of the total discriminative
power of MSPA emerged from first-derivative slope (42.9%)
and second-derivative concavity (30.5%) rather than zero-order
reflectance amplitude (26.6%).
The internal module ablation (Table 5 and Fig. 15A) clari-
fies how the three blocks of MSPA cooperate to capture this
contiguous structure. Module 3 (z can ∈R 24) serves as the
primary discriminative backbone because it projects the cou-
pled multi-order state space [X(0), X(1), X(2)] weighted by inter-
to intra-class Fisher discriminability (ωq) and regularized via
Ledoit–Wolf-type trace-scaled covariance shrinkage (γ=0.35)
[47, 21]. Unlike unsupervised PCA, whose leading eigenvec-
tors are dominated by broad NIR albedo variations, and unlike
standard PLS-DA, whose discriminative rank is bounded by
C− 1, causing subsequent latent variables in binary or ternary
tasks to fit residual intra-class noise (as evidenced by the per-
formance decay for K> 16 in Fig. 9B) [49], Module 3 extracts
regularized discriminant axes separately across each differential
order and completes the orthogonal basis according to the root
singular-energy decay (
q
σ2k/σ21) [73, 74].
Combining Module 2 (zwin∈R 24) with Module 3 consistently
improved accuracy across the foliar datasets (Table 5). Because
Module 2 extracts rank-1 singular modes of Order 0, Order
1, and Order 2 independently within eight contiguous radia-
tive sub-windows, its strictly block-diagonal loading structure
(Fig. 4C,D) acts as a local scale-protection mechanism: nar-
row pigment or SWIR absorption features of small absolute
amplitude are preserved without being masked by high-variance
NIR scattering plateaus [51, 68]. Finally, incorporating Mod-
ule 1 (z bio∈R 16) yielded the highest overall performance by
introducing nonlinear functionals that cannot be represented by
fixed linear projections (Xv). Specifically, maximum continuum-
removed band depth ( Dmax = maxλ d(λ)) and the continuous
red-edge inflection wavelength centroid (λRE) capture horizontal
wavelength shifts (blue-shiftsandred-shiftsdriven by chloro-
phyll degradation and cellular stress) [62, 4, 64], enabling MSPA
to jointly encode linear subspace projections and nonlinear
absorption-band kinematics.
5.2 Architectural Synergy With Tabular Foundation Models
and Classical Classifiers
Evaluating ten classifiers across six spectral representations
demonstrated that the impact of dimensionality reduction is gov-
erned by four distinct mathematical decision mechanisms (Ta-
bles 3 and 4). The first mechanism comprises column-attention
Tabular Foundation Models (TabPFN v3 [44] and TabICL v2
[45]), which achieved the highest accuracy directly on the
unreduced Full Spectrum ( D =1 , 580). Because their two-
dimensional induced attention and Set-Transformer encoders
learn label-conditioned feature aggregations directly across input
channels, pre-filtering the spectrum through conventional reduc-
ers (PCA, PLS-DA, UMAP, or SOM) strips away low-variance
bands and local derivative cues, causing systematic accuracy
degradation. By supplying compact variables that explicitly en-
code local amplitude, slope, concavity, and continuum-removed
absorption contrast, MSPA allowed TabPFN v3 and TabICL
v2 to outperform the Full Spectrum in Cotton, Soybean, and
Savanna Pastures, while cutting predictive Shannon entropy by
more than half and shifting both models into real-time inference
regimes (up to 69.5× and 126.5× faster; Fig. 14B). Crucially,
the computational overhead of the MSPA operator itself is negli-
gible: fitting all three modules ( fit) required 0.22–0.48 s per
fold on CPU, and out-of-sample projection (transform) added
only 0.04–0.07 ms/sample, preserving the end-to-end pipeline
speedup.
The second mechanism encompasses architectures governed
by pairwise metric distances, kernel margins, and in-context
neighborhood retrieval (k-NN, SVM-RBF, TabDPT v1.2 [46],
and TabFM). Across 1, 580 collinear bands, high-dimensional
distance concentration and redundant scattering plateaus dilute
the contribution of diagnostic wavelengths in Euclidean and
cosine metrics, causing severe performance collapse on the Full
Spectrum in multiclass regimes. Projecting the spectra into the
64 orthogonal, Fisher-weighted coordinates of MSPA restored

Preprint– Morpho-SpectralProjectionAnalysis(MSPA)forPlantSpectroscopy18
Table 6:Comparison of probabilistic uncertainty (Normalized Shannon EntropyH norm), selective error detection capability (AUROCErr), and
per-sample inference latency (ms) across representations (mean across the five plant spectroscopy datasets).
Metric Dimension Model Architecture Full Spectrum (1580) PCA (10–50) PLS-DA (8–21) SOM (64–100) MSPA (64) MSPA vs. Full Spectrum Effect
Normalized Entropy (Hnorm↓)
TabPFN v30.2031 0.2187 0.1645 0.23360.0870−57.2% uncertainty
TabICL v20.2419 0.2903 0.2034 0.25840.0876−63.8% uncertainty
TabDPT v1.20.5126 0.4106 0.2636 0.35240.1433−72.0% uncertainty
XGBoost0.4536 0.5041 0.4653 0.53180.3251−28.3% uncertainty
TabFM0.5329 0.6502 0.3808 0.39170.3798−28.7% uncertainty
Random Forest0.6388 0.7521 0.6589 0.69220.5202−18.6% uncertainty
Error Detection (AUROCErr↑)
TabPFN v30.9117 0.9026 0.8514 0.86980.9240+1.23 pp gain
TabICL v20.8853 0.8684 0.8191 0.84470.8932+0.79 pp gain
TabDPT v1.20.8194 0.8489 0.8164 0.79810.8630+4.36 pp gain
Random Forest0.7837 0.7711 0.7761 0.72890.8171+3.34 pp gain
TabFM0.7858 0.7317 0.7399 0.76150.8164+3.06 pp gain
XGBoost0.7847 0.7894 0.7850 0.69930.8119+2.72 pp gain
Inference Latency (ms/sample↓)
TabPFN v3115.7484 3.77863.50129.1245 4.186127.6×faster(up to 69.5×)
TabICL v2106.3057 1.85901.31662.2412 2.099350.6×faster(up to 126.5×)
TabDPT v1.21.1529 0.39010.37520.4533 0.39562.9×faster(up to 3.9×)
SVM-RBF(Pastures) 1.7516 0.20450.15120.2418 0.19479.0×faster
k-NN(Pastures) 0.0703 0.00520.00410.0074 0.005512.7×faster
XGBoost0.0322 0.0200 0.0182 0.01950.0158 2.0×faster
TabPFN v3 TabICL v2TabDPT v1.2 TabFMRandom Forest XGB

论文原文与出版信息

arXiv
查看论文原文

继续探索