Generalizability and transferability of machine learning models - using hyperspectral reflectance data for maize traits

Aug 4, 2026

Hyperspectral reflectance enables rapid, non-destructive phenotyping of plant leaves and supports machine learning for trait prediction, but transfer across environments remains a challenge. The study pairs hyperspectral reflectance with 25 anatomical, gas exchange, and chlorophyll fluorescence traits from 320 recombinant inbred lines across three seasons. Using nested cross-validation, it benchmarks PLSR versus SVR, compares performance under single, repeated, and calibration strategies, and tests aggregation effects on accuracy. Trait-specific generalizability varies by trait type, with structural/biochemical traits transferring best while physiological traits show reduced transferability. Results provide a rigorous benchmark and clarify opportunities and limitations for robust model generalization.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages31

Reading Time1 h 18 min

AccessFree to download

Abstract

Keywords

Introduction

Hyperspectral reflectance for high-throughput phenotyping

Machine learning for trait prediction from HSR

Challenges in generalizability and modeling choices

1 Generalizability and transferability of machine learning models 2 using hyperspectral reflectance data for maize traits 3 Rudan Xu 1,2,†, John Ferguson3,†, Matthieu Breil-Aubert4, Johannes Kromdijk4,*, and Zoran 4 Nikoloski 1,2, * 5 1Bioinformatics Department, Institute of Biochemistry and Biology, University of Potsdam, 6 Potsdam, Germany 7 2 Systems Biology and Mathematical Modelling Group, Max Planck Institute of Molecular 8 Plant Physiology, Potsdam, Germany 9 3 School of Life Sciences, University of Essex, Colchester, UK 10 4Department of Plant Sciences, University of Cambridge, Cambridge, UK 11 †These authors contributed equally. 12 *Corresponding authors. 13 14 Email address: 15 [xu2@uni-potsdam.de](xu2@uni-potsdam.de) 16 [jfergu@essex.ac.uk](jfergu@essex.ac.uk) 17 [mb2651@cam.ac.uk](mb2651@cam.ac.uk) 18 [jk417@cam.ac.uk](jk417@cam.ac.uk) 19 [nikoloski@mpimp-golm.mpg.de](nikoloski@mpimp-golm.mpg.de)[ ](nikoloski@mpimp-golm.mpg.de)20 21 22 Keywords 23 Hyperspectral reflectance, machine learning, Zea mays, model generalizability, anatomical traits, 24 chlorophyll fluorescence, gas exchange. 25 27 Abstract 28 Hyperspectral reflectance provides rapid, non-destructive phenotyping of plant leaves. These data 29 have been used to develop machine learning models for predicting diverse plant traits, yet key 30 challenges remain. We collected hyperspectral reflectance data together with 25 anatomical, gas 31 exchange, and chlorophyll fluorescence traits from 320 recombinant inbred lines grown over three 32 seasons. Using these data, we systematically (1) compare the performance of PLSR and SVR 33 across a wide range of traits, including also slow fluorescence kinetics, (2) assess model 34 generalizability and transferability, and (3) investigate how different aggregation strategies affect 35 predictive accuracy. Based on a nested cross-validation framework, single cross-validation with 36 MSE as metric performed comparably to repeated cross-validation or PRESS-based calibration. 37 Optimal performance of trait-specific predictions was found to be dependent on the combination 38 of model and data aggregation levels. Structural and biochemical traits showed the best 39 generalizability and transferability, whereas physiological traits, particularly those derived from 40 gas exchange and fluorescence kinetics, exhibited markedly reduced transferability. Together, 41 these results provide a rigorous benchmark for evaluating machine learning models for trait 42 prediction from hyperspectral reflectance data, and highlight both the opportunities and limitations 43 for achieving robust generalization across diverse environments and genotypes. 44 45 Abbreviations 46 HSR, hyperspectral reflectance; partial least squares regression, PLSR; number of component, 47 NoC; support vector regression, SVR; A-Ci curve, photosynthetic rate versus internal CO2 48 concentration curve; phosphoenolpyruvate carboxylase, PEPC; maximum rate of carboxylation by 49 PEPC, Vpmax ; asymptote of theA-Ci curve, Vmax ; photosystem II, PSII; quantum efficiency of PSII, 50 Fv/Fm; non-photochemical quenching, NPQ; PSII operating efficiency, ΦPSII; relative C 51 content,%C; relative N content,%N; stable carbon isotope ratio, δ 13C; stable nitrogen isotope 52 ratio, δ 15N; intrinsic water use efficiency, iWUE; stomatal conductance, gsw; coefficient of 53 determination, R2; cross-validation, CV; coefficients of variation, CVs; standard deviation, SD. 54 Introduction 55 Efforts to develop climate-resilient crops go hand-in-hand with advances in high-throughput 56 phenotyping technologies that can be deployed at scale in accelerated breeding 1,2 . Urgent 57 innovations in high-throughput profiling are needed given that many breeding targets involve 58 molecular and biochemical traits that are difficult to measure across large breeding populations3. 59 Hyperspectral refl

FAQs

What data were collected for training and evaluating the machine learning models?

Hyperspectral reflectance data from maize leaves were collected together with 25 traits covering anatomical traits, gas exchange, and chlorophyll fluorescence from 320 recombinant inbred lines across three seasons.

How were PLSR and SVR model performance compared?

The study benchmarks PLSR and SVR using a nested cross-validation framework and evaluates performance with mean squared error, comparing single cross-validation, repeated cross-validation, and PRESS-based calibration.

What factors most affected transferability of predictions?

Transferability depended on the combination of model type and the data aggregation level. Structural and biochemical traits showed better generalizability, while physiological traits—especially those derived from gas exchange and fluorescence kinetics—exhibited reduced transferability.

Aug 4, 2026

Upload by Valentina