Near-infrared spectroscopy is increasingly used as a rapid, non-destructive chemical sensing technology for the analysis of food, pharmaceutical, biological, and environmental samples. However, the practical deployment of NIR sensors still depends on calibration models able to handle high-dimensional, collinear spectra, limited sample sizes, preprocessing dependence, spectral outliers, and extrapolation beyond the calibration domain. Here, we evaluate whether tabular foundation models can provide a new calibration strategy for NIR chemical sensing. We benchmark TabPFN on 66 NIR datasets covering 54 regression and 12 classification tasks, and compare direct inference on raw spectra with preprocessing-optimized inference against PLS/PLS-DA, Ridge, Catboost, and one-dimensional convolutional neural networks. The study uses a unified validation framework in which preprocessing and model selection are performed exclusively on calibration data before external test evaluation. In regression, preprocessing-optimized TabPFN achieves the best overall average rank and significantly outperforms PLS, CatBoost, TabPFN on raw spectra, and CNN-1D, while remaining statistically comparable to Ridge. In classification, TabPFN applied directly to raw spectra provides the best average rank, with performance close to the optimized variant. Robustness analyses show that TabPFN provides strong average predictive performance but that its advantage decreases on spectral outliers and extrapolated samples, where classical chemometric models remain competitive. These results suggest that tabular foundation models can complement established chemometric workflows for NIR chemical sensing, especially in small- to medium-sized calibration settings, while highlighting the need for spectroscopy-specific priors and uncertainty-aware deployment strategies.

Preprocessing screening is often the most expensive part of a near-infrared spectroscopy calibration workflow. It works because smoothing, derivatives, detrending and related filters change the spectral directions seen by partial least squares (PLS) or Ridge regression, but a full external search repeatedly refits nearly the same linear model. This paper studies the case where that search can be collapsed into one calibration step. For a strict linear preprocessing operator A acting on row spectra as XA^T, the transformed PLS cross-covariance satisfies (XA^T)^T Y = A X^T Y, and Ridge regression depends on the operator-induced kernel X A^T A X^T. These identities let a finite operator bank be screened inside the model while retaining original-wavelength coefficients, and the same identity extends to cheaply evaluated linear operator chains. Sample-adaptive or fitted corrections such as SNV, MSC, EMSC and ASLS are not strict linear; we prove the boundary and keep them as fold-local branches. The cohort has 61 regression and 17 classification rows, with a strict paired regression denominator of N=32 for the eight paper variants. There, AOM-PLS reaches median RMSEP ratios of 0.991/0.990 (simple) and 0.985/1.002 (best) against PLS-default/PLS-HPO, and AOM-Ridge reaches 0.974/0.984 (simple) and 0.918/0.966 (best) against Ridge-default/Ridge-HPO. The operator-adaptive classifier AOM-PLS-DA improves balanced accuracy by a median 0.159 on N=13 datasets (12/13 wins). The practical result is the runtime gap: PLS-HPO takes a median 710.81 s per run, whereas AOM-PLS takes 1.18-1.63 s – 436 to 602 times less PLS fitting time. Linear operator-adaptive calibration thus gives prediction quality comparable to exhaustive preprocessing screening, with orders-of-magnitude less fitting time for PLS.

A published near-infrared calibration is rarely a fully reproducible one. Its reported error depends on analyst decisions about signal representation, scatter correction, replicate handling, validation design, latent-variable selection, aggregation and model family. These decisions usually survive as scripts and output files rather than as one re-executable calibration object. We present nirs4all, an open-source Python system for making that object explicit. The contribution is a computer technique for chemometric reproducibility, not a new calibration algorithm. nirs4all records preprocessing and latent-variable search as canonical workflow declarations, links predictions and fitted artefacts back to the exact fold, chain and dataset hashes that produced them, and makes validation semantics such as signal type, replicate-grouped splitting and specimen-level aggregation first-class. A controller layer lets standard Python estimators and near-infrared operators share this workflow, provenance and deployment contract. On public datasets we show deterministic representation and verifiable provenance under a pinned environment, held-out partial-least-squares reporting isolated from selection, auditable replicate policy, and bounded operational cost. This is reproducibility of the declaration and lineage; strict bitwise numerical replay still requires a pinned environment or container. nirs4all is available at https://github.com/GBeurier/nirs4all (pip install nirs4all, AGPL-3.0-or-later primary public licence).

While the green revolution adapted a handful of crops to homogenous and high-input industrialized agriculture, much of the global population still relies on local food production from low-input smallholder farms that grow highly variable crop cultivars. The high diversity of the grain and bioenergy crop sorghum [1][1]–[4][2], and many other crops that were not homogenized during the green revolution [5][3], not only provides the raw materials for breeders to make substantial gains in cultivar improvement, but also constrains breeding efforts due to highly specialized locally adapted plant phenotypes [6][4]. Here, we construct a 33-member pangenome and identify trait-associated variants in 1,988 cultivars and landraces. We then apply these resources to explore the complex interplay between historical contingency, ongoing adaptation, and the potential for future gains through climate-aware genome-enabled breeding. Specifically, our analyses conclusively demonstrate that multiple nested, deeply diverged, and previously uncharacterized structural variants in the domestication gene SHATTERING1 distinguish the previously established multicentric origin of sorghum. We then apply landscape genomics tests to reveal how gene flow, adaptation, and secondary contact created the complex genetic mosaic in current global breeding networks. Further analysis of climate-gene associations highlights candidate loci underlying adaptation, including the biosynthetic gene cluster for the cyanogenic glucoside dhurrin. Combined, the pangenome-informed variants developed here will enable both trait discovery and subsequent marker assays to accelerate breeding and provide a framework for similar applications in other diverse and non-model crops. ### Competing Interest Statement The authors have declared no competing interest. United States Department of Energy, DE-AC02-05CH11231 [1]: #ref-1 [2]: #ref-4 [3]: #ref-5 [4]: #ref-6

Why do we observe some plant phenotypes but not others? The multivariate phenotypic space occupied by individuals or species often reveals both limits and phenotypes strikingly deviating from main syndromes. These observations are usually thought to indicate, respectively, inviable trait combinations and unique phenotypes adapted to specific environments. However, the evolutionary drivers underlying trait covariations often remain unclear. Here, we characterized the phenotypic space of Arabidopsis thaliana by comparing 713 wild accessions collected across the globe with 2,544 artificially-created recombinant individuals. This, combined with the detection of adaptive processes operating within species, allowed us to elucidate the roles of natural selection as a driver of phenotypic (co)variations within A. thaliana. We found that the phenotypic space of this species is constrained and driven by varying levels of divergent and stabilizing selection across different traits. Moreover, at the margins of the European geographic range, strong directional selection favored outlier phenotypes characterized by very late flowering and variation in a WRKY transcription factor gene. Genome analyses revealed that these extreme phenotypes may be explained by hybridization between ancestral and modern lineages of A. thaliana. Our findings demonstrate how interplays between population history and natural selection shape phenotypic diversity in a plant species.

Near-infrared spectroscopy (NIRS) has become a popular tool for investigating phenotypic variability in plants. We developed the Shiny NIRSpredict application to get predictions of 81 Arabidopsis thaliana phenotypic traits, including classical functional traits as well as a large variety of commonly measured chemical compounds, based from near-infrared spectroscopy values based on deep learning. It is freely accessible at the following URL: https://shiny.cefe.cnrs.fr/NirsPredict/ . NIRSpredict has three main functionalities. First, it allows users to submit their spectrum values to get the predictions of plant traits from models built with the hosted A. thaliana database. Second, users have access to the database of traits used for model calibration. Data can be filtered and extracted on user's choice and visualized in a global context. Third, a user can submit his own dataset to extend the database and get part of the application development. NIRSpredict provides an easy-to-use and efficient method for trait prediction and an access to a large dataset of A. thaliana trait values. In addition to covering many of functional traits it also allows to predict a large variety of commonly measured chemical compounds. As a reliable way of characterizing plant populations across geographical ranges, NIRSpredict can facilitate the adoption of phenomics in functional and evolutionary ecology.

IntroductionPredicting the performance (yield or other integrative traits) of cultivated plants is complex because it involves not only estimating the genetic value of the candidates to selection, the interactions between the genotype and the environment (GxE) but also the epistatic interactions between genomic regions for a given trait, and the interactions between the traits contributing to the integrative trait. Classical Genomic Prediction (GP) models mostly account for additive effects and are not suitable to estimate non-additive effects such as epistasis. Therefore, the use of machine learning and deep learning methods has been previously proposed to model those non-linear effects.MethodsIn this study, we propose a type of Artificial Neural Network (ANN) called Convolutional Neural Network (CNN) and compare it to two classical GP regression methods for their ability to predict an integrative trait of sorghum: aboveground fresh weight accumulation. We also suggest that the use of a crop growth model (CGM) can enhance predictions of integrative traits by decomposing them into more heritable intermediate traits.ResultsThe results show that CNN outperformed both LASSO and Bayes C methods in accuracy, suggesting that CNN are better suited to predict integrative traits. Furthermore, the predictive ability of the combined CGM-GP approach surpassed that of GP without the CGM integration, irrespective of the regression method used.DiscussionThese results are consistent with recent works aiming to develop Genome-to-Phenotype models and advocate for the use of non-linear prediction methods, and the use of combined CGM-GP to enhance the prediction of crop performances.

Data from functional trait databases have been increasingly used to address questions related to plant diversity and trait-environment relationships. However, such databases provide intraspecific data that combine individual records obtained from distinct populations at different sites and, hence, environmental conditions. This prevents distinguishing sources of variation ( e.g ., genetic-based variation vs . phenotypic plasticity), a necessary condition to test for adaptive processes and other determinants of plant phenotypic diversity. Consequently, individual traits measured under common growing conditions and encompassing within-species variation across the occupied geographic range have the potential to leverage trait databases with valuable data for functional and evolutionary ecology. Here, we recorded 16 functional traits and leaf hyperspectral reflectance (NIRS) data for 721 widely distributed Arabidopsis thaliana natural accessions grown in a common garden experiment. These data records, together with meteorological variables obtained during the experiment, were assembled to create the AraDiv dataset. AraDiv is a comprehensive dataset of A. thaliana ’s intraspecific variability that can be explored to address questions at the interface of genetics and ecology.

BACKGROUND:Yam (Dioscorea alata L.) is the staple food of many populations in the intertropical zone, where it is grown. The lack of phenotyping methods for tuber quality has hindered the adoption of new genotypes from breeding programs. Recently, near-infrared spectroscopy (NIRS) has been used as a reliable tool to characterize the chemical composition of the yam tuber. However, it failed to predict the amylose content, although this trait is strongly involved in the quality of the product. RESULTS:This study used NIRS to predict the amylose content from 186 yam flour samples. Two calibration methods were developed and validated on an independent dataset: partial least squares (PLS) and convolutional neural networks (CNN). To evaluate final model performances, the coefficient of determination (R2), the root mean square error (RMSE), and the ratio of performance to deviation (RPD) were calculated using predictions on an independent validation dataset. The tested models showed contrasting performances (i.e., R2 of 0.72 and 0.89, RMSE of 1.33 and 0.81, RPD of 2.13 and 3.49 respectively, for the PLS and the CNN model). CONCLUSION:According to the quality standard for NIRS model prediction used in food science, the PLS method proved unsuccessful (RPD < 3 and R2 < 0.8) for predicting amylose content from yam flour but the CNN model proved to be reliable and efficient method. With the application of deep learning methods, this study established the proof of concept that amylose content, a key driver of yam textural quality and acceptance, can be predicted accurately using NIRS as a high throughput phenotyping method. © 2023 The Authors. Journal of The Science of Food and Agriculture published by John Wiley & Sons Ltd on behalf of Society of Chemical Industry.

Background and Aims Determining within-species large-scale variation in phenotypic traits is central to elucidate the drivers of species’ ranges. Intraspecific comparisons offer the opportunity to understand how trade-offs and biogeographical history constrain adaptation to contrasted environmental conditions. Here we test whether functional traits, ecological strategies and phenotypic plasticity in response to abiotic stress vary along a latitudinal or a center-margins gradient within the native range of Arabidopsis thaliana . Methods The phenotypic outcomes of plant adaptation at the center and margins of its geographic range were experimentally examined in 30 accessions from southern, central and northern Europe. The variation of traits related to stress tolerance, resource use, colonization ability as well as survival and fecundity was determined in response to high temperature (34°C) or frost (- 6°C), in combination with response to water deficit. Key Results Both evidence for a latitudinal and a center-margins differentiation was found. Traits related to the acquisitive/conservative strategy trade-off varied along a latitudinal gradient. Northern accessions presented a greater survival to stress than central and southern accessions. Traits related to a colonization-competition trade-off followed a center-margin differentiation. Central accessions presented a higher phenotypic plasticity and trait values associated with a higher colonization ability than northern and southern accessions which instead had a higher competition ability. Conclusions Intraspecific phenotypic variation helps us understand how the distribution range has evolved in Arabidopsis thaliana , which is shaped both by climate and the population migratory history. We advocate to consider intraspecific trait variation in species range studies instead of species means only as classically done in macroecology.

The trait-based approach in plant ecology aims at understanding and classifying the diversity of ecological strategies by comparing plant morphology and physiology across organisms. The major drawback of the approach is that the time and financial cost of measuring the traits on many individuals and environments can be prohibitive. We show that combining near-infrared spectroscopy (NIRS) with deep learning resolves this limitation by quickly, non-destructively, and accurately measuring a suite of traits, including plant morphology, chemistry, and metabolism. Such an approach also allows to position plants within the well-known CSR triangle that depicts the diversity of plant ecological strategies. The processing of NIRS through deep learning identifies the effect of growth conditions on trait values, an issue that plagues traditional statistical approaches. Together, the coupling of NIRS and deep learning is a promising high-throughput approach to capture a range of ecological information on plant diversity and functioning and can accelerate the creation of extensive trait databases.

Meeting food demand for the growing population will require an increase to crop production despite climate changes and, more particularly, severe drought episodes. Sorghum is one of the cereals most adapted to drought that feed millions of people around the world. Valorizing its genetic diversity for crop improvement can benefit from extensive phenotyping. The current methods to evaluate plant biomass, leaves area and plants height involve destructive sampling and are not practical in breeding. Phenotyping relying on drone based imagery is a powerful approach in this context. The objective of this study was to develop and validate a high throughput field phenotyping method of sorghum growth traits under contrasted water conditions relying on drone based imagery. Experiments were conducted in Bambey (Senegal) in 2018 and 2019, to test the ability of multi-spectral sensing technologies on-board a UAV platform to calculate various vegetation indices to estimate plants characteristics. In total, ten (10) contrasted varieties of West African sorghum collection were selected and arranged in a randomized complete block design with three (3) replicates and two (2) water treatments (well-watered and drought stress). This study focused on plant biomass, leaf area index (LAI) and the plant height that were measured weekly from emergence to maturity. Drone flights were performed just before each destructive sampling and images were taken by multi-spectral and visible cameras. UAV-derived vegetation indices exhibited their capacity of estimating LAI and biomass in the 2018 calibration data set, in particular: normalized difference vegetative index (NDVI), corrected transformed vegetation index (CTVI), seconded modified soil-adjusted vegetation index (MSAVI2), green normalize difference vegetation index (GNDVI), and simple ratio (SR) (r2 of 0.8 and 0.6 for LAI and biomass, respectively). Developed models were validated with 2019 data, showing a good performance (r2 of 0.92 and 0.91 for LAI and biomass accordingly). Results were also promising regarding plant height estimation (RMSE = 9.88 cm). Regression plots between the image-based estimation and the measured plant height showed a r2 of 0.83. The validation results were similar between water treatments. This study is the first successful application of drone based imagery for phenotyping sorghum growth and development in a West African context characterized by severe drought occurrence. The developed approach could be used as a decision support tool for breeding programs and as a tool to increase the throughput of sorghum genetic diversity characterization for adaptive traits.

Despite the importance of yam (Dioscorea spp.) tuber quality traits, and more precisely texture attributes, high-throughput screening methods for varietal selection are still lacking. This study sets out to define the profile of good quality pounded yam and provide screening tools based on predictive models using near infrared reflectance spectroscopy. Seventy-four out of 216 studied samples proved to be moldable, i.e. suitable for pounded yam. While samples with low dry matter (<25%), high sugar (>4%) and high protein (>6%) contents, low hardness (<5 N), high springiness (>0.5) and high cohesiveness (>0.5) grouped mostly non-moldable genotypes, the opposite was not true. This outline definition of a desirable chemotype may allow breeders to choose screening thresholds to support their choice. Moreover, traditional near infrared reflectance spectroscopy quantitative prediction models provided good prediction for chemical aspects (R-2 > 0.85 for dry matter, starch, protein and sugar content), but not for texture attributes (R-2 < 0.58). Conversely, convolutional neural network classification models enabled good qualitative prediction for all texture parameters but hardness (i.e. an accuracy of 80, 95, 100 and 55%, respectively, for moldability, cohesiveness, springiness and hardness). This study demonstrated the usefulness of near infrared reflectance spectroscopy as a high-throughput way of phenotyping pounded yam quality. Altogether, these results allow for an efficient screening toolbox for quality traits in yams.

The integrative capacity of crop models is of great value to identify in-silico optimal combinations of traits (ideotypes) and traits x cultu ral practices in targeted agro-environments. This approach becomes even more challenging when considering the multiple environ mental factors constituting future agro-climatic scenarios : increasing stress frequency and severity (e.g., heat, drought, wind, flooding); enhanced atmospheric C02 concentration (e-COzl, and also adoption of more sustainable and resilient cultural practices (agroecology) combining of productivity with ecosystem services. Several studies reported the weaknesses of crop models in predicting crop performance in response to climate change. While these limitations were until now mainly explained by poor simulation by crop models of physiological responses to heat and drought, crop model shortcomings of the representation of Carbon (C) source-sink relations and interactions, involving phenotypic plasticity of both source and sink, should explain this limitations and have received less attention (Chang and Zhu 2017). Recent studies reported a down-regulation of C source capacity (i.e. photosynthesis) in C3 crops under e-C02 by sink lim itation in the afternoon, involving low TPU levels (Triose Phosphate Utilization) (rice: (Fabre et al. 2019). (Fabre, in prep) indicated that high constitutive source-sink ratios increase photosynthesis under e-C02 • Finally, (Kikuchi et al. 2017) demonstrated in a FACE trial that rice plants with high adaptive plasticity of tillering and panicle size respond better to e-C02 • Although particularly relevant to C3 crops that respond strongly to e-C02 , this also applies to C4 crops when they are C-sink limited (Oszvald et al., 2018). Therefore, Carbon source-sink relationships and their physiological and morphological adaptability (feedbacks) are pivotai in predicting crop ideotypes in a climate change context. ln addition, the agro-ecological transition needs better crop models to design solutions for improving (1) crop energy and carbon use, (2) resilience under abiotic stresses, and (3) ecosystem services such as channeling assimilates into the roots/soil (4p1000 initiative: C sequestration and soil improvement). This implies to further model plantplant and/or plant-soil interactions and related impacts on C source-sink relationships and competitions for (light) resources. Crop models should indeed be able to predict trade-offs among several crop performance objectives such as multiple production purposes (e.g., grain and biomass), between potential productivity and adaptation, and between productivity and ecosystem services such as C sequestration into the soil. For this, we will provide examples of analytical and modeling concepts. More quantitative, extrapolatable evidence is needed to understand the importance of C source and sink traits, their adaptive plasticity as they interact during plant development, and their impact on crop performance under anticipated agro-climatic conditions. This requires a dialogue between experimental and modeling research for which we are presenting concepts here. Our laboratory focuses on rice (C3) and sorghum (C4) model cereals using crop models simulating sink- and source driven phenotypic plasticity, namely SAMARA (Kumar et al., 2016) and Ecomeristem (Larue et al., 2019). Sorne (experimental, modelling) preliminary resu lts will be shown but a broader dynamics is needed. Once further improved, the models will be used to (i) estimate in silico the prediction errors caused by ignoring source-sink feedbacks and plasticity; (ii) predict the potential of improved trait combinations and plasticity on the performance of future crops; and (iii) propose how existing, generic crop models should be improved and what type of data will be needed for that.

Crop growth modelling formalizes the biological processes at which genotype X environment interactions (GxE) are expressed. It has the potential to evaluate, in silico, the effect of elementary traits and related genetic factors on phenotype and yield elaborations. Recent studies showed that, by driving the value of crop model parameters using a genomic selection model, yield was predicted more accurately than by a classical genetic model. However these studies dealt with few, integrative parameters and a narrow genetic diversity, i.e., a reduced number of molecular markers. This contrasts with the necessity to make crop models more responsive to climate change variables and thus increase the number of physiological parameters, while studying wider genetic diversity to seek for adaptive markers. With this respect, methods that reduce the dimensionality of the problem are needed. The autoencoder, a semi-supervised machine learning method, can reduce the number of predictors (markers) without prior information, by compressing input data into an encoded neural network layer. Applied to genetic diversity, it should ease modelling and predicting the genetic value of crop model parameters underlying GxE and yield variability, compared to classical regression methods. This study aimed at testing the autoencoding of the genetic data (ca. 1.5M markers) within a West-African sorghum diversity panel of 200 individuals. It evaluated then the relevance of autoencoded data to predict the genetic value of the dozen of crop/plant parameters controlling growth and plasticity in Ecomeristem model, estimated using data from an experiment in the Phenoarch platform.

Comparative analysis of high throughput sequencing data between multiple conditions often involves mapping of sequencing reads to a reference and downstream bioinformatics analyses. Both of these steps may introduce heavy bias and potential data loss. This is especially true in studies where patient transcriptomes or genomes may vary from their references, such as in cancer. Here we describe a novel approach and associated software that makes use of advances in genetic algorithms and feature selection to comprehensively explore massive volumes of sequencing data to classify and discover new sequences of interest without a mapping step and without intensive use of specialized bioinformatics pipelines. We demonstrate that our approach called GECKO for GEnetic Classification using k-mer Optimization is effective at classifying and extracting meaningful sequences from multiple types of sequencing approaches including mRNA, microRNA, and DNA methylome data.

BACKGROUND AND AIMS:Plant modelling can efficiently support ideotype conception, particularly in multi-criteria selection contexts. This is the case for biomass sorghum, implying the need to consider traits related to biomass production and quality. This study evaluated three modelling approaches for their ability to predict tiller growth, mortality and their impact, together with other morphological and physiological traits, on biomass sorghum ideotype prediction.METHODS:Three Ecomeristem model versions were compared to evaluate whether tillering cessation and mortality were source (access to light) or sink (age-based hierarchical access to C supply) driven. They were tested using a field data set considering two biomass sorghum genotypes at two planting densities. An additional data set comparing eight genotypes was used to validate the best approach for its ability to predict the genotypic and environmental control of biomass production. A sensitivity analysis was performed to explore the impact of key genotypic parameters and define optimal parameter combinations depending on planting density and targeted production (sugar and fibre).KEY RESULTS:The sink-driven control of tillering cessation and mortality was the most accurate, and represented the phenotypic variability of studied sorghum genotypes in terms of biomass production and partitioning between structural and non-structural carbohydrates. Model sensitivity analysis revealed that light conversion efficiency and stem diameter are key traits to target for improving sorghum biomass within existing genetic diversity. Tillering contribution to biomass production appeared highly genotype and environment dependent, making it a challenging trait for designing ideotypes.CONCLUSIONS:By modelling tiller growth and mortality as sink-driven processes, Ecomeristem could predict and explore the genotypic and environmental variability of biomass sorghum production. Its application to larger sorghum genetic diversity considering water deficit regulations and its coupling to a genetic model will make it a powerful tool to assist ideotyping for current and future climatic scenario.

La calibration d'un nouveau modele de prediction base sur des spectres proche infrarouge, requiert une bonne dose de savoir expert et beaucoup de patience. En effet la grande diversite de pretraitements des spectres et des modeles de calibration offre un nombre tres eleve de combinaisons a tester. Et meme si le savoir expert peut aider a orienter les choix de combinaisons, l'identification du meilleur modele reste tres empirique et repose sur la comparaison d'un grand nombre de tentative. Traditionnellement, la construction des modeles et leur comparaison est realisee manuellement par des experts. Cependant, la democratisation des spectrometres et leur vulgarisation amene un nombre grandissant d'utilisateurs provenant de disciplines souvent eloignee de la chimiometrie (e.g. ameliorateur, agronome). Leur objectif n'est pas toujours d'identifier le meilleur modele mais bien de construire, sans savoir expert et avec un minimum d'investissement en temps, un modele predictif remplissant une exigence de performance connue prealablement. Dans ce contexte particulier, et afin de tirer le meilleur parti des avancees recentes en science des donnees et en capacites de calcul, cette communication presente les arguments theoriques et methodologiques justifiant la construction d'un pipeline de calibration performant mais surtout suffisamment generique pour s'appliquer a une grande diversite d'analytes, de spectrometres, de taille de jeux de donnees et de produits. Le pipeline doit permettre de combiner les avantages des approches traditionnelles et les techniques modernes d'apprentissage profond. Ainsi les algorithmes spatio-temporels dedies a l'analyse de signaux 2D et 3D seront utilises afin de mieux integrer les informations de position des spectres. En outre, les ameliorations apportees aux algorithmes de recherche heuristique offrent la possibilite d'etudier un plus vaste espace de combinaisons de pretraitements des spectres, de type de modeles et d'hyper-parametres. Afin de tirer parti de cette diversite d'approches, nous utiliserons des techniques d'assemblage de modeles (i.e. stacking). Ces methodes permettront de capitaliser les informations complementaires apportees par les multiples modeles de base au sein d'un meta-modele. Enfin, la genericite d'un tel pipeline offrira la possibilite d'effectuer une calibration directement sur un vecteur d'analytes plutot que de les traiter un par un. Cette approche permettra de mieux gerer les valeurs aberrantes et le bruit de fond, et d'utiliser les possibles liens entre variables (e.g. covariance). Cette communication presente la structure du pipeline et les premiers resultats appliques dans le cadre du projet RTBfoods (BMGF), axes sur les cultures des plantes a racines et tubercules et des bananes en Afrique subsaharienne.