A reproducible and interpretable workflow for leaf hyperspectral phenotyping under limited sample conditions with hierarchical validation and cross-stage evaluation

Abstract Leaf hyperspectral phenotyping is widely used for non-destructive trait estimation in plants, but datasets with limited biological samples relative to high-dimensional spectral variables remain methodologically challenging because of strong band redundancy, limited robustness under developmental shift, and weak interpretability of feature-importance analyses. Here we present a reproducible and interpretable workflow for SPAD-based leaf chlorophyll status phenotyping from hyperspectral reflectance, using oat ( Avena sativa L.) as a representative use case. The workflow integrates spectral preprocessing benchmarking, two-stage band selection (Lasso screening followed by successive projections refinement), tabular deep learning optimized for structured spectral predictors (TabM-v2 with Mixup and stochastic weight averaging), and a multi-level interpretation scheme spanning feature, region, and dependency analyses. Using 1,142 retained ROI-level modeling samples derived from 200 sampled oat leaves (125 bands, 400–1000 nm) and five independent growth-stage validation batches, we show that the workflow achieves strong predictive performance ( $$R^2 = 0.8575 \pm 0.0021$$ R 2 = 0.8575 ± 0.0021 ) while reducing spectral dimensionality by 80%. To address the hierarchical structure of ROI-level measurements, we further performed strict leaf-grouped validation in which all ROIs from the same reconstructed leaf identity were assigned exclusively to either training or testing sets; performance remained comparable to ROI-level random splitting, indicating that within-leaf pseudo-replication did not substantially inflate reported accuracy. Comparative benchmarking indicated that, under the present limited-sample structured spectral regime, TabM-v2 achieved the highest in-distribution accuracy, whereas the compact Lasso+SPA+SVR route provided a more parsimonious alternative with stronger cross-stage transferability at intermediate growth stages, although all models showed limited transferability at the maturity stage ( $$R^2 < 0.15$$ R 2 0.15 ), defining a critical operational boundary. Interpretation analyses consistently identified green, red-edge, and near-infrared regions as dominant contributors, with the red-edge acting as the central dependency hub. Independent cross-stage batch validation defined the operational boundary of the workflow, with strongest transferability at intermediate growth stages and reduced robustness under marked developmental shift. Although the reference trait was SPAD-based chlorophyll status rather than absolute biochemical chlorophyll concentration, the present study provides a reproducible methodological workflow for leaf hyperspectral phenotyping under limited sample conditions, with a clearly defined validation boundary, with broader validation across genotypes, sites, years, and devices remaining necessary.