Efficient VNIR Hyperspectral Models for Estimating Pigment and Fluorescence Traits in Vertically Farmed Lettuce

Nondestructive estimation of pigment content and chlorophyll fluorescence traits is valuable for vertical farming. However, hyperspectral models for fluorescence traits have rarely been tested on vertically farmed crops. This study evaluated a VNIR hyperspectral lettuce dataset with a fully auditable workflow-based on white-normalized and leaf-masked reflectance cubes acquired between 400 and 1000 nm. The final analysis used 113 matched leaf samples from 57 plants grown in nutrient film technique and aeroponic systems. Model evaluation was performed with five-fold grouped cross-validation by plant, so that leaves from the same plant were never split between training and test folds. Models were selected only within the training folds, and the whole procedure was repeated ten times. A growing system holdout was used as a robustness check. The study focuses on four targets selected by the agronomical team as the main variables of interest: chlorophyll, anthocyanins, PItotal (overall photosynthetic performance), and ET0/RC (electron transport flux per reaction center). The main scientific claim is based on spectral-only models, while spectral plus metadata models are reported separately as operational extensions. The most consistent result was obtained for ET0/RC with an interval partial least squares model built on leaf core reflectance in the 580–620 nm region, reaching grouped out-of-fold R2 = 0.76 and between growing system R2 = 0.46 and 0.59. Anthocyanins showed a promising within-dataset result, although transfer between growing systems was limited, with the best spectral-only model obtained from a core-only leaf summary (R2 = 0.55). Chlorophyll and PItotal were only moderately predictable (R2 = 0.38 and 0.41, respectively) and are presented as supporting results rather than headline findings. For all four targets, spectral models performed better than models using only cultivation data. These results suggest that careful preprocessing, plant-grouped validation, and compact latent variable models can recover useful VNIR signals for selected lettuce traits, although predictive performance remains strongly target dependent and should be confirmed on other cultivars and growing systems.