Four numbers, one axis: deep learning models reveal what leaf spectrum constrains about Farquhar-von Caemmerer-Berry photosynthesis

Abstract

Leaf reflectance spectra are emerging as a viable substitute for gas-exchange measurements of photosynthetic capacity, with a community benchmark reporting that a spectrum accurately recovers most Farquhar-von Caemmerer-Berry (FvCB) parameters. This study re-scores the recovery under dataset-blocked, species-blocked, and leave-one-dataset-out designs, measuring the split-half reliability of each curated parameter. We constructed a convolutional encoder that maps a spectrum to the four parameters through a fixed, differentiable FvCB decoder trained on measured assimilation. A conspecific of 97.4% of held-out leaves were present in the training set, and accuracy is lost along the dataset axis but not along the species axis. Under blocked evaluation, a spectrum constrains a single capacity axis. Jmax25 retains only 17% of its recovery when Vcmax25 is held constant, and the Jmax25:Vcmax25 ratio is not predicted above a median null. The curated values of TPU25 are not reproducible, whereas those of Rday25 are well determined, but its recovery fails due to the loss. The published study measures interpolation rather than transfer, and spectra constrain less of the FvCB parameter space than assumed, including the carboxylation to electron transport balance. Routing predictions through explicit biochemistry makes identifiability measurable, although it does not improve prediction accuracy.