High-Throughput Estimation of Photosynthetic Phenotypic Parameters Using Hyperspectral Data

Mengqi Zhang, Jiaren Zhou, Man Zhang, Minjuan Wang; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 7169-7177

Photosynthetic efficiency is a key determinant of agricultural productivity, quantifiable through various photosynthetic phenotypic parameters (PPPs). A comprehensive understanding of a plant's physiological status, however, requires the simultaneous monitoring of multiple, interrelated PPPs, a task traditional hyperspectral methods and deep learning approaches have rarely addressed. To overcome this limitation, we developed SpectralNet, an end-to-end deep learning model for the simultaneous, nondestructive estimation of five key rice PPPs: chlorophyll a, chlorophyll b, SPAD values, net photosynthetic rate, and leaf nitrogen content. This parameter suite was chosen to provide a holistic evaluation of photosynthetic capacity. SpectralNet achieved a robust average R2 of 0.743, with individual parameter R2 values all exceeding 0.630. The superior accuracy and robustness of the model were validated through ablation studies and comparative experiments, with a statistical analysis (Mann-Whitney test, p>0.05) confirming the reliability of its estimations. Using these accurate PPP estimations, we evaluated the impact of different rice cultivars and nitrogen management practices on photosynthetic performance. This study underscores the potential of proximal hyperspectral imaging and end-to-end deep learning as a valuable high-throughput tool for PPP acquisition, offering significant advantages for breeding high-yield rice cultivars.