📈 Leaf Spectral Reflectance Prediction Using Multi-Head Attention

Parastoo Farajpoor, Alireza Pourreza, Mohammadreza Narimani, Ashraf El-Kereamy, and Matthew W. Fidelibus

Proc. SPIE 13475 · 134750V · 28 May 2025 · Autonomous Air and Ground Sensing Systems for Agricultural Optimization and Phenotyping X

Digital Agriculture Laboratory · University of California, Davis · University of California, Riverside

Read the SPIE paper

Trait-to-Spectra, Not Spectra-to-Trait

Most phenotyping models estimate chemistry from reflectance. This SPIE paper reverses the mapping: it predicts full leaf spectra from 16 biochemical and structural traits, so grapevine-specific radiative behavior can replace a generic PROSPECT-PRO forward run when species matter.

Leaves were collected from Sunpreme, Flame Seedless, and Solbrio at Kearney Agricultural Research and Extension Center (2021–2023): 2,305 leaves (992 unique samples) across canopy heights and phenology. Spectra (SVC HR-1024i, 400–2500 nm) were interpolated to 1 nm and Savitzky–Golay smoothed. Missing Nstruct, anthocyanins, carotenoids, and some chlorophyll values were filled with inverse PROSPECT-PRO after a chlorophyll scale of 0.75.

Core story: a grapevine-specific multi-head attention network predicts leaf reflectance from traits with mean R² 0.84 and NRMSE 1.52%, beating PROSPECT-PRO especially in NIR and SWIR.
992Unique leaf samples
16Input leaf traits
0.84Mean R² (5-fold)
1.52%NRMSE
3Grapevine varieties
SPIE 2025Vol. 13475, 134750V
Mohammadreza Narimani mean grapevine leaf spectral reflectance and quantile envelope from 400 to 2500 nm for SPIE 2025 trait-to-spectra modeling
Figure 1. Mean and 1%–99% quantile envelope of grapevine leaf reflectance used to train the network.

Inverse PROSPECT-PRO chlorophyll was systematically high. A slope-based correction (scale 0.75) aligned estimated chlorophyll with wet-lab values before those traits entered the forward model.

Mohammadreza Narimani PROSPECT-PRO chlorophyll versus laboratory grapevine leaf chlorophyll before bias correction
Figure 2a. PROSPECT-PRO chlorophyll versus laboratory values before the bias correction.
Mohammadreza Narimani adjusted PROSPECT-PRO chlorophyll versus laboratory grapevine leaf chlorophyll after bias correction
Figure 2b. Adjusted PROSPECT-PRO chlorophyll versus laboratory values after the bias correction.

Attention Architecture and Error Versus PROSPECT

A nutrient-imputation CNN supplied missing minerals where needed. The forward model then maps the 16-trait vector through dense layers, reshape, multi-head attention, layer normalization, and 1-D convolutions to a full reflectance curve. Evaluation used stratified five-fold cross-validation.

Mohammadreza Narimani imputation-model R-squared and NRMSE for grapevine leaf nutrients phosphorus potassium calcium magnesium zinc manganese iron copper and boron
Figure 3. Nutrient imputation skill (R² and NRMSE) that completes the 16-trait input vector.
Mohammadreza Narimani multi-head attention neural network architecture for grapevine leaf spectral reflectance prediction from biochemical traits
Figure 4. Multi-head attention architecture: traits → dense/reshape → attention → Conv1D → full spectrum.
Mohammadreza Narimani mean absolute error by wavelength comparing the multi-head attention grapevine model with PROSPECT-PRO forward simulation
Figure 5. Mean absolute error versus wavelength: attention model (blue) versus PROSPECT-PRO (orange).

Compared with PROSPECT-PRO in forward mode, the neural network has lower mean absolute error, with the largest gains in the near-infrared and shortwave-infrared, where grapevine structure and biochemistry depart from a generic leaf library. The intended uses are canopy trait retrieval, vineyard monitoring, and species-aware remote-sensing simulations.

Paper and Citation

Parastoo Farajpoor, Alireza Pourreza, Mohammadreza Narimani, et al. "Leaf spectral reflectance prediction using multihead attention neural networks", Proc. SPIE 13475, Autonomous Air and Ground Sensing Systems for Agricultural Optimization and Phenotyping X, 134750V (28 May 2025); https://doi.org/10.1117/12.3061298

Please cite the publisher version when you use this work. The summary on this page is a research card, not a substitute for the article.

Cite and read the paper

Contact

Mohammadreza Narimani
PhD Candidate, University of California, Davis

📧 mnarimani@ucdavis.edu  |  🎓 Google Scholar