ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Phenologically Contextualized Deep Learning for Multi-Nutrient Prediction in Grapevine Leaves from In-field Hyperspectral Imaging

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2601349.(doi:10.13031/aim.202601349)
Authors:   Atif Bilal Asad, Dawood Ahmed, Nataliya Shcherbatyuk, Pierre Davadant, Manoj Karkee, Markus Keller
Keywords:   Ablation Study, Artificial Intelligence, Convolutional-Transformer, Deep Learning, Grapevine leaves, Hyperspectral Imaging, Multi-nutrient modeling, Precision Viticulture, Spectral modeling.

Abstract. Conventional laboratory-based foliar tissue sampling for grapevine leaf nutrient chemical analysis is destructive, temporally insufficient, and spatially sparse for precision viticulture nutrition management decisions. This study developed and evaluated a Convolutional-Transformer architecture for simultaneous prediction of eleven foliar nutrients from ground-based hyperspectral imaging within the 400–1,000nm range in Chardonnay (Vitis vinifera L.) grapevines across two phenological stages (bloom and veraison) and three growing seasons (2022–2024). The architecture integrates a one-dimensional convolutional stem for local spectral feature extraction, a transformer self-attention encoder for modeling long-range inter-band dependencies, and learned phenological growth-stage embeddings that provide contextual information about nutrient concentration dynamics across the growing season. An ablation study using stratified cross-validation benchmarked the full Conv-Transformer against Conv-Only and Transformer-Only single-branch variants, each independently optimized under architecture-specific hyperparameter search. The Conv-Transformer model achieved mean R² = 0.68, MAPE = 12.18%, and RPD = 2.03 across all eleven foliar nutrients, with macronutrients averaging an RPD of 2.21 and S and B having higher RPD values than 2.90. The largest prediction gains were observed for S, K, and Fe, where the Conv-Transformer's non-local spectral attention and phenological stage embeddings recovered indirect biochemical co-variation signals that neither single-branch model could capture alone. Paired Wilcoxon signed-rank tests confirmed the Conv-Transformer model significantly outperformed both single-branch variants across all three scale-independent metrics (p ≤ 0.01). These results showed the Conv-Transformer to be a viable architecture for simultaneous multi-nutrient foliar monitoring in commercial vineyards using non-destructive spectral sensing, enabling timely nutrition management decisions without laboratory analysis.

(Download PDF)    (Export to EndNotes)