1Digital Agriculture Laboratory, Department of Biological and Agricultural Engineering, University of California, Davis, CA, USA
2Department of Botany and Plant Sciences, University of California, Riverside, CA, USA
3Department of Viticulture and Enology, University of California, Davis, CA, USA
| Received 09 May 2025 |
Accepted 11 Nov 2025 |
Published 18 Dec 2025 |
Analysis of leaf hemispherical radiative properties for retrieval of its biochemical and mineral nutrients could lead to a powerful monitoring approach for precise farm management. This study explores the potential of leaf spectral modeling techniques for estimation of key biochemical and nutritional traits in grapevine leaves. Hyperspectral data spanning the 400–2500 nm range were collected from around 1000 leaf grapevine leaf samples across three growing seasons. Certain traits, including leaf structural parameter (Nstruct), anthocyanins, carotenoids, and chlorophyll, were imputed using the PROSPECT-PRO radiative transfer model in the inverse mode to enrich the dataset. An imputation model was developed to address missing labels for part of the dataset, employing a Convolutional Neural Network (CNN) with 23 principal components derived from the spectral data as inputs. This model enabled the completion of the dataset by predicting missing trait values, providing a comprehensive foundation for subsequent modeling efforts. For the primary trait prediction models, the spectral data were then reduced from 2101 bands to 204 bands through band merging based on pairwise correlations. Two predictive modeling approaches were evaluated: a single-trait model, where each trait is predicted independently, and a multi-trait model, where all traits are predicted simultaneously. Both models employed a hybrid of CNN and Long Short-Term Memory (LSTM) networks designed to capture spatial and sequential patterns in spectral data. The single-trait model utilized CNN-LSTM architecture with a single output node, requiring independent training for each trait. In contrast, the multi-trait model employed the same architecture but featured 16 output nodes, enabling the simultaneous prediction of all traits. A weighting strategy was implemented to balance the influence of fully measured and imputed samples during training, ensuring reliable predictions. The multi-trait model demonstrated superior predictive performance across most traits, achieving a higher coefficient of determination (R2) and RPD (Residual Predictive Deviation), and lower normalized root mean squared error (NRMSE) values than the single-trait models. Some traits, such as nitrogen, phosphorus, Nstruct, and manganese benefited significantly in the multi-trait model with R2 values of 0.42, 0.81, 0.90, and 0.62, respectively, compared to 0.26, 0.64, 0.25, and 0.30 in single-trait models. The results highlight the advantages of multi-trait modeling in leveraging shared spectral information and inter-trait dependencies, offering an efficient and accurate approach to predicting grapevine traits.
