Assessment of grapevine water status from hyperspectral imaging of leaves
This work reports preliminary results on a novel application of neural networks to the assessment of grapevine leaf water content using hyperspectral data and the influence of the grapevine cultivar, clone and side of imaging on the spectral results. Hyperspectral images were collected for 299 indiv...
| Autores: | , , , , , |
|---|---|
| Tipo de documento: | artigo |
| Estado: | Versão publicada |
| Data de publicação: | 2014 |
| País: | España |
| Recursos: | Universidad de La Rioja (UR) |
| Repositório: | RIUR. Repositorio Institucional de la Universidad de La Rioja |
| OAI Identifier: | oai:portal.dialnet.es:doc/5bbc6967b750603269e81a85 |
| Acesso em linha: | https://investigacion.unirioja.es/documentos/5bbc6967b750603269e81a85 |
| Access Level: | Acceso aberto |
| Palavra-chave: | Abaxial Adaxial Clone Infrared spectra Neural network Water stress |
| Resumo: | This work reports preliminary results on a novel application of neural networks to the assessment of grapevine leaf water content using hyperspectral data and the influence of the grapevine cultivar, clone and side of imaging on the spectral results. Hyperspectral images were collected for 299 individual leaves of four different clones for each of the three Vitis vinifera L. cultivars, 'Cabernet Sauvignon', 'Grenache' and 'Tempranillo'. The hyperspectral camera used operated in the infrared region, from 900 to 1,700 nm and was positioned approximately 40 cm away from the leaves. For each leaf, three discs of 2 cm of diameter were taken and these were imaged with the hyperspectral camera on the abaxial and adaxial sides, at two different leaf water contents. Reference values of relative water content (RWC) were determined by measuring leaf-discs weights totally hydrated and after drying out in an oven at 60°C until constant weight. RWC averaged 95% in the former case and 75% in the latter case. The neural network was trained with 156 leaves using the Levenberg-Marquardt algorithm and early stopping, while validation was done with 143 leaves. The mean of the absolute error between values of RWC obtained by the neural network and the reference measurements was 2.77% for the validation leaves. Percentiles 75 and 90 of the error were 4.74 and 6.21. The spectral response to the change in RWC was shown to be cultivar, clone and leaf side dependent. The reduction in RWC of leaves led to a bigger reflectance enhancement in 'Grenache' and 'Cabernet Sauvignon' when images were taken on the adaxial side, whereas for 'Tempranillo' on the abaxial side. All these factors should be taken into account when the grapevine leaf water status is to be assessed by hyperspectral imagery, either proximal or remote. |
|---|