On-the-go assessment of the grapevine trunk's diameter: a comparison of different convolutional neural networks
The digital techniques, spreading across the agriculture sector, allow access to helpful information from fields, crops, and routine operations. Additionally, artificial intelligence is essential to automate and optimise laborious activities. The present study is focused on estimating the diameter o...
| Autores: | , , , , , , |
|---|---|
| Tipo de recurso: | artículo |
| Estado: | Versión publicada |
| Fecha de publicación: | 2025 |
| País: | España |
| Institución: | Consejo Superior de Investigaciones Científicas (CSIC) |
| Repositorio: | DIGITAL.CSIC. Repositorio Institucional del CSIC |
| OAI Identifier: | oai:digital.csic.es:10261/380589 |
| Acceso en línea: | http://hdl.handle.net/10261/380589 https://api.elsevier.com/content/abstract/scopus_id/85211027671 |
| Access Level: | acceso abierto |
| Palabra clave: | Deep learning Precision viticulture Proximal sensing Trunk's diameter estimation |
| Sumario: | The digital techniques, spreading across the agriculture sector, allow access to helpful information from fields, crops, and routine operations. Additionally, artificial intelligence is essential to automate and optimise laborious activities. The present study is focused on estimating the diameter of the grapevine (Vitis vinifera L. cv. Tempranillo) trunks through several artificial intelligence approaches. Then, benefits and constraints were investigated for the in-field application. Several RGB images of vines were acquired under different conditions from two fields planted with vines of different ages. The same high-resolution camera was used to take RGB images in two acquisition modes, manually and on-the-go. Additionally, different camera angles were compared to detect any distortion effect on the analysis. Finally, the impact of some common disturbance factors was compared. The trunk measurement through image analysis followed two phases. First, a YOLOv4 algorithm was set up to detect all the trunks. Then, two main approaches using deep learning were considered to estimate the diameter of each trunk. Four convolutional neural networks were trained to estimate the trunk's diameter through a regression model. On the other hand, two semantic segmentation models were trained to localise the pixels of the trunk, and a tape in the trunk marked the diameter measurement. The uncertainty analysis detected the most deleterious disturbance factors. The actual and predicted diameter regression was compared among the approaches. Xception was proved as the most accurate architecture for the estimation through a regression model. The semantic segmentation model showed a higher R-squared value than the Xception, which resulted in 0.842 and 0.619, respectively. The semantic segmentation model's normalised root mean square error was lower than the Xception-based regression one, 0.071 and 0.098, respectively. Moreover, the light condition, the vine age, and the acquisition mode showed relevant interference for the trunk diameter estimation. The trunk diameter is a key factor for monitoring the vines’ vigour and the reserve stocks. An automated trunk diameter estimation would be essential for mapping the vineyard variability at a large scale. |
|---|