Applications of computer vision techniques in precision viticulture
Precision viticulture is a technique that aims at improving grapevine production and quality while reducing the environmental impact by optimising resource use. For its implementation, the correct, georeferenced, precise measurement of the vine status which represent the inter- and intra-field varia...
Guardado en:
| Autor principal: | |
|---|---|
| Otros Autores: | |
| Formato: | text (thesis) |
| Lenguaje: | eng |
| Publicado: |
Universidad de La Rioja (España)
2017
|
| Materias: | |
| Acceso en línea: | https://dialnet.unirioja.es/servlet/oaites?codigo=122699 |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| Sumario: | Precision viticulture is a technique that aims at improving grapevine production and quality while reducing the environmental impact by optimising resource use. For its implementation, the correct, georeferenced, precise measurement of the vine status which represent the inter- and intra-field variability is mandatory. The development of the geo-positioning systems and sensing technologies, capable of monitoring vine status in a non-invasive, fast and reliable way has stimulated the development and implementation of precision viticulture. Image analysis techniques are currently of increased interest to agricultural monitoring. Their low costs and wide range of applications make them ideal for crop status evaluation.
The main goal of this PhD thesis is to provide new reliable, objective and simple methodologies for vineyard status monitoring using image analysis. To this end, different procedures have been developed to do so: i) assessment of flower number per inflorescence; ii) estimation of the yield before harvest; and iii) evaluation of canopy status. The use of different capturing procedures (manual, smartphone based and on-the-go) was also taken into account, tested and analysed.
The algorithm developed for the assessment of flower number per inflorescence provided estimations with over 90% precision for all the studied varieties. When an improved version of this algorithm was implemented for use in an Android smartphone, the precision rose to the 94%. The new version identified 84% of the flowers present in the image correctly. The number of flower that were visible per image (not all the flowers are visible in the image due to occlusions) was used to estimate the total flower number using a non-lineal model with a root mean square error (RMSE) of 37.1.
The yield assessment before harvest was carried out using two approaches: firstly, a series of vine images were captured manually using a white screen as background, resulting in a classification performance of 98% for clusters and 92% for leaves, this allowed the assessment of the yield with R2 = 0.73. Not all the berries are visible in a vine image due to occlusion from clusters or other parts of the vine. Secondly, the use of a Boolean model was used to reduce the error associated to the occlusion and segmentation errors, resulting in an error in the yield estimation of RMSE = 203g per vine from images captured on-the-go.
Canopy status assessment was carried out with a multi-site experiment conducted in New Zealand, Croatia and Spain. The comparison between the reference method (point quadrat analysis) and the results obtained by analysis of manually captured images (taken on the field using a white screen as background) yielded a determination coefficient over 0.90 on every evaluated site and R2=0.93 when all the data was analysed together. The following experiment was carried out using a modified all-terrain vehicle (ATV) for the automatic image capture at a speed of approximately 7 km/h. This setup permitted high sampling rate data capture and thus vine status map generation. The correlations obtained for the canopy porosity and exposed leaves showed a R2>0.85 and R2>0.71 respectively. Finally, the pruning wood weight is a classic vine vigour indicator. The use of manually captured images (with white screen as background) resulted in RMSE=87.7g and R2=0.91. When the images were captured with a modified ATV the precision slightly dropped to RMSE=115.7 and R2=0.85 but with a significant reduction in the capturing effort.
The results show how computer vision can provide valuable information on vineyard status for precision viticulture. The low cost of the sensor, its non-destructive and fast capturing process offers a great advantage over classical manual reference methods. Image analysis showed high precision in the assessment of flower number per inflorescence, yield estimation and canopy status assessment. The possibility to capture the images on-the-go greatly increases its applicability reducing the effort for data capturing and allowing map generation |
|---|