A carregar...
Publicação
Development of a new non-invasive vineyard yield estimation method based on image analysis
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 3.16 MB | Adobe PDF |
Resumo(s)
Predicting vineyard yield with accuracy can provide several advantages to the whole vine and wine industry. Today this is majorly done using manual and sometimes destructive methods, based on bunch samples. Yield estimation using computer vision and image analysis can potentially perform this task extensively, automatically, and non-invasively. In the present work this approach is explored in three main steps: image collection, occluded fruit estimation and image traits conversion to mass. On the first step, grapevine images were collected in field conditions along some of the main grapevine phenological stages. Visible yield components were identified in the image and compared to ground truth. When analyzing inflorescences and bunches, more than 50% were occluded by leaves or other plant organs, on three cultivars. No significant differences were observed on bunch visibility after fruit set. Visible bunch projected area explained an average of 49% of vine yield variation, between veraison and harvest. On the second step, vine images were collected, in field conditions, with different levels of defoliation intensity at bunch zone. A regression model was computed combining canopy porosity and visible bunch area, obtained via image analysis, which explained 70-84% of bunch exposure variation. This approach allowed for an estimation of the occluded fraction of bunches with average errors below |10|%. No significant differences were found between the model’s output at veraison and harvest. On the last step, the conversion of bunch image traits into mass was explored in laboratory and field conditions. In both cases, cultivar differences related to bunch architecture were found to affect weight estimation. A combination of derived variables which included visible bunch area, estimated total bunch area, visible bunch perimeter, visible berry number and bunch compactness was used to estimate yield on undisturbed grapevines. The final model achieved a R2 = 0.86 between actual and estimated yield (n = 213). If performed automatically, the final approach suggested in this work has the potential to provide a non-invasive method that can be performed accurately across whole vineyards.
Descrição
Doutoramento em Engenharia Agronómica / Instituto Superior de Agronomia. Universidade de Lisboa
Palavras-chave
grapevine yield prediction bunch occlusion image analysis non-invasive bunch features regression model
Contexto Educativo
Citação
Victorino, G.F.S. - Development of a new non-invasive vineyard yield estimation method based on image analysis. Lisboa: ISA, 2022, 130 p.