Estimation of the flower buttons per inflorescences of grapevine (Vitis vinifera L.) by image auto-assessment processing

Abstract

The aim of this research is to develop a reliable tool by a special method, combining image processing based on watershed algorithm, and a predictive model to estimate automatically the flowers number per inflorescence. Eighty images of Vitis vinifera L. inflorescence (the Cardinal cultivar) were processed. Watershed algorithm was used for the image processing and this was followed by statistical analysis that provides robust predictive estimation of the flower button number. The results show a robust estimation, compared to manual flowers counting, with strong correlation. The developed algorithm shows that the watershed algorithm was able to provide an automatic assessment of the flower button number in the inflorescence. The method used is more robust and provides a more significant level compared with recent studies. In the applied research in viticulture, it is crucial to improve knowledge of yield forecasting and to study the fruit set rates estimation. The technique is used to determine, with a higher significance level, the fruitiness rate of grapevine at the early stage of flowering.

Cultivated Grapevine Displays a Great Diversity for Reproductive Performance Variables

“…As a result, a case-by-case validation step is required for each variety, which is unpractical for wide research studies. Similar problems are expected with other methods for estimating the number of flowers in grapevine inflorescences based on 2D image analysis that have been recently proposed[30,32,46]. They are surely useful for certain purposes in the varieties tested, but not in this case because the very different type of inflorescences in the different varieties studied.For that reason, we opted for an alternative method in 2016 and 2017, based on the direct counting of the calyptras.…”

“…Instead, counting the number of flower caps or calyptras (fused petals) collected in a bag is a non-destructive valid method, because each flower releases one calyptra at blooming [9,17,18,23], which can be delayed. On the other hand, new image analyses tools have been recently developed for estimating the number of flowers from 2D images of the inflorescences, although they have usually been tested in a limited set of varieties [28][29][30][31][32][33].…”

Computer Vision and Deep Learning for Precision Viticulture

“…The method proposed by [45] improves the method of [47] by using specific morphological operators to remove the natural background (an artificial background is therefore no longer necessary) and detect potential flower buds. The method proposed by [49] used a similar approach in the binary image only, where a morphological operator and a watershed algorithm were used to delineate the flower buds. Although it achieves a good counting correlation of R 2 = 0.99, it does not compute the location of the flowers in the image.…”

“…Several weaknesses in the previous studies have been identified by [50]. The methods proposed by [45][46][47]49] all depend on manually chosen parameters (the size of the morphological operators, for example). Therefore, these methods are sensitive to the color and the apparent size of the flowers.…”