Why counting grapevine leaves?

Leaf number is a key indicator of vegetative organogenesis in grapevine. Leaf emission rate (or vegetative organogenesis rhythm) depends on the variety and can be impacted by the environment (water availability in the soil, air temperature, air humidity, radiation).

There are several uses of leaf counting such as:

  • modeling organogenesis in grapevine and improve existing models of leaf emission rate as a function of time or thermal time (e.g. Pallas et al., 2011).
  • characterising genetic diversity of this leaf emission rate is a major issue to better understand the existing variability (which could be used to adapt variety choices to local climate) and its genetic control (region of the genome – QTL or genes) (e.g. Houel et al., 2015).
  • in genetic studies which simultaneously study thousands of plants, such as the LEPSE experiments carried out into the PhenoArch phenotyping platform (Coupel-Ledru et al., 2014 and Coupel-Ledru et al., 2016), knowing the development stage (leaf number) of all plants on a given day is essential to avoid bias in the measurement of other variables. For example, if transpiration of the whole progeny is measured, ideally it must be done at a similar developmental stage for all plants.

However, counting leaves of thousands of plants is a laborious and time-consuming task, which is hard to manage in parallel with other measurements (as was the case for experiments into the PhenoArch platform: while in 2012 and 2013 leaves have been counted weekly for all 1680 plants, this was not possible in 2014).

Within the BigDataGrapes project, the National Research Council of Italy (CNR) developed a machine-learned pipeline aiming at counting leaves from side- view grapevine images taken into the imaging cabin of the PhenoArch platform (see picture below).


Figure 1: A photo of a grapevine from the PhenoArch phenotyping platform.

How does the machine-learned pipeline aiming at counting leaves work?

The new counting pipeline exploits deep learning techniques based on artificial neural networks to infer the number of leaves from each grapevine image. In details, CNR exploits a combination of Convolutional and Feedforward neural networks to solve the counting problem as a regression task, i.e., given a single side-view image of the grapevine, the neural network predicts the number of its leaves.

Picture2.pngFigure 2: Architecture of the deep neural network model used to solve the counting leaves problem.

In Figure 2 (from left to right) the image of the grapevine goes through several blocks (six) of convolutional and max-pooling layers followed by two feedforward layers aimed to produce the final prediction, i.e., the actual number of leaves of the plant in the image. Deep neural network (DNN) solutions are well-known to be “data hungry” in order to effectively train complex models. Here, the application of DNN is made possible due to the availability of a dataset consisting of more than 250,000 images (~3 TeraBytes of data). Preliminary results prove the validity of the proposed approach as the solution allows to achieve an average error of less than one leaf per plant.

Future work

As future work, an adaptation of this leaf counting tool could be developed, particularly for precision agriculture. Indeed, image analysis from drones or other technologies could enable a specific and local watering or spraying in function of each grapevine needs in the field. It is a relevant tool to help the determination of subfield areas (or delineation of management zones) within the vineyard in order to decrease input costs and increase the production value. From a technical point of view, we aim at extending the DNN architecture to use as input all the available side-view images of the grapevine instead of only one.

Authors: Aude Coupel-Ledru (INRA), Llorenç Cabrera-Bosquet (INRA), Coraline Damasio (INRA), Franco Maria Nardini (CNR), Pascal Neveu (INRA), Raffaele Perego (CNR) and Salvatore Trani (CNR)


Coupel-Ledru, A., Lebon, É., Christophe, A., Doligez, A., Cabrera-Bosquet, L., Péchier, P., ... & Simonneau, T. (2014). Genetic variation in a grapevine progeny (Vitis vinifera L. cvs Grenache× Syrah) reveals inconsistencies between maintenance of daytime leaf water potential and response of transpiration rate under drought. Journal of Experimental Botany, 65(21), 6205-6218.

Coupel-Ledru, A., Lebon, E., Christophe, A., Gallo, A., Gago, P., Pantin, F., ... & Simonneau, T. (2016). Reduced nighttime transpiration is a relevant breeding target for high water-use efficiency in grapevine. Proceedings of the National Academy of Sciences, 113(32), 8963-8968.

Houel, C., Chatbanyong, R., Doligez, A., Rienth, M., Foria, S., Luchaire, N., ... & Pellegrino, A. (2015). Identification of stable QTLs for vegetative and reproductive traits in the microvine (Vitis vinifera L.) using the 18 K Infinium chip. BMC plant biology, 15(1), 205.

Pallas, B., Loi, C., Christophe, A., Cournède, P. H., & Lecoeur, J. (2010). Comparison of three approaches to model grapevine organogenesis in conditions of fluctuating temperature, solar radiation and soil water content. Annals of botany, 107(5), 729-745.