Thinning of pome and stone fruit is an important horticultural practice that is used to enhance fruit set and quality by removing excess floral buds. As it is still mostly conducted through manual labor, thinning comprises a large part of a grower’s production costs. Various thinning machines developed in recent years have clearly demonstrated that mechanization of this technique is both feasible and cost effective. Generally, these machines still lack sufficient selectivity to take into account the specific fruit bearing capacity of each tree. Furthermore, the current devices often cause damage to shoots, leaves and fruitlets which makes the trees more susceptible to dangerous diseases such as fire blight (Erwinia amylovora) and cankers.

To address these issues, we investigated a new non-contact way of thinning using pulses of compressed air in combination with a sensor capable of detecting the floral bud distribution. This way, the thinning efficiency can be improved by providing real time information of the floral bud distribution. We focused on the early phenological stages (until bloom) of the pear cultivar Conference, for which there are few chemical thinning alternatives.

The forces required to remove a floral bud were measured in a laboratory test bench. These required forces change as a function of bud development. A pneumatic setup was built and tested during a two-year trial in an orchard to determine the effects of air pressure, nozzle type, distance and phenology on the attainable removal efficiency. Hereafter, a performance model was built using stepwise logistic regression modeling. Thinning grades as high as 93.13 % and 74.52 % could be achieved for, respectively, a dry and a wet season. Furthermore, pneumatic thinning was observed to reduce tree damage to a minimum since floral buds were removed at their natural breaking point, i.e. the pedicel abscission layer.

Besides this, we developed a multispectral vision sensor capable of detecting floral pear buds during the phenological stages before bloom. During two flowering seasons, scenes were captured in the orchard at six distinct optical wavebands in the visible and near infrared region of the spectrum. Measurements were conducted under controlled illumination. Using canonical correlation analysis, a spectral discrimination model was built that recognizes pixels originating from floral buds. Hereafter, an image analysis technique was developed to translate the pixel classification to object recognition. This algorithm was able to recognize more than 80 % of the floral buds that were captured under proper illumination. Therefore, the multispectral sensor can be used to increase the efficiency of pneumatic thinning or other thinning machines. Furthermore, it can as well be used independently for early-season yield estimation.