Counting gerbera buds with deep learning

May 3, 2019

By applying deep learning to camera images of gerberas, it is possible to make a harvest prognosis. This is the conclusion of the research in the PPS Smart Greenhouse Horticulture, where a 'gerbera scout' is being developed: a mobile platform with cameras that run alongside the crop. The study was commissioned by the Club of 100 by the Greenhouse Horticulture Business Unit at Wageningen University & Research.

Within the 'Smart Greenhouse Horticulture' project, WUR is working on detection systems for diseases, pests and harvest forecasts. The Gerberas Scout is a good example: it rides on the pipe rail system in a greenhouse and makes images of the crop with different cameras. Based on input from gerbera growers, the scout is tested for harvest prediction and for the detection of mildew and whitefly.

Deep learning

For the harvest forecast, an RGB camera takes pictures from the top of the crop. Image recognition is needed to calculate how many buds and flowers of more than 2 centimeters are present. 'Normal' image recognition does not recognize all the buds and flowers (around 85%). That is why deep learning has been implemented. The computer now learns to recognize a bud based on shapes and colors. The detection based on deep learning detects more than 95% of the total number of buds and flowers.

Hyperspectral camera

The Gerbera Scout is not yet a good weapon in the fight against the white fly - a pest that often occurs in the gerbera. As soon as a camera detects the white fly, the damage has already been done: the infection is already too far advanced.

The RGB cameras only measure three colors: red, green and blue. That is too little to recognize mildew. A hyperspectral camera divides light into smaller snacks than just three colors, and has a larger spectral range. With the hyperspectral camera, the Gerberas Scout can easily recognize mildew, but the scout still gives a false warning relatively often. Work is therefore being done on refining the system to make detection more robust.

This project is funded by Club of 100 and the Ministry of LNV.