The 'Know what you're growing' (Weet wat er leeft) project aims to combine different monitoring data into one overview of species present in greenhouses. One of these data streams comes from automatic image recognition. Bruce Schoelitsz, HAS green academy student, shares developments made over the past year.
One of the problems HAS green academy students encountered last year was the effect of coloured LED lighting in greenhouses on the performance of the image recognition model. Insects in pictures of yellow sticky plates taken with phones were no longer recognised. A new group of students tackled this and investigated how to solve that problem. As it turned out, both using a strong external flash, and adjusting the colour of the photo in post-processing, ensured that the insects were recognised just as well as under white light. This worked better than training a new model with the purple light.
Image recognition model in practice
In addition, the image recognition model was tested in practice. The results of the model were disappointing in greenhouses of rose, pot orchid, chrysanthemum and cucumber. The number of individuals of the pest species the model was trained on was very low. Furthermore, several greenhouses contained species that were not present in the pictures the model trained on. As a result, they were often wrongly recognised as pest species. In other greenhouses, the density of other insects was so high that, for example, thrips could not be easily seen among all the legs and wings.
The photos taken in the field will be used after the summer to retrain the model. This is expected to greatly improve the model and enable it to determine trends in numbers of thrips, whiteflies, aphids and parasitic wasps. To easily analyse the photos and make the data easily understandable, the app developed by imec will be further developed. The data will be made insightful within a LetsGrow module. For proper further development, user experiences are important and growers already involved in the project will be contacted.
Leaf material
Besides photos of sticky trays, the Edapholog was also used. This is a funnel into which leaf material can be placed that is heated with a lamp. Crawling insects and mites walk down and then land on a small platform that is filmed. These films can be used to train recognition models and these have now been made by Edaphone for Swirskii, Montdo, the feeding mite Carpoglyphus lactis and thrips larvae. The next step is to determine the accuracy and improve the models.
Workgroup
The Vision workgroup of 'Weet wat er leeft' consists of HAS green academy, imec/OnePlanet, WUR and Koppert Biological Systems. Thanks to students from HAS green academy: Sjoerd Barten, Cas Beekhuizen, Jacoline Boer, Sarah Botman, Stijn van Boxtel, Ilse Geurts, Nienke Haaring, Alex Houtmans, Karlijn Nissen, Lucas Smit, Jan Paul Veenema, Wisse op 't Veld, Tom Stals; and imec/OnePlanet: Nicholas Termote.
Source: Glastuinbouw Nederland