lu.se

Matematikcentrum

Lunds universitet

Denna sida på svenska This page in English

Calender

Using Machine Learning to Detect Grapevine Disease in Wine Production

The figure shows three images of leaves without and with mildew.
From Ivar Vänglunds Master's thesis - Using Machine Learning to Detect Grapevine Disease in Wine Production

Seminarium

Tid: 2021-04-23 15:15 till 16:15
Plats: https://lu-se.zoom.us/j/63277711935
Kontakt: kalle [at] maths [dot] lth [dot] se
Spara händelsen till din kalender


Ivar Vänglund presenterar sitt examensarbete Using Machine Learning to Detect Grapevine Disease in Wine Production Fredagen den 23 april kl 15:15 på zoom https://lu-se.zoom.us/j/63277711935



Abstract:
When growing grapevines there are commonly risks for disease outbreaks when the pathogen has remained from earlier growing cycles and the weather and soil conditions work in the diseases’ favour. The diseases may spread a lot before being highly noticeable, and then the harvests of grapes for wine production are in jeopardy. This thesis concerns a branch of a, from the Swedish Board of Agriculture funded, project with the special focus for the thesis is to use machine learning tools such as convolutional neural networks to detect the most common grapevine diseases downy mildew and powdery mildew. By using an image inventory from the project staff the purpose is to give decision support regarding the mentioned diseases of an input image using neural networks and image analysis. As being able to notice different types and stages of the grapevine diseases is desired, the network was sought to have a 5 class outputs but also testing to have fewer (3 and 2). The resulting model, after training to reach a good validation accuracy, reached an accuracy of ∼ 65% when predicting to 5 labels, but with lower recall on the infected classes with fewer image samples. Training for predicting 3 and 2 different outputs gave higher results of ∼ 70% and ∼ 75% respectively and better tendencies of finding non-healthy leaves. It is believed that a larger amount of training images could improve the performance of the network.