Deep learning for in-field image-based grapevine downy mildew identification

J Boulent, M Beaulieu, PL St-Charles… - Precision …, 2019 - wageningenacademic.com
J Boulent, M Beaulieu, PL St-Charles, J Théau, S Foucher
Precision agriculture'19, 2019wageningenacademic.com
Since the early 2010's, great progress has been made in image classification through deep
learning and convolutional neural networks (CNNs). This allows us to consider the
development of automatic high-precision scouting tools for agricultural use. The objective of
this project is the automatic identification of grapevine downy mildew-Plasmopara viticola.
To assess the potential of a commercial prototype, several wine-growing regions and grape
varieties were studied. Images were acquired under in-field conditions. The selected CNN …
Since the early 2010’s, great progress has been made in image classification through deep learning and convolutional neural networks (CNNs). This allows us to consider the development of automatic high-precision scouting tools for agricultural use. The objective of this project is the automatic identification of grapevine downy mildew - Plasmopara viticola. To assess the potential of a commercial prototype, several wine-growing regions and grape varieties were studied. Images were acquired under in-field conditions. The selected CNN architecture used for developing this model was ResNet. Multiple depths of ResNet have been tested to determine which one achieves the best score for this study case and for the amount of data collected. The accuracies reached by five different depths were surprisingly close. The best accuracy is achieved by a ResNet with 18 layers, with 95.48% on an independent test set.
Wageningen Academic
以上显示的是最相近的搜索结果。 查看全部搜索结果