Image-based Counting of Leaf Number of Direct Seeding Onions (Allium cepa L.) during Early Growth Stages by Applying Deep Learning Techniques

Abstract

In this study, we developed a convolutional neural network (CNN)-based classification model to count the number of leaves of direct-seeded onions using images captured by a 100 million pixel aerial camera mounted on a drone in an outdoor field. Firstly, since there were several other types of vegetation, such as weeds, in outdoor fields, we evaluated the effect of removing non-onions from the training data on the classification model. It was demonstrated that it was not necessary to define classes such as weeds, as they appear infrequently, when performing CNN using images captured in outdoor fields. Secondly, in the validation of classification model-2, it was determined that the model could not be sufficiently modified to extract features due to the small sample size, which led to biased results. Therefore, in order to correct the imbalance between the classes of development data in this study and include a wide range of features in each class, we split the data set into two classes based on the number of leaves. As a result, the classification of the model was improved and it was concluded that this modelling approach could be used to detect seedlings with fewer than four leaves in onion cultivation.