Deep learning-based detection of seedling development from controlled environment to field

In this communication, we study the possibility of transferring knowledge from indoor to field conditions for automatic classification of the early stages of seedling development. We have recently demonstrated that using simulated outdoor images from indoor images and fine-tuning the model with a small greenhouse data set can improve the classification results. Here, we confirm these results for a field outdoor data set with a significant average 10% improvement of detection performance thanks to the transfer from indoor knowledge. This establishes the possibility of benefiting from data sets obtained in a controlled environment that can be collected throughout the year to classify field images that are strongly influenced by seasonality. Moreover, image annotation is a very costly task. Therefore, we could gain time for annotation by this approach since the annotation process is still more complicated on outdoor images than on indoor ones.