Abstract
In this paper, we illustrate the technological innovations we implemented in a test-bed field to automate the bug scouting process. Our work is motivated by the invasive global pest Halyomorpha halys (HH), whose damages have a huge economic impact for fruit orchards. We propose the automation of the time- and labor-intensive process of the HH scouting, traditionally performed by phytosanitary operators. We then describe the selection criteria that led to the hardware architecture designed consisting of a UAV, an RGB vision chip, a new ad hoc trap, and micro-climate stations. We also look for recognition algorithms based on deep learning models that can learn to recognize the HH after a training based on a dataset of images. Our very preliminary results show that the performances of UAV deep learning algorithms trained on artificial datasets are not satisfactory when tested on real images. However, very satisfactory results were obtained from the stationary ad hoc trap monitoring system running on the edge.