The main goal of the project was the development of a smart trap, called xTrap, for the recognition, counting, and modelling of the phenology of the target insects. Main activities and achievements during the project have been:
• Definition of technical specifications for xTrap Delta, xTrap Color and xTrap Stink to monitor target insects.
• Building xTrap prototypes to test and optimization of the hardware component to improve accuracy and efficiency of the system.
• Design of the service infrastructure, including database, framework for front-end and back-end development, and communication protocol between xTrap and the cloud application.
• Design and development of the dataset for collecting pictures used to train and test computer vision algorithms for insect detection and counting. Few Shot Learning (FSL) techniques have been used when necessary. Images from our test traps and public dataset have been collected to build our own dataset for target insects. Using data augmentation to generate larger dataset
• Design and development of algorithms for insect detection and counting. An analysis of the state-of-the-art of algorithms used for generic object detection and for pest insect detection specifically has been done. Finally, we chose to use the YOLOv5 framework showing the best performance in the literature. This selection simplifies the model development to extend the target insects.
• Design and development of algorithms for phenology prediction. An analysis of the state-of-the-art literature on models for phenology and dynamics of insects was carried out to identify the best model. We selected the class of phenological models, based on climatic factors, that we can get from our sensors in the x-traps. These models are based on the Growing Degree Days (GDD) which are strongly correlated to the insect´s development. For each target insect we have defined the parameters of the model to get GDD thresholds for each phenological stage.
• The test infrastructure for the on-field tests has been designed as well as the test protocols for testing the performance of the traps and identification and counting of target insects. The system showed excellent performance for Tuta absoluta reaching 98% of accuracy, 93% on Lobesia Botrana, 100% on Halyomorpha halys and 70% on Ostrinia nubilalis. The recognition of Scaphoideus titanus, showed difficulties related to the presence of midges, common during the ripening period of the grapes.
• In-field tests carried out during the pilot season for testing the performance of the algorithms for phenology and insect dynamics. Results show the effectiveness of xTrap in the automated monitoring of pests and the prediction of insects' dynamics through GDD analysis and field detection. The accuracy of the system made possible to identify new generations of insects at an early stage, enabling more targeted phytosanitary interventions and integrated pest management. The reduction of chemical interventions through precise management of insect populations is a significant step towards a more sustainable agriculture, with tangible economic and environmental benefits.
