Periodic Reporting for period 3 - xTrap (Real-time automatic monitoring and ML-based prediction of pest insects)
Période du rapport: 2024-02-01 au 2025-01-31
Nowadays, insect pest monitoring is costly, time-consuming and inaccurate. The use of pesticides has withdraws such as biodiversity loss and less resistance against pests and pathogens. Additionally, the global warming exacerbates the problem, causing a 40% of food lost due to pests valued in about 200B€ yearly economic damage. X-Trap is an innovative smart trap solution that will transform pest management operations in farms, fostering the transition towards the Agriculture 4,0 era. It combines real-time automatic monitoring of pest insects with prediction based on machine-learning models. xTrap is the 1st predictive solution for pest life-cycle events (over 95% accuracy) on the market. It accurately identifies crop pests and diseases in time, allowing farmers to prevent them. xTrap includes 2 main functions: automatic insect detection through smart traps equipped with sensors and high-resolution cameras to count and identify insects by means of ML algorithms, and pest prediction through algorithms based on environmental data and phenology of insects combined with the use of AI. The output from the algorithms and the predictive models will be available to farmers through our digital platform already commercialized by xFarm Technologies.
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.
• 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.
Insect pests interfere with the growth and cause serious damage to cultivated plants, and this results in plants dying or failing to reach their genetic potential. Therefore, the quantity and quality of harvests are negatively affected, and global food security as a whole can be affected. xTrap represents a significant leap forward in pest management, which translate into less uncertainty in the food industry derived from agricultural activities. Ultimately, xTrap contributes to ensure food supply and security addressing the second UN Sustainable Development Goal of “Zero Hunger”. The efficient control of pests can ensure sufficient food production and in this scenario xTrap is a valuable tool to reduce yield losses in agriculture, thus helping to fight the threat of hunger. Additionally, higher availability makes food less expensive and more accessible. On the other hand, pest control protects the livelihood of farmers and their families, preventing the ruin of local economies and communities. Moreover, it also enables to reduce the use of agrochemicals, hazardous to human health and the environment, having a direct impact in citizen´s life. xTrap is an innovative solution for pest monitoring and management that can be fully integrated into xFarm's ecosystem, fostering the transition towards the Agriculture 4.0 era