AI-based robotic trap for real-time insect detection, monitoring and population prediction

Timely, accurate, and efficient pest detection is essential in agricultural sectors to mitigate economic losses and environmental impact, mainly in insect spreads. Today, AI-based robotic traps are used for insect monitoring and forecasting population growth, enabling rapid responses in pest management to reduce pesticide use and achieve sustainable goals. The core of this challenge is to build fast and accurate Artificial Intelligence (AI) models that promptly alert farmers for immediate action. In this study, we demonstrate the innovative potential of a robust AI-based trap using a robust dataset created in real settings consisting of 217 images and 6787 total annotations. The SpyFly robotic trap was used for continuous real-time monitoring and detection of the invasive insect Tuta absoluta. The detection system leverages Deep Learning (DL) models, with the best-performing model (YOLOv8x) achieving high accuracy, reaching a mean Average Precision (mAP50) of 97.6%. Additionally, environmental data such as temperature, humidity, and barometric pressure captured by the trap are used to forecast insect population growth utilizing various Machine Learning (ML) models. The ARIMAX model demonstrates significant accuracy in anticipating pest outbreaks, with a Mean Square Error (MSE) of 36.69 for weekly averaged predictions and 309.67 for daily predictions. This translates to an offset of 6 insects for weekly predictions and 17 insects for daily predictions, indicating the average deviation between the predicted and actual insect counts. This integrated approach supports early intervention and sustainable pest management, highlighting the potential of AI and DL in smart agriculture enhancing productivity and reducing pesticide usage. Dataset is available for download at https://doi.org/10.5281/zenodo.13134110.