Energy Neutral Machine Learning Based IoT Device for Pest Detection in Precision Agriculture

Abstract

Apples are among the topmost fruit crops of the world, and apple orchards are widely expanding in many regions and countries. The most common problem for these crops is the attack of the codling moth, which is a dangerous parasite for apples. IoT sensing devices can nowadays run near sensor machine learning algorithms, thus giving not only the possibility of collecting data over wide coverage but even featuring immediate data analysis and anomaly detection. Near sensor neural network algorithms can automatically detect the codling moth: the system takes a picture of the trap, preprocesses it, crops each insect for classification, and eventually sends a notification to the farmer if any codling moth is detected. The application is developed on a low-energy platform powered by a solar panel of a few hundred square centimeters, realizing an energy autonomous system capable of operating unattended continuously over low power wide area networks. An insightful aspect of this IoT solution is the low power platform for a machine learning algorithm used for IoT fast prototyping. The hardware is based on the Raspberry Pi3 board and the Intel Movidius Neural Compute Stick, responsible for the preprocessing technique and the neural network implementation, respectively. The network model has been analyzed in detail, showing parameter settings and the limitations for the specific hardware constraints. The performance of the proposed system is assessed, and remarks on power consumption are discussed for achieving the zero energy balance of the system.