REAL-TIME PEST DETECTION USING DEEP LEARNING ON EDGE DEVICES FOR PRECISION FARMING

Abstract

Precision agriculture needs better pest monitoring. Current traps send many images to the cloud. This uses much energy and needs human review. The paper presents a smart trap that runs deep learning on the node. The trap uses a Raspberry Pi and an Intel Neural Compute Stick. It captures images inside pheromone traps. It runs neural networks to detect codling moths. Three models were trained and compared: LeNet-5, VGG16 and MobileNetV2. The system selects models that balance accuracy and power use. The trap only sends small alerts not full images. This reduces data transfer and power needs. The device uses a solar harvester and a battery. Energy harvesting lets the trap run for long periods. The system works without human intervention. Training used 4400 images and augmentation steps. The models reached high accuracy in tests. VGG16 showed the highest precision. LeNet gave the best energy trade-off on Raspberry Pi3. The detection pipeline uses sliding windows and ROI extraction. Non-maximum suppression filters overlapping detections. Preprocessing steps include color correction and edge extraction. Training used pruning to reduce model size and keep accuracy. Models were optimized with node merging and batch normalization. The design supports low bitrate radios like LoRa. This enables wide area deployment across orchards. The approach helps reduce pesticide use and save resources. It speeds pest detection and farmer response. It lowers labor and inspection costs. The work combines embedded hardware, edge AI and power management. It moves pest monitoring toward automated, energy-neutral systems.