SmartEdge PestNet: An Edge Adaptive On-Device Neural System for RealTime Pest Recognition

Abstract

Agriculture remains the backbone of food security, yet crop productivity is severely affected by pest infestations, leading to significant yield losses and economic damage. Traditional pest identification methods rely heavily on manual inspection by farmers or experts, which is time-consuming, subjective, and often incapable of providing early warnings. As a result, pests are frequently detected only after substantial crop damage has occurred, causing losses of up to 30–80% in major crops such as cotton, rice, and vegetables. To address these limitations, this research proposes an Edge Adaptive Neural Classifier for On-Device Pest Recognition using image processing and deep learning techniques. The proposed system utilizes an image-based pest recognition framework deployed on edge devices to enable real- time and early pest detection directly in the field. Pest images are collected through cameras or sensor-enabled devices and stored as an image dataset. Each image is preprocessed to extract pixellevel information, including RGB values and visual features essential for accurate classification. The dataset is divided into training and testing subsets, typically following an 80:20 ratio, to ensure reliable model evaluation. A neural network–based learning model is trained on the processed dataset to learn pest patterns and visual characteristics. Multiple model configurations and algorithms are evaluated to achieve optimal accuracy, and the best-performing model is selected based on performance metrics such as accuracy and prediction reliability. Once trained, the selected model is deployed on an edge device, enabling on-device inference without continuous internet connectivity. This edge- adaptive approach reduces latency, improves responsiveness, and ensures scalability in rural and remote agricultural environments. During the testing phase, the model predicts pest presence from real-time input images, allowing early alerts and timely preventive actions. compared to the existing manual pest detection system, the proposed solution offers faster prediction, higher accuracy, reduced dependency on human expertise, and proactive pest management. By enabling early pest recognition, the system supports farmers in minimizing crop damage, optimizing pesticide usage, and improving overall agricultural productivity. This research demonstrates that integrating image processing with neural networks on edge devices provides an efficient, low-cost, and practical solution for modern precision agriculture.