AGRIBOT EDGEVISION: A LIGHTWEIGHT DEEP LEARNING APPROACH FOR AUTONOMOUS WEED DETECTION IN CHILLI FIELDS

Abstract

Managing weeds in chilli (Capsicum annuum) fields is a major problem because young chilli plants look very similar to common weeds like Alternanthera caracasana and Oci-mum tenuiflorum, etc. These weeds compete for nutrients, sunlight, and water, reducing crop yield if not controlled properly. Manual weeding requires a significant amount of time and effort, whereas cloud-based systems for weed detection often encounter issues such as slow processing and poor connectivity in farms. To address these challenges, this study introduces AgriBot-ChilliWeed, an AI-powered robot designed to detect and treat weeds in real-time. The robot uses a Lightweight Hybrid Convolutional Neural Network (LH-CNN) built on a Raspberry Pi 4B with a Coral USB Accelerator for fast image processing. A high-resolution camera captures live images of the field, and weeds are identified within 45 milliseconds per frame. When a weed is detected, a micro-sprayer mounted on the robot sprays a small, precise amount of herbicide directly on the weed, avoiding damage to the chilli plants. Field tests carried out in Karnataka and other regions showed promising results. The system achieved a detection accuracy of 93.2%, with a precision of 94.1%, a recall of 92.7%, and an F1-score of 93.4%. The robot consumed only 6.5–6.8 watts of power and could cover about one acre per battery cycle. Comparisons with other lightweight AI models like YOLOv8-nano, MobileNetV3, and Efficient-Net-lite showed that the proposed method offers a better balance between speed and accuracy. Overall, the findings show that AgriBot-ChilliWeed is a low-cost, energy-efficient, and practical solution for weed management in chilli farms. It reduces labour needs, lowers herbicide use, and supports sustainable farming practices for small and medium-scale farmers.