An Automated Weed-Detection Approach Using Deep Learning in Agriculture System

Abstract

Weed management is a critical factor in ensuring agricultural productivity, as weeds compete with crops for essential resources, significantly reducing yield quality and quantity. Traditional methods, such as manual weeding and indiscriminate herbicide application, are not only labour-intensive but also pose environmental hazards and escalate production costs. This paper introduces an advanced automated weed detection system designed to support precision agriculture by leveraging deep learning technologies. The system integrates DenseNet for high-accuracy binary classification and YOLO for real-time object detection and localization of weeds. DenseNet achieved a validation accuracy of 96%, effectively distinguishing between crop and weed images, while YOLO demonstrated a mean Average Precision (mAP) of 92%, accurately localizing weeds using bounding boxes. This dual-model approach provides an efficient, scalable, and environmentally sustainable solution for targeted weed control. Experimental results highlight its potential to significantly reduce herbicide dependency, optimize resource utilization, and facilitate automated farming practices. Future work will focus on enhancing dataset diversity, improving real-time deployment on drones and robots, and expanding system functionality to include crop health monitoring.