Automated weed monitoring and control: enhancing detection accuracy using a YOLOv7-AlexNet fusion network.

The agricultural sector is crucial to global sustainability, but it still faces challenges, particularly from weed invasions that severely compromise crop yields. Although considerable efforts have been made to address the weed problem using computer vision detection methods, the technology is still limited. Weedy sites or their crop hosts share many perceptual features, making it difficult to detect with confidence. Most weed detection methods used today suffer from several problems: the inability to distinguish crops from similar-looking weeds, inconsistent performance across weed growth stages, and sensitivity to operational constraints. Previous methods have employed models such as YOLOv5, ResNet, and Faster R-CNN, but have suffered from issues with accuracy, estimation times, and the ability to detect small weeds in dense stands. In this study, we present a hybrid deep learning system that utilizes YOLOv7 for weed detection and AlexNet for weed species classification. YOLOv7 was used due to its fast recognition capabilities and ability to discriminate with better granularity when detecting grass in dense environments. It was found that using AlexNet to classify weed species accurately increases the specificity of the system. Experimental results of the hybrid model demonstrated improvements over previous methods, achieving a precision, recall, F1 score, mAP@0.50, and mAP@0.5:0.95 of 0.80, 0.85, 0.87 0.89, and 0.50, respectively. The field test detection capability showed that AlexNet achieved precision, recall, and F1 scores of 95%, 97%, and 94%, respectively. Thus, these results indicate that the YOLOv7-AlexNet hybrid model provides both robust and efficient real-time detection and classification of weeds in agriculture. The next step is to expand the dataset to include a wider variety of weed species and environmental conditions, and to validate the developed model by deploying the YOLOv7-AlexNet hybrid model on field computers, thereby expanding its practical application in production environments.