Development of intelligent equipment for weed identification and variable spraying in lettuce fields based on instance segmentation framework

Abstract

Weeds in the field compete with crops for nutrients, water and sunlight, hindering the early growth of crops. If not controlled in time, weeds may adversely affect crop growth and yield. Although chemical weed control is low cost, efficient and widely applicable, excessive use of chemical agents may lead to herbicide residues and environmental pollution. In this study, an instance segmentation-based intelligent equipment was developed for weed recognition and targeted variable-rate spraying in lettuce fields. The You-Only-Look-Once version 8 segmentation (YOLOv8-seg) model was optimized through three key enhancements. Initially, Depthwise Separable Convolution (DSConv) was adopted to replace standard convolutional layers, effectively reducing model complexity, and improving computational efficiency. After that, a novel Faster Implementation of Cross Stage Partial Bottleneck with 2 Convolutions-Star shaped Convolutional (C2f_Star) module was proposed, which integrated the StarBlock from the Star-shaped Convolutional Neural Network (StarNet) into the existing structure, thereby enhancing the feature extraction capabilities of the model. Finally, the Simple Attention Module (SimAM), a parameter-free attention mechanism, was introduced to improve the model's attention to relevant features without increasing the number of parameters. These improvements led to the development of the YOLOv8n-seg model, which achieved a mean Average Precision (mAP) of 90.15 % at 0.5 Intersection over Union (IoU), with 2,281,702 parameters and an inference speed of 15.7 ms per frame. Compared with the original model, the average precision and inference speed increased by 2.65 % and 4.3 %, respectively, while the number of parameters was reduced by 30 %. By combining this model with post-processing algorithms, a precision variable spraying algorithm and equipment were developed. Laboratory experiments at three different weed density levels demonstrated that the system achieved an average recognition accuracy of 95.2 % and a target spraying success rate of 97.2 % for weeds in lettuce fields. Herbicide dosage was reduced by 88.42 %, 65.25 %, and 37.30 % at the three density levels, respectively. This research provides essential theoretical and technical support for the development of precision spraying and weeding robots.