Research on Indirect Identification Method and Performance Optimization of Green Vegetable Weeds Based on Lightweight Model

To address the challenge of difficult weed management in large-scale agriculture, this study presents an indirect weed identification method for green vegetables. By constructing a morphological-spectral feature database of green vegetables and integrating lightweight models (GhostNet/ResNet50/ShuffleNet), binary classification of “crop-non-crop” is realized. Experimental results show that this method reduces the weed misjudgment rate to $1.8 \%$, improves recognition efficiency by $32 \%$ compared with traditional algorithms, and ensures a $98.2 \%$ crop protection rate under field conditions. Model comparisons demonstrate that GhostNet (overall accuracy: $98.95 \%$, missed detection rate: $0.60 \%$) exhibits the best comprehensive performance, while ResNet50 (missed detection rate: $2.1 \%$) shows advantages in small-target recognition. This research validates the feasibility of indirect weed identification and removal for green vegetables, as well as the effectiveness of the “strengthening crop features-removing background” strategy. Future work will focus on integrating multi-modal data and optimizing network architectures to enhance recognition stability in complex scenarios.