LeafNet-CBAM: a lightweight attention-enhanced network for weed image classification

Abstract

Accurate weed identification is essential for precision cotton farming, as early-stage weed infestation significantly affects crop yield and resource efficiency. This study proposes LeafNet (L-Net), a lightweight and stable convolutional neural network (CNN) designed for cotton-weed image classification under real-field conditions. L-Net integrates depth wise separable convolutions with Convolutional Block Attention Module (CBAM) and residual connections to enhance discriminative feature learning while maintaining low computational complexity. A cotton field image dataset was collected under varying illumination and background conditions and augmented to improve robustness. Experimental results demonstrate that L-Net outperforms EfficientNet-B0, ResNet-50, and MobileNet, achieving an accuracy of 98.99%, a recall of 99.17%, and an F1-score of 99.00%, with minimal false classifications. Ablation and confusion matrix analyses confirm the effectiveness of attention and residual learning in improving model stability and generalization. The proposed L-Net architecture offers a reliable and efficient solution for real-time and edge-based weed detection in precision agriculture.