Enhancing weed detection in wheat with hybrid attention mechanism CBAM: a comparative study with deep learning approaches

Introduction Weed infestation is a major constraint on wheat growing, yet traditional detection methods are hampered by the close visual similarity between weeds and crops under field conditions. Although convolutional neural networks (CNN) have a strong potential for automated weed detection, their efficacy remains limited by challenges in feature selection and computational demands. Existing attention-based approaches apply channel and spatial attention in sequence, creating inter-relationship dependencies that limit the quality of the representation of features. This study addresses this gap by proposing a parallel configuration of the CBAM (Convolution Block Attention Module) for the classification of wheat and weed. Methods A dataset of 1 260 field images (784 wheat; 476 grass) was collected in the village of Berli Khurd in Rewari, Haryana, India, under different lighting conditions and supplemented to 6634 images to resolve the class imbalance and improve the overall coverage. Four widely accepted CNN architectures - ResNet-50, ResNet-152, VGG-16 and VGG-19 - and their variants enhanced by CBAM have been systematically assessed for the classification of wheat and weed. The proposed parallel CBAM configuration process will simultaneously focus both the channel and the spatial attention on the same input feature map, allowing independent and complementary refinement of the features. Models were trained for 30 epochs with AdamW optimization (learning rate = 0.0001) and a batch size of 16 at Google TPU (V2-8) runtime. Results CBAM integration significantly improved classification performance across all architectures. CBAM-ResNet-50 and CBAM-ResNet-152 both achieved near-perfect validation accuracy of 99.5%, compared to 89.2% and 98.0% for their baseline counterparts representing improvements of 10.3% and 1.5%, respectively. CBAM-ResNet-50 attained precision of 99.1%, recall of 98.5%, and an F1-score of 99.2%; CBAM-ResNet-152 achieved precision of 99.4%, recall of 98.6%, and an F1-score of 99.3%. VGG-16 and VGG-19 also improved substantially, from 87.0% to 97.3% and from 88.2% to 98.4%, respectively. Conclusion The parallel configuration of the CBAM consistently improved the robustness of the extraction of features, reduced misclassification and improved coverage across all four CNN architectures. Attention-based gains were largest in the residual networks, indicating that skipping the link increases the benefits of parallel hybrid attention. These findings demonstrate the practical value of targeted residual networks for reliable and field-based weed identification in precision agriculture.