Automated weed classification using attention-embedded ConvNeXtV2 architecture

Autonomous weed identification presents a critical challenge in precision agriculture, where early detection is essential to prevent crop failure and enhance yield. Using publicly available datasets, various deep learning approaches have been explored for image-based weed detection across different crops. In this study, we present a novel attention-based ConvNeXtTiny network designed to detect six different weed classes in wheat crop and healthy wheat using an image dataset collected from natural field conditions on a farmer’s field in India. Each ConvNeXt block integrates a Simple Attention Module (SimAM) to selectively extract critical features while minimizing the influence of less relevant regions, thereby improving classification accuracy. Our model achieved outstanding performance, with training and testing accuracies of 99.97% and 98.39%, respectively, and an average F1 score of 98.75% with the lowest mean squared error (MSE). The attention layers are parameter-free, ensuring the model remains lightweight and computationally efficient. This makes it suitable for real-time deployment in weed identification systems on mobile or portable devices with limited computational resources or embedded systems, facilitating timely and effective weed management.