Weed Identification in Rice Cultivation Employing Semantic Segmentation

Abstract Weed detection is a crucial task in precision agriculture, as it enables farmers to take targeted measures to control weeds and reduce crop losses. In this paper, an attention-enhanced SegNet model is used in a deep learning-based weed detection method. Images of paddy fields are used to train the network to segregate weeds at the pixel level. In order to identify weeds, this research compares the SegNet architecture with an improved SegNet model that incorporates attention techniques. SegNet serves as a baseline encoder–decoder network, while the SegNet-Attention model incorporates attention gates to emphasize relevant spatial features and suppress background noise. Standard performance criteria, including as accuracy, precision, recall, intersection over union (IoU), and dice coefficient, were used to train and assess both models on a dataset of weed images. According to experimental data, the SegNet-Attention model consistently outperforms the standard SegNet by enhancing weed boundary delineation and segmentation accuracy, especially in complicated field scenarios with overlapping crops and different lighting conditions. In comparison to traditional SegNet designs, experimental research shows that the attention-enhanced SegNet model delivers improved segmentation accuracy and robustness, especially in complicated image settings. These findings demonstrate the applicability of the suggested method for high-precision semantic segmentation applications by showing increased segmentation accuracy and robustness when compared to traditional encoder–decoder models.