Multiclass weed identification using semantic segmentation: An automated approach for precision agriculture

Abstract

Accurate identification and categorization of numerous weed species are critical for implementing effective control measures and management methods in precision agriculture. Manual weed treatment is time-consuming, labor-intensive, and poses risks of human pesticide exposure. Therefore, the development of automated weed management systems is highly desirable. This study aims to propose an automated approach for multiclass weed identification using semantic segmentation, with the goal of improving weed control techniques, reducing pesticide usage, and enhancing crop yields in a sustainable manner. To address the research objective, we created a novel multiclass weed dataset, focusing on two weed categories found in a brinjal farm located in Gorakhpur, Uttar Pradesh, India during the 2022 field seasons. The dataset covers various developmental phases and was captured under ambient lighting conditions. Leveraging transfer learning, we evaluated four advanced deep learning models to establish a benchmark for weed identification. Among the evaluated models, the U-Net-based Inception-ReseNetV2 achieved the highest F1-score of 96.78%, while the other three models attained F1-scores above 91.0%. These findings demonstrate the efficacy of the proposed approach in accurately identifying and categorizing weeds in agricultural fields. The results of this research provide a foundation for further investigations on weed detection and localization in field environments. The use of semantic segmentation for multiclass weed identification can significantly enhance the efficiency and effectiveness of weed management operations, resulting in reduced pesticide usage and improved crop yields. By adopting automated weed management systems, farmers can minimize labor requirements, save time, and mitigate the risks associated with human pesticide exposure.