Early weed identification based on deep learning: A review

Abstract

Weeds were one of the most destructive constraints on crop production and posed a significant threat to agricultural productivity. The increasing development of smart agriculture promoted the innovation and development of precise weed control techniques. With the application of deep learning in agriculture, more and more emerging technologies have been applied to weed identification. This paper reviewed recent emerging technologies based on deep learning in weed detection. First, the definition, development, and application of technologies such as transfer learning, neural architecture search, domain adaptation, knowledge distillation, and generative adjunctive neural networks are effectively summarized and explained. Next, specific cases of new technologies to solve the challenges in the field of early weed identification are provided. Finally, the technical challenges and possible future roadmap for emerging technologies are discussed and proposed. Collaboration between emerging technologies would become a more attractive development direction than upgrading single technologies in this roadmap. This review would provide a timely field survey and attract more researchers to address this interdisciplinary research issue.

Keywords

• Transfer learning
• Neural architecture search
• Weed identification
• Deep learning
• Smart agriculture