Data quality assessment and analysis for pest identification in smart agriculture

Deep learning has played a crucial role in the field of smart agriculture and been widely used in various applications. However, the deep learning models are constrained by data quality, which means poor data quality and unreliable data annotation will seriously restrict the performance of smart applications. In this paper, we proposed two methods to assess data quality, named Bound -DE and Multi-Branch. In experiments, the IP06 dataset and the ResNet-18 backbone network were adopted. The results show that the redundancy of the used public dataset is so large that about 50% of the data can achieve the similar test accuracy. Furthermore, we also analyzed the high contributive samples and summarized the rules of those selected informative samples, which is significant for the design of high-efficiency datasets. In summary, this study guides and promotes the following data-centric research in the field of smart agriculture.