Combining generative adversarial networks and agricultural transfer learning for weeds identification

Abstract

In recent years, automatic weed control has emerged as a promising alternative for reducing the amount of herbicide applied to the field, instead of conventional spraying. The use of artificial intelligence through the implementation of deep learning for early weeds identification has been one of the engines to boost this progress. However, these techniques usually need very large datasets coping with real-world conditions, which are scarce in the agricultural domain. To address the lack of such datasets, this paper proposes a methodology that combines the use of agricultural transfer learning and the creation of artificial images by generative adversarial networks (GANs). Several architectures and configurations have been evaluated on a dataset containing images of tomato and black nightshade. The best configuration was a combination of GANs creating plausible synthetic images and the Xception network, with a performance of 99.07% on the test set and 93.23% on a noisy version of the same set. Other architectures, such as Inception or DenseNet have also been evaluated, and they obtained promising results by using GANs. According to the results, the combination of advanced transfer learning and data augmentation techniques through GANs should be deeply studied in the future with more complex datasets. • Weeds identification was performed using transfer learning techniques. • Weeds identification was performed using Generative Adversarial Networks. • The methodology presented was evaluated with performances of 99.07%. • A comparison of several Generative Adversarial Networks is provided. • The methodology presented shows an overall better efficiency.