Enhancing Precision Weed Management with Federated Learning-Based GAN and Transfer Learning

Abstract

Precision weed management is critical to sustainable cotton production because overuse of herbicides can have detrimental environmental and health effects and increase production costs. The problem of weed recognition from field images is still difficult even with complex background, intra-class variation, class imbalance and limited number of images labeled for training. In this regard, this study proposes FedAgriGAN, a federated learning-based generative adversarial network to generate augmented cotton weed images and classify the images into the Cotton Weed ID15 dataset. The proposed framework uses one generator and three discriminators to collaboratively encode structural, textural and semantic features of field images, in order to generate a variety of realistic and diverse synthetic samples. The generated images are used to increase the size of the training set and to remove class imbalance. FedAgriGAN is compared with representative GAN models, such as InfoGAN, DCGAN, ViTGAN, and WGAN, both in terms of image qualities and classification performance. Experimental results show that FedAgriGAN outperforms the other models with an average FID score of 287.45. Moreover, the Vision Transformer (ViT) classifier achieves 98.57% testing accuracy, 97.81% testing precision, 97.33% testing recall, and 97.57% testing F1-score using the enhanced dataset. The improvements observed are statistically significant (p < 0.01) proving the efficiency of the proposed framework for precision weed management applications.