SugarFedNet Federated Learning Model with Ant Lion Optimization for High-Accuracy Sugarcane Leaf Disease Identification

This study proposes a Federated Learning-based SugarFedNet framework for sugarcane leaf disease identification. In this framework, a hybrid deep learning approach is designed to train the local model, which integrates a residual network, a Modified Swin Transformer, and a convolutional block attention model. The ResNet model is applied to derive the low- and mid-level spatial features, while the Modified Swin Transformer is employed to capture the long-range dependencies. Further, the Convolutional Block Attention module is deployed to highlight the disease-relevant regions in the leaf images. In addition to that, Ant Lion Optimization with the Dynamic Opposite Strategy approach is utilized to tune the hyperparameters and to optimize the local model. The proposed SugarFedNet framework is integrated with Federated Learning to preserve the security and data privacy of various clients. The various experimental evaluations are performed on the sugarcane leaf disease classification dataset, sugarcane disease dataset, and sugarcane leaf disease dataset to estimate the effectiveness of the proposed SugarFedNet framework. The proposed SugarFedNet framework achieved a superior accuracy of about 98.78%, F1-score of about 98.31%, and a specificity of about 97.97%, over existing models. These attained values exhibit the effectiveness of the SugarFedNet framework in the detection of sugarcane leaf disease.