SugarcaneGAN: A U-RSwinT and StyleGAN3-Driven Deep Learning Framework for High-Accuracy Sugarcane Leaf Disease Detection and Classification
Abstract
Early and accurate diagnosis of sugarcane leaf diseases is essential for sustainable agriculture, improved crop productivity, and minimizing industry losses. Manual inspection methods are often inefficient, labor-intensive, and prone to inconsistency, especially in large-scale farming contexts. SugarcaneGAN refers to our custom generative model designed to augment the dataset with high-quality synthetic diseased leaf images, addressing class imbalance and improving model generalization. This study introduces SugarcaneGAN, a comprehensive deep learning framework that integrates advanced image preprocessing, segmentation, generative modeling, and classification techniques to deliver high-accuracy sugarcane leaf disease detection. The pipeline begins with the acquisition of RGB images, followed by Contrast Limited Adaptive Histogram Equalization (CLAHE) to enhance contrast, median filtering for noise suppression, and a transformation to HSV color space to emphasize disease-specific chromatic patterns. For precise segmentation of sugarcane leaves in cluttered environments, the framework utilizes U-RSwinT a U-Net architecture enhanced with a Residual Swin Transformer backbone leveraging hierarchical design and self-attention mechanisms to capture both global context and local features. The Segment Anything Model (SAM) is optionally incorporated for zero-shot segmentation, thereby improving flexibility and robustness. To address dataset limitations such as class imbalance and low variability, a custom generative module named SugarcaneGAN is developed using StyleGAN3 and DiffusionGAN to synthesize high-fidelity diseased leaf images. These augmented samples enhance model generalization during training. Feature extraction is conducted via the ConvNeXt architecture, further enriched by Gray Level Co-occurrence Matrix (GLCM) texture descriptors that capture lesion structure. The classification stage utilizes vision transformer-based models, including MaxViT, DeiT, and ViT-B16, to identify disease type, assess severity, and localize affected regions. This integrated approach enables a scalable, real-time solution for intelligent sugarcane disease monitoring, aligning with the broader goals of precision agriculture, smart farming, and global food security. Experimental evaluation demonstrates SugarcaneGAN achieving over 98% accuracy and F1-score across six disease classes. The model also records a highest AUC of 0.97 and average precision (AP) of 0.90, confirming its superior generalization and reliability under class imbalance.
