Optimized Crop Disease Segmentation with High Accuracy Using Efficient U-Net

Aim: In smart agriculture, timely segmentation of potato diseased leaves is vital for disease control. Tiny disease spots often result in data loss and low segmentation accuracy. EfficientUNet-B3, introduced as a novel segmentation framework, enhances parameter efficiency and feature extraction to accurately capture disease details. Its UNet decoder combines contextual and spatial data, enabling precise segmentation of small disease patches. Material and Methods: Group 1, RS-UNet, obtained a Dice coefficient as high as 88.86% and MIoU as 79.8%, thus establishing its effectiveness in the segmentation of regions of potato leaf disease with 1000 samples. For example, the Dice coefficient of Group 2, EfficientUNet, was 90.86%, while the MIoU was 83.25%, thanks to improvements in feature extraction and the optimized architecture of EfficientNet with 1000 samples. Result: Compared to RS-UNet, the baseline architecture, EfficientUNet-B3 improves mean Intersection over Union (MIoU) from 79.8% to 83.25% and the Dice coefficient from 88.86% to 90.86%, outperforming ResNet50 by 3.45% and 2%, respectively. Integrating the attention mechanism further enhances focus on disease-specific regions and suppresses irrelevant data. Conclusion: EfficientUNet-B3 demonstrates superior Dice coefficient and generalization, making it a promising solution for high-precision segmentation tasks in potato leaf disease prevention. This work highlights deep learning's potential in promoting intelligent farming and sustainable crop health monitoring.