Semantic segmentation of microbial alterations based on SegFormer.

Area of Science:

• Agricultural Science
• Computer Vision
• Plant Pathology

Background:

• Accurate identification of strawberry diseases is crucial for effective management and yield preservation.
• Semantic segmentation models offer potential for automated disease detection in agricultural settings.

Purpose of the Study:

• To evaluate the performance of SegFormer models with different Mix Transformer (MiT) encoders for precise semantic segmentation of strawberry diseases.
• To compare the efficacy of MiT-B0, MiT-B3, and MiT-B5 encoders in detecting various strawberry diseases under natural conditions.

Main Methods:

• Utilized SegFormer with MiT-B0, MiT-B3, and MiT-B5 encoders for semantic segmentation of strawberry diseases.
• Trained and evaluated models on a dataset of 2,450 raw and 4,574 augmented images.
• Employed Segment Anything Model within Roboflow for efficient data annotation.

Main Results:

• MiT-B0 showed balanced but slightly overfitting performance.
• MiT-B3 demonstrated rapid adaptation and consistent performance.
• MiT-B5 provided efficient learning and robust performance, with MiT-B3 and MiT-B5 outperforming MiT-B0, and MiT-B5 achieving the most precise segmentation.

Conclusions:

• MiT-B3 and MiT-B5 are recommended for strawberry disease detection applications due to their superior segmentation accuracy.
• The SegFormer approach shows promise for broader applications in crop disease analysis and automated agriculture.