DWTFormer: a frequency-spatial features fusion model for tomato leaf disease identification.

Area of Science:

• Agricultural Science
• Computer Science
• Plant Pathology

Background:

• Tomato leaf diseases present significant identification challenges due to high inter-class similarity and intra-class variability.
• Existing identification models often struggle with accuracy because of these inherent complexities.

Purpose of the Study:

• To develop a novel and accurate tomato leaf disease identification model.
• To address the limitations of current models in distinguishing between similar diseases and variations within a single disease type.

Main Methods:

• Proposed DWTFormer model utilizing frequency-spatial feature fusion.
• Incorporated a Bneck-DSM module for shallow feature extraction.
• Developed a dual-branch feature mapping network (DFMM) with 2D discrete wavelet transform for frequency features and multi-scale convolution with Pyramid Vision Transformer (PVT) for spatial features.
• Employed dynamic cross-attention for fusing frequency-spatial features.

Main Results:

• DWTFormer achieved 99.28% accuracy on a tomato leaf disease dataset.
• Demonstrated high performance on external datasets (AI Challenger 2018: 96.18%, PlantVillage: 99.89%).
• In-field tests showed 97.22% accuracy with a 0.028-second inference time.

Conclusions:

• The DWTFormer model effectively mitigates the impact of inter-class similarity and intra-class variability in tomato leaf disease identification.
• Offers a scalable solution for rapid and precise disease identification in real-world agricultural settings.