Few-Shot Image Classification of Crop Diseases Based on Vision-Language Models.

Area of Science:

• Agricultural Science
• Computer Vision
• Artificial Intelligence

Background:

• Accurate crop disease classification is vital for food security and agricultural productivity.
• Current methods often rely on single image types and extensive datasets, limiting their practical application.
• Existing unimodal approaches face challenges with data scarcity and complex disease recognition.

Purpose of the Study:

• To develop an effective crop disease classification method that overcomes data limitations and improves accuracy.
• To enhance multimodal synergy for superior crop disease identification compared to traditional unimodal techniques.
• To provide a pragmatic tool for agricultural pathology and smart farming surveillance.

Main Methods:

• Utilized pre-trained Vision-Language Models (VLMs) for multimodal crop disease classification.
• Employed Qwen-VL to generate detailed textual descriptions from clustered disease images for prompt text generation.
• Integrated cross-attention and SE (Squeeze-and-Excitation) Attention into training-free (VLCD) and training-required (VLCD-T) modes for enhanced classifier weights.

Main Results:

• Demonstrated significantly improved classification effectiveness in few-shot crop disease scenarios.
• The multimodal approach, using VLM-generated prompts, outperformed traditional unimodal methods.
• Enhanced prompt quality by capturing fine-grained, image-specific information.

Conclusions:

• The proposed VLM-based method effectively addresses data limitations in crop disease classification.
• This approach offers a robust solution for agricultural pathology and smart farming.
• The study reinforces the potential of multimodal AI in enhancing agricultural surveillance and productivity.