Fusing Perceptual Vision Experts with Multimodal LLMs for Explainable Plant Disease Diagnosis: From Benchmark Imagery to Real-World Robotic Field Validation

Explainability matters for real-world deployment: The system generates not just disease diagnoses but also explains risk levels, treatment urgency, and financial exposure in structured formats. This transparency is crucial for farmers and agricultural professionals who need to understand why a recommendation is being made before acting on it.

AI systems can disagree, and that's useful information: When two different AI models (EfficientNet-B3 and ConvNeXt-Tiny) conflicted on diagnoses, a large language model was better equipped to arbitrate and reach accurate conclusions. On the most conflicted subset of images, accuracy improved by 7.6 percentage points, suggesting that acknowledging uncertainty and using multiple perspectives strengthens decision-making rather than weakening it.

Different AI models have different failure modes that matter: While Gemma's critical-risk errors were minimal (0.14-0.5 points), Qwen significantly over-flagged risks by 3.5-14.4 points. For responsible deployment, this means careful calibration and testing of specific models is essential—a system that cries wolf too often could lead farmers to waste resources or lose trust in the technology.

Real-world performance can differ dramatically from benchmark results: The system achieved 68.5% accuracy on standard benchmark imagery but 98.9-99.3% accuracy on actual robot-collected field data. This gap highlights the importance of testing AI systems in genuine operational conditions before deployment, as controlled datasets may not reflect real-world complexity.

Continuous field validation is necessary for agricultural AI: The researchers used two large, continuously captured datasets from robots working in uncontrolled field conditions. This approach to validation—rather than one-time testing—better ensures the system remains reliable as environmental conditions, crop varieties, and disease presentations change over time.