LLM-assisted multichannel similarity fusion versus supervised deep learning for raman-based early detection of citrus huanglongbing with metabolomic validation

This is the authors' abstract. We don't add key points for this paper.

Timely detection of citrus Huanglongbing (HLB) during its prolonged asymptomatic phase remains critical for epidemic management, yet current molecular diagnostics require destructive sampling and extended turnaround. This study presents a non-destructive diagnostic framework integrating handheld Raman spectroscopy, targeted metabolomics, and a large language model-assisted multichannel spectral similarity fusion (LAMSSF) framework benchmarked against supervised deep learning for few-shot early diagnosis. Raman spectra from 1030 Thai pomelo and 480 Wogan mandarin leaves reveal consistent spectral patterns across healthy, early HLB, symptomatic HLB, and nutrient deficiency groups. High-performance liquid chromatography-mass spectrometry validation identifies progressive accumulation of p-coumaric acid, glucose, and L-glutamate as the biochemical basis for these signatures, reflecting phenylpropanoid defense activation, phloem transport impairment, and nitrogen metabolic reprogramming, respectively. While a convolutional neural network (CNN)-Transformer achieves the highest accuracy of 94.8% via cross-cultivar fine-tuning, it requires extensive quantitative PCR-annotated data. In contrast, the proposed LAMSSF framework attains comparable accuracy of 92.2% at equivalent data scales while generating transparent diagnostic reports. Critically, LAMSSF maintains above 90% accuracy with only 24 Thai pomelo and 12 Wogan samples as the training data, a data regime where the CNN-Transformer requires 8.3-fold and 10.0-fold more annotations to match performance. This study demonstrates that while deep learning yields marginal accuracy gains given abundant data, the proposed LAMSSF framework enables scalable, few-shot early HLB diagnosis with minimal annotation requirements and significantly greater field deployment potential.