Answers are grounded in this paper and cite their sources.
This study explores Few-Shot Learning (FSL) to improve a pre-trained Transformer model for classifying avocado ripeness using Near-Infrared (NIR) spectral data from a portable, low-cost scanner. The goal is to allow users in the field to customise the model with little local data, decreasing reliance on centralised retraining while keeping predictive accuracy intact. A dataset was created with spectral scans from the AS7265x sensor, capturing 18 wavelengths for each sample. The population consists of Buccaneer avocado samples at three ripeness stages: raw, aged, and ripe. Subsets of different sizes were extracted for simulation to represent various k-shot support scenarios, from 2% to 50% of the total labelled data per class. A few-shot Transformer architecture classified samples in episodic learning conditions. We assessed performance through average F1-score, ROC AUC, and confusion matrices using 5-fold cross-validation. Findings show that with as little as 20% labelled support data per class, the model achieves over 80% F1-score, demonstrating a clear performance gain through minimal shot expansion. Classification accuracy keeps improving until it levels off around 40–50% shot ratios. This study advances embedded agricultural AI by demonstrating the effective use of few-shot adaptation in portable NIR scanners for localised model improvement. Limitations involve possible sensitivity to class imbalance and noise in low-shot scenarios, necessitating strong sampling and calibration.
No takes yet. Share an insight, caveat, or question.
Tipauksorn et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: