Residual Derivative-Guided Spectral Fusion Module for Few-Shot Classification of Soybean Seed Varieties Using...

Area of Science:

• Agricultural science
• Spectroscopy
• Machine learning

Background:

• Accurate soybean seed variety identification is crucial for quality control and germplasm management.
• Few-shot classification of soybean seeds is challenging due to similar hyperspectral signatures and limited labeled data.
• Existing methods struggle with distinguishing between closely related soybean varieties using spectral data.

Purpose of the Study:

• To develop an improved spectral representation module for few-shot classification of soybean seed varieties.
• To enhance the accuracy of soybean seed variety identification under limited sample conditions.
• To evaluate the effectiveness of the proposed module when integrated with different meta-learning algorithms.

Main Methods:

• Proposed a Residual Derivative-Guided Spectral Fusion (RDSF) module utilizing raw spectra and their first and second-order derivatives.
• Integrated the RDSF module as a plug-and-play component into Prototypical Network (ProtoNet), Relation Network (RelationNet), and Model-Agnostic Meta-Learning (MAML).
• Evaluated the module using hyperspectral data from 11,000 soybean seeds across 11 varieties under known-class and unseen-class protocols.

Main Results:

• RDSF significantly improved meta-test accuracy in few-shot settings, notably increasing RelationNet accuracy from 0.8898 to 0.9184 in a known-class scenario.
• The module consistently enhanced ProtoNet and RelationNet performance under unseen-class protocols, with a notable accuracy increase from 0.8848 to 0.9094 for ProtoNet in the 5-shot setting.
• The effectiveness of RDSF was partly dependent on the underlying meta-learning mechanism, showing varied improvements across ProtoNet, RelationNet, and MAML.

Conclusions:

• The Residual Derivative-Guided Spectral Fusion (RDSF) module offers an effective approach for enhancing spectral representation in metric-based few-shot classification.
• RDSF demonstrates significant improvements in soybean seed variety identification accuracy, particularly under limited-sample and unseen-class conditions.
• The study highlights the complementary nature of spectral derivatives and the advantages of bounded residual fusion for robust spectral feature extraction.