基于无监督模型转移的苹果跨产地研究

Research on Apple Cross-Producing Area Detection Based on Unsupervised Model Transfer

摘要: 针对苹果近红外光谱可溶性固形物(SSC)检测中，不同产地光谱系统性偏移导致模型跨产地泛化性能严重衰减，传统无监督模型转移方法校正精度不足、易过校正的问题，本文开展基于滤波学习的无监督模型转移方法(uCTRL)的苹果跨产地无标样品质检测研究。该方法通过学习线性滤波器投影矩阵，将源域与目标域光谱映射至共享特征子空间，在最大化保留SSC预测有效特征的前提下，最小化产地间光谱分布差异，实现无监督条件下的跨域模型校正。以北京平谷、新疆阿克苏192个红富士苹果样本为对象开展验证，结果表明，经uCTRL校正后，模型双向跨产地预测RMSEP最大降幅达62.90%，RPD最高提升至2.3432，满足定量检测要求，性能显著优于uDOP、di-PLS经典方法，且无过校正风险，为苹果品质现场规模化无标样检测提供了可靠技术方案。

Abstract: Aiming at the problems that the systematic offset of spectra from different producing areas leads to serious attenuation of the cross-producing area generalization performance of the model in apple soluble solids content (SSC) detection by near-infrared spectroscopy, and the traditional unsupervised model transfer methods have insufficient calibration accuracy and are prone to over-calibration, this paper carries out research on label-free cross-producing area quality detection of apples based on Unsupervised Calibration Transfer via Representation Learning (uCTRL). This method maps the spectra of source domain and target domain to a shared feature subspace by learning a linear filter projection matrix, minimizes the spectral distribution difference between producing areas on the premise of maximally retaining the effective features for SSC prediction, and realizes cross-domain model calibration under unsupervised conditions. Verification was carried out with 192 Red Fuji apple samples from Pinggu of Beijing and Aksu of Xinjiang as objects. The results show that after calibration by uCTRL, the maximum reduction of the model’s bidirectional cross-producing area prediction RMSEP reaches 62.90%, and the highest RPD is increased to 2.3432, which meets the requirements of quantitative detection. Its performance is significantly better than that of classical methods such as uDOP and di-PLS without over-calibration risk, which provides a reliable technical scheme for on-site large-scale label-free detection of apple quality.