Dual-Level Probability Consistency FS-PFCE for Robust Transfer Classification of Rice Seed Vigor Using Near-Infrared Hyperspectral Imaging

Rice seed vigor is an important indicator for evaluating seed quality and germination capability. Near-infrared hyperspectral imaging technology has shown great potential for seed vigor assessment. However, significant spectral distribution differences often exist among different rice varieties, resulting in degraded transfer classification performance. Moreover, under small-sample conditions, models have difficulty in sufficiently learning shared discriminative information among different varieties, leading to limited transfer generalization capability and restricting their application in cross-variety agricultural scenarios. In this paper, we propose a novel Dual-Level Probability Consistency FS-PFCE (DPC-FS-PFCE) model to achieve robust transfer classification of rice seed vigor across multiple varieties using near-infrared hyperspectral imaging. A total of 900 seed samples from three rice varieties were used, with 600 Chunyou83 samples serving as the master dataset and 150 samples each from Zhongzu53 and Zhongzao39 serving as slave datasets; transfer performance was evaluated on Zhongzu53, Zhongzao39, and their mixed dataset (MIX). Different probability consistency loss strategies were investigated to achieve joint utilization of global and local probability information. Furthermore, a multi-class weight direction consistency constraint was introduced to enhance parameter alignment and knowledge transfer between the master and slave models. The results demonstrated that the master model established using the Chunyou83 dataset achieved classification accuracies of 93.33%, 88.89%, and 87.78% on the Zhongzu53, Zhongzao39, and mixed MIX datasets, respectively, through the DPC-FS-PFCE transfer model. Compared with the Direct Transfer method, the proposed approach improved classification accuracy by 11.11%, 13.33%, and 8.89%, respectively. In addition, the proposed model exhibited more stable transfer classification performance under different master dataset proportions and complex data scenarios. The proposed DPC-FS-PFCE method effectively improves the cross-variety transfer classification capability for rice seed vigor assessment and provides a promising technical approach for nondestructive quality evaluation of crop seeds.