Accurate and non-destructive prediction of lettuce quality traits is essential for variety identification, germplasm utilization, and intelligent breeding. However, existing approaches relying on handcrafted features or purely data-driven models face limitations under small-sample conditions, including constrained prediction accuracy, weak interpretability, and an increased risk of overfitting. To address these challenges, we propose a knowledge-guided feature tokenizer transformer (KG-FT-Transformer) framework for hyperspectral quality trait prediction and fingerprint analysis. This framework integrates domain prior knowledge with data-driven learning, significantly improving prediction accuracy while enhancing biological interpretability. The KG-FT-Transformer employs a Transformer-based architecture integrating multi-head attention (MHA) with a gated feed-forward network (GFFN), enabling nonlinear spectral modeling and rich feature interactions. We evaluated its performance on three key quality traits: relative chlorophyll content (SPAD), soluble solids content (SSC), and moisture content (MC). The model achieved R 2 values of 0.9534, 0.9185, and 0.9226, with corresponding residual predictive deviation (RPD) values of 4.63, 3.50, and 3.60, outperforming all baseline models and demonstrating stable and consistent prediction performance. Moreover, pixel-wise predictions were used to construct quality trait fingerprints through pseudo-color mapping, intuitively visualizing the spatial distribution and varietal specificity of traits. SPAD and SSC exhibited visually consistent central aggregation patterns, while MC revealed distinct spatial variations among cultivars. These fingerprint-based representations provide spatially informed, qualitative references that may assist the interpretation of DUS-related (Distinctness, Uniformity, Stability) trait characteristics. Overall, this study demonstrates the potential of integrating hyperspectral prediction with quality fingerprinting for non-destructive quality assessment and breeding-oriented analysis, and provides a complementary perspective for germplasm identification and crop improvement. • A knowledge-guided feature tokenizer Transformer (KG-FT-Transformer) is proposed for lettuce quality trait prediction under small-sample conditions. • The proposed model enables accurate and stable prediction of SPAD, SSC, and MC, supporting rapid and non-destructive assessment of lettuce quality traits. • Pixel-wise prediction results are visualized as quality trait fingerprints, intuitively revealing spatial distribution patterns and providing complementary references for quality visualization, DUS-related interpretation, and germplasm analysis.
Qiu et al. (Wed,) studied this question.