关键词: 冬小麦, 冠层高光谱, 全生育期, 适应性评价, 产量预测, 集成学习

Abstract:

【Objective】Aiming at the problem that traditional single-temporal spectroscopy is difficult to accurately acquire time-series dynamic phenotyping throughout the whole wheat growth period and has insufficient capability for evaluating high-yield and stable-yield potential, this study integrates the full-growth-period time-series hyperspectral features with the Stacking ensemble learning algorithm to construct a method for evaluating and selecting wheat germplasm with stable yield potential, which takes into account both qualitative classification and quantitative yield estimation.【Method】This study selected 100 winter wheat genotypes as test materials. Canopy hyperspectral images were collected at 16 key growth stages over two consecutive years (2022-2023 and 2023-2024). The temporal mean (cumulative photosynthetic potential), standard deviation (stay-green stability) and maximum value (maximum growth potential) of the time-series spectral vegetation indices were calculated. A 21-dimensional feature space containing temporal statistical features and interaction features was constructed for dimensionality reduction and information enrichment. After applying within-year standardization to eliminate annual sensor and environmental drift, the K-Means algorithm (K=3) was applied to classify 195 valid variety-year samples under an unsupervised framework. On this basis, a stacking ensemble regression model was trained on 92 samples from the first year to predict the yield of 98 unseen genotypes in the second year (Leave-One-Year-Out validation), with Variance Inflation applied to address the regression toward the mean problem.【Result】The principal components analysis (PCA) results showed that the first two PCs successfully interpreted cumulative photosynthetic potential and growth stability, with a cumulative variance contribution rate exceeding 85%. The subsequent K-means clustering classified the varieties into three categories: high-yield and stable-yield, general adaptability and low-yield sensitive. Subgroup-level yield variance analysis showed highly significant differences in measured plot yield among the three groups (F=17.83, P=7.91×10^-8), and the average plot yield of the high-yield stable-yield group was 17.61% higher than that of the low-yield sensitive group. Evaluated against within-year yield tertiles as ground truth, the overall classification accuracy (OA) reached 50.77% (Kappa=0.263), and the safety retention rate for high-yielding genotypes reached 93.94%, confirming that this unsupervised framework could effectively identify and eliminate low-performing germplasm. Feature analysis showed that whole-growth-period GNDVI and OSAVI contributed highly to evaluating the high-yielding groups. This physiologically explained that green and red-edge bands had stronger canopy penetration than the red band, effectively overcoming the saturation limitation of traditional NDVI under high canopy closure in the mid-to-late growth stages. In the cross-year independent validation, the predicted yields were significantly correlated with the measured yields (r=0.51, P=1.99×10^-7, RMSE=0.38 t·hm^-2), and the prediction slope was successfully adjusted to 0.31.【Conclusion】This study verified the effectiveness of time-series dynamic canopy spectroscopy in identifying stable and high-yielding wheat genotypes, and proposed a breeding decision-making framework combining qualitative screening with quantitative ranking, which provided a methodological reference for germplasm selection.

Key words: winter wheat, canopy hyperspectral, whole growth period, adaptability evaluation, yield prediction, ensemble learning