Hyperspectral imaging combined with texture features for maize hybrid purity detection: a multi-model comparison based on machine learning and SHAP interpretability study.

Abstract

Hybrid maize performance depends strongly on the genetic purity of hybrid seeds, but female self-pollinated seeds and target hybrid seeds are difficult to distinguish by conventional visual inspection because of their highly similar phenotypes. This study developed a nondestructive and interpretable maize hybrid purity detection framework by integrating hyperspectral imaging and RGB-derived texture features. Hyperspectral images were acquired from the embryo and endosperm sides of five female parents, one common male parent, and their corresponding hybrids. Texture-based, full-band spectral, characteristic-band spectral, and texture-spectral fusion models were systematically constructed and compared. Competitive Adaptive Reweighted Sampling (CARS), Successive Projections Algorithm (SPA), and Synchronous Two-Dimensional Correlation Spectroscopy (Sync2D) were used for characteristic wavelength selection. The results showed that the embryo side provided more stable and discriminative spectral information than the endosperm side. Texture-only models showed limited ability to distinguish hybrids from female self-pollinated seeds, whereas embryo-side texture-spectral fusion models combined with CARS or SPA and Support Vector Machine (SVM) or Partial Least Squares Discriminant Analysis (PLS-DA) achieved average test accuracies of 0.99-1.00, meeting the national maize hybrid seed purity requirement of 97%. In the optimal low-dimensional models, the retained high-dimensional spectral variables were compressed to 28-69 key features, corresponding to a dimensionality reduction ratio of approximately 88%-95%. SHAP analysis identified mean saturation and seed size as important texture features; among spectral intervals, the 450-462 nm region appeared among the top-ranked embryo-side SHAP features in all five maize lines and showed the highest embryo-side mean absolute SHAP magnitude (0.0107 ± 0.0028). Overall, the proposed framework provides a high-throughput, low-dimensional, and interpretable solution for maize hybrid seed purity detection.

Plant phenotyping relevance

ハイパースペクトル画像、RGBテクスチャ、特徴選択、機械学習を統合し、トウモロコシ種子のハイブリッド純度を非破壊推定する手法の開発・比較が研究の中心である。

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.