Deep learning and XAI-enabled prediction of maize seed vigor phenotypes and physiological trait with spectral feature association analysis

Abstract

Seed vigor is a key indicator of seed quality, directly influencing plant growth and yield. This study proposes a novel deep learning framework for the qualitative and quantitative assessment of maize seed vigor. First, the Multi-scale residual gated recurrent unit network (MS-ResGRU-Net) was developed for maize spectral vigor detection, achieving an accuracy of 94.35 %. Second, a two-stage model optimization strategy was employed, transferring deep spectral features from MS-ResGRU-Net to the ensemble learning model, further improving the vigor detection accuracy to 95.48 %. The model facilitated quantitative analysis of vigor-related phenotypic traits and physiological indicator, achieving Pearson correlation coefficients of 0.8130 for root length, 0.8057 for root weight, and 0.7876 for physiological indicator between predicted and true values. Furthermore, Explainable artificial intelligence (XAI) was utilized to elucidate the relationships among model features, spectral features, and seed vigor traits, providing clearer insights into the model’s decision-making process. This study presents a non-destructive, efficient method for detecting maize seed vigor, offering a novel approach for assessing the vigor of other crop seeds.

Plant phenotyping relevance

トウモロコシ種子の活力と根長・根重などの表現型を非破壊スペクトル測定と深層学習で推定する手法を開発・評価しており、表現型取得・抽出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.