Phenotyping maize stay green traits via in situ leaf hyperspectral reflectance sensing

Abstract

Advancements in stay-green phenotyping are increasingly utilizing hyperspectral sensing technology to assess crop response under extreme environmental conditions. Yet, the effectiveness of different spectral features in explaining stay green remains to be fully elucidated. This includes identifying which bands and spectral indices are more effective in capturing the genotypic differences in stay-green traits. The main objective of this study was to evaluate hyperspectral leaf reflectance as a means to estimate stay-green visual scores (SGVS) as an indicator of drought tolerance and to further understand whether chlorophyll absorption-band spectral indices can differentiate SGVS classifications during post-flowering stages of maize. The experiment was conducted over two growing seasons in Germany, comprising 18 maize genotypes under two contrasting water availability conditions. We measured leaf hyperspectral reflectance using a spectroradiometer in the second, fourth, and sixth week after flowering, along with stay-green traits measurements. We employed raw spectral reflectance, hyperspectral vegetation indices (VIs) in combination with random forest (RF) and ANN models to predict SGVS. Results showed that drought stress significantly affected stay-green-related traits and led to a 43.5% decrease in grain yield in the inbred lines. The grain dry yield (GDY) was positively correlated with stay-green visual scores (SGVS), with higher SGVS associated with higher GDY. Stay-green traits were correlated with various VIs, with the best correlation observed for the Chl_NDI (r = 0.91). Stay-green groups were successfully classified using the selected VIs, with the water-absorption band VIs performing better than the chlorophyll-absorption band VIs and other VIs. Similarly, for predicting the SGVS, the water absorption band indices (R² = 0.79 ± 0.04 and RMSE = 0.12 ± 0.01) outperformed the chlorophyll absorption band indices when using RF. Leave-one-out-location/year cross-validation revealed pronounced variation in model transferability driven by environmental and temporal domain shifts. RF consistently outperformed ANN, showing greater robustness to inter-site heterogeneity and interannual variability, whereas performance degraded most in spectrally distinct environments or atypical seasons. Interestingly, RDIS_3b (1280, 1250, 1180 nm), NDIS_2b (2190, 1510 nm), and NDWI2 (860, 1241 nm) were identified as the most critical predictors in the RF models, across merged and separated datasets. These findings demonstrate the potential of spectral signatures, particularly water-absorption band spectral indices, for quantitative phenotyping of stay-green as a proxy for drought tolerance in maize breeding programs; however, multisite, multiyear calibration is needed to enhance generalizability.

Plant phenotyping relevance

トウモロコシのstay-green形質を対象に、葉のハイパースペクトル反射を用いた形質推定・分類モデルを評価し、交差検証で転移性と頑健性も検証しているため、センサー型表現型計測手法が中心である。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.