Fusarium head blight detection from spectral measurements in a field phenotyping setting — A pre-registered study

Abstract

Spectroscopic methods can contribute to addressing the field phenotyping bottleneck problem in crop breeding programs. In disease resistance phenotyping, spectral signatures can be analysed to derive infection severity scores and to screen breeding lines. Hyperspectra of winter wheat spikes were acquired in a Fusarium head blight phenotyping trial at the milk- and wax-ripening phenological phases. Disease severity ratings were simultaneously performed by an expert on a 9-point visual scale. Ordinal support vector machine models were then trained to assign hill plots to the individual severity levels. The predictive models' performance was evaluated for data collection timing, spectral pre-processing and permitted rating-error tolerance. The models trained to spectra acquired at the milk-ripening phase were sufficiently accurate to reliably distinguish between low, medium and high symptom severity; with accuracy approaching 100% for two-point error tolerance. However, deterioration in prediction quality was noted for the wax-ripening campaign, presumably due to spike-drying. After aggregation of the spectra using the median function no gain could be associated with further pre-processing. Modest performance improvements obtained with two schemes do not justify the additional data acquisition costs involved, but standard normal variate could be advantageous for some scenarios with mean-aggregated spectra. In addition to phenotyping, the results are discussed in relation to large-scale farming applications. Elevated infection risk detection prior to anthesis is recommended for fungicide treatment, considering the pathogen biology. The study is accompanied by a publicly-available dataset and the computational scripts employed to obtain the results.

Plant phenotyping relevance

スペクトル測定と機械学習によりコムギ赤かび病の感染重症度を推定し、収集時期・前処理・誤差許容度を評価しているため、植物表現型取得法の検証・応用が中心です。

Code and data availability

The authors deposited the paper's spectral phenotyping dataset (hyperspectra of winter wheat spikes, visual symptom scores) together with the computational analysis scripts and a GNU Guix environment specification in a public Zenodo repository, explicitly excluding only the unused hyperspectral image cubes.