Identification of fungal diseases in strawberry by analysis of hyperspectral images using machine learning methods.

Abstract

Leaf spot, leaf scorch and phomopsis leaf blight are the most common fungal diseases of strawberry in Western Siberia, which significantly reduce its yield and quality. Accurate, fast and non-invasive diagnosis of these diseases is important for strawberry production. This article explores the ability of hyperspectral imaging to detect and differentiate symptoms caused to strawberry leaves by pathogenic fungi Ramularia tulasnei Sacc., Marssonina potentillae Desm. and Dendrophoma obscurans Anders. The reflection spectrum of leaves was acquired with a Photonfocus MV1-D2048x1088-HS05-96-G2-10 hyperspectral camera under laboratory conditions using the line scanning method. Five machine learning methods were considered to differentiate between healthy and diseased leaf areas: Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Linear Discriminant Analysis (LDA), Partial Least Squares Discriminant Analysis (PLS-DA), and Random Forest (RF). In order to reduce the high dimensionality of the extracted spectral data and to increase the speed of their processing, several subsets of optimal wavelengths were selected. The following dimensionality reduction methods were explored: ROC curve analysis method, derivative analysis method, PLS-DA method, and ReliefF method. In addition, 16 vegetation indices were used as features. The support vector machine method demonstrated the highest classification accuracy of 89.9 % on the full range spectral data. When using vegetation indices and optimal wavelengths, the overall classification accuracy of all methods decreased slightly compared to the classification on the full range spectral data. The results of the study confirm the potential of using hyperspectral imaging methods in combination with machine learning for differentiating fungal diseases of strawberries.

Code and data availability

The paper describes hyperspectral imaging of strawberry leaves with machine learning classification, but no public phenotype dataset, image collection, analysis code, or trained model is deposited. The only linked resource is the journal's supplementary PDF containing figures (S1–S4) of wavelength-selection results, i.

No evidence-backed public reproduction asset is currently recorded.