Hyperspectral Non-Imaging Measurements and Perceptron Neural Network for Pre-Harvesting Assessment of Damage Degree Caused by Septoria/Stagonospora Blotch Diseases of Wheat

Abstract

The detection and identification of plant diseases is a fundamental task for sustainable crop production. Septoria tritici and Stagonospora nodorum blotch (STB and SNB) are two of the most common diseases of cereal crops that cause significant economic damage. Both pathogens are difficult to identify at early stages of infection. Determining the degree of the disease at a late infection stage is useful for assessing cereal crops before harvesting, as it allows the assessment of potential yield losses. Hyperspectral sensing could allow for automatic recognition of Septoria harmfulness on wheat in field conditions. In this research, we aimed to collect information on the hyperspectral data on wheat plants with different lesion degrees of STB&SNB and to create and train a neural network for the detection of lesions on leaves and ears caused by STB&SNB infection at the late stage of disease development. Spring wheat was artificially infected twice with Septoria pathogens in the stem elongation stage and in the heading stage. Hyperspectral reflections and brightness measurements were collected in the field on wheat leaves and ears on the 37th day after STB and the 30th day after SNB pathogen inoculation using an Ocean Insight “Flame” VIS-NIR hyperspectrometer. Obtained non-imaging data were pre-treated, and the perceptron model neural network (PNN) was created and trained based on a pairwise comparison of datasets for healthy and diseased plants. Both statistical and neural network approaches showed the high quality of the differentiation between healthy and damaged wheat plants by the hyperspectral signature. A comparison of the results of visual recognition and automatic STB&SNB estimation showed that the neural network was equally effective in the quality of the disease definition. The PNN, based on a neuron model of hyperspectral signature with a spectral step of 6 nm and 2000–4000 value datasets, showed a high quality of detection of the STB&SNB severity. There were 0.99 accuracy, 0.94 precision, 0.89 recall and 0.91 F-score metrics of the PNN model after 10,000 learning epochs. The estimation accuracy of diseased/healthy leaves ranged from 88.1 to 97.7% for different datasets. The accuracy of detection of a light and medium degree of disease was lower (38–66%). This method of non-imaging hyperspectral signature classification could be useful for the identification of the STB and SNB lesion degree identification in field conditions for pre-harvesting crop estimation.