Seed purity assessment by means of spectral imaging

In this work, we propose a technique for identifying impurity grains from spectral images using neural networks that is able to analyze a heap of seeds, grouping grains with similar spectral and morphological characteristics and optimizing the main stages of forming a training sample of a neural network model, recording and processing data. An architecture of the neural network model is proposed based on sequentially running LSTM layers and fully connected layers of neurons. Approaches are proposed for choosing the training sample size, the number and position of central wavelengths of video spectrometer channels used in analysis, and a method for segmenting spectral images to form a training sample. The developed methodology is distinguished by the ability to analyze a heap of seeds and the ease of replenishing the database of distinguished crops and impurities. Testing of the method on wheat and barley seeds showed high classification accuracy (over 99 %) even for grains with very similar spectral and morphological characteristics. The proposed approach increases the accuracy, productivity and objectivity of assessing the purity of seed material, does not require the involvement of experienced personnel and, thus, may be expected to facilitate the introduction of video spectrometers when addressing research and production problems of the agro-industrial complex.