Processing and Analysis of Spectra Using Machine Learning and Neural Networks

The dissertation is devoted to the application of spectral analysis together with machine learning and neural network methods for the diagnosis of plant conditions, the asymptomatic detection of plant diseases, which is an important topic for increasing agricultural productivity and sustainability. In the modern agro-industrial complex, one of the most important tasks is the fight against plant diseases, which can significantly reduce the yield and quality of crops. The standard diagnosis of plant diseases consists of visual examinations and laboratory tests, which takes a lot of time, and also often does not allow detecting diseases at an early stage. The advent of spectral analysis technologies combined with the power of machine learning or neural networks represents a fundamental solution to these problems. The integration of machine learning and neural networks with spectral analysis for early detection of plant diseases is considered. This approach is very important in agriculture because it makes it possible to carry out a quick, accurate diagnosis and take measures to cure or eliminate the plant before the appearance of visible symptoms. The aim of the study is to develop and validate machine learning and neural network models that will be able to classify plants as healthy or sick based on their spectral data with high accuracy. Preprocessing of spectral data was carried out to ensure consistency, the use of dimensionality reduction methods to increase efficiency, the use of machine learning models for disease prediction, the development of a neural network architecture and the selection of optimal hyperparameters for this architecture for the best classification of plants. The results of the study indicate a high accuracy in detecting diseases. The best achieved neural network with optimized hyperparameters produces an accuracy of 89.59%. Machine learning model was able to produce a better result than the neural network, SVM model, used in combination with LDA reduction and StandardScaler pretreatment, achieved test accuracy of almost 99.7%, and cross-validation was performed, which confirmed that the model was not retrained, which highlights the potential of this technology in the revolutionary fight against plant diseases.