No abstract

This paper describes an experimental study on the effect of reducing time series collected from IoT electrical agro-sensors through approximation techniques, in time series classification tasks, for plant stress detection. From large sets of real data, stored in time series format, experiments were carried out to analyze: (i) performance of mathematical methods to reduce the dimensionality of time series - PAA, SAX and MCB; and (ii) Whether the application of these techniques influences the performance of time series classification models for plant stress detection, using machine learning algorithms KNN, SVM and ANN. Both in terms of data volume reduction and time series classification, the experiment showed significant improvements in terms of compression rate and accuracy, with the best result found in the use of PAA+SAX techniques for reduction and SVM for classification.

The recognition of patterns in the electrical activities of plants (electromes, in time series format) has gained prominence in recent years. The use of Internet of Things (IoT) devices and Machine Learning (ML) techniques has automated and enhanced data collection and classification, helping researchers identify behaviors and classify them to detect plant stress. However, processing this information means dealing with large amounts of data, which is a major challenge from a computer science perspective. Thus, in this work, we propose an approach for reduction and classification of time series representing plant electromes to balance the trade-off between reduction and data quality, without compromising the classification task. We investigated the use of three time series approximation techniques (PAA, SAX, and MCB) in combination with ML algorithms, such as ANN, KNN, and SVM, in order to find the most suitable approach for this scope. The results validated the proposed approach, with the best performance obtained with the PAA+SAX techniques combined with the SVM algorithm, achieving good data reduction and improving stress detection, without compromising data quality. The main challenges in these tasks and future research directions are also discussed.