Análise comparativa do pré-processamento de dados na classificação de sementes

Abstract

Defining the type and quality of a seed before planting is vital for the harvest, as the use of low-quality seeds can result in low productivity, even under favorable growing conditions. This task has traditionally been performed manually, making it arduous, time-consuming, and prone to errors. Machine learning techniques can solve this problem by using classification algorithms to label seeds into different classes, reducing difficulty, time consumption, and the number of errors. However, these techniques are subject to the quality of the data provided to them, such as the presence of missing data, outliers, and unnormalized data, which directly impact the performance of the classification, as it is generally sensitive to these characteristics in the datasets provided. To increase the quality of the data provided to the classifiers, data preprocessing techniques can be applied. This paper investigates the impact of data preprocessing techniques on the performance of machine learning models in the classification of dry bean seeds. Using a dataset from the UCI Machine Learning Repository, derived from an experiment by Koklu and Ozkan (KOKLU; OZKAN, 2020), various preprocessing techniques, such as missing value imputation, outlier removal, and data normalization, were applied. kNearest Neighbors (kNN) and Multi-Layer Perceptron (MLP) classification methods were used to evaluate the effectiveness of these preprocessing techniques. Additionally, an enhanced MLP model with optimized parameters, including the learning rate and hidden layer configuration, was proposed. The experimental results demonstrate the critical role of data preprocessing, especially data normalization. The enhanced MLP model significantly outperformed the baseline model, highlighting the importance of using an optimized neural network model to solve the problem.