Uso de inteligência computacional na fenotipagem de soja
Loading...
Date
Journal Title
Journal ISSN
Volume Title
Publisher
Universidade Federal de Viçosa
Abstract
A soja (Glycine max (L.)) é uma fonte valiosa de alimento humano, animal e como matéria prima industrial. Para atender às demandas crescentes, a soja enfrenta uma série de desafios complexos. Esses desafios estão intrinsecamente ligados à busca por variedades mais produtivas, resistentes a doenças e adaptadas a condições ambientais variáveis. A fenotipagem de alto rendimento emerge como uma ferramenta crucial nesse processo, acelerando o desenvolvimento eficiente de novas variedades mais produtivas, capazes de enfrentar os desafios ambientais e sociais em constante evolução. Esta oferece vantagens quando na comparação com a fenotipagem tradicional, uma vez que a partir do uso de imagens e sensores, auxiliados com softwares e algoritmos de inteligência computacional, otimizam o processo de medição, possibilita maior fenotipagem em escala e reduz a variabilidade da mensuração humana. A utilização de índices de vegetação está entre os principais meios utilizados na fenotipagem de alto rendimento. Com a utilização destes, é possível realizar uma variedade de estudos, incluindo a avaliação do teor de nitrogênio nas folhas, a determinação de características físicas como biomassa, altura da planta e área foliar, a análise da heterogeneidade das plantas no campo, a estimativa do teor de clorofila, a avaliação do teor de água nas plantas, a quantificação do teor de lignina, e a detecção de danos causados por pragas e doenças. Com o uso de técnicas de fenotipagem em larga escala, torna-se crescente volume e a complexidade dos dados obtidos, aumentando a exigência para abordagens inovadoras para análise dos dados com foco em seleção eficiente. Nesse contexto, a aplicação de inteligência computacional tem emergido como uma ferramenta essencial para transformar a forma como se aborda a fenotipagem e o melhoramento de plantas. Assim, este trabalho tem como objetivo avaliar o desempenho de algoritmos de Perceptron Multicamadas (PMC) e Random Forest (RF) na predição de características fenotípicas de soja a partir dos seguintes índices de vegetação: Índice de vegetação de Diferença Normalizada (NDVI), Índice de Vegetação Ajustado ao Solo (SAVI), Índice de Vegetação Ajustado ao Solo Modificado (MSAVI), Índice de vegetação de Diferença Normalizada Verde (GNVDI), Índice de Vegetação Aprimorado (EVI), Índice de Diferença Normalizada de Borda Vermelha (NDRE) e Índice Simplificado de Conteúdo de Clorofila do Dossel (SCCCI). Além disso, busca-se investigar o impacto dos parâmetros das redes neurais no desempenho de um PMC e como esses parâmetros influenciam sua eficácia na predição de características fenotípicas de soja. Constatou- se que, para os modelos de PMC, conforme a quantidade de neurônios aumentava até 10, e o número de folds na validação cruzada aumentava até 15, os modelos apresentavam resultados progressivamente melhores. Por outro lado, o aumento do número de épocas resultava em um aumento nos valores de R² dos modelos, alcançando um limite de 30000, após o qual os resultados começavam a diminuir. Ademais, o acréscimo no número de camadas ocultas ocasionava uma redução nos valores dos coeficientes de determinação, indicando que o melhor número de camadas foi um. Além disso, houve destaque na predição das variáveis AIV e AP, especialmente em relação aos modelos de Random Forest avaliados. Palavras-chave: Predição; Aprendizado de Máquina; Índices de Vegetação; Glycine max (L).
Soybean (Glycine max (L.)) is a valuable source of human, animal, and industrial food. To meet growing demands, soybeans face a series of complex challenges. These challenges are intrinsically linked to the search for more productive varieties, resistant to diseases, and adapted to variable environmental conditions. High-throughput phenotyping emerges as a crucial tool in this process, accelerating the efficient development of new, more productive varieties capable of addressing constantly evolving environmental and social challenges. It offers advantages compared to traditional phenotyping, as it uses images and sensors aided by software and computational intelligence algorithms, optimizing the measurement process, enabling greater phenotyping on scale, and reducing human variability. The use of vegetation indices is among the main methods used in high-throughput phenotyping. With their use, a variety of studies can be conducted, including the evaluation of nitrogen content in leaves, determination of physical characteristics such as biomass, plant height, and leaf area, analysis of plant heterogeneity in the field, estimation of chlorophyll content, evaluation of plant water content, quantification of lignin content, and detection of damage caused by pests and diseases. With the use of large-scale phenotyping techniques, the volume and complexity of the obtained data are increasing, increasing the demand for innovative approaches to data analysis focused on efficient selection. In this context, the application of computational intelligence has emerged as an essential tool for transforming the way plant phenotyping and improvement are approached. Thus, this work aims to evaluate the performance of Multilayer Perceptron (MLP) and Random Forest (RF) algorithms in predicting soybean phenotypic characteristics from the following vegetation indices: Normalized Difference Vegetation Index (NDVI), Soil-Adjusted Vegetation Index (SAVI), Modified Soil-Adjusted Vegetation Index (MSAVI), Green Normalized Difference Vegetation Index (GNVDI), Enhanced Vegetation Index (EVI), Normalized Difference Red Edge Index (NDRE), and Simplified Canopy Chlorophyll Content Index (SCCCI). Additionally, it seeks to investigate the impact of neural network parameters on the performance of an MLP and how these parameters influence its effectiveness in predicting soybean phenotypic characteristics. It was found that, for MLP models, as the number of neurons increased up to 10, and the number of folds in cross-validation increased up to 15, the models showed progressively better results. On the other hand, an increase in the number of epochs resulted in an increase in the R² values of the models, reaching a limit of 30000, after which the results began to decrease. Furthermore, an increase in the number of hidden layers led to a reduction in the values of the determination coefficients, indicating that the optimal number of layers was one. Moreover, there was a highlight in the prediction of the variables AIV and AP, especially regarding the evaluated Random Forest models. Keywords: Prediction; Machine Learning; Vegetation Indices; Glycine max (L).
Soybean (Glycine max (L.)) is a valuable source of human, animal, and industrial food. To meet growing demands, soybeans face a series of complex challenges. These challenges are intrinsically linked to the search for more productive varieties, resistant to diseases, and adapted to variable environmental conditions. High-throughput phenotyping emerges as a crucial tool in this process, accelerating the efficient development of new, more productive varieties capable of addressing constantly evolving environmental and social challenges. It offers advantages compared to traditional phenotyping, as it uses images and sensors aided by software and computational intelligence algorithms, optimizing the measurement process, enabling greater phenotyping on scale, and reducing human variability. The use of vegetation indices is among the main methods used in high-throughput phenotyping. With their use, a variety of studies can be conducted, including the evaluation of nitrogen content in leaves, determination of physical characteristics such as biomass, plant height, and leaf area, analysis of plant heterogeneity in the field, estimation of chlorophyll content, evaluation of plant water content, quantification of lignin content, and detection of damage caused by pests and diseases. With the use of large-scale phenotyping techniques, the volume and complexity of the obtained data are increasing, increasing the demand for innovative approaches to data analysis focused on efficient selection. In this context, the application of computational intelligence has emerged as an essential tool for transforming the way plant phenotyping and improvement are approached. Thus, this work aims to evaluate the performance of Multilayer Perceptron (MLP) and Random Forest (RF) algorithms in predicting soybean phenotypic characteristics from the following vegetation indices: Normalized Difference Vegetation Index (NDVI), Soil-Adjusted Vegetation Index (SAVI), Modified Soil-Adjusted Vegetation Index (MSAVI), Green Normalized Difference Vegetation Index (GNVDI), Enhanced Vegetation Index (EVI), Normalized Difference Red Edge Index (NDRE), and Simplified Canopy Chlorophyll Content Index (SCCCI). Additionally, it seeks to investigate the impact of neural network parameters on the performance of an MLP and how these parameters influence its effectiveness in predicting soybean phenotypic characteristics. It was found that, for MLP models, as the number of neurons increased up to 10, and the number of folds in cross-validation increased up to 15, the models showed progressively better results. On the other hand, an increase in the number of epochs resulted in an increase in the R² values of the models, reaching a limit of 30000, after which the results began to decrease. Furthermore, an increase in the number of hidden layers led to a reduction in the values of the determination coefficients, indicating that the optimal number of layers was one. Moreover, there was a highlight in the prediction of the variables AIV and AP, especially regarding the evaluated Random Forest models. Keywords: Prediction; Machine Learning; Vegetation Indices; Glycine max (L).
Description
Citation
VIEIRA NETTO, João Amaro Ferreira. Uso de inteligência computacional na fenotipagem de soja. 2024. 82 f. Dissertação (Mestrado em Genética e Melhoramento) - Universidade Federal de Viçosa, Viçosa. 2024.
