Fenotipagem de alto rendimento e aprendizado de máquina no melhoramento da soja
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidade Federal de Viçosa
Abstract
A soja (Glycine max (L.) Merrill) é uma cultura agrícola global de grande importância e tem impulsionado o desenvolvimento de tecnologias de fenotipagem de alto rendimento para otimizar a seleção de genótipos promissores. A integração da fenotipagem de alto rendimento com técnicas avançadas de aprendizado de máquina representa um avanço significativo na eficiência dos programas de melhoramento genético. Essa inovação utiliza índices de vegetação (IVs) para predizer a produtividade de grãos e o ciclo da cultura, uma característica relacionada à maturidade, geneticamente controlada e influenciada pelo fotoperíodo. Essa característica é essencial para classificar genótipos de soja. Assim, novas tecnologias são necessárias para otimizar o trabalho dos melhoristas na etapa de seleção, como o aumento da acurácia, da precisão e do rendimento operacional nas avaliações. Este estudo teve como objetivo avaliar o uso de um sistema aéreo não tripulado (VANT), equipado com um sensor multiespectral e RGB, para analisar características agronômicas de maturidade e produtividade de grãos em genótipos de soja, utilizando modelos de aprendizado de máquina e IVs derivados de bandas espectrais. Os IVs NDVI (Normalized Difference Vegetation Index), GNDVI (Green Normalized Difference Vegetation Index), NDRE (Normalized Difference Red Edge Index), e NGRDI (Normalized Green Red Difference Index) foram gerados a partir de mapas de refletância no software Pix4D®. As características avaliadas incluíram o número de dias da semeadura até o florescimento e à maturidade (estágio R8), a produtividade de grãos (kg ha-1) e os IVs correlacionados. Foram utilizadas linhagens da EMBRAPA para este estudo. A análise, realizada com modelos mistos e um modelo espacial de linhas e colunas no pacote SpATS, revelou correlações significativas entre os IVs e as características avaliadas. Além disso, os resultados aprimoraram as estratégias de seleção indireta, como a escolha do melhor dia de voo considerando o estádio fenológico da cultura. Os resultados sugerem a viabilidade dos IVs, destacando a eficácia do índice RGB NGRDI, que se mostrou comparável aos sensores multiespectrais NDVI e GNDVI. O índice RGB demonstrou alta precisão na predição da maturidade (R² > 0,9) e viabilidade econômica para a predição de genótipos de soja. A aplicação de modelos de aprendizado de máquina, como XGBoost e Random Forest, obteve bons resultados nos problemas de classificação dos grupos de maturidade e predição do ciclo da cultura, com predições mais precisas para a maturidade aos 90, 95 e 100 DAS, correspondendo aos estádios fenológicos R5, R6 e R7. A inclusão do aprendizado profundo com o modelo de redes neurais convolucionais (CNN) explorou a classificação das classes de maturidade, evidenciando o potencial do aprendizado profundo para futuras otimizações. Portanto, este estudo contribui para o avanço do melhoramento genético, oferecendo novas perspectivas para a seleção otimizada de genótipos de soja e ressaltando a importância de abordagens multidisciplinares para maximizar o ganho genético. Palavras-chave: índices de vegetação; imagens de drone; maturidade de plantas; fenômica; Glycine max ; aprendizado de máquina.
Soybean (Glycine max (L.) Merrill) is a major global crop and has driven the development of high-throughput phenotyping technologies to optimize the selection of promising genotypes. The integration of high-throughput phenotyping with advanced machine learning techniques represents a significant advance in the efficiency of breeding programs. This innovation uses vegetation indices (VIs) to predict grain yield and the crop cycle, a maturity-related trait that is genetically controlled and influenced by photoperiod. This trait is essential for classifying soybean genotypes. Thus, new technologies are needed to optimize breeders' work in the selection stage, including improvements in accuracy, precision, and operational performance in evaluations. This study aimed to evaluate the use of an unmanned aerial system (UAS), equipped with a multispectral and RGB sensor, to analyze agronomic traits such as maturity and grain yield in soybean genotypes, using machine learning models and vegetation indices (VIs) derived from spectral bands. The vegetation indices: NDVI (Normalized Difference Vegetation Index), GNDVI (Green Normalized Difference Vegetation Index), NDRE (Normalized Difference Red Edge Index), and NGRDI (Normalized Green Red Difference Index) were generated from reflectance maps using Pix4D® software. The evaluated traits included the number of days from sowing to flowering and maturity (R8 stage), grain yield (kg ha- 1), and their correlated VIs. EMBRAPA breeding lines were used in this study. The analysis, conducted using mixed models and a row-column spatial model in the SpATS package, revealed significant correlations between the VIs and the evaluated traits. Additionally, the results improved indirect selection strategies, such as identifying the optimal flight day based on the crop’s phenological stage. The findings support the viability of the VIs, highlighting the effectiveness of the RGB-derived NGRDI index, which performed comparably to multispectral indices like NDVI and GNDVI. The RGB index showed high accuracy for predicting maturity (R² > 0.9) and proved economically viable for predicting soybean genotypes. Machine learning models such as XGBoost and Random Forest yielded strong results in classifying maturity classes and predicting the crop cycle, with more accurate predictions for maturity at 90, 95, and 100 DAS, corresponding to phenological stages R5, R6, and R7. Incorporating deep learning via a convolutional neural network (CNN) model further explored the classification of maturity classes, demonstrating the potential of deep learning for future optimization efforts. Therefore, this study contributes to the advancement of plant breeding, offering new perspectives for the optimized selection of soybean genotypes and reinforcing the importance of multidisciplinary approaches to maximize genetic gain. Keywords: vegetation indices; drone imagery; plant maturity; phenomics; Glycine max; machine learning.
Soybean (Glycine max (L.) Merrill) is a major global crop and has driven the development of high-throughput phenotyping technologies to optimize the selection of promising genotypes. The integration of high-throughput phenotyping with advanced machine learning techniques represents a significant advance in the efficiency of breeding programs. This innovation uses vegetation indices (VIs) to predict grain yield and the crop cycle, a maturity-related trait that is genetically controlled and influenced by photoperiod. This trait is essential for classifying soybean genotypes. Thus, new technologies are needed to optimize breeders' work in the selection stage, including improvements in accuracy, precision, and operational performance in evaluations. This study aimed to evaluate the use of an unmanned aerial system (UAS), equipped with a multispectral and RGB sensor, to analyze agronomic traits such as maturity and grain yield in soybean genotypes, using machine learning models and vegetation indices (VIs) derived from spectral bands. The vegetation indices: NDVI (Normalized Difference Vegetation Index), GNDVI (Green Normalized Difference Vegetation Index), NDRE (Normalized Difference Red Edge Index), and NGRDI (Normalized Green Red Difference Index) were generated from reflectance maps using Pix4D® software. The evaluated traits included the number of days from sowing to flowering and maturity (R8 stage), grain yield (kg ha- 1), and their correlated VIs. EMBRAPA breeding lines were used in this study. The analysis, conducted using mixed models and a row-column spatial model in the SpATS package, revealed significant correlations between the VIs and the evaluated traits. Additionally, the results improved indirect selection strategies, such as identifying the optimal flight day based on the crop’s phenological stage. The findings support the viability of the VIs, highlighting the effectiveness of the RGB-derived NGRDI index, which performed comparably to multispectral indices like NDVI and GNDVI. The RGB index showed high accuracy for predicting maturity (R² > 0.9) and proved economically viable for predicting soybean genotypes. Machine learning models such as XGBoost and Random Forest yielded strong results in classifying maturity classes and predicting the crop cycle, with more accurate predictions for maturity at 90, 95, and 100 DAS, corresponding to phenological stages R5, R6, and R7. Incorporating deep learning via a convolutional neural network (CNN) model further explored the classification of maturity classes, demonstrating the potential of deep learning for future optimization efforts. Therefore, this study contributes to the advancement of plant breeding, offering new perspectives for the optimized selection of soybean genotypes and reinforcing the importance of multidisciplinary approaches to maximize genetic gain. Keywords: vegetation indices; drone imagery; plant maturity; phenomics; Glycine max; machine learning.
Description
Citation
FÉLIX, Marco Renan. Fenotipagem de alto rendimento e aprendizado de máquina no melhoramento da soja. 2024. 65 f. Tese (Doutorado em Genética e Melhoramento) - Universidade Federal de Viçosa, Viçosa. 2024.
