Predição de mofo-branco do feijoeiro-comum utilizando técnicas de aprendizagem de máquinas
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidade Federal de Viçosa
Abstract
O feijão é uma importante fonte alimentar em várias regiões do mundo, fornecendo proteínas, carboidratos e minerais essenciais. No entanto, os produtores enfrentam desafios na condução das lavouras, sendo a doença denominada mofo-branco (causado pelo fungo Sclerotinia sclerotiorum), um dos problemas mais sérios. Esta doença pode resultar em perdas de até 90% na produtividade. A detecção precoce do mofo-branco é dificultada pelas características tanto do fungo quanto da cultura. Técnicas emergentes, como o uso de aprendizado de máquina, têm se mostrado promissoras para a identificação dessa doença. Essas técnicas têm sido eficazes em várias aplicações agrícolas, incluindo a previsão de condições ambientais, contagem automática de frutos e avaliação da severidade de doenças em culturas. Os objetivos deste estudo foram: (i) detectar a incidência e severidade do mofo-branco, além de prever a produtividade do feijoeiro-comum em condições de infecção, utilizando imagens multiespectrais obtidas por câmeras acopladas a Aeronaves Remotamente Pilotadas (RPAs) e técnicas de aprendizado de máquina; e (ii) desenvolver um modelo preditivo para estimar o risco de ocorrência do mofo-branco, com base em dados de campo sobre a doença e a produtividade da cultura. Foram aplicados quatro algoritmos de aprendizado de máquina: Cubist, Kernel K Nearest Neighbors (KKNN), Random Forest (RF) e Support Vector Machine (SVM). Os modelos foram treinados com dados espectrais obtidos de câmeras multiespectrais acopladas a RPAs em experimentos de campo, utilizando 51 índices de vegetação calculados a partir das imagens adquiridas, além de considerar o espaçamento da cultura em cada experimento. Esses dados foram utilizados para prever o risco de ocorrência do mofo-branco e a produtividade do feijoeiro-comum. Os algoritmos Cubist e KKNN se destacaram nas previsões, com o Cubist mostrando melhor desempenho na detecção da incidência e severidade do mofo-branco, utilizando principalmente as refletâncias nas bandas verde e azul aos 38, 52 e 59 dias após a emergência (DAE), além dos índices de vegetação MSR (Razão Simples Modificada) e EVI (Índice de Vegetação Melhorado). Na predição de produtividade, o KKNN, juntamente com o RF, proporcionou bons resultados, sendo o modelo de predição de produtividade mais preciso do que os modelos de predição do mofo- branco. Além disso, foram aplicados modelos de aprendizado de máquina para prever a produtividade e o risco de ocorrência do mofo-branco, com base em variáveis como densidade de plantio, tipo de cultivar, fungicida utilizado, precipitação e temperatura (mínima, média e máxima). Os algoritmos SVM, RF e Linear Model (LM) apresentaram desempenhos promissores. Palavras-chave: agricultura digital; sensoriamento remoto; Phaseolus vulgaris; Sclerotinia sclerotiorum; dados multiespectrais
Common bean is a major food source in several regions of the world, providing proteins, carbohydrates, and essential minerals. However, growers face significant challenges in crop management, with white mold (caused by the fungus Sclerotinia sclerotiorum) being one of the most serious problems. This disease can lead to yield losses of up to 90%. Early detection of white mold is hindered by the biological characteristics of both the pathogen and the crop. Emerging techniques, such as the use of machine learning, have shown promise for disease identification. These approaches have proven effective in various agricultural applications, including forecasting environmental conditions, automatic fruit counting, and assessing disease severity in crops. The objectives of this study were: (i) to detect the incidence and severity of white mold and to predict common bean yield under infection conditions using multispectral images acquired by cameras mounted on remotely piloted aircraft (RPAs) combined with machine learning techniques; and (ii) to develop a predictive model to estimate the risk of white mold occurrence based on field data on disease incidence and crop yield. Four machine learning algorithms were applied: Cubist, Kernel K Nearest Neighbors (KKNN), Random Forest (RF), and Support Vector Machine (SVM). The models were trained with spectral data obtained from multispectral cameras mounted on RPAs in field experiments, using 51 vegetation indices calculated from the acquired images, along with crop spacing in each experiment. These data were used to predict both the risk of white mold occurrence and common bean yield. Among the algorithms, Cubist and KKNN stood out in prediction accuracy. Cubist performed best in detecting white mold incidence and severity, primarily using reflectance in the green and blue bands at 38, 52, and 59 days after emergence (DAE), as well as the Modified Simple Ratio (MSR) and Enhanced Vegetation Index (EVI). For yield prediction, KKNN and RF produced robust results, with the yield prediction model proving more accurate than the white mold prediction models. Additionally, machine learning models were applied to predict both yield and white mold risk based on variables such as planting density, cultivar type, fungicide use, precipitation, and temperature (minimum, mean, and maximum). The SVM, RF, and Linear Model (LM) algorithms also showed promising performance. Keywords: digital agriculture; remote sensing; Phaseolus vulgaris; Sclerotinia sclerotiorum; multispectral data.
Common bean is a major food source in several regions of the world, providing proteins, carbohydrates, and essential minerals. However, growers face significant challenges in crop management, with white mold (caused by the fungus Sclerotinia sclerotiorum) being one of the most serious problems. This disease can lead to yield losses of up to 90%. Early detection of white mold is hindered by the biological characteristics of both the pathogen and the crop. Emerging techniques, such as the use of machine learning, have shown promise for disease identification. These approaches have proven effective in various agricultural applications, including forecasting environmental conditions, automatic fruit counting, and assessing disease severity in crops. The objectives of this study were: (i) to detect the incidence and severity of white mold and to predict common bean yield under infection conditions using multispectral images acquired by cameras mounted on remotely piloted aircraft (RPAs) combined with machine learning techniques; and (ii) to develop a predictive model to estimate the risk of white mold occurrence based on field data on disease incidence and crop yield. Four machine learning algorithms were applied: Cubist, Kernel K Nearest Neighbors (KKNN), Random Forest (RF), and Support Vector Machine (SVM). The models were trained with spectral data obtained from multispectral cameras mounted on RPAs in field experiments, using 51 vegetation indices calculated from the acquired images, along with crop spacing in each experiment. These data were used to predict both the risk of white mold occurrence and common bean yield. Among the algorithms, Cubist and KKNN stood out in prediction accuracy. Cubist performed best in detecting white mold incidence and severity, primarily using reflectance in the green and blue bands at 38, 52, and 59 days after emergence (DAE), as well as the Modified Simple Ratio (MSR) and Enhanced Vegetation Index (EVI). For yield prediction, KKNN and RF produced robust results, with the yield prediction model proving more accurate than the white mold prediction models. Additionally, machine learning models were applied to predict both yield and white mold risk based on variables such as planting density, cultivar type, fungicide use, precipitation, and temperature (minimum, mean, and maximum). The SVM, RF, and Linear Model (LM) algorithms also showed promising performance. Keywords: digital agriculture; remote sensing; Phaseolus vulgaris; Sclerotinia sclerotiorum; multispectral data.
Description
Citation
SANTOS, Flávio Souza. Predição de mofo-branco do feijoeiro-comum utilizando técnicas de aprendizagem de máquinas. 2025. 113 f. Tese (Doutorado em Engenharia Agrícola) - Universidade Federal de Viçosa, Viçosa. 2025.
