Análise de imagens, espectroscopia e aprendizado de máquina na predição do potencial fisiológico e da deterioração de sementes de algodão
Loading...
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidade Federal de Viçosa
Abstract
Sementes de algodão estão sujeitas a danos ao longo do processo produtivo e, por serem oleaginosas, apresentam maior suscetibilidade à deterioração. Devido ao seu alto valor comercial, o controle de qualidade eficiente dos lotes é essencial. Nesse contexto, métodos rápidos, automatizados e, não destrutivos são importantes para complementar os testes tradicionais. O objetivo foi avaliar a eficiência das análises de imagens radiográficas e imagens de plântulas aliadas ao aprendizado interativo de máquina, e das técnicas espectrais para a classificação do potencial fisiológico de sementes de algodão. Buscou-se, ainda, relacionar análises espectrais e alterações bioquímicas e fisiológicas em sementes submetidas à deterioração. Foram utilizadas sementes da cultivar 21065TLP, que foram avaliadas quanto ao potencial fisiológico por meio dos testes de germinação e vigor. Ensaio I - Foi realizado o teste de raios X, sendo as sementes radiografadas e submetidas ao teste de germinação. As imagens foram processadas pelo software ImageJ, obtendo-se as variáveis área, perímetro, preenchimento, densidade relativa e integrada e escala de cinza. A análise de imagens de plântulas foi realizada aos dois, três e quatro dias de germinação, processadas por meio do equipamento GroundEye, obtendo-se as variáveis: comprimento do hipocótilo, da raiz e total das plântulas, índices de crescimento, uniformidade e vigor. Por meio do software Ilastik, foram gerados dados de porcentagem de plântulas normais fortes, normais fracas, anormais e sementes mortas. Ensaio II - Foram realizadas leituras de 200 sementes por lote em espectofotômetro FT-NIR, na faixa de comprimento de onda de 1.000 a 2.500 nm. Imagens multiespectrais também foram obtidas das mesmas sementes nos comprimentos de 395, 460, 520, 585, 620, 740, 850 e 940 nm, utilizando-se um protótipo equipado com câmera digital e LEDs. Em seguida, estas mesmas sementes foram submetidas ao teste de germinação. Os lotes foram classificados em três categorias de acordo com a germinação: alto ( 91%), médio (85–90%) e baixo potencial fisiológico ( 84%). Os dados espectrais NIR originais foram pré- processados utilizando-se os métodos: Standard Normal Variate (SNV); Multiplicative Scatter Correction (MSC); derivadas de Savitzky-Golay (SG) de 1ª e 2ª ordem. Para os dados multiespectrais, foram utilizados SNV e MSC. Os algoritmos de classificação testados foram: Análise Discriminantepor Mínimos Quadrados Parciais (PLS-DA), Random Forest (RF) e Redes Neurais (NN). Ensaio III - Sementes de um lote foram envelhecidas artificialmente por 0 (controle), 24, 48, 72, 96 e 120 h. Foram então realizados testes de germinação, vigor e análises bioquímicas (catalase - CAT, superóxido dismutase – SOD e peroxidase – POX; peróxido de hidrogênio- H2O2 e malonaldeído – MDA). Foram também realizadas análises espectrais FT-NIR e multiespectrais, utilizando-se algoritmos de pré-processamento e classificação conforme descrito acima. Foram estabelecidas as classes de germinação: alto ( 90%), médio (80-89%) e baixo potencial fisiológico ( 79%). Concluiu-se que, para a técnica de raios X, houve relação entre as variáveis fisiológicas e as variáveis morfobiométricas de densidade relativa, densidade integrada e média de cinza, permitindo identificar sementes bem formadas e a integridade dos tecidos. Aos três dias, variáveis de crescimento de plântulas foram eficientes na classificação dos lotes. O uso do Ilastik apresentou resultados comparáveis aos testes convencionais. Verificou-se, ainda, que a espectroscopia NIR mostrou alta eficiência (até 98% de acurácia), especialmente com PLS-DA. Para imagens multiespectrais, o Random Forest teve melhor desempenho (até 81%). No estudo de deterioração, houve redução da germinação, vigor e atividade enzimática (SOD, CAT, POX), além de aumento de HO e MDA. As técnicas espectrais permitiram classificar sementes em diferentes níveis de deterioração. Palavras-chave: ft-nir; imagem multiespectral; raios x ; inteligência artificial; qualidade fisiológica.
Cotton seeds are subject to damage throughout the production process and, being oilseeds, exhibit greater susceptibility to deterioration. Due to their high commercial value, efficient quality control of seed lots is essential. In this context, rapid, automated, and non-destructive methods are important to complement traditional tests. The objective was to evaluate the efficiency of radiographic image analysis and seedling images combined with interactive machine learning, and spectral techniques for classifying the physiological potential of cotton seeds. Furthermore, the study sought to relate spectral analyses to biochemical and physiological changes in seeds subjected to restrictions. Seeds of the cultivar 21065TLP were used, and their physiological potential was evaluated through germination and vigor tests. Experiment I – The X-ray test was performed, with the seeds radiographed and subjected to the germination test. The images were processed using ImageJ software, obtaining the variables of area, perimeter, filling, relative and integrated density, and grayscale. Seedling image analysis was performed on days two, three, and four after germination, processed using GroundEye equipment, obtaining the following variables: hypocotyl length, root length, and total seedling length, growth indices, uniformity, and vigor. Using the Ilastik software, data on the percentage of strong normal seedlings, weak normal seedlings, abnormal seedlings, and dead seeds were generated. Experiment II – Readings were taken from 200 seeds per batch using an FT-NIR spectrophotometer, in the wavelength range of 1,000 to 2,500 nm. Multispectral images were also obtained from the same seeds at wavelengths of 395 to 940 nm, using a prototype equipped with a digital camera and LEDs. Subsequently, these same seeds were subjected to the germination test. The seed lots were classified into three categories according to germination: high ( 91%), medium (85–90%), and low physiological potential ( 84%). The original NIR spectral data were pre-processed using the following methods: Standard Normal Variate (SNV); Multiplicative Scatter Correction (MSC); and 1st and 2nd order Savitzky-Golay (SG) derivatives. For the multispectral data, SNV and MSC were used. The classification algorithms tested were: Partial Least Squares Discriminant Analysis (PLS-DA), Random Forest (RF), and Neural Networks (NN). Experiment III – Seeds from one lot were subjected to acceleratedaging for 0 (control), 24, 48, 72, 96, and 120 h. Germination, vigor, and biochemical analyses (catalase – CAT, superoxide dismutase – SOD, and peroxidase – POX; hydrogen peroxide – HO and malondialdehyde – MDA) were then performed. FT-NIR and multispectral spectral analyses were also carried out, using the pre-processing and classification algorithms as described above. Germination classes were conditional: high ( 90%), medium (80-89%), and low physiological potential ( 79%). It was concluded that, for the X-ray technique, there was a relationship between physiological variables and morphobiometric variables of relative density, integrated density, and grayscale average, allowing the identification of well-formed seeds and tissue integrity. At three days, seedling growth variables were efficient in classifying the lots. The use of Ilastik presented results comparable to conventional tests. It was also found that NIR spectroscopy showed high efficiency (up to 98% accuracy), especially with PLS-DA. For multispectral images, Random Forest performed best (up to 81%). In the interference study, there was a reduction in germination, vigor, and enzymatic activity (SOD, CAT, POX), as well as an increase in HO and MDA. The spectral techniques allowed the seeds to be classified into different amplitude levels. Keywords: ft-nir; multispectral imaging; x-ray; artificial intelligence; physiological quality.
Cotton seeds are subject to damage throughout the production process and, being oilseeds, exhibit greater susceptibility to deterioration. Due to their high commercial value, efficient quality control of seed lots is essential. In this context, rapid, automated, and non-destructive methods are important to complement traditional tests. The objective was to evaluate the efficiency of radiographic image analysis and seedling images combined with interactive machine learning, and spectral techniques for classifying the physiological potential of cotton seeds. Furthermore, the study sought to relate spectral analyses to biochemical and physiological changes in seeds subjected to restrictions. Seeds of the cultivar 21065TLP were used, and their physiological potential was evaluated through germination and vigor tests. Experiment I – The X-ray test was performed, with the seeds radiographed and subjected to the germination test. The images were processed using ImageJ software, obtaining the variables of area, perimeter, filling, relative and integrated density, and grayscale. Seedling image analysis was performed on days two, three, and four after germination, processed using GroundEye equipment, obtaining the following variables: hypocotyl length, root length, and total seedling length, growth indices, uniformity, and vigor. Using the Ilastik software, data on the percentage of strong normal seedlings, weak normal seedlings, abnormal seedlings, and dead seeds were generated. Experiment II – Readings were taken from 200 seeds per batch using an FT-NIR spectrophotometer, in the wavelength range of 1,000 to 2,500 nm. Multispectral images were also obtained from the same seeds at wavelengths of 395 to 940 nm, using a prototype equipped with a digital camera and LEDs. Subsequently, these same seeds were subjected to the germination test. The seed lots were classified into three categories according to germination: high ( 91%), medium (85–90%), and low physiological potential ( 84%). The original NIR spectral data were pre-processed using the following methods: Standard Normal Variate (SNV); Multiplicative Scatter Correction (MSC); and 1st and 2nd order Savitzky-Golay (SG) derivatives. For the multispectral data, SNV and MSC were used. The classification algorithms tested were: Partial Least Squares Discriminant Analysis (PLS-DA), Random Forest (RF), and Neural Networks (NN). Experiment III – Seeds from one lot were subjected to acceleratedaging for 0 (control), 24, 48, 72, 96, and 120 h. Germination, vigor, and biochemical analyses (catalase – CAT, superoxide dismutase – SOD, and peroxidase – POX; hydrogen peroxide – HO and malondialdehyde – MDA) were then performed. FT-NIR and multispectral spectral analyses were also carried out, using the pre-processing and classification algorithms as described above. Germination classes were conditional: high ( 90%), medium (80-89%), and low physiological potential ( 79%). It was concluded that, for the X-ray technique, there was a relationship between physiological variables and morphobiometric variables of relative density, integrated density, and grayscale average, allowing the identification of well-formed seeds and tissue integrity. At three days, seedling growth variables were efficient in classifying the lots. The use of Ilastik presented results comparable to conventional tests. It was also found that NIR spectroscopy showed high efficiency (up to 98% accuracy), especially with PLS-DA. For multispectral images, Random Forest performed best (up to 81%). In the interference study, there was a reduction in germination, vigor, and enzymatic activity (SOD, CAT, POX), as well as an increase in HO and MDA. The spectral techniques allowed the seeds to be classified into different amplitude levels. Keywords: ft-nir; multispectral imaging; x-ray; artificial intelligence; physiological quality.
Description
Citation
LIMÃO, Marcelo Augusto Rocha. Análise de imagens, espectroscopia e aprendizado de máquina na predição do potencial fisiológico e da deterioração de sementes de algodão. 2026. 110 f. Tese (Doutorado em Fitotecnia) - Universidade Federal de Viçosa, Viçosa. 2026.
