Predições de características estruturais e valor nutritivo de Urochloa decumbens por meio de imagens aéreas
Loading...
Date
Journal Title
Journal ISSN
Volume Title
Publisher
Universidade Federal de Viçosa
Abstract
O uso da inteligência artificial, aliado a técnicas de agricultura e zootecnia de precisão, está transformando o monitoramento e a gestão da agropecuária. O sensoriamento remoto, em particular, oferece uma solução inovadora para monitorar a variabilidade espaço-temporal em pastagens, fornecendo informações cruciais sobre diversos parâmetros biofísicos. Nesse contexto, o objetivo com este trabalho foi desenvolver modelos preditivos para a massa de forragem verde e seca, altura do dossel, densidade e concentração de matéria seca (%MS) e proteína bruta (%PB) de pastos de Urochloa decumbens (Stapf) R.D. Webster, utilizando aprendizado de máquinas e imagens multiespectrais obtidas por aeronave remotamente pilotada. O experimento foi conduzido na Universidade Federal de Viçosa, entre 2019 e 2020, com diferentes doses de nitrogênio aplicadas após cada corte, com o objetivo de promover variações nas características do capim. Foram avaliados o desempenho dos seguintes modelos: regressão linear simples (RLS), regressão linear múltipla (RLM), e randon forest regression (RFR). Para isso, dados multiespectrais foram extraídos de imagens aéreas e combinados à dados meteorológicos para alimentar os modelos preditivos. Os modelos foram avaliados por meio de validação cruzada “Leave-one-out”, utilizando métricas como coeficiente de determinação (R²), erro quadrático médio (RMSE) e erro absoluto médio (MAE). As melhores predições foram observadas para massa de forragem verde (R² = 0,77, RMSE = 3.281,29 kg ha-1, MAE = 2.361,80 kg ha-1) e seca (R² = 0,71, RMSE = 657,46 kg ha-1, MAE = 507,53 kg ha-1) com RLM, e altura do dossel com RLS (R² = 0,59, RMSE = 8,72 cm, MAE = 7,26 cm). As predições para % MS e % PB não apresentaram desempenho satisfatório para nenhum dos modelos avaliados. Conclui-se que a combinação de imagens aéreas multiespectrais com aprendizado de máquinas oferece uma ferramenta eficaz para prever variáveis como massa verde, massa seca e altura em pastos de capim-brachiaria, podendo contribuir significativamente para a gestão e o manejo das pastagens. Contudo, são necessárias mais pesquisas para otimizar os modelos desenvolvidos. Palavras-chave: sensoriamento remoto; multiespectrais; índices de vegetação aprendizado de máquina; imagens.
The use of artificial intelligence, combined with precision agriculture and livestock techniques, is transforming the monitoring and management of agricultural systems. Remote sensing offers an innovative solution for monitoring spatiotemporal variability in pastures, providing crucial information on various biophysical parameters. In this context, the objective of this study was to develop predictive models for fresh and dry forage mass, canopy height, density and dry matter (%DM) and crude protein concentrations (%CP) of signalgrass [Urochloa decumbens (Stapf) R.D. Webster] pastures, using machine learning and multispectral images obtained by an unmanned aerial vehicle (UAV). The experiment was conducted at the Federal University of Viçosa between 2019 and 2020, with different nitrogen doses applied after each cut to promote variations in grass characteristics. The performance of the following models was evaluated: simple linear regression (SLR), multiple linear regression (MLR), and random forest regression (RFR). Multispectral images were extracted and combined with meteorological data to feed the predictive models, which were evaluated using Leave-one-out cross-validation, and metrics such as the coefficient of determination (R²), root mean square error (RMSE) and mean absolute error (MAE). The best predictions were obtained for fresh forage mass (R² = 0.77, RMSE = 3,281.29 kg ha-1, MAE = 2,361.80 kg ha-1) and dry forage mass (R² = 0.71, RMSE = 657.46 kg ha-1, MAE = 507.53 kg ha-1) using MLR, and for canopy height using SLR (R² = 0.59, RMSE = 8.72 cm, MAE = 7.26 cm). Predictions for % DM and % CP were unsatisfactory. The findings demonstrate that combining multispectral imagery with machine learning effectively predicts forage mass and canopy height of signalgrass pastures, supporting improved grazing management, though further refinement of models is needed. Keywords: remote sensing; machine learning; multispectral imagery; vegetation indices.
The use of artificial intelligence, combined with precision agriculture and livestock techniques, is transforming the monitoring and management of agricultural systems. Remote sensing offers an innovative solution for monitoring spatiotemporal variability in pastures, providing crucial information on various biophysical parameters. In this context, the objective of this study was to develop predictive models for fresh and dry forage mass, canopy height, density and dry matter (%DM) and crude protein concentrations (%CP) of signalgrass [Urochloa decumbens (Stapf) R.D. Webster] pastures, using machine learning and multispectral images obtained by an unmanned aerial vehicle (UAV). The experiment was conducted at the Federal University of Viçosa between 2019 and 2020, with different nitrogen doses applied after each cut to promote variations in grass characteristics. The performance of the following models was evaluated: simple linear regression (SLR), multiple linear regression (MLR), and random forest regression (RFR). Multispectral images were extracted and combined with meteorological data to feed the predictive models, which were evaluated using Leave-one-out cross-validation, and metrics such as the coefficient of determination (R²), root mean square error (RMSE) and mean absolute error (MAE). The best predictions were obtained for fresh forage mass (R² = 0.77, RMSE = 3,281.29 kg ha-1, MAE = 2,361.80 kg ha-1) and dry forage mass (R² = 0.71, RMSE = 657.46 kg ha-1, MAE = 507.53 kg ha-1) using MLR, and for canopy height using SLR (R² = 0.59, RMSE = 8.72 cm, MAE = 7.26 cm). Predictions for % DM and % CP were unsatisfactory. The findings demonstrate that combining multispectral imagery with machine learning effectively predicts forage mass and canopy height of signalgrass pastures, supporting improved grazing management, though further refinement of models is needed. Keywords: remote sensing; machine learning; multispectral imagery; vegetation indices.
Description
Citation
SOUZA, Iuly Francisca Rodrigues de. Predições de características estruturais e valor nutritivo de Urochloa decumbens por meio de imagens aéreas. 2024. 53 f. Dissertação (Mestrado em Zootecnia) - Universidade Federal de Viçosa, Viçosa. 2024.
