Publicação
Predicting oenological attributes using machine learning models
| Resumo: | The potential of hyperspectral images combined with machine learning algorithms to predict anthocyanin concentration, pH index and sugar content in grapes is presented as a starting point do develop flexible models with large generalization capacity to estimate oenological parameters. In this context, in order to evaluate the generalization capacity of the machine learning procedures, a comparison with current state of the art approaches and between three different methods, Neural Networks (NNs), Decision Trees (DTs) and Support Vector Regression (SVR), when combined with hyperspectral images, was performed to predict the anthocyanin concentration, pH index and sugar content and support the adequate monitoring of wine quality. The models were trained with six whole grape berries for each sample, using different approaches of cross-validation and data pre-processing. The oenological parameters were estimated using models trained with the spectra of 2012, 2013 and 2014 samples from the Touriga Franca variety, and the generalization capacity was tested using 2013 samples of the Tinta Barroca and Touriga Nacional varieties. The results suggest that combining hyperspectral images with appropriate data analysis tools achieve accurate predictions. The machine learning methods were able to predict the values of oenological parameters without significant differences, improving the state of the art results. Good indicators were obtained in the generalization capacity of the models, suggesting that a robust model capable of predicting oenological parameters on different varieties and harvest years of wine grapes can be obtained without additional training. An environmentallyfriendly, fast and low-cost approach is therefore achievable and should be the subject of future testing. |
|---|---|
| Autores principais: | Silva, Rui Manuel Machado |
| Assunto: | Hyperspectral Imaging Neural Networks |
| Ano: | 2017 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Trás-os-Montes e Alto Douro |
| Idioma: | inglês |
| Origem: | Repositório da UTAD |
| Resumo: | The potential of hyperspectral images combined with machine learning algorithms to predict anthocyanin concentration, pH index and sugar content in grapes is presented as a starting point do develop flexible models with large generalization capacity to estimate oenological parameters. In this context, in order to evaluate the generalization capacity of the machine learning procedures, a comparison with current state of the art approaches and between three different methods, Neural Networks (NNs), Decision Trees (DTs) and Support Vector Regression (SVR), when combined with hyperspectral images, was performed to predict the anthocyanin concentration, pH index and sugar content and support the adequate monitoring of wine quality. The models were trained with six whole grape berries for each sample, using different approaches of cross-validation and data pre-processing. The oenological parameters were estimated using models trained with the spectra of 2012, 2013 and 2014 samples from the Touriga Franca variety, and the generalization capacity was tested using 2013 samples of the Tinta Barroca and Touriga Nacional varieties. The results suggest that combining hyperspectral images with appropriate data analysis tools achieve accurate predictions. The machine learning methods were able to predict the values of oenological parameters without significant differences, improving the state of the art results. Good indicators were obtained in the generalization capacity of the models, suggesting that a robust model capable of predicting oenological parameters on different varieties and harvest years of wine grapes can be obtained without additional training. An environmentallyfriendly, fast and low-cost approach is therefore achievable and should be the subject of future testing. |
|---|