Publicação
Deep Learning Methods with Limited Data Sources
| Resumo: | As the viticulture market becomes global and competitiveness levels rise, producers face a serious challenge in continuing to improve the winemaking process to maintain prominence. New methodologies are introduced to raise this process to new standards, with a special focus on methods for grape ripeness assessment based on the evaluation of oenological parameters, which help determine the optimal point of maturity for grapes’ harvesting and in the selection of grapes according to the desired characteristics of the final product. In this context, the precision viticulture research area has grown into a strong alternative to the traditional laboratorial analysis, due to its’ ability to perform a faster, non-destructive, and sustainable assessment of grape maturity with resource to image-based systems: among these systems, hyperspectral imaging is a cost-effective technique that allows the collection of both spatial and spectral information about an object, with recent studies showing that, paired with data analysis tools, this technology provides the means to achieve accurate predictions of oenological parameters with a small error rate. The downside to this technology is that powerful data analysis tools are required to extract the necessary information from the spectra, since it has complex patterns and a large volume of variability: consequently, the overall objective of this dissertation was to develop models, with resource to machine and deep learning, that provide accurate insights about the oenological parameters from hyperspectral images of wine grape berries. Sugar content and pH index were the parameters considered for grape ripeness assessment, due to their strong connection with flavour and overall maturity stage. The biggest concern when developing deep learning models to operate in this context is the generalization ability, since these models should exhibit a robust prediction capacity across different varieties and vintages of wine grape berries, while also maintaining a strong performance on small-sized data sets, due to the inherent difficulties of acquiring higher volumes of data: this represents a highly complex problem due to the high variability in the samples (oenological parameters are affected by several factors such as climate, different soil quality, terrain inclination, sun exposition, the plants’ water stress, and ripeness stage), the number of wine grape varieties (which introduce even more sample variability due to the extremely high number of different varieties), and the nonexistence of public large data sets (while traditional deep learning solutions use millions of examples for training, in this problem only hundreds of samples are available to use). Hence, this dissertation intends to contribute to the progress in knowledge regarding the application of deep learning methods in association with imaging spectroscopy to problems with limited data sources, and to achieve this goal it was divided into four main segments of development: the first, where an extensive study and implementation of dimensionality reduction methods was conducted to verify their impact on reducing storage space, computation time, and removing redundant features from a data set, which helps alleviate the known difficulties in processing the complex patterns in hyperspectral data; the second, where a technique capable of retaining both the local and global structure of the data in a lower dimension was studied, constituting itself as a strong alternative to the traditionally used linear methods; the third, where some of the most recent deep learning networks were implemented and combined with dimensionality reduction techniques to boost their performance; and the fourth, where recent deep learning architectures are investigated while attempting to develop a set of best practices and a unified methodology to perform a proper evaluation of generalization ability for the particular case of predicting oenological parameters from wine grape berries. Finally, this dissertation represents a step forward in terms of grape ripeness assessment with cost-effective and non-destructive technologies, with the conducted study allowing to understand the major impact of sample variability in several deep learning models’ performances: with a thorough research work already developed for dimensionality reduction and advanced data analysis tools, the attention should now shift to feature engineering, with the creation of new sample descriptors to use alongside the hyperspectral images that can help the models overcome the lack of examples for learning, which will lead to even more improvements in generalization capacity. |
|---|---|
| Autores principais: | Silva, Rui Manuel Machado |
| Assunto: | Deep Learning Hyperspectral Images |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | tese de doutoramento |
| Tipo de acesso: | acesso restrito |
| Instituição associada: | Universidade de Trás-os-Montes e Alto Douro |
| Idioma: | inglês |
| Origem: | Repositório da UTAD |
| Resumo: | As the viticulture market becomes global and competitiveness levels rise, producers face a serious challenge in continuing to improve the winemaking process to maintain prominence. New methodologies are introduced to raise this process to new standards, with a special focus on methods for grape ripeness assessment based on the evaluation of oenological parameters, which help determine the optimal point of maturity for grapes’ harvesting and in the selection of grapes according to the desired characteristics of the final product. In this context, the precision viticulture research area has grown into a strong alternative to the traditional laboratorial analysis, due to its’ ability to perform a faster, non-destructive, and sustainable assessment of grape maturity with resource to image-based systems: among these systems, hyperspectral imaging is a cost-effective technique that allows the collection of both spatial and spectral information about an object, with recent studies showing that, paired with data analysis tools, this technology provides the means to achieve accurate predictions of oenological parameters with a small error rate. The downside to this technology is that powerful data analysis tools are required to extract the necessary information from the spectra, since it has complex patterns and a large volume of variability: consequently, the overall objective of this dissertation was to develop models, with resource to machine and deep learning, that provide accurate insights about the oenological parameters from hyperspectral images of wine grape berries. Sugar content and pH index were the parameters considered for grape ripeness assessment, due to their strong connection with flavour and overall maturity stage. The biggest concern when developing deep learning models to operate in this context is the generalization ability, since these models should exhibit a robust prediction capacity across different varieties and vintages of wine grape berries, while also maintaining a strong performance on small-sized data sets, due to the inherent difficulties of acquiring higher volumes of data: this represents a highly complex problem due to the high variability in the samples (oenological parameters are affected by several factors such as climate, different soil quality, terrain inclination, sun exposition, the plants’ water stress, and ripeness stage), the number of wine grape varieties (which introduce even more sample variability due to the extremely high number of different varieties), and the nonexistence of public large data sets (while traditional deep learning solutions use millions of examples for training, in this problem only hundreds of samples are available to use). Hence, this dissertation intends to contribute to the progress in knowledge regarding the application of deep learning methods in association with imaging spectroscopy to problems with limited data sources, and to achieve this goal it was divided into four main segments of development: the first, where an extensive study and implementation of dimensionality reduction methods was conducted to verify their impact on reducing storage space, computation time, and removing redundant features from a data set, which helps alleviate the known difficulties in processing the complex patterns in hyperspectral data; the second, where a technique capable of retaining both the local and global structure of the data in a lower dimension was studied, constituting itself as a strong alternative to the traditionally used linear methods; the third, where some of the most recent deep learning networks were implemented and combined with dimensionality reduction techniques to boost their performance; and the fourth, where recent deep learning architectures are investigated while attempting to develop a set of best practices and a unified methodology to perform a proper evaluation of generalization ability for the particular case of predicting oenological parameters from wine grape berries. Finally, this dissertation represents a step forward in terms of grape ripeness assessment with cost-effective and non-destructive technologies, with the conducted study allowing to understand the major impact of sample variability in several deep learning models’ performances: with a thorough research work already developed for dimensionality reduction and advanced data analysis tools, the attention should now shift to feature engineering, with the creation of new sample descriptors to use alongside the hyperspectral images that can help the models overcome the lack of examples for learning, which will lead to even more improvements in generalization capacity. |
|---|