Pedro Melo-Pinto

University of Trás-os-Montes and Alto Douro, Engineering Department

Centre for the Research and Technology of Agro-Environmental and Biological Sciences

Quinta de Prados. 5000-801 Vila Real, Portugal

pmelo@utad.pt

Machine Learning based methodologies for grape and wine production

Abstract

Wine production is a major area in Portuguese economy. The sector accounts for approximately 2.6% of Portugal’s GDP, and is crucial to the development of inland regions. Wine quality depends on grape quality, which is a result of different factors such as harvesting at the optimal point of maturity, or selecting the grapes according to quality features. The growing importance of precision agriculture, demands novel methods to make the most out of the data generated.

Machine learning approaches are particularly suitable for complex processes’ modelling, where too many arguments are involved. Many of the agricultural processes such as yield estimation or ripening, fall into this category. However, several challenges may arise: even if models are available for particular cases, general models are difficult to come by, largely due to the high variability involved; with such models, there is a need for huge amounts of data, in a way which will not always be easy to obtain. This is particularly true in the case of grape production and the wine industry, where the number of grape varieties, climate, soil, etc., introduce a degree of variability that is uncommon in other sectors.

One way of obtaining more information in such cases has been the use of images, whether RGB, multi-spectral or hyperspectral. The use of images introduces additional challenges, such as data size and lighting conditions, which underscores the need for data-driven modelling methodologies. The use of machine learning (ML) techniques combined with images has been establishing itself as a viable alternative predicting yields or quality parameters in grapes, and assist on harvesting critical decisions. Several ML approaches have been proposed to handle such data characteristics, but selecting a suitable methodology that best address the problem under study and make sure it generalizes well, is, most of the times, a cumbersome task.

We will address three wine and grape production problems, yield estimation, grape quality assessment and illness detection, presenting different difficulties.

Yield estimation is a major aspect in the wine production industry. The number of berries is one of the most relevant yield aspects, and can be estimated using computer vision, namely image segmentation algorithms. Lighting is one of the key factors in computer vision approaches, and occlusions—which are common in this problem—can be minimized, but require estimation models to achieve reliable results.

Illness early detection is, of course, of the utmost importance, with an impact on production and quality. Detecting early symptoms and distinguishing them from other conditions is a difficult task, which makes the use of machine learning algorithms advisable, using automatic visual inspection of the leaves, addressed as a pixel-based classification problem. However, the available datasets are limited and complex to compile, as different diseases typically require very specific timeframes and locations. The use of generative models to overcome this obstacle, in conjunction with robust classification models, lead to models that are intended to be as general as possible.

The evaluation of the grape’s maturation profile based on the evolution of their oenological parameters over time is one of the most important aspects for ensuring wine quality, playing a decisive role in the definition of the optimal time for harvesting. The use hyperspectral imaging combined with learning algorithms represents a faster and non-destructive grape ripeness assessment method. We focus in two fundamental aspects to address the resulting data variability: a possible wavelength bands selection (with the purpose of reducing the dimensionality of data without losing predictive power) and the generalization ability of the ML model under such demanding conditions.

Short Bio

Pedro Melo-Pinto graduated from Univ. do Porto in 1984 and received a Ph.D. in Computer Science from the Univ. de Trás-os-Montes e Alto Douro (UTAD) in 1998. He is a Full Professor at UTAD where he served as Pro-Rector, as Director and Vice-Director of R\&D units, and as member of UTAD’s General Council. He also served as a member of The Portuguese Strategic Infrastructures (for the Digital Areas) Monitoring Committee and usually serves as a member (or co-chair) of the evaluation panel for different Portuguese and international R\&D projects and grants competitions. He is currently Associate Editor of Computers and Electronics in Agriculture journal. He was Chair of the EuroFuse 2011 conference and the NATO 2001 Advanced Research Workshop, and served as member of the organization committee for more than 40 international conferences/workshops/special sessions. He is the head of the Image-Based Systems Lab from the Centre for the Research and Technology of Agro-Environmental and Biological Sciences.

His scientific interests include Computer Vision (CV) and Machine Learning (ML) and, in the last 15 years, their applications to the Agro-forestry area, namely to grape oenological parameters assessment. He authored/co-authored more than 100 scientific papers. He participated in many Portuguese and international R\&D projects, predominantly in the areas mentioned above.