Artificial Intelligence in Digital Agriculture. Towards In-Field Grapevine Monitoring using Non-invasive Sensors

Abstract

Agriculture seeks for a reduction of costs and environmental impact, better sustainability and to increase crop yield and quality. It is necessary to deliver useful applications for farmers and industries, to help for greater efficiency and sustainability. To achieve this in digital viticulture, useful information about the vineyard is necessary so better decisions can be taken. Advances in non-invasive sensing technologies allow the acquisition of high amounts of data from the vineyard, but these data alone are not enough to be used when decisions need to be made, it needs to be transformed into information. Artificial intelligence is a revolution at different social, work and industrial levels to deal with data. Within artificial intelligence, machine learning has evolved greatly during the last decades providing tools to make computers learn, and these algorithms are used in many different fields due to their high versatility for many data-related tasks, generating knowledge and information, and improving the decision-making process. Therefore, the combination of non-invasive sensors and artificial intelligence needs to be explored to meet the requirements needed to apply digital agriculture, the data-driven agriculture. The main objective of this PhD Thesis is the combination of machine learning and non-invasive sensing technologies for the assessment of relevant agronomical, physiological and qualitative traits in digital agriculture and viticulture. Specifically, the following objectives have been pursued: i) to make use of different machine learning algorithms on data from spectroscopy for in-field grapevine phenotyping and monitoring; ii) the application of ensemble data analysis techniques for vineyard water status assessment with thermal imaging; and iii) to deploy hyperspectral imaging in the field, supported by intensive machine learning combinations, for the monitoring of different crop traits. The first objective, covered in Chapter 3, was the combination of machine learning algorithms and near-infrared spectroscopy for vineyard monitoring and phenotyping. A handheld spectrometer was used for two goals: the classification of grapevine varieties, from several vineyard plots and vintages; and water status assessment, using the same spectral signal. Accurate models were developed for both goals. The results allow to open new ways in digital viticulture for the quick grapevine phenotyping under field conditions, an useful tool for several actors in the wine industry. The application of ensemble machine learning algorithms to in-field thermal images acquired on-the-go for vineyard water status monitoring, the second objective, is addressed in Chapter 4. A thermal camera was mounted on an all-terrain vehicle for continuous acquisition. A combination of rotation forests and decision trees was used for the training of prediction models. The outcomes provided by the machine learning algorithms support the use of thermal imaging for fast, reliable estimation of a vineyard water status, even suppressing the necessity of supervised acquisition of reference temperatures. The new developed on-the-go method can be very useful in the grape and wine industry for assessing and mapping vineyard water status. The last objective was the use of on-the-go hyperspectral imaging under field conditions, modelled with machine learning techniques, and it is discussed in Chapter 5. Hyperspectral imaging is a powerful technology, but it has been classically used under laboratory conditions. Very few attempts on in-field hyperspectral imaging have been reported in the literature, due to the difficulties, like natural, irregular illumination or unknown a priori sample positioning in the recorded scene, that it is necessary to face. For this reason, a considerable amount of the work developed in this PhD Thesis has been focused on surpassing the challenges that come from deploying a hyperspectral camera in the field for the on-the-go vineyard monitoring. Also, as hyperspectral imaging involves the management of a high amount of data, advanced machine learning algorithms become appealing to be applied in this scenario. Three different applications were developed: varietal classification, grape composition assessment and yield estimation. In all of them, it was designed a mechanism for the automated identification of the different grapevine organs. Potent models were obtained for the monitoring of different key viticulture and agriculture parameters. The results suggest that machine learning and hyperspectral imaging can be used to accurately estimate several traits in vineyards and other crops, becoming a powerful and accurate tool in the decision making process. The results from the research work carried out in this PhD Thesis, also published in several scientific articles, demonstrated that artificial intelligence techniques are able to exploit the potential of data acquired using non-invasive sensing technologies for the monitoring and phenotyping of key crop traits. This can be of utmost importance in digital agriculture and viticulture as new solutions can be developed as decision support tools.