Viticulture suitability for specific oenological objectives through machine learning integration in a multicriteria analysis: the case of Cannonau terroir in Sardinia (Italy)

The projected warmer temperatures, together with the expected increase in seasonal dryness, frequency, and intensity of extreme climate events during sensitive phenological phases, may have strong effects on the regions suitability for grapevine cultivation determining a shift from currently suitable areas toward new ones. Furthermore, shortening phenological advancement is expected to affect the ripening period negatively, by affecting biochemical and physiological processes and thus impacting berry sugar-acid and flavonoid levels, colour, and aroma, especially for early ripening varieties. In this research, multiple climate, soil, topography, and land use data are analyzed and integrated into a multi-criteria evaluation (MCE) to classify suitable areas for grapevine according to FAO classification under actual and future climate conditions. In particular, through the adoption of machine learning techniques, some specific qualitative targets (BRIX, acidity, polyphenol content), functional to obtaining specific oenological objectives will be analyzed. The analysis is focused on the Cannonau terroir, in the region of Sardinia (Italy), and in particular the qualitative target data for land suitability model calibration and validation will be acquired from three wine cellars collecting production from single farmers located in three bioclimatic areas that can be considered as representatives of the whole Sardinia region (Nurra, Barbagia and Parteolla, located respectively in North-west, Center and South of Sardinia). A set of 8 bioclimatic, 5 pedological, and 3 topographic indicators with 1 land cover classification was selected and then divided into a range of values, according to the literature, each of which was associated with a suitability class (FAO). Bioclimatic indicators are obtained by the analysis of current and future climate scenarios from the regionalized climate models downscaled for the whole of Italy at 2.2 km spatial resolution. Considering main and secondary relevant and explanatory criteria with a hierarchical structure, after statistical autocorrelation analysis, different weights will be assigned, calculated, and associated with each factor using the analytical hierarchy (AHP) process and machine learning methods, depending on the importance of each factor in achieving specific production targets according to expert knowledge and literature. The performance of machine learning and statistical inference to define suitability as a function of environmental and bioclimatic characteristics (ANN, Random Forest, MaxEnt, Support Vector Machines), will be subsequently compared to GIS-based results to assess its applicability. The field measurements will be carried out in the pilot sites located in the north, center, and south of Sardinia and will be useful for obtaining pedological, phenological, and qualitative data for the calibration and validation of the model. This work aims to provide an assessment of the spatial variability of the environmental factors that drive terroir distribution, to preserve vineyard production and quality in a changing climate. The research is also a methodological contribution, with the integration of a machine learning approach to the multicriterial land suitability analysis techniques.