Giesco
21st GiESCO International Meeting: ‘A Multidisciplinary Vision towards Sustainable Viticulture’ DATA MINING APPROACHES FOR TIME SERIES DATA ANALYSIS IN VITICULTURE. POTENTIAL OF THE BLiSS (BAYESIAN FUNCTIONAL LINEAR REGRESSION WITH SPARSE STEP FUNCTIONS) METHOD TO IDENTIFY TEMPERATURE EFFECTS ON YIELD POTENTIAL
Cécile LAURENT1,2,3,* Meïli BARAGATTI4, James TAYLOR1, Bruno TISSEYRE1, Aurélie METAY2, Thibaut 3 SCHOLASCH 1 2 ITAP, Univ. Montpellier, Montpellier SupAgro, Irstea, France SYSTEM, Univ Montpellier, CIHEAM-IAMM, CIRAD, INRA, Montpellier SupAgro, France 3 Fruition Sciences, Montpellier, France 4 MISTEA, Univ Montpellier Montpellier SupAgro, INRA, France *Corresponding author : cecile@fruitionsciences.com
Abstract: Context and purpose of the study – Vine development, and hence management, depends on dynamic factors (climate, soil moisture, cultural practices etc.) whose impact can vary depending upon their temporal modalities (timing, duration, threshold, eventually trajectory and memory effects). Therefore, understanding the effect of the temporal variation of these factors on grapevine physiology would be of strategic benefit in viticulture, for example in establishing yield potential. Today many estates own data that can support temporal analyses, while the emergence of precision viticulture allows management at higher spatial and temporal resolutions. These data are a great opportunity to advance knowledge about the dynamics of grapevine physiology and production, and promote an improved precision of vineyard practices. The exploitation of these data needs analytical methods that fully explore time series data. However, current methods tend to only focus on a few key phenological stages or time steps. Such approaches do not fully address the potential information captured by continuous temporal measurements because they introduce limitations : i) they rely on choices of variables and timing, ii) they often require suppressing data or analysing only parts of a time series and iii) data correlation over time is not taken into account. A new approach is explored in this paper, using a Bayesian functional Linear regression with Sparse Steps functions (BLiSS method). The BLiSS method overcomes the mentioned limitations and leads to a more complete and objective analysis of time series data. Based on the identification of climatic periods affecting yield, the objective of the study is to evaluate the potential of the BLiSS method. Materials and method ‐ Minimum and maximum daily temperatures during the year preceding the harvest year were regressed against the number of clusters per vine using the BLiSS method on one block of a commercial vineyard in the Bordeaux region over 11 years. The reliability and pertinence of the BLiSS method to reveal already reported, ignored or underestimated temperature effects on the number of clusters per vine are tested by comparison with literature results. Results ‐ The BLiSS method allowed the detection of periods when temperature influenced the number of clusters per vine during the year preceding the harvest year. Some of the detected periods of influence had already been reported in literature. However, the BLiSS outcomes suggested that some of those known periods may have a different duration or several effects, thus challenging actual knowledge. Finally, some new periods of influence were identified by the BLiSS method. These results confirmed the potential of the BLiSS method to undertake a fuller exploration of time series data in the case of climate influence on grape yield.
Keywords: climate, functional analysis, temporal variability, cluster number June 23 - 28, 2019 | Thessaloniki | Greece GiESCO Thessaloniki | 603 21st GiESCO International Meeting: ‘A Multidisciplinary Vision towards Sustainable Viticulture’
Abstract: Context and purpose of the study – Vine development, and hence management, depends on dynamic factors (climate, soil moisture, cultural practices etc.) whose impact can vary depending upon their temporal modalities (timing, duration, threshold, eventually trajectory and memory effects). Therefore, understanding the effect of the temporal variation of these factors on grapevine physiology would be of strategic benefit in viticulture, for example in establishing yield potential. Today many estates own data that can support temporal analyses, while the emergence of precision viticulture allows management at higher spatial and temporal resolutions. These data are a great opportunity to advance knowledge about the dynamics of grapevine physiology and production, and promote an improved precision of vineyard practices. The exploitation of these data needs analytical methods that fully explore time series data. However, current methods tend to only focus on a few key phenological stages or time steps. Such approaches do not fully address the potential information captured by continuous temporal measurements because they introduce limitations : i) they rely on choices of variables and timing, ii) they often require suppressing data or analysing only parts of a time series and iii) data correlation over time is not taken into account. A new approach is explored in this paper, using a Bayesian functional Linear regression with Sparse Steps functions (BLiSS method). The BLiSS method overcomes the mentioned limitations and leads to a more complete and objective analysis of time series data. Based on the identification of climatic periods affecting yield, the objective of the study is to evaluate the potential of the BLiSS method. Materials and method ‐ Minimum and maximum daily temperatures during the year preceding the harvest year were regressed against the number of clusters per vine using the BLiSS method on one block of a commercial vineyard in the Bordeaux region over 11 years. The reliability and pertinence of the BLiSS method to reveal already reported, ignored or underestimated temperature effects on the number of clusters per vine are tested by comparison with literature results. Results ‐ The BLiSS method allowed the detection of periods when temperature influenced the number of clusters per vine during the year preceding the harvest year. Some of the detected periods of influence had already been reported in literature. However, the BLiSS outcomes suggested that some of those known periods may have a different duration or several effects, thus challenging actual knowledge. Finally, some new periods of influence were identified by the BLiSS method. These results confirmed the potential of the BLiSS method to undertake a fuller exploration of time series data in the case of climate influence on grape yield.
Keywords: climate, functional analysis, temporal variability, cluster number June 23 - 28, 2019 | Thessaloniki | Greece GiESCO Thessaloniki | 603 21st GiESCO International Meeting: ‘A Multidisciplinary Vision towards Sustainable Viticulture’