Comparison of imputation methods in long and varied phenological series. Application to the Conegliano dataset, including observations from 1964 over 400 grape varieties

Abstract

A large varietal collection including over 1700 varieties was maintained in Conegliano, ITA, since the 1950s. Phenological data on a subset of 400 grape varieties including wine grapes, table grapes, and raisins were acquired at bud break, flowering, veraison, and ripening since 1964. Despite the efforts in maintaining and acquiring data over such an extensive collection, the data set has varying degrees of missing cases depending on the variety and the year. This is ubiquitous in phenology datasets with significant size and length. In this work, we evaluated four state-of-the-art methods to estimate missing values in this phenological series: k-Nearest Neighbour (kNN), Multivariate Imputation by Chained Equations (mice), MissForest, and Bidirectional Recurrent Imputation for Time Series (BRITS). For each phenological stage, we evaluated the performance of the methods in two ways. 1) On the full dataset, we randomly hold-out 10% of the true values for use as a test set and repeated the process 1000 times (Monte Carlo cross-validation). 2) On a reduced and almost complete subset of varieties, we varied the percentage of missing values from 10% to 70% by random deletion. In all cases, we evaluated the performance on the original values using normalized root mean squared error. For the full dataset we also obtained performance statistics by variety and by year. MissForest provided average errors of 17% (3 days) at budbreak, 14% (4 days) at flowering, 14.5% (7 days) at veraison, and 17% (3 days) at maturity. We completed the imputations of the Conegliano dataset, one of the world’s most extensive and varied phenological time series and a steppingstone for future climate change studies in grapes. The dataset is now ready for further analysis, and a rigorous evaluation of imputation errors is included.