Adaptive Imputation for Daily Meteorological Time Series Under Bimodal Seasonal Regimes

Abstract

Daily meteorological time series data constitute a critical scientific infrastructure for climate monitoring, trend analysis, and decision-making in many industrial fields, particularly in applied meteorology. This study proposes and validates an adaptive imputation approach for daily meteorological time series under bimodal seasonality, aimed at selecting a reconstruction method based on the size of the gap and the local climate structure. This approach was organized into three phases. In the first phase, an exploratory characterization of the missing data was conducted, yielding 539 observed gaps, which led to the evaluation of 80 synthetic scenarios defined by two-time windows, four variables, and five length categories. The second phase involved constructing a controlled benchmark that included 10 local and global reconstruction methods evaluated using a multiplicative score (“Climate Score”). In this phase, the results showed a clear pattern of non-dominance: Prophet outperformed in micro-gaps, while seasonal jitter and seasonal KNN dominated in small to very large gaps. In the third phase, the final imputation showed high preservation of the global statistics, with standard deviation ratios ranging from 0.9898 to 0.9971 and very low KS statistics for five of the six variables. These results demonstrate that imputation should not be applied using a single method for all cases; instead, the characteristics of the data must be considered. This concept aligns with data-centric AI, where it is essential to understand and preserve the statistical and seasonal structure of the series.