Monitoring the state of reservoirs is an important environmental task. The Kaniv Reservoir, the second in the Dnipro cascade, is subject to regular hydroecological monitoring and is at the same time a typical example of a small dataset, in which missing values make an adequate imputation method a precondition for any further processing. Using the monitoring data for 2021–2025, five regression models (Linear Regression, ElasticNet, Bayesian Ridge, Random Forest, SVR) and three reference methods (mean imputation, kNN, MissForest) are compared for five target parameters. The quality of imputation is assessed by leave-one-out cross-validation with bootstrap confidence intervals and the Wilcoxon signed-rank test, and checked against real measurements of the previously missing values. The selected models reduce the mean absolute error by 3–42 % compared with mean imputation, yet the reduction is confirmed statistically for only two parameters out of five, ammonium and nitrite, and even there the significance is lost after correction for multiple comparisons. For three of the five parameters the coefficient of determination is positive and comparable to values published for river monitoring of a much larger volume. For suspended solids the result is borderline, while for nitrate and inorganic phosphorus mean imputation remains acceptable. In practical terms, regression imputation is worth applying selectively, only to parameters with a confirmed error reduction, and the choice of method has to be verified on one’s own data rather than carried over from other studies.