Comparison of machine learning and spatial interpolation methods for hourly rainfall data imputation in the Oliwski Stream catchment

This is the authors' abstract. We don't add key points for this paper.

Rainfall is a key factor in hydrological analysis, yet precipitation time series are often incomplete due to malfunctioning of instruments. This study evaluates methods for imputing missing precipitation data at the Oliwa gauging station in Gdańsk, Poland. Two approaches were compared: spatial interpolation (SI) and machine learning (ML). Data from nearby stations, obtained by Gdańskie Wody, were used as predictors. A three-year period with complete records across all stations was selected. ML models, Multilayer Perceptron (MLP), Extreme Gradient Boosting (XGBoost), Random Forest (RF), and Multiple Linear Regression (MLR), while SI methods included Mean, Inverse Distance Weighting (IDW), and Correlation Coefficient Weighted Method (CCWM). Two missing-data scenarios were simulated: Missing Completely At Random (MCAR) and Missing Not At Random (MNAR). Model performance was evaluated using R 2 , RMSE, and MAE. Results showed that ML methods outperformed SI methods. For MCAR scenarios (5%, 10%, and 15% missing data), MLP achieved R 2 values between 0.789 and 0.959, while XGBoost performed lowest among ML models (0.737–0.899). Among SI methods, IDW performed best (R 2 : 0.735–0.968), while the Mean method performed worst (0.634–0.933). Model accuracy decreased as the proportion of missing data increased. For MNAR scenarios, the model performed better when imputing missing values during the driest month compared to the wettest. Later, using the Osowa station as a predictor applying ML models only, MLP again performed best with (R 2 = 0.923). These findings support the use of ML methods for improving rainfall datasets in hydrological studies.