Machine learning imputation of missing Mesonet temperature observations

Abstract Uninterrupted and reliable weather data is a necessary foundation for agricultural decision making, required for models based on accumulated growing degree days (GDD), chill units, and evapotranspiration. When a weather station experiences a mechanical or communications failure, a replacement (imputed) value should be substituted for any missing data. This study introduces a machine learning, network-based approach to imputing missing 15-minute and daily maximum/minimum air temperature observations from 8.5 years of air temperature, relative humidity, wind, and solar radiation observations at 134 AgWeatherNet (AWN) stations in Washington State. A random forest imputation model trained on temperature and humidity observations from the full network predicted 15-minute, daily maximum, and daily minimum temperature values with mean absolute errors of 0.43 °C, 0.53 °C, and 0.70 °C, respectively. Sensitivity experiments determined that imputation skill was related a number of external factors including volume and type of training data, proximity of surrounding stations, and regional topography. In particular, nocturnal cold air flows in the upper Yakima Valley of south-central Washington caused temperature to be less correlated with surrounding stations in the overnight hours. In a separate experiment, the imputation model was used to predict base- 10 °C GDD on 2020 July 1 trained entirely on 15-minute station data from previous years. Even with the entire season of observations missing, the model predicted the GDD value within an average error 1.4% with 125 of 134 stations within 5% of observations. Since missing data can typically be resolved within a timeframe of a few days, the network-based imputation model is a sufficient substitute for short periods of missing observational weather data. Other potential applications of an imputation model are briefly discussed.