A comparative analysis of machine learning approaches to gap filling meteorological datasets

Observational data of the Earth's weather and climate at the level of ground-based weather stations are prone to gaps due to a variety of causes. These gaps can inhibit scientific research as they impede the use of numerical models for agricultural, meteorological and climatological applications as well as introducing analytic biases. In this research, different machine learning techniques are evaluated together with traditional approaches to gap filling automated weather station data. When filling gaps for a specific data stream, data from neighbouring weather stations are used in addition to reanalysis data from the European Centre for Medium-Range Weather Forecasts atmospheric reanalyses of the global climate, ERA-5 Land. A novel gap creation method is introduced that provides 100% coverage in sampling the dataset while ensuring that the sampled data are randomly distributed. Gap filling across a range of different gap lengths and target variables are compared using a range of error functions. The variables selected for modelling are mean air temperature, dew point, mean relative humidity and leaf wetness. Our results show that models perform best on gap-filling temperature and dew point with worst performance on leaf wetness. As expected, model performance decreases with increasing gap length. Comparison between machine learning and reanalysis approaches show very promising results from a number of the machine learning models.