Data triangulation and machine learning: a hybrid approach to fill missing climate data

Historical data in climatology is important for recognizing patterns and discovering trends. However, data gaps often occur in some weather station time series. This paper presents a framework consisting of machine learning techniques combined with triangulation methods to identify missing meteorological data. Our approach is based on using data from neighboring weather stations as input for triangulation methods in combination with machine learning techniques. The current focus of the proposed framework is on filling missing temperature data and it was applied in ten different regions of Brazil, each with a different climatic configuration. Furthermore, a statistical study was conducted to estimate the best configuration which showed that a popular mathematical triangulation model combined with neural networks produced the most satisfactory results in predicting missing data, outperforming the results of traditional triangulation methods.