Temperature reconstruction methods

The surface air temperature of Earth is historically measured at a collection of individual weather stations. Deriving regional averages and long-term trends from these point sources require sophisticated mathematical analysis and algorithm development. To understand historical temperature, non-climatic artefacts written into the record must first be identified and removed. These artefacts include everything not attributable to the true climate, including changes in the environment surrounding weather stations, and changes in measuring methods and instrumentation. Data quality also varies and in some cases temperature measurements are missing or incorrect. The historical temperature record must hence be reconstructed from a collection of individual time series. Reconstruction involves quality control to remove suspicious data, infilling to recover missing data, and spatial interpolation to estimate temperature series for locations where no measurements were ever taken. From the reconstructed record, averages and trends can be determined for individual locations and larger regions. This report describes temperature measurement and reconstruction with a focus on Australia over 1910 to 2018. The official homogenised reconstruction of the Bureau of Meteorology is considered, and two alternative methods are presented with some comparisons. The alternative methods include nearest-neighbour infilling and an approach based on artificial neural networks. An overview of existing spatial interpolation techniques is also given, aiming to identify suitable benchmarks for evaluating alternative reconstruction techniques. The nearest-neighbour technique was used to infill all raw monthly mean temperature data available for Australia. Average anomalies were calculated from the infilled record, which show a good match with the official reconstruction of the Bureau, especially for an area where stations are distributed approximately uniformly. The neural network method is illustrated conceptually in this report, with results indicating improved performance compared with existing benchmark methods, when using the technique to estimate temperature data for unsampled locations. Further development of the neural network method is required to improve performance and to estimate trends and averages of larger regions.