Automatic correction and completion of weather and climate time series

Scientific observations show that Earth’s climate has rapidly changed in the last 100 years. The human-induced global warming and natural variations from year to year and decade to decade, shape our climate. Therefore, the World Meteorological Organization has defined rules for creating climate normals that represent the average over a period of 30 years and are updated every 10 years. These climate normals are then used to compare shorter-term data at local, national or global levels. According to the official recommendations of the WMO, from the year 2021, all the meteorological services in Europe and in the world will launch a recalculation of the climate normals over the period from January 1, 1990 to December 31, 2020. However, unexpected events for example, data logger overload, renovation of the weather station will cause missing values in time series, making the data much harder to be utilized. Errors in the datasets can also occur, such as incorrect calibration of the measuring instruments, overloading of the thermometer shelter, etc. In this master's thesis, we compare the performance of several matrix completion methods for the problem of missing data imputation, the error detection and correction using some own-made benchmark. We compare two state of the arts methods with new matrix completion methods, among them graph regularized matrix factorization techniques. We are also proud to have found methods giving better results than the state of the art for the data completion problem. Although the results are good, there are still areas for improvement that can be considered for the future.