Abstract

Climate reanalysis data describe how large-scale climate features evolve in time and space. They can be used for training machine learning models to automatically identify specific weather features. Such models have to handle complex spatial structures, and temporal variability, as well as class imbalance and small numbers of representative data. This work proposes to use ViT in combination with TCN to recognize different weather phenomena in reanalysis data. The structures of the data used in the work are analyzed, and climatological features are removed. Finally, long range temporal sequences of the data are constructed, and the models are trained to impute missing observations. ViT is used to model spatial structures in the data, and TCN to model temporal structures.

References

Download PDF
Back