Code for my MSc thesis "Developing a deep learning forecasting system for short-term and high-resolution prediction of sea ice concentration" written in collaboration with MET Norway for the program Computational Science: Geoscience offered the University of Oslo
If the entire dataset is used, variables containing 1/1 or 0/1 SeaIceConcentartion will dominate, thus skewing the targetpool of the model. As a counter, try to develop a "belt" of variables along the MIZ to reduce the number of 1s and 0s.
The belt can be dynamical, i.e. following the MIZ with a thresholding distance calculated from ice-free areas and fast-ice areas.
(Include patch if close to the MIZ and 0.1-0.9, else remove)
A static belt, essentially reducing the total area-coverage in the dataset.
If the entire dataset is used, variables containing 1/1 or 0/1 SeaIceConcentartion will dominate, thus skewing the targetpool of the model. As a counter, try to develop a "belt" of variables along the MIZ to reduce the number of 1s and 0s. The belt can be dynamical, i.e. following the MIZ with a thresholding distance calculated from ice-free areas and fast-ice areas. (Include patch if close to the MIZ and 0.1-0.9, else remove) A static belt, essentially reducing the total area-coverage in the dataset.