Sea surface temperaure (SST) is crucial for accurate weather forecasting.
However, clouds above the sea are preventing sensors on satellites to mea-
sure SST beneath them. Therefore, some measurements are missing. Current
state of the art method DINCAE2 consists of autoencoder with refine-
ment step. It reconstructs the missing data with bilinear convolution. We
proposed new method for SST reconstruction, which is based on fast Fourier
convolution (FFC). In spectral space, every frequency covers the en-
tire image. Consequently, using convolution in spectral space, the receptive
field covers the entire image in the first layer of network. We have analysed
the effect of different number of FFC blocks, skip connections, refinement
step, feature fusion module and different loss functions. Our best method
is AEFFC, which is an autoencoder with 9 FFC blocks without refinement
step. State of the art method DINCAE2 has 0.5% lower error on the entire
sea surface. Nevertheless, AEFFC has 2% lower error on the reconstructed
surface.
|