Sea surface temperature (SST) plays an important role in weather forecasting models, which require a dense set of measurements in some area as input. However, satellite measurements are often imperfect as parts of the sea are obscured by clouds through which satellites cannot penetrate. Due to the widespread use of SST measurements, several reconstruction methods that fill in missing data based on existing measurements have been developed. The current most successful method, DINCAE2, is based on a deep convolutional neural network. An architecture based on transformers has not yet been tested for reconstruction of SST measurements even though in recent years, transformers have proven to be very powerful neural networks in the field of computer vision. In this thesis, we propose a new method called PGD (Prior-Guided Decoder for SST reconstruction) with a two-part structure that combines a transformer and a refinement decoder. The transformer calculates a low-resolution reconstruction which the refinement decoder upscales and fills in by taking into account the original measurements, thus estimating the final full-resolution reconstruction. The PGD method was analyzed and compared to the current best method for SST reconstruction—DINCAE2. On a challenging dataset a variant of the PGD method achieves a 21% smaller reconstruction error than DINCAE2.
|
