20.500.12556/RUL-149499 Transformerska nevronska mreža za rekonstrukcijo temperature na morski gladini A transformer-based neural network for sea-surface temperature reconstruction Temperatura na gladini morja (angl. sea surface temperature, SST) igra pomembno vlogo v vremenskih napovednih modelih, ki potrebujejo za vhod gosto množico meritev na nekem področju. Satelitske meritve pa so pogosto pomanjkljive, saj dele morja zastirajo oblaki, skozi katere sateliti ne morejo prodreti. Zaradi razširjene uporabe meritev SST so se razvile številne rekonstrukcijske metode, ki manjkajoče podatke dopolnijo na podlagi obstoječih meritev. Trenutno najuspešnejša metoda, DINCAE2, temelji na globoki konvolucijski nevronski mreži. Za rekonstrukcijo meritev SST pa še ni bila preizkušena arhitektura, ki bi temeljila na transformerjih, četudi so se ti v zadnjih letih na področju računalniškega vida izkazali kot zelo zmogljive nevronske mreže. V nalogi predlagamo novo metodo, imenovano PGD (Prior-Guided Decoder for SST reconstruction), z dvodelno strukturo, ki kombinira transformer in izpopolnjevalni dekodirnik. Transformer izračuna nizkoresolucijsko rekonstrukcijo, ki jo izpopolnjevalni dekodirnik poveča in izpopolni z upoštevanjem originalnih meritev ter tako oceni končno rekonstrukcijo v polni resoluciji. Metodo PGD smo analizirali in primerjali s trenutno najboljšo metodo za rekonstrukcijo SST – DINCAE2. Na zahtevni podatkovni množici različica PGD dosega do 21% manjšo rekonstrukcijsko napako kot DINCAE2. 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. transformer rekonstrukcija temperatura na gladini morja transformer reconstruction sea surface temperature true false false Slovenski jezik Angleški jezik Diplomsko delo/naloga 2023-09-07 11:18:36 2023-09-07 11:18:37 2023-11-24 11:43:39 0000-00-00 00:00:00 2023 0 0 0000-00-00 NiDoloceno NiDoloceno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 36689 163956739 Frank_Karmen_-_Transformerska_nevronska_mreza_za_rekonstrukcijo_temperature_na_morski_gladini.pdf Frank_Karmen_-_Transformerska_nevronska_mreza_za_rekonstrukcijo_temperature_na_morski_gladini.pdf 1 2DCAB05873E8BEA839ADC2605136A0C8 772b40e9891797208f519e08f96ce16369fba400e1095fbcf839cb58c4a1aa02 865ea75f-4d5f-11ee-9934-0050569b8976 https://repozitorij.uni-lj.si/Dokument.php?lang=slv&id=173983 Fakulteta za računalništvo in informatiko Fakulteta za matematiko in fiziko 0 0 0