Rekonstrukcija temperaturnih polj s Fourierjevo transformacijo in globokim učenjem

Vogrin, Domen (author), Kristan, Matej (mentor)

Površinska temperatura morja (SST) ima močan vpliv na vreme in podnebje. Težavo pri beleženju SST predstavljajo oblaki, ki prekrivajo posamezne predele površine morja in s tem preprečujejo meritve na teh predelih. Posledično so meritve SST nepopolne, njihova uporaba za kratkoročno in dolgoročno napovedovanje vremena pa otežena. Za reševanje te težave so bili predlagani različni pristopi rekonstrukcije, med njimi tudi metoda DINCAE2, ki temelji na dveh zaporednih samokodirnikih. V nalogi predlagamo pristop, ki metodo DINCAE2 nadgradi z uporabo Fourierove transformacije. Metodi DINCAE2 dodamo vzporedni samokodirnik, ki uporablja Fourierovo transformacijo, nato pa rezultata obeh vej združimo. S tem smo uspeli zmanjšati RMSE rekonstrukcije SST za 34 % v primerjavi z metodo DINCAE2. S prilagojeno funkcijo izgube smo dosegli še dodatno zmanjšanje RMSE za 40 % v primerjavi z metodo DINCAE2.

• Language: Slovenian
• Keywords: Fourierova transformacija, površinska temperatura morja, rekonstrukcija, globoko učenje
• Work type: Bachelor thesis/paper
• Typology: 2.11 - Undergraduate Thesis
• Organization: FRI - Faculty of Computer and Information Science ; FMF - Faculty of Mathematics and Physics
• Year: 2024
• PID: 20.500.12556/RUL-159393
• COBISS.SI-ID: 202292995
• Publication date in RUL: 09.07.2024

Secondary language

• Language: English
• Title: Reconstruction of temperature fields using a Fourier transform and deep learning

Abstract

Sea Surface Temperature (SST) has a significant impact on weather and climate. A problem in recording SST is presented by clouds that cover certain areas of the sea surface, thus preventing measurements in these areas. As a result, SST measurements are incomplete, making their use for short- and long-term weather forecasting more difficult. Various reconstruction approaches have been proposed to solve this problem, including the DINCAE2 method based on two consecutive autoencoders. In this thesis, we propose an approach that enhances the DINCAE2 method by using Fourier transformation. We add a parallel autoencoder to the DINCAE2 method that uses Fourier transformation, and then combine the results of both branches. With this, we managed to lower the RMSE of SST reconstruction by 34 % compared to the DINCAE2 method. With the adapted loss function, we reduced RMSE by 40 % compared to the DINCAE2 method.

• Keywords: Fourier transform, sea surface temperature, reconstruction, deep learning