Korekcija kratkoročne vremenske prognoze radarskim snimcima i strojnim učenjem

master's thesis

• Title (english): Weather nowcasting using radar imaging and machine learning
• Author: Sara Orlić
• Mentor: Vinko Lešić (mentor)
• Committee member: Vinko Lešić (član povjerenstva)
• Committee member: Mario Vašak (član povjerenstva)
• Committee member: Branimir Novoselnik (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2025-02-21, Croatia

Abstract

Ovaj rad istražuje primjenu konvolucijskih mreža s dugom kratkoročnom memori-
jom (ConvLSTM) za kratkoročnu prognozu oborina korištenjem radarskih snimaka. Mo-
del objedinjuje konvolucijske neuronske mreže (CNN) za izdvajanje prostornih značajki
i mreže s dugom kratkoročnom memorijom (LSTM) koja je sposobna uočiti dugoročne
vremenske ovisnosti. Model je temeljen na arhitekturi enkoder-dekoder. Poseban nagla-
sak u radu stavljen je na usporedbu dvaju pristupa prognoziranju: teacher-forcing i auto-
regresivno predviđanje. Teacher-forcing model tijekom treniranja u svakom vremenskom
koraku kao ulaz prima stvarne vrijednosti iz skupa podataka. Nasuprot tome, autoregre-
sivni model koristi vlastita prethodno predviđena stanja kao ulaz za sljedeći vremenski
korak, oponašajući stvarne uvjete prognoze. Glavni izazov je neuravnoteženost poda-
taka, jer oborine zauzimaju samo mali dio radarskih snimaka. Kako bi se to riješilo, pri-
mjenjuje se funkcija gubitka Weighted Mean Squared Error (MSE). Treniranje podataka
se provodilo na superračunalu Supek, uz visoke računalne performanse koje omogućuju
zaobilaženje standardnih hardverskih ograničenja. Rezultati pokazuju da ConvLSTM
učinkovito predviđa kratkoročne oborine, postižući veliku prostornu i vremensku toč-
nost. Tehnike poput adaptive learning rate scheduling i early stopping dodatno pobolj-
šavaju učinkovitost treniranja, čime dokazujemo da je duboko učenje dobar pristup za
poboljšanje vremenske prognoze za kratkoročno predviđanje vremena (nowcasting).

Abstract (english)

This study investigates the use of Convolutional Long Short-Term Memory (ConvL-
STM) networks for precipitation nowcasting using radar images. The model integrates
Convolutional Neural Networks (CNNs) for spatial feature extraction and Long Short-
Term Memory (LSTM) networks for capturing temporal dependencies. The model is
based on an encoder-decoder architecture. A key focus of this study is the comparison
between two forecasting approaches: teacher-forcing and autoregressive prediction. The
teacher-forcing model receives the actual ground truth as input at each step during train-
ing. In contrast, the autoregressive model generates forecasts by feeding its own previous
predictions as input, mimicking real-world conditions. A key challenge is the data im-
balance, where precipitation covers only a small portion of radar images. To address this,
a Weighted Mean Squared Error (MSE) loss function is applied. Training was performed
on the Supek supercomputer, using high-performance computing to overcome standard
hardware limitations. The results show that ConvLSTM effectively predicts short-term
precipitation, achieving high spatial and temporal accuracy. Techniques like adaptive
learning rate scheduling and early stopping further optimize training efficiency, demon-
strating the potential of deep learning for improving weather nowcasting.

• Keywords: duboko učenje; vremenska prognoza; strojno učenje; ConvLSTM; napredno računanje
• Keywords (english): deep learning; weather prediction; machine learning; ConvLSTM; high-performance computing
• Language: english
• URN:NBN: urn:nbn:hr:168:279240
• Study programme: Title: Information and Communication Technology; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka informacijske i komunikacijske tehnologije (magistar/magistra inženjer/inženjerka informacijske i komunikacijske tehnologije)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Open access
• Created on: 2025-03-12 11:15:19