Penerapan Deep Learning Untuk Proses Downscaling Dan Koreksi Bias Pada Dataset Meteorologi Spatiotemporal

Marzuqi, Mochammad Zharif Asyam (2026) Penerapan Deep Learning Untuk Proses Downscaling Dan Koreksi Bias Pada Dataset Meteorologi Spatiotemporal. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025221163-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (8MB) | Request a copy

Abstract

Rendahnya resolusi data Global Climate Model membuat analisis cuaca dan iklim lokal kurang memadai untuk merepresentasikan dinamika secara presisi. Secara konvensional, keterbatasan data beresolusi rendah ini ditangani melalui proses downscaling menggunakan metode statistik seperti Quantile Mapping. Dalam bidang visi komputer, downscaling memiliki prinsip yang hampir sama dengan teknik super resolution yang bertugas merekonstruksi data beresolusi rendah menjadi tinggi. Namun, berbeda dengan citra visual biasa, downscaling meteorologi memproses grid data yang terikat oleh hukum fisis dunia nyata. Penelitian ini bertujuan menerapkan dan mengevaluasi metode deep learning untuk downscaling serta koreksi bias pada variabel temperatur, presipitasi, dan angin di wilayah Asia Tenggara. Bias mengacu pada kesalahan sistematis model global terhadap data referensi. Pengujian menggunakan dua pendekatan: 1 dataset berupa input resolusi rendah dan target dari European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v5 atau ERA5 serta 2 dataset yaitu input prakiraan ECMWF Integrated Forecasting System (IFS) dengan target ERA5. Empat arsitektur dievaluasi, yaitu UNet, ResNet18, ConvNeXt, dan cGAN, menggunakan teknik pre-upsampling dan post-upsampling. Hasilnya, pre-upsampling menunjukkan performa yang lebih tinggi. Pada skenario 1 dataset, UNet meraih performa terbaik dengan nilai Korelasi Pearson 0,9689 (Angin U) dan 0,9622 (Angin V). Pada tantangan nyata 2 dataset, UNet konsisten unggul dengan Root Mean Square Error (RMSE) 0,5555 dan Mean Absolute Error (MAE) 0,4329 pada temperatur 2 meter, menghasilkan metrik yang lebih baik dibandingkan Quantile Mapping.
===============================================================================================================================
The low resolution of Global Climate Model data renders local weather and climate analysis inadequate for precisely representing meteorological dynamics. Conventionally, the limitations of this low-resolution data are addressed through a downscaling process using statistical methods such as Quantile Mapping. In the field of computer vision, downscaling shares almost the same principles as super-resolution techniques, which are tasked with reconstructing low-resolution data into high resolution. However, unlike standard visual images, meteorological downscaling processes data grids bound by real-world physical laws. This research aims to apply and evaluate deep learning methods for the downscaling and bias correction of temperature, precipitation, and wind variables in the Southeast Asia region. Bias refers to the systematic deviation of a global model from the reference data. The evaluation utilizes two approaches: a 1-dataset scenario low-resolution input and target from European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v5 or ERA5 and a 2-dataset scenario forecast input from the ECMWF Integrated Forecasting System (IFS) with ERA5 as the target. Four architectures were evaluated, namely UNet, ResNet18, ConvNeXt, and cGAN, using pre-upsampling and post-upsampling techniques. The results indicate that pre-upsampling yields higher performance. In the 1-dataset scenario, UNet achieved the best performance with Pearson Correlation values of 0.9689 (U-Wind) and 0.9622 (V-Wind). In the real-world challenge of the 2-dataset scenario, UNet consistently excelled with a Root Mean Square Error (RMSE) of 0.5555 and a Mean Absolute Error (MAE) of 0.4329 for 2-meter temperature, producing better metrics compared to Quantile Mapping.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Asia Tenggara, Deep Learning, Downscaling, Koreksi Bias, Meteorologi, Spatiotemporal Bias Correction, Deep Learning, Downscaling, Meteorology, Southeast Asia, Spatiotemporal, UNet
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Mochammad Zharif Asyam Marzuqi
Date Deposited:	17 Jun 2026 07:08
Last Modified:	17 Jun 2026 07:09
URI:	http://repository.its.ac.id/id/eprint/133813

Actions (login required)

View Item