R2U-Net3+: Recurrent Residual Convolutional Network U-Net3+ untuk Segmentasi Radio Frequency Interference (RFI) dalam Proses Restorasi Citra Radar Cuaca C-Band

Alfarisy, Alfan (2023) R2U-Net3+: Recurrent Residual Convolutional Network U-Net3+ untuk Segmentasi Radio Frequency Interference (RFI) dalam Proses Restorasi Citra Radar Cuaca C-Band. Masters thesis, Institute Teknologi Sepuluh Nopember.

Text 6025221052-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only until 1 April 2026. Download (4MB) | Request a copy

Abstract

Penelitian ini mengatasi masalah gangguan Radio Frequency Interference (RFI) dalam data radar cuaca yang sangat vital untuk prediksi kondisi cuaca. Gangguan tersebut umumnya berasal dari perangkat WIFI yang beroperasi di frekuensi C-Band, sama dengan radar cuaca. Kami mengusulkan sebuah solusi inovatif yang menggabungkan teknik segmentasi semantik dan image inpainting dengan penerapan Convolutional Neural Networks (CNN). Solusi ini diwujudkan dalam bentuk arsitektur baru, Recurrent Residual Convolutional Network U-Net3+ (R2U-Net3+), yang mengintegrasikan U-Net3+ full scale skip connections dengan Recurrent Residual Convolutional Layer (RRCL). Hasil penelitian menunjukkan bahwa R2U-Net3+ memberikan peningkatan signifikan dalam deteksi area RFI dibandingkan dengan model-model segmentasi gambar sebelumnya seperti U-Net, U-Net++, R2U-Net, R2U++ dan U-Net3+. R2U-Net3+ mendapatkan nilai Dice Coefficient sebesar 95,352% dan IoU 91.409%. Evaluasi lebih lanjut menunjukkan bahwa penyesuaian parameter time- step pada recurrent berdampak signifikan terhadap akurasi dan efektivitas segmentasi RFI.
Dalam proses pemulihan citra pasca penghapusan RFI, metode inpainting Telea berhasil memulihkan nilai-nilai yang hilang dengan akurasi tinggi, terindikasi dari nilai MAE dan MSE yang rendah. Hasil ini membuktikan efektivitas metode yang kami kembangkan dalam menghasilkan data radar cuaca yang lebih akurat dan bebas dari gangguan RFI.
=============================================================================================================================
This research addresses the problem of Radio Frequency Interference (RFI) in vital weather radar data for weather condition prediction. The interference typically originates from WIFI devices operating on C-Band frequencies, the same as weather radar. We propose an innovative solution that combines semantic segmentation and image inpainting techniques using Convolutional Neural Networks (CNN). This solution is realized in the form of a new architecture, Recurrent Residual Convolutional Network U-Net3+ (R2U-Net3+), integrating U- Net3+ full-scale skip connections with Recurrent Residual Convolutional Layer (RRCL).
The research results show that R2U-Net3+ significantly improves RFI area detection compared to previous image segmentation models such as U-Net, U- Net++, R2U-Net, R2U++, and U-Net3+. R2U-Net3+ achieved a Dice Coefficient of 95.352% and an IoU of 91.409%. Further evaluation shows that adjusting the time-step parameter in the recurrent significantly impacts the accuracy and effectiveness of RFI segmentation. In the post-RFI removal image restoration process, the Telea inpainting method successfully recovered missing values with high accuracy, indicated by the low MAE and MSE values. These results prove the effectiveness of our developed method in producing more accurate and RFI-free weather radar data.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Radar Cuaca, Radio Frequency Interference, Segmentasi Semantik, U-Net3+, Recurrent Residual Convolutional Layer (RRCL), Weather Radar, Semantic Segmentation
Subjects:	T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55101-(S2) Master Thesis
Depositing User:	Alfan Alfarisy
Date Deposited:	09 Feb 2024 23:35
Last Modified:	09 Feb 2024 23:35
URI:	http://repository.its.ac.id/id/eprint/106686

Actions (login required)

View Item