Image-to-Image Translation Citra SAR ke Citra Optik Menggunakan CycleGAN dan Hybrid Pix2Pix-CycleGAN

Imdad, Ahmad (2026) Image-to-Image Translation Citra SAR ke Citra Optik Menggunakan CycleGAN dan Hybrid Pix2Pix-CycleGAN. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6003241006-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (32MB) | Request a copy

Abstract

Perkembangan teknologi komputasi dan kecerdasan buatan telah mendorong kemajuan signifikan dalam analisis data penginderaan jauh, khususnya pada integrasi berbagai jenis sensor. Citra synthetic aperture radar (SAR) dan citra optik merupakan dua sumber data utama yang banyak dimanfaatkan dalam pemantauan penggunaan lahan dan kebencanaan. Citra optik memiliki representasi visual yang intuitif, namun sangat dipengaruhi kondisi cuaca dan tutupan awan. Sebaliknya, citra SAR mampu beroperasi dalam segala kondisi cuaca, tetapi memiliki karakteristik visual yang sulit diinterpretasikan. Oleh karena itu, translasi citra SAR ke citra optik menjadi solusi untuk menghasilkan representasi visual yang lebih mudah dipahami tanpa kehilangan keunggulan operasional SAR. Penelitian ini bertujuan untuk mengkaji penerapan metode image-to-image translation berbasis Generative Adversarial Networks (GAN), yaitu CycleGAN dan hybrid Pix2Pix-CycleGAN, untuk mentranslasikan citra SAR ke citra optik di wilayah Sumatra pada periode banjir. Kajian algoritma menunjukkan bahwa perbedaan utama kedua metode terletak pada inisialisasi bobot generator. Hasil eksperimen menunjukkan bahwa skenario hybrid menghasilkan Fréchet Inception Distance (FID) rata-rata sebesar 193,231, lebih rendah dibandingkan skenario baseline sebesar 230,209. Skenario hybrid lebih andal secara semantik karena tidak mengalami semantic inconsistency berupa pembalikan tutupan lahan yang teramati pada skenario baseline. Pengujian ketahanan terhadap speckle noise menunjukkan bahwa kedua skenario mempertahankan kualitas translasi yang baik hingga tingkat speckle sedang, kemudian mengalami degradasi pada tingkat speckle berat.
============================================================================================================================================
Advancements in computational technology have significantly enhanced remote sensing data analysis, particularly in the integration of multi-sensor data. Synthetic Aperture Radar (SAR) imagery and optical imagery are two primary data sources widely used in land-use monitoring and disaster management. Optical imagery provides intuitive visual representations; however, it is highly dependent on weather conditions and cloud cover. In contrast, SAR imagery can operate under all weather conditions but has visual characteristics that are difficult to interpret. Therefore, translating SAR imagery into optical imagery offers a solution for generating a more easily understandable visual representation without losing the operational advantages of SAR. This study aims to examine the application of Generative Adversarial Networks (GAN) image-to-image translation methods, specifically CycleGAN and the hybrid Pix2Pix-CycleGAN to translate SAR imagery into optical imagery in the Sumatra region during the flood period. An analysis of the algorithms shows that the main difference between the two methods lies in the initialization of the generator weights. Experimental results show that the hybrid scenario yields an average Fréchet Inception Distance (FID) of 193.231, which is lower than the baseline scenario’s 230.209. The hybrid scenario exhibits greater semantic reliability, as it does not suffer from semantic inconsistency in the form of land cover reversal, which was observed in the baseline scenario. Robustness testing against speckle noise indicates that both scenarios maintain satisfactory translation quality up to moderate speckle levels, beyond which degradation occurs at heavy speckle levels.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Citra Optik, CycleGAN, Pix2pix, Remote Sensing, SAR, CycleGAN, Optik, Pix2pix, Remote Sensing, SAR
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Mathematics and Science > Statistics > 49101-(S2) Master Thesis
Depositing User:	Ahmad Imdad
Date Deposited:	04 Aug 2026 08:33
Last Modified:	04 Aug 2026 08:33
URI:	http://repository.its.ac.id/id/eprint/143607

Actions (login required)

View Item