Implementasi Generative Adversarial Networks Dengan Arsitektur Multi-Temporal Resunet Dan Conditional Patch Discriminator Untuk Restorasi Citra Satelit Berawan

Rafi, Muhammad Abdul (2026) Implementasi Generative Adversarial Networks Dengan Arsitektur Multi-Temporal Resunet Dan Conditional Patch Discriminator Untuk Restorasi Citra Satelit Berawan. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5003221130-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Citra Sentinel-2 menyediakan informasi multispektral untuk pengamatan permukaan bumi, tetapi kualitasnya sering terganggu oleh tutupan awan. Penelitian ini bertujuan mengimplementasikan dan membandingkan tiga konfigurasi model untuk restorasi citra Sentinel-2 berawan, yaitu Baseline Multi-Temporal ResUNet, Generative Adversarial Networks (GAN) dengan Custom ResNet-like Discriminator, dan GAN dengan Conditional Patch Discriminator. Data yang digunakan berasal dari dataset SEN12MS-CR-TS subset asiaWest_n sebanyak 1.920 sampel. Setiap sampel terdiri atas empat timestep dan disusun menjadi tensor masukan 64 kanal yang mencakup 52 kanal Sentinel-2, empat kanal probabilitas awan, serta delapan kanal Sentinel-1 SAR polarisasi VV dan VH. Target model berupa citra Sentinel-2 13 pita yang dibentuk sebagai pseudo ground truth melalui komposit temporal. Ketiga model menggunakan generator Multi-Temporal ResUNet, sedangkan Conditional Patch GAN diinisialisasi dari checkpoint Baseline terbaik melalui adversarial fine-tuning. Evaluasi dilakukan pada 192 sampel uji, dengan 141 sampel memiliki area masker awan valid, menggunakan Mean Absolute Error (MAE), Peak Signal-to-Noise Ratio (PSNR), dan Structural Similarity Index Measure (SSIM) pada representasi RGB serta keseluruhan 13 pita, baik secara global maupun mask-only. Hasil pengujian menunjukkan bahwa blended output Conditional Patch GAN memperoleh kinerja terbaik pada seluruh metrik utama. Model tersebut menghasilkan RGB global MAE sebesar 0,006835 dan RGB mask-only MAE sebesar 0,013959, dibandingkan Baseline sebesar 0,007019 dan 0,015089 serta GAN ResNet-like sebesar 0,007660 dan 0,017445. Pada evaluasi 13 pita, Conditional Patch GAN memperoleh global MAE sebesar 0,008721 dan mask-only MAE sebesar 0,017113. Analisis visual menunjukkan bahwa Conditional Patch GAN menghasilkan warna dan intensitas yang relatif lebih konsisten terhadap pseudo ground truth, sedangkan GAN ResNet-like lebih banyak menunjukkan perubahan warna dan artefak lokal. Meskipun demikian, residual awan tipis dan kabut masih ditemukan, terutama pada area dengan nilai soft cloud mask rendah. Dengan demikian, Conditional Patch GAN memberikan blended output terbaik pada konfigurasi penelitian ini, tetapi belum menghasilkan penghilangan awan secara sempurna.
=====================================================================================================================================
Sentinel-2 imagery provides multispectral information for Earth surface observation, but its quality is frequently affected by cloud cover. This study aims to implement and compare three model configurations for restoring cloudy Sentinel-2 imagery: a Baseline Multi-Temporal ResUNet, a Generative Adversarial Network (GAN) with a Custom ResNet-like Discriminator, and a GAN with a Conditional Patch Discriminator. The data were obtained from the asiaWest_n subset of the SEN12MS-CR-TS dataset, comprising 1,920 samples. Each sample consists of four timesteps and is arranged into a 64-channel input tensor containing 52 Sentinel-2 channels, four cloud-probability channels, and eight Sentinel-1 SAR channels with VV and VH polarizations. The target is a 13-band Sentinel-2 image constructed as a temporal-composite pseudo ground truth. All three models employ a Multi-Temporal ResUNet generator, while the Conditional Patch GAN is initialized from the best Baseline checkpoint through adversarial fine-tuning. Evaluation was conducted on 192 test samples, of which 141 contained valid cloud-mask regions, using Mean Absolute Error (MAE), Peak Signal-to-Noise Ratio (PSNR), and Structural Similarity Index Measure (SSIM) for both RGB representation and all 13 spectral bands, under global and mask-only evaluation settings. The results show that the blended output of the Conditional Patch GAN achieved the best performance across all primary evaluation metrics. The model obtained an RGB global MAE of 0.006835 and an RGB mask-only MAE of 0.013959, compared with 0.007019 and 0.015089 for the Baseline, and 0.007660 and 0.017445 for the ResNet-like GAN, respectively. In the 13-band evaluation, the Conditional Patch GAN achieved a global MAE of 0.008721 and a mask-only MAE of 0.017113. Visual analysis indicates that the Conditional Patch GAN produced colors and intensities that were relatively more consistent with the pseudo ground truth, whereas the ResNet-like GAN exhibited more pronounced color shifts and localized artifacts. Nevertheless, residual thin clouds and haze remained, particularly in areas with low soft cloud-mask values. Therefore, the Conditional Patch GAN produced the best blended output under the experimental configuration used in this study, although it did not achieve complete cloud removal.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Adversarial fine-tuning, Conditional Patch Discriminator, Generative Adversarial Networks, Restorasi citra berawan, Sentinel-2, Adversarial fine-tuning, Cloudy image restoration, Conditional Patch Discriminator, Generative Adversarial Networks, Sentinel-2
Subjects:	G Geography. Anthropology. Recreation > G Geography (General) > G70.5.I4 Remote sensing Q Science > QA Mathematics > QA336 Artificial Intelligence T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing.
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Statistics > 49201-(S1) Undergraduate Thesis
Depositing User:	Muhammad Abdul Rafi
Date Deposited:	05 Aug 2026 04:32
Last Modified:	05 Aug 2026 04:32
URI:	http://repository.its.ac.id/id/eprint/141731

Actions (login required)

View Item