Estimasi Kedalaman Perairan Dangkal Berbasis Convolutional Neural Networks dengan Strategi Transfer Learning Melalui Integrasi Citra Sentinel-1 dan Sentinel-2

Akbar, Rifky Maulana (2026) Estimasi Kedalaman Perairan Dangkal Berbasis Convolutional Neural Networks dengan Strategi Transfer Learning Melalui Integrasi Citra Sentinel-1 dan Sentinel-2. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6016241013-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Data batimetri yang akurat merupakan kebutuhan mendasar bagi Indonesia sebagai negara maritim, namun lebih dari 60% perairan dangkal Indonesia belum memiliki data batimetri yang memadai, sementara survei hidrografi konvensional (MBES) memerlukan biaya dan waktu besar. Satellite-Derived Bathymetry (SDB) berbasis citra optik seperti Sentinel-2 rentan menurun akurasinya pada kondisi kekeruhan tinggi dan tutupan awan, sementara SAR Sentinel-1 terbatas jika digunakan secara tunggal. Kajian yang membandingkan beberapa arsitektur CNN secara konsisten pada konfigurasi input tunggal maupun fusion Sentinel-1/Sentinel-2, serta pada variasi kekeruhan perairan di wilayah tropis, masih terbatas. Penelitian ini mengevaluasi efektivitas integrasi Sentinel-1 dan Sentinel-2 untuk estimasi kedalaman perairan dangkal di pesisir Pekalongan menggunakan empat arsitektur CNN, dilatih melalui pre-training berbasis pseudo-label hasil interpolasi spasial dan fine-tuning data in-situ melalui skema 5-fold cross-validation. Hasil menunjukkan Arsitektur 4 (CNN-2D) mencapai RMSE global terendah (0,3905 m), namun Arsitektur 1 (YNet-2D PPM) dipilih sebagai model utama karena stabilitas antar-fold lebih tinggi dan desain dual-branch yang unggul untuk fusion. Sentinel-2 lebih informatif secara individual, namun fusion keduanya tetap menghasilkan akurasi tertinggi, membuktikan Sentinel-1 berperan komplementer. Kondisi keruh menghasilkan akurasi terbaik (RMSE 0,2877 m), lebih dipengaruhi distribusi rentang kedalaman sampel dibandingkan tingkat kekeruhan itu sendiri. Penelitian ini membuktikan integrasi multi-sensor SAR-optik dengan CNN dua tahap efektif untuk estimasi batimetri pesisir tropis dengan turbiditas bervariasi.
=====================================================================================================================================
Accurate bathymetric data is essential for Indonesia as a maritime nation, yet morethan 60% of Indonesia’s shallow waters lack adequate bathymetric data, whileconventional hydrographic surveys (MBES) are costly and time-consuming.Satellite-Derived Bathymetry (SDB) based on optical imagery such as Sentinel-2 isprone to reduced accuracy under conditions of high turbidity and cloud cover, whileSentinel-1 SAR is limited when used alone. Studies comparing various CNNarchitectures both in single-input configurations and Sentinel-1/Sentinel-2 fusionsetups, as well as across varying water turbidity levels in tropical regions remainlimited. This study evaluates the effectiveness of integrating Sentinel-1 andSentinel-2 for estimating shallow water depths along the Pekalongan coast usingfour CNN architectures, trained via pre-training based on pseudo-labels derivedfrom spatial interpolation and fine-tuning with in-situ data using a 5-fold cross-validation scheme. The results show that Architecture 4 (CNN-2D) achieved thelowest global RMSE (0.3905 m), but Architecture 1 (YNet-2D PPM) was selectedas the primary model due to its higher inter-fold stability and superior dual-branchdesign for fusion. Sentinel-2 is more informative on its own, but fusing both satellitesources still yields the highest accuracy, proving that Sentinel-1 plays acomplementary role. Turbid conditions produce the best accuracy (RMSE 0.2877m), which is more influenced by the distribution of the depth range. Sentinel-2 ismore informative on its own, but fusing the two still yields the highest accuracy,proving that Sentinel-1 plays a complementary role. Turbid conditions produce thebest accuracy (RMSE 0.2877 m), which is influenced more by the distribution ofthe sample depth range than by the turbidity level itself. This study demonstratesthat the integration of multi-sensor SAR-optical data using a two-stage CNN iseffective for estimating bathymetry in tropical coastal areas with varying turbiditylevels.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Batimetri, Multi-sensor Fusion, Optik, Pemetaan Pesisir, SAR, Satellite Derived Bathymetry, Bathymetry, Multi-sensor Fusion, Optic, Coastal Mapping, SAR, Satellite Derived Bathymetry
Subjects:	G Geography. Anthropology. Recreation > G Geography (General) > G70.212 ArcGIS. Geographic information systems. G Geography. Anthropology. Recreation > G Geography (General) > G70.217 Geospatial data G Geography. Anthropology. Recreation > G Geography (General) > G70.5.I4 Remote sensing T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T57.74 Linear programming
Divisions:	Faculty of Civil, Planning, and Geo Engineering (CIVPLAN) > Geomatics Engineering > 29101-(S2) Master Thesis
Depositing User:	Rifky Maulana Akbar
Date Deposited:	28 Jul 2026 02:11
Last Modified:	28 Jul 2026 02:11
URI:	http://repository.its.ac.id/id/eprint/138086

Actions (login required)

View Item