Estimasi Batimetri Menggunakan Altimetry-Derived Gravity Pada Perairan Laut Banda Berbasis Model Multi-Channel U-Net

Nugraha, Aditya Raka (2026) Estimasi Batimetri Menggunakan Altimetry-Derived Gravity Pada Perairan Laut Banda Berbasis Model Multi-Channel U-Net. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5016221043-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (9MB) | Request a copy

Abstract

Ketersediaan data batimetri resolusi tinggi di perairan dalam Laut Banda masih terbatas akibat kendala biaya dan waktu survei kapal (ship-borne sounding). Inversi berbasis satelit altimetri bisa menjadi alternatif, tetapi metode inversi konvensional terkendala dalam memodelkan hubungan non-linear dan mengintegrasikan multi-variabel parameter geofisika. Penelitian ini menerapkan model Deep learning Multi-Channel U-Net dengan mengintegrasikan tujuh variabel turunan altimetri (Free Air Gravity Anomaly, Vertical Gravity Gradient, Deflection of the Vertical Utara-Timur, Mean Dynamic Topography, Slope, dan Topo 25.1) menggunakan target data in situ Multibeam echosounder (MBES). Model dilatih pada resolusi 1 menit busur dan 0,25 menit busur kemudian dievaluasi menggunakan matriks Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Standard Deviation Error (STDe) dan Koefisien Korelasi Person (r). Model resolusi 1 menit busur menghasilkan RMSE sebesar 53,396 m, MAE sebesar 36,937 m, STDe 53,398 m, dan r 0,998 . Sementara itu, resolusi 0,25 menit busur, menghasilkan RMSE 67,497 m, MAE 43,230 m, STDe 65,968 m, dan r 0,997 . Hasil kuantitatif membuktikan bahwa arsitektur MC U-Net mengungguli model global secara signifikan (GEBCO, BATNAS, dan TOPO 25.1). Kinerja terbaik dicapai pada zona laut dalam (1.500 - 3.000 m) dengan RMSE 16,893 m. Pada zona abyssal dengan kedalaman melebihi -3.000 m, model menunjukkan stabilitas luar biasa dengan Relative RMSE kurang dari 2%. Pemetaan spasial non-linear pada data geofisika dengan MC U-Net berhasil menyajikan batimetri regional high-fidelity yang superior pada medan laut dalam yang ekstrem. Namun, analisis komparasi menunjukkan model Multi-Channel dikalahkan oleh Single-Channel (VGG dan Topo 25.1) akibat redundansi data input. Selain itu, terdapat keterbatasan fisis berupa over-smoothing teridentifikasi pada area dangkal, yang memotong puncak gunung laut hingga melampaui kedalaman 400 meter. Disimpulkan bahwa Multi-Channel U-Net efektif merekonstruksi morfologi batimetri makro dan mengisi kekosongan data laut dalam pada skala pelatihannya.
================================================================================================================================
The availability of high-resolution bathymetric data in the deep waters of the Banda Sea remains limited due to the costs and time constraints associated with ship-borne sounding surveys. Satellite altimetry-based inversion offers an alternative, but conventional inversion methods face challenges in modelling non-linear relationships and integrating multiple geophysical parameters. This study applies a Multi-Channel U-Net deep learning model by integrating seven altimetry-derived variables (Free Air Gravity Anomaly, Vertical Gravity Gradient, North-East Vertical Deflection, Mean Dynamic Topography, Slope, and Topo 25.1) using in situ Multibeam echosounder (MBES) data as the target. The model was trained at resolutions of 1 arcminute and 0.25 arcminutes and subsequently evaluated using the Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Standard Deviation of Error (STDe) and Pearson’s correlation coefficient (r). The 1-minute arc resolution model produced an RMSE of 53.396 m, an MAE of 36.937 m, an STDe of 53.398, and an r of 0.998. Meanwhile, the 0.25-minute-of-arc resolution model yielded an RMSE of 67.497 m, an MAE of 43.230 m, an STDe of 65.968 m, and an r of 0.997. The quantitative results demonstrate that the MC U-Net architecture significantly outperforms global models (GEBCO, BATNAS, and TOPO 25.1). The best performance was achieved in the deep-sea zone (1,500–3,000 m) with an RMSE of 16.893 m. In the abyssal zone, at depths exceeding -3,000 m, the model demonstrated exceptional stability with a Relative RMSE of less than 2%. Non-linear spatial mapping of geophysical data using the MC U-Net successfully produced superior high-fidelity regional bathymetry in extreme deep-sea environments. However, comparative analysis showed that the Multi-Channel model was outperformed by Single-Channel models (VGG and TOPO 25.1) due to redundancy in the input data. Furthermore, a physical limitation in the form of over-smoothing was identified in shallow areas, which cut off the peaks of seamounts beyond a depth of 400 metres. It was concluded that the Multi-Channel U-Net is effective at reconstructing macro-bathymetric morphology and filling in deep-sea data gaps within its training scale.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deep learning, Batimetri, U-Net, Altimetri, Anomali Gayaberat, Bathymetry, U-Net, Altimetry, Gravity Anomaly
Subjects:	G Geography. Anthropology. Recreation > G Geography (General) > G70.5.I4 Remote sensing R Medicine > R Medicine (General) > R858 Deep Learning
Divisions:	Faculty of Civil, Planning, and Geo Engineering (CIVPLAN) > Geomatics Engineering > 29202-(S1) Undergraduate Thesis
Depositing User:	Aditya Raka Nugraha
Date Deposited:	25 Jun 2026 00:54
Last Modified:	25 Jun 2026 00:54
URI:	http://repository.its.ac.id/id/eprint/134043

Actions (login required)

View Item