Integrasi Citra Sentinel-1, Sentinel-2, Dan DEMNAS Untuk Deteksi Area Banjir Di Sumatra Berbasis Progressive Cross-Attention Network

Putra, Tito Fauzan (2026) Integrasi Citra Sentinel-1, Sentinel-2, Dan DEMNAS Untuk Deteksi Area Banjir Di Sumatra Berbasis Progressive Cross-Attention Network. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5003221164-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Banjir merupakan bencana yang sering terjadi di Pulau Sumatra dan membutuhkan pemetaan wilayah terdampak secara cepat dan akurat. Pemetaan banjir berbasis citra satelit masih menghadapi kendala berupa tutupan awan pada sensor optis, gangguan speckle dan ambiguitas hamburan balik pada citra radar, serta kondisi topografi yang dapat menyebabkan kesalahan klasifikasi. Penelitian ini bertujuan membangun dan membandingkan U-Net dan Progressive Cross-Attention Network (ProCANet) untuk segmentasi area banjir melalui integrasi Sentinel-1, Sentinel-2, dan Digital Elevation Model Nasional. Data input terdiri atas tujuh channel, yaitu VV dan VH Sentinel-1, Hue, Saturation, dan Value hasil transformasi Sentinel-2, serta Slope dan Height Above Nearest Drainage dari DEMNAS. Model dilatih dan divalidasi menggunakan 5-fold spatial cross-validation, kemudian checkpoint terpilih dievaluasi pada wilayah uji independen Aceh Utara menggunakan label referensi UNOSAT. Hasil penelitian menunjukkan bahwa wilayah uji independen Aceh Utara, ProCANet memperoleh IoU sebesar 0,854, Dice/F1 sebesar 0,921, akurasi sebesar 0,973, precision sebesar 0,950, recall sebesar 0,894, dan specificity sebesar 0,990. U-Net memperoleh IoU sebesar 0,851, Dice/F1 sebesar 0,919, akurasi sebesar 0,972, precision sebesar 0,944, recall sebesar 0,896, dan specificity sebesar 0,989. Hasil tersebut menunjukkan bahwa ProCANet memiliki performa keseluruhan yang sedikit lebih tinggi dan lebih efektif menekan false positive, sedangkan U-Net sedikit lebih sensitif dalam mendeteksi piksel banjir. Pada subset out-of-fold dengan Sentinel-2 kosong atau hampir kosong, U-Net dan ProCANet masing-masing memperoleh IoU sebesar 0,4250 dan 0,4309. Nilai tersebut lebih rendah daripada hasil pada Aceh Utara, tetapi perbedaannya tidak dapat dikaitkan hanya dengan kekosongan Sentinel-2 karena karakteristik wilayah, cakupan piksel, dan distribusi kelas yang dievaluasi juga berbeda. Badan air permanen tetap menjadi sumber kesalahan positif bagi kedua model. Dengan demikian, ProCANet memberikan hasil terbaik secara keseluruhan pada wilayah uji independen, tetapi selisih performanya terhadap U-Net relatif kecil dan generalisasi hasil penelitian masih terbatas pada Aceh Utara.
=======================================================================================================================================
Flooding is a frequent disaster on Sumatra Island and requires rapid and accurate mapping of affected areas. Satellite based flood mapping continues to face several challenges, including cloud cover in optical imagery, speckle noise and backscatter ambiguity in radar imagery, and topographic conditions that may lead to classification errors. This study aims to develop and compare U-Net and the Progressive Cross-Attention Network (ProCANet) for flood area segmentation through the integration of Sentinel-1, Sentinel-2, and the Indonesian National Digital Elevation Model (DEMNAS). The input data consist of seven channels: Sentinel-1 VV and VH, Hue, Saturation, and Value derived from Sentinel-2 transformation, and Slope and Height Above Nearest Drainage derived from DEMNAS. The models were trained and validated using five-fold spatial cross-validation. The selected checkpoints were subsequently evaluated in the independent test area of North Aceh using UNOSAT reference labels. In the independent North Aceh test area, ProCANet achieved an IoU of 0.854, a Dice/F1 score of 0.921, an accuracy of 0.973, a precision of 0.950, a recall of 0.894, and a specificity of 0.990. U-Net achieved an IoU of 0.851, a Dice/F1 score of 0.919, an accuracy of 0.972, a precision of 0.944, a recall of 0.896, and a specificity of 0.989. These results indicate that ProCANet achieved slightly better overall performance and was more effective in reducing false positive predictions, whereas U-Net was marginally more sensitive in detecting flood pixels. On the out-of-fold subset with missing or nearly missing Sentinel-2 data, U-Net and ProCANet achieved IoU values of 0.4250 and 0.4309, respectively. These values were lower than those obtained in North Aceh. However, the difference cannot be attributed solely to the absence of Sentinel-2 data because the evaluated regions also differed in their characteristics, pixel coverage, and class distributions. Permanent water bodies remained a source of false positive predictions for both models. Therefore, ProCANet provided the best overall performance in the independent test area, although its performance advantage over U-Net was relatively small and the generalizability of the findings remains limited to North Aceh.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Banjir Sumatra, DEMNAS, Flood, ProCANet, Sentinel-1, Sentinel-2
Subjects:	G Geography. Anthropology. Recreation > GB Physical geography > GB1399.9 Floods Q Science > QA Mathematics > QA336 Artificial Intelligence
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Statistics > 49201-(S1) Undergraduate Thesis
Depositing User:	Tito Fauzan Putra
Date Deposited:	04 Nov 2026 08:47
Last Modified:	04 Nov 2026 08:47
URI:	http://repository.its.ac.id/id/eprint/146028

Actions (login required)

View Item