Pengembangan Model RSAM-Seg Adaptif Untuk Segmentasi Semantik Pada Citra Satelit Tutupan Lahan

Pertiwi, Ardita Rahastri (2026) Pengembangan Model RSAM-Seg Adaptif Untuk Segmentasi Semantik Pada Citra Satelit Tutupan Lahan. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6002241006-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (17MB) | Request a copy

Abstract

Indonesia merupakan negara dengan area tutupan lahan yang sangat luas, sehingga aspek ini memiliki peran vital dalam menjaga keseimbangan ekosistem dan tata ruang nasional. Kondisi ini menuntut adanya mekanisme pengelolaan dan pemantauan lahan yang lebih efektif melalui pemanfaatan citra satelit sebagai sumber informasi spasial. Dalam menganalisis citra tersebut, dibutuhkan teknik segmentasi semantik berbasis deep learning. Namun, model deep learning konvensional sangat bergantung pada ketersediaan dataset berlabel dalam jumlah besar, sedangkan untuk membentuk dataset berlabel itu melalui proses anotasi piksel yang terkendala biaya tinggi dan waktu lama. Penelitian ini bertujuan mengembangkan model RSAM-Seg Adaptif yang dirancang agar tetap andal pada skema pembelajaran dengan ketersediaan data yang sedikit. Model ini memodifikasi arsitektur RSAM-Seg dengan mengintegrasikan mekanisme Channel Attention untuk meningkatkan selektivitas fitur dan Dilated Convolution untuk memperluas jangkauan spasial. Evaluasi dilakukan pada dataset DeepGlobe dengan variasi proporsi data latih (10%, 40%, 70%, dan 100%). Hasil penelitian menunjukkan bahwa RSAM-Seg Adaptif secara konsisten mengungguli model pembanding (U-Net, DeepLabV3+, SegFomer, dan RSAM-Seg) pada seluruh skenario. Temuan signifikan mencatat bahwa model RSAM-Seg Adaptif dengan 70% data latih mampu melampaui kinerja seluruh model pembanding yang menggunakan 100% data. Dengan demikian model RSAM-Seg Adaptif ini menawarkan solusi yang robust dan dan mampu mereduksi data untuk pelatihan untuk segmentasi semantik pada citra satelit tutupan lahan otomatis di tengah keterbatasan data berlabel.
====================================================================================================================================
Indonesia is a country with a very large land cover area, so this aspect has a vital role in maintaining the balance of ecosystems and national spatial planning. This condition demands a more effective land management and monitoring mechanism through the use of satellite imagery as a source of spatial information. In analyzing these images, a semantic segmentation technique based on deep learning is required. However, conventional deep learning models are highly dependent on the availability of large amounts of labeled datasets. meanwhile, forming those labeled datasets through a pixel annotation process is constrained by high costs and long times. This research aims to develop an Adaptive RSAM-Seg model designed to remain reliable in learning schemes with low data availability. This model modifies the RSAM-Seg architecture by integrating a Channel Attention mechanism to improve feature selectivity and Dilated Convolution to expand spatial reach. Evaluations were conducted on the DeepGlobe dataset with varying proportions of training data (10%, 40%, 70%, and 100%). Research results show that Adaptive RSAM-Seg consistently outperforms comparison models (U-Net, DeepLabV3+, SegFormer, and RSAM-Seg) across all scenarios. Significant findings noted that the Adaptive RSAM-Seg model with 70% training data was able to exceed the performance of all comparison models using 100% of the data. Thus, this Adaptive RSAM-Seg model offers a robust solution capable of reducing training data requirements for automated land cover semantic segmentation on satellite imagery, even amidst the limitations of labeled data.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	citra satelit, deep learning, RSAM-Seg Adaptif, segmentasi semantik, tutupan lahan, Adaptive RSAM-Seg, deep learning, land cover, satellite imagery, semantic segmentation
Subjects:	Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Mathematics > 44101-(S2) Master Thesis
Depositing User:	Ardita Rahastri Pertiwi
Date Deposited:	29 Jan 2026 02:26
Last Modified:	29 Jan 2026 02:26
URI:	http://repository.its.ac.id/id/eprint/130721

Actions (login required)

View Item