Klasifikasi Few-Shot Citra Hiperspektral Menggunakan Contrastive Learning Untuk Pemetaan Tumpahan Minyak

Karami, Fathan Abi (2025) Klasifikasi Few-Shot Citra Hiperspektral Menggunakan Contrastive Learning Untuk Pemetaan Tumpahan Minyak. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5025211156-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (10MB) | Request a copy

Abstract

Insiden tumpahan minyak sangat berbahaya bagi lingkungan dan masyarakat di sekitar wilayah terdampak. Oleh karena itu, pemetaan tumpahan minyak penting dilakukan untuk memantau persebarannya. Citra hiperspektral menyediakan informasi spektral yang kaya sehingga bermanfaat dalam proses pemantauan tumpahan minyak di lautan. Namun, sebagian besar penelitian terdahulu berfokus pada metode klasifikasi supervised untuk mendeteksi tumpahan minyak dari citra hiperspektral. Padahal metode supervised memerlukan banyak data training yang proses pelabelannya memerlukan waktu dan tenaga yang banyak, sehingga metode ini kurang cocok diterapkan untuk mengatasi permasalahan tumpahan minyak. Untuk mengatasi permasalahan tersebut, diperlukan suatu metode klasifikasi yang cocok untuk jumlah data training yang sangat kecil (Few-Shot Classification). Oleh karena itu, penelitian ini mengimplementasikan metode klasifikasi Few-Shot menggunakan strategi contrastive learning untuk memetakan tumpahan minyak dari citra hiperspektral, yang membutuhkan data pelatihan jauh lebih sedikit dibandingkan dengan metode supervised. Pengembangan model dilakukan melalui beberapa tahap, yaitu tahap pre-processing, tahap pre-training, tahap fine-tuning, dan tahap testing. Pertama, tahap pre-processing, di mana sampel training akan di-generate dari dataset. Pada tahap ini, sampel training juga akan diaugmentasi menggunakan transformasi spektral dan spasial untuk menambah jumlah sampel training. Kedua, tahap pre-training, di mana deep learning network dilatih menggunakan strategi contrastive learning simple siamese network (SimSiam). SimSiam sendiri adalah siamese network yang tidak menggunakan pasangan negatif, batch ukuran besar, dan momentum encoders. Pada tahap fine-tuning, classification layer ditambahkan ke model yang telah dilatih pada tahap pre-training, lalu layer tersebut dilatih menggunakan sampel training yang telah diaugmentasi dengan menggunakan metode supervised learning. Terakhir, tahap pengujian dilakukan untuk mengukur performa model menggunakan semua dataset. Hasil uji coba menunjukkan bahwa model yang diusulkan menghasilkan performa yang paling tinggi jika dibandingkan dengan model lain, seperti LR, RF, SVM, 1DCNN, 2DCNN, dan HybridSN. Model hasil penelitian berhasil mencapai AUC sebesar 99,31 ± 0,11 pada salah satu dataset. Penelitian ini juga menunjukkan bahwa jumlah parameter pada model berpengaruh terhadap performa klasifikasi.
================================================================================================================================
Oil spill incidents are very dangerous for the environment and communities around the affected areas. Therefore, oil spill mapping is important to monitor the distribution of oil spills. Hyperspectral image provides rich spectral information, making it useful in the process of monitoring oil spills in the ocean. However, most previous studies have focused on supervised classification methods to detect oil spills from hyperspectral imagery. In fact, the supervised method requires a lot of training data whose labeling process requires a lot of time and effort so that this method is less applicable to solve oil spill problems. In order to solve that problem, a classification method for small training data (Few-Shot Classification) is needed. This study implements the Few-Shot classification method using a contrastive learning strategy to map oil spills from hyperspectral images, which requires much less training data compared to supervised methods. The development of the Few-Shot classification is carried out through several stages, namely the pre-processing stage, the pre-training stage, the fine-tuning stage, and the testing stage. First, the pre-processing stage, where the samples are generated from a dataset. At this stage, the training samples will also be augmented using spectral and spatial transformations to increase the number of training samples. Second, the pre-training stage, where deep learning network are trained using a contrastive learning strategy simple siamese network (SimSiam). SimSiam is a siamese network that are not using negative pairs, large batches, and momentum encoders. Third, the fine-tuning stage, where the classification layer is added to the trained model, then the layer is trained using training samples using supervised learning method. Finally, the testing stage is carried out to measure the performance of the model using the whole dataset. Experiment shows the resulting model achieved the highest performance compared to other models, such as LR, RF, SVM, 1DCNN, 2DCNN, and HybridSN. The resulting model achieved an AUC score of 99,31 ± 0,11 in one of the dataset. This study also shows that size of parameters in model is influential to classification performance.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Citra Hiperspektral, Klasifikasi Few-Shot, Contrastive Learning, VGG-16, Pemetaan Tumpahan Minyak, SimSiam, Hyperspectral Image, Few-Shot Classification, Contrastive Learning, VGG-16, Oil Spills Mapping, SimSiam.
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. 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 Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Fathan Abi Karami
Date Deposited:	23 Jul 2025 05:48
Last Modified:	23 Jul 2025 05:48
URI:	http://repository.its.ac.id/id/eprint/120481

Actions (login required)

View Item