การสร้างลาเบลสังเคราะห์สำหรับการจำแนกภาพถ่ายดาวเทียมผ่านการเรียนรู้แบบมีผู้สอนด้วยการแทนค่าด้วยตัวเอง
Year: 2024
Type: article
Abstract: Supervised deep learning techniques are state-of-the-art methods in satellite image analysis; however, this technique requires a large, labeled dataset. The acquisition of such datasets is expensive in terms of both manpower and resources. Conversely, there is an abundance of raw satellite images, available for both commercial and academic purposes. To address the scarcity of labeled data in satellite image classification, this study presents a novel method to utilize these unlabeled data. It uses self-supervised learning technique to create synthetic labels that act as a training dataset for supervised learning models. Experimental results show that models trained with synthetic labels perform comparably to those trained with real labels, using 9 times fewer labeled data, achieving 75% accuracy on the palm oil plantation dataset and 86% accuracy on the Amazon rainforest dataset. Additionally, the process of generating synthetic labels yields versatile and knowledge-transferable visual representation vectors.
Institution: Thammasat University
Language: Thai
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W7218314751