MSc Thesis Defense: Amna Amir Butt, MULTI-LABEL CONTENT-BASED REMOTE SENSING IMAGE RETRIEVAL
MULTI-LABEL CONTENT-BASED REMOTE SENSING IMAGE RETRIEVAL
Amna Amir Butt
Computer Science and Engineering, MSc. Thesis, 2025
Thesis Jury
Prof. Erchan Aptoula (Thesis Advisor)
Assoc. Prof. Öznur Taştan
Assoc. Prof. Alp Ertürk
Date & Time: December 11th , 2025 – 02:40 PM
Place: FENS L067
Keywords : Contrastive Learning, Multi-Label Images, Content-Based Image Retrieval, Optical Remote Sensing.
Abstract
This thesis aims to address the fact that the rapid expansion of Earth observation technologies has been accompanied by the creation of massive remote sensing image archives, leading to a demand for efficient and reliable retrieval systems. Single-label Content-Based Image Retrieval (CBIR) methods are often criticized for their limitation in handling complex multi-label scenes with diverse, overlapping and co-occurring land-cover types. To address these challenges, in this thesis the use of contrastive learning for multi-label CBIR is investigated to obtain discriminative and semantically meaningful feature representations. Besides an extensive comparison of reported approaches, a new Multi-Label Adaptive Contrastive Loss function is introduced. It incorporates two adaptive mechanisms, referred to as pairwise label reweighting and dynamic temperature scaling, to promote balanced and relationship-aware feature learning that captures inter-class co-occurrence relations. The influence of rare or informative labels is increased by the first mechanism, while similarity sensitivity is adjusted by the second according to the semantic overlap between samples. These adaptive processes are designed to enable more accurate modeling of inter-class relationships, ensuring that both frequent and rare categories are allowed to contribute effectively. Experiments on multiple multi-label remote sensing datasets show consistent improvements across evaluation metrics and underline the effectiveness of the proposed framework for large-scale, semantically complex image retrieval.