MSc Thesis Defense: Alaa Almouradi, MULTI-LABEL DOMAIN-GENERALIZATION FRAMEWORK FOR REMOTE SENSING, Date & Time: 21 July, 2026 – 10:00 AM, Place: FENS L067

MULTI-LABEL DOMAIN-GENERALIZATION FRAMEWORK FOR

REMOTE SENSING

Alaa Almouradi
Computer Science and Engineering, MSc Thesis, 2026

Thesis Jury

     Prof. Erchan Aptoula (Thesis Advisor)

  Assoc. Prof. Hüseyin Özkan

  Assoc. Prof. Alp Ertürk

Date & Time: 21st July, 2026 – 10:00 AM

Place: FENS L067

Keywords : domain generalization, multi-label classification, remote sensing, style augmentation, attention

Abstract

Remote sensing produces large amounts of data, and the need for automated analysis has driven the creation of many annotated datasets. However, automated systems often face performance degradation when generalizing to unseen data. One reason is the covariate shift arising from datasets that are not universally representative and that use different sensors and acquisition methods. Thus, the need arises for domain generalization techniques that alleviate the domain shift effect. Existing style-augmentation methods for domain generalization operate on global feature maps that often represent images with a single focus. In remote sensing scenes, images are often multi-label, and global statistics would entangle the texture distributions of semantically distinct objects, therefore corrupting the intended domain regularization and weakening cross-domain transfer. To address this, I propose a multi-label domain generalization framework that decouples style augmentation at the per-label level through two localization strategies. The first employs a trainable label-localization attention module, which decomposes feature maps into per-label spatial regions via a temperature-sharpened spatial softmax and an auxiliary diversity loss that enforces spatial orthogonality between co-occurring labels. The second grounds localization in gradient-weighted class activation maps derived from the classifier itself, with maps cached in a periodically refreshed bank so that localization quality improves alongside classifier accuracy. Using these two strategies, style augmentation can be applied per label to preserve semantics. Experiments across multi-label remote sensing benchmarks under leave-one-domain-out evaluation show that this framework is a viable method to convert global style augmentation into label-specific augmentation that generalizes better in multi-label settings.