Práce měla za cíl navržení vhodného řešení pro fůzi multimodálních dat z pohledu detekčních schopností nad unifikovanými daty pro hledání změn jako jsou například stavy vodní hladiny při povodních.

Práce byla konzultována a realizována průběžně.

Evaluation level: zadání splněno

The submitted thesis fulfills all points of the assignment. The student thoroughly studied the possibilities of using multiple remote sensing modalities for anomaly and change detection in satellite imagery and provided a comprehensive overview of the relevant sensing technologies and multimodal learning approaches. Based on this analysis, the student proposed a custom architecture for multimodal representation learning focused on reducing the heterogeneity gap between SAR and multispectral data through disentangled representations.

A suitable dataset was selected and appropriately processed for the conducted experiments, including preprocessing, normalization, clipping of SAR intensities, dataset splitting, and generation of specialized validation subsets. The proposed solution was fully implemented using modern machine learning frameworks and experimentally validated on benchmark datasets.

The thesis also satisfies the requirement of comparison with existing approaches. The proposed architecture was evaluated against multiple benchmark models on the Sen1Floods11 dataset.

Minor shortcomings can be found mainly in the limited scope of ablation studies and the relatively constrained evaluation on a single benchmark domain. Nevertheless, these limitations do not significantly reduce the overall quality of the thesis and are understandable given the complexity of the assignment.

The technical report is generally well structured and understandable. The individual chapters follow a logical order. The proposed approach and its motivation are explained clearly, and the reader can follow the overall objectives of the work.

However, some parts of the theoretical background are presented in a rather point-by-point manner, resembling a collection of related topics rather than a continuously connected narrative. For example, the sections on image fusion categorization, common fusion objectives, image fusion challenges, and image registration provide relevant information but could be linked more explicitly to the proposed method and research objectives. Stronger transitions between these topics and more discussion of their relevance to the selected architecture would improve the overall coherence of the text.  I would also expect generally more information in form of legend, axes and colorbars in figures 5.1., 5.2. and 5.3. to better present the data  and results.

Despite these minor shortcomings, the report remains readable and technically sound. Therefore, I evaluate the presentation level with Grade B.

The thesis is based on a broad and highly relevant set of academic sources covering multimodal machine learning, remote sensing, change detection, and disentangled representation learning. The bibliography includes both foundational works and recent state-of-the-art publications, including journal articles, conference papers, surveys, and preprints.

A minor limitation is the strong reliance on survey literature in some parts of the related work. Additionally, minor inconsistencies in citation formatting are present. Nevertheless, the overall quality of literature work is very high and corresponds to grade A.

The results demonstrate practical relevance, particularly for remote sensing-based disaster monitoring tasks such as flood detection. The work provides a meaningful extension of existing research with practical applicability.