Multi-Source Remote Sensing Image Change Analysis

Multi-source remote sensing change detection (CD) compares observations acquired at different times and often by different sensors. Differences in imaging mechanism, radiometric response, spatial resolution, and geometric correspondence may produce apparent changes that do not correspond to actual changes. Dense pixel-level change masks provide direct spatial supervision, but they are costly to produce for heterogeneous image pairs. This thesis studies how change information can be learned under different supervision conditions using dense masks, image-derived structural cues, and sparse scribble annotations. It develops supervised, unsupervised, and weakly supervised methods for multi-source remote sensing binary change detection (BCD) and reviews the development of remote sensing change analysis from binary change localization to broader change understanding (CU). The first contribution is DenseChangeDINO, a dense foundation-model adaptation framework for supervised BCD. It selectively fine-tunes the upper transformer blocks and uses lightweight adapters to extract dense features from three backbone depths. Multi-cue temporal comparison converts the pretrained features into change-sensitive representations, which are fused by a lightweight decoder to produce pixel-wise change maps. The framework is evaluated on optical–optical, optical–synthetic aperture radar (SAR), and optical–digital surface model (DSM) image pairs. Results on six benchmarks show that satellite-pretrained dense representations can be effectively adapted to change localization under different sensor configurations. The second contribution introduces two label-free frameworks for unsupervised BCD. ETD-CD generates pseudo-change maps from modality-invariant structural cues and uses them to supervise a Transformer-enhanced encoder–decoder. G2LMNet achieves graph-guided pixel refinement by coupling region-level graph modeling with dense pixel-level representation learning. The graph branch combines region-level representation learning with weighted graph smoothing to produce a graph guidance map, while the pixel branch produces dense predictions and refines uncertain regions through graph guidance and boundary-aware regularization. The two methods indicate that structural cues or priors can help mitigate cross-modal discrepancies without manual annotation and achieve competitive performance against other unsupervised methods on heterogeneous benchmarks. The third contribution presents three scribble-supervised frameworks for BCD that use sparse labels as an intermediate form of supervision requiring only localized annotated pixels. SG-SRF converts scribbles into distance-weighted hypergraph priors. SHIFT transforms them into learnable Gaussian attention maps for iterative deep feature tracking and pseudo-label refinement. SGCA injects scribble-derived priors into cross-modal attention for end-to-end guidance of cross-source feature interaction. The three methods incorporate sparse spatial supervision at different stages of the modeling process and collectively show that sparse scribble supervision can improve change localization in several difficult heterogeneous scenes. The thesis also includes a survey chapter that places the methodological studies within the broader evolution of multi-source remote sensing change analysis. It traces the transition from binary change localization toward semantic, temporal, and language-driven CU, reviews representative benchmarks and methods across four task categories and four methodological directions, examines the growing role of visual foundation models and vision–language models, and identifies open research directions in multi-source alignment, annotation scarcity, and trustworthy reasoning. Overall, this thesis develops a coherent family of supervision-aware methods for multi-source remote sensing change analysis: 1) dense labels support accurate adaptation of pretrained visual representations; 2) structural priors support annotation-free learning when cross-modal correspondence can be inferred from the input image pair; and 3) sparse scribble annotations provide localized change information in regions where automatic structural priors are uncertain or incomplete. The survey places these methodological results within the broader evolution of multi-source change analysis, clarifying how annotation availability, sensor heterogeneity, task formulation, and emerging foundation models influence change representation and detection performance in multi-source Earth observation data.