Remote Sensing Image Change Analysis

• RSICA is the analysis of multitemporal remote sensing data to detect pixel-level, semantic, and object-based changes on Earth's surface.
• It integrates methods such as binary change detection, semantic mapping, and change captioning to provide detailed spatial and temporal insights.
• Key challenges include handling seasonal variations, heterogeneous sensor data, and registration errors, driving innovations in deep learning and domain alignment.

Remote Sensing Image Change Analysis (RSICA) denotes the analysis of changes on the Earth’s surface from multitemporal remote sensing data. In the literature considered here, it spans binary change detection, semantic change detection, change localization, object-based change analysis, and higher-level interpretation tasks such as change captioning and instruction-guided question answering. A canonical pixel-level formulation maps two co-registered images $I_A, I_B \in \mathbb{R}^{H \times W \times C}$ to a binary change mask $M \in \{0,1\}^{H \times W}$, while more expressive formulations generate semantic maps, captions, or interactive textual responses conditioned on the same bi-temporal imagery (Dong et al., 19 Jan 2025, Yu et al., 2024, Deng et al., 30 Jul 2025).

1. Conceptual scope and task structure

RSICA is commonly organized around several related problem formulations. Traditional formulations include binary change detection, which classifies pixels as changed or unchanged; semantic change detection, which identifies both changed regions and their semantic transitions; and change localization, which emphasizes where changes occur in the scene. Survey and task-oriented papers further include object-based change maps, damage assessment, and higher-level interpretation as part of the same analytic continuum (Yu et al., 2024, Liu et al., 2024).

A standard supervised change-detection formulation is

$f_\theta : (I_A, I_B) \mapsto M,$

where $M_{ij}=1$ if the land surface at pixel $(i,j)$ has changed between $t_A$ and $t_B$, and $0$ otherwise. This formulation underlies many Siamese encoder-decoder systems and segmentation-style architectures (Dong et al., 19 Jan 2025). Heterogeneous change detection generalizes the setting to cases where the two images are acquired by different sensors, with pixel vectors $\boldsymbol{x}_m \in \mathbb{R}^P$ and $\boldsymbol{y}_m \in \mathbb{R}^Q$ living in different feature spaces; in that setting, direct differencing is no longer meaningful, and change analysis becomes a domain-alignment problem (Luppino et al., 2018).

Recent work broadens RSICA beyond mask prediction. Remote Sensing Image Change Captioning (RSICC) requires natural-language descriptions of changed object categories, spatial relations, and temporal dynamics, making it a high-level structured change analysis task. Interactive RSICA goes further by defining the model as $M \in \{0,1\}^{H \times W}$0, where $M \in \{0,1\}^{H \times W}$1 is an instruction or question and $M \in \{0,1\}^{H \times W}$2 is a generated answer; this setting unifies captioning, binary classification, quantification, localization, open-ended question answering, and multi-turn dialogue (Liu et al., 2024, Deng et al., 30 Jul 2025). Change-Agent makes a closely related distinction by using the term Remote Sensing Image Change Interpretation (RSICI) for a broader goal that explicitly combines pixel-level localization with semantic explanation (Liu et al., 2024).

A recurring theme is that these formulations are nested rather than disjoint. RSICC, for example, must implicitly localize changes to describe “lower right corner” or “near the road,” must identify object categories such as buildings or roads, and must infer temporal operations such as added, removed, or expanded. This suggests that high-performing language-based systems presuppose strong RSICA capabilities even when their final output is text rather than a mask (Liu et al., 2024).

2. Data modalities, acquisition conditions, and failure modes

RSICA operates on bi-temporal or multi-temporal imagery acquired by a range of sensors and at multiple resolutions. The survey literature lists optical RGB, multispectral imagery such as Sentinel-2 and Landsat, SAR such as Sentinel-1, and, in some settings, LiDAR, hyperspectral data, map data, or text metadata. Spatial resolution ranges from very high resolution imagery used for building and road monitoring to medium-resolution data used for land-cover and agricultural change analysis (Yu et al., 2024).

The apparent simplicity of comparing two images masks several failure modes. Seasonal variation is a canonical example: vegetation color and texture can differ dramatically between winter and summer, leading PCAKM, difference images, thresholding, and even CNN-based detectors to flag vegetation as changed when no land-cover change occurred. The style-translation literature therefore treats seasonal variation as a style discrepancy that must be normalized before downstream change detection (Zhang et al., 2021). A related problem appears in self-supervised and captioning settings, where pseudo-changes may arise from illumination, seasonal effects, sensor noise, or minor misregistration rather than genuine object or land-cover transitions (Wang et al., 14 May 2026).

Heterogeneous change detection introduces an additional difficulty: the pre-event and post-event images may be acquired by different sensors with different numbers of channels and different imaging physics. In that case, assumptions valid in homogeneous change detection—shared feature space, comparable spectral signatures, or direct band differencing—no longer hold. Regression-based heterogeneous change detection explicitly frames the problem as learning mappings between sensor domains and using residuals as change indicators (Luppino et al., 2018).

Geometric and scale factors also matter. Side-looking imagery, parallax, and view-angle differences can create mismatches even when semantics remain stable, as highlighted by datasets such as S2Looking and by methods designed to handle strong inter-image correlations and slight misalignment. Multi-resolution settings further complicate matters: object scale relative to pixel size changes across high-, medium-, and low-resolution imagery, and this can shift the balance between false alarms at detailed boundaries and missed detections of small objects (Dong et al., 19 Jan 2025, Jiang et al., 2021).

Most methods therefore assume co-registration or approximate co-registration. Diffusion-based change detection explicitly starts from “two co-registered images taken at different times,” weakly supervised temporal learning assumes orthorectified and aligned imagery from mapping agencies, and several models note that significant misregistration or parallax may degrade performance (Kiruluta et al., 2024, Bou et al., 5 Jan 2026, Liu et al., 2024). This suggests that RSICA is not only a modeling problem but also a data-conditioning problem, where acquisition geometry and nuisance variability are often decisive.

3. Methodological paradigms

The methodological landscape of RSICA includes classical differencing, supervised deep architectures, self-supervised and weakly supervised learning, explicit graph and state-space interaction, generative modeling, and style-normalization pipelines. The following representative paradigms are all documented in the cited literature.

Paradigm	Representative mechanism	Example paper
Classical change analysis	Image differencing, ratioing, CVA, PCA, post-classification comparison	(Yu et al., 2024)
Heterogeneous domain alignment	Bidirectional regression and residual distance images	(Luppino et al., 2018)
Siamese encoder-decoder CD	Shared backbone, bi-temporal feature fusion, pixel-level mask prediction	(Dong et al., 19 Jan 2025)
Difference-aware deep modeling	Channel-spatial difference weighting and layer exchange	(Dong et al., 19 Jan 2025)
Graph-based interaction	Bitemporal graphs, unified self-focus, graph interaction module	(Liu, 2023)
State-space modeling	SD-SSM, TT-SSM, stacked CaMa layers with linear complexity	(Liu et al., 2024)
Diffusion-based CD	Generate $M \in \{0,1\}^{H \times W}$3 from $M \in \{0,1\}^{H \times W}$4, then derive SSIM-based change map	(Kiruluta et al., 2024)
Style-normalized CD	Seasonal image translation via SRM and ISD before CD	(Zhang et al., 2021)
Self-/weak supervision	Pixel-wise contrastive learning, multi-view SSL, weak temporal supervision	(Chen et al., 2021, Chen et al., 2021, Bou et al., 5 Jan 2026)

Classical approaches compute differences or ratios, apply PCA or clustering, or compare independently derived class labels. These methods are computationally simple and historically important, but the survey literature emphasizes their sensitivity to noise, registration errors, and limited spatial context (Yu et al., 2024). In heterogeneous settings, regression-based alignment replaces direct differencing: one learns $M \in \{0,1\}^{H \times W}$5 and $M \in \{0,1\}^{H \times W}$6 to map one sensor domain into the other, computes residual distance images $M \in \{0,1\}^{H \times W}$7 and $M \in \{0,1\}^{H \times W}$8, normalizes and combines them, and thresholds the result to obtain a change map (Luppino et al., 2018).

Deep learning replaces hand-crafted features with learned hierarchical representations. LENet, for example, frames change detection as pixel-wise segmentation with a Siamese Swin Transformer V2 encoder, ChangeFPN, and a Layer-Exchange Decoder. Its Channel–Spatial Difference Weighting module computes cosine-similarity-derived weights in both channel and spatial dimensions, while the Layer-Exchange Decoder enhances interaction between temporal streams during decoding (Dong et al., 19 Jan 2025). WRICNet addresses multi-resolution RSICA with a Weighted Rich-scale Inception module for shallow multi-scale features, a Weighted Rich-scale Coder for deep multi-scale features, and weighted scale blocks that emphasize edge information and reduce both false alarms and missed alarms (Jiang et al., 2021).

Several recent methods make temporal interaction explicit. BGINet-CD constructs graphs from bitemporal feature maps by soft-clustering pixels into graph vertices, then applies a unified self-focus mechanism and graph interaction module to enhance information coupling between the two times while suppressing task-irrelevant interference (Liu, 2023). RSCaMa introduces state-space modeling into change captioning and, by extension, into RSICA design more broadly: SD-SSM performs difference-aware spatial modeling with a bidirectional scan over flattened spatial tokens, while TT-SSM interleaves temporal tokens in a cross-wise sequence to model before-after dependencies with global receptive field and linear complexity (Liu et al., 2024).

Generative approaches form another line of work. Diffusion-based change detection uses a Stable Diffusion model to generate an approximation $M \in \{0,1\}^{H \times W}$9 from $f_\theta : (I_A, I_B) \mapsto M,$0, then computes an SSIM-based change map by comparing $f_\theta : (I_A, I_B) \mapsto M,$1 to the real $f_\theta : (I_A, I_B) \mapsto M,$2; lower structural similarity indicates likely change (Kiruluta et al., 2024). Seasonal normalization via image translation addresses a different nuisance factor: a CycleGAN-style model with a Style-Based Recalibration Module and Improved Style Discriminator translates winter images into summer style or vice versa, after which a conventional detector such as PCAKM or GETNET can operate in a season-consistent domain (Zhang et al., 2021).

Annotation-efficient paradigms are increasingly central. Self-supervised pixel-wise contrastive learning trains a Siamese ResUnet to align pixel features from shifted positive pairs, uses vector quantization for feature augmentation, and derives change maps by thresholding feature dissimilarity, with an uncertainty mechanism to improve temporal robustness (Chen et al., 2021). Earlier multi-view self-supervised change detection uses pseudo-Siamese ResNet-34 branches, explicit contrastive loss for heterogeneous data, and BYOL-style implicit contrastive learning for homogeneous data, then computes multi-scale feature regression errors and thresholds them (Chen et al., 2021). Weak temporal supervision extends single-date semantic datasets with additional temporal observations, assumes real pairs are mostly unchanged, constructs fake change pairs from different locations, uses object-aware sIoU-based change maps, and iteratively filters noisy “real” pairs to obtain strong zero-shot and low-data performance (Bou et al., 5 Jan 2026).

4. Semantic, open-vocabulary, and interactive extensions

A major development in RSICA is the movement from binary change masks toward semantic and linguistic interpretation. RSICC formalizes change understanding as natural-language generation from bi-temporal images. In RSCaMa, two CLIP-encoded image streams are refined through multiple CaMa layers and then decoded into captions that describe changed object categories, locations, and temporal dynamics; the same work explicitly states that RSICC can be viewed as a high-level, structured change analysis task requiring localization, semantic understanding, and temporal reasoning (Liu et al., 2024).

Semantic change detection makes these semantics explicit at pixel level. Semantic-CD predicts a binary change mask $f_\theta : (I_A, I_B) \mapsto M,$3 and two semantic maps $f_\theta : (I_A, I_B) \mapsto M,$4, using a bi-temporal CLIP visual encoder, an open semantic prompter that forms semantic cost volume maps from text embeddings, a binary change detection decoder, and a semantic change detection decoder. Its design is fully decoupled: BCD and SCD are trained in separate stages to reduce task interference, and CLIP’s vision-language priors are used to move SCD toward an open-vocabulary setting (Zhu et al., 12 Jan 2025).

HiSem extends RSICC by arguing that changed and unchanged image pairs have intrinsically different semantic granularities and should not be processed under a unified modeling strategy. Its Bidirectional Differential Attention Modulation module enhances cross-temporal interactions using discrepancy-aware attention, while the Hierarchical Adaptive Semantic Disentanglement module first routes samples at image level into changed versus unchanged paths and then applies token-level Mixture-of-Experts modeling for heterogeneous changed samples (Wang et al., 14 May 2026). This implies a hierarchical RSICA design principle in which coarse change-existence perception and fine semantic interpretation are distinct but coupled stages.

Interactive systems make this hierarchy available to end users. Change-Agent combines an MCI model, which jointly performs pixel-level change detection and semantic-level change captioning, with a LLM that selects tools, executes analysis code, and returns answers about change detection, change captioning, object counting, and change cause analysis (Liu et al., 2024). DeltaVLM defines RSICA itself as a multi-turn, instruction-guided paradigm. It introduces ChangeChat-105k, a large-scale instruction-following dataset covering six interaction types, and an architecture with a bi-temporal EVA-ViT-g/14 encoder, a visual difference perception module with Cross-Semantic Relation Measuring, and an instruction-guided Q-former aligned to a frozen Vicuna-7B decoder (Deng et al., 30 Jul 2025).

These developments alter the meaning of “change analysis.” In the binary setting, RSICA asks whether or where change occurred. In semantic and interactive settings, it additionally asks what changed, how many changed objects exist, where they are in coarse or fine spatial terms, and what explanation can be given in natural language. A plausible implication is that future RSICA systems will increasingly be evaluated not only by segmentation fidelity but also by semantic specificity and instruction-following behavior.

5. Datasets, metrics, and empirical evaluation

RSICA research is grounded in a diverse benchmark ecosystem. Binary building-change datasets include LEVIR-CD, WHU-CD, SYSU-CD, CDD, and GZ-CD; semantic change detection uses SECOND; captioning relies on LEVIR-CC and WHU-CDC; interactive and joint interpretation work uses LEVIR-MCI and ChangeChat-105k; and annotation-efficient studies introduce extended datasets such as b-FLAIR, b-IAILD, and b-FLAIR-spot (Yu et al., 2024, Liu et al., 2024, Deng et al., 30 Jul 2025, Bou et al., 5 Jan 2026). Individual papers provide finer details: LEVIR-CC contains 10,077 image pairs and 50,385 captions, SECOND contains six primary classes for semantic change detection, and GZ-CD consists of suburban Guangzhou imagery collected over 2006–2019 (Liu et al., 2024, Zhu et al., 12 Jan 2025, Liu, 2023).

Evaluation protocols follow task type. Pixel-level binary or semantic detection typically uses true positives, false positives, false negatives, and true negatives to derive IoU, precision, recall, F1, overall accuracy, and, in some studies, Kappa or Separated Kappa. LENet reports IoU as the primary metric, with gains across CLCD, PX-CLCD, LEVIR-CD, and S2Looking; BGINet-CD reports precision, recall, and F1 on WHU and GZ-CD; self-supervised studies report overall accuracy and Kappa in addition to F1 (Dong et al., 19 Jan 2025, Liu, 2023, Chen et al., 2021, Chen et al., 2021). Captioning evaluation uses BLEU-1/2/3/4, ROUGE$f_\theta : (I_A, I_B) \mapsto M,$5, METEOR, and CIDEr-D, and several RSICC papers additionally report the composite score

$f_\theta : (I_A, I_B) \mapsto M,$6

which summarizes sentence-level overlap and semantic consensus (Liu et al., 2024, Wang et al., 14 May 2026). Interactive RSICA adds task-specific metrics such as accuracy and F1 for binary yes/no classification, MAE and RMSE for change quantification, and Jaccard similarity or subset accuracy for coarse-grid localization (Deng et al., 30 Jul 2025).

Representative empirical results illustrate both specialization and expansion of scope. On LEVIR-CC, RSCaMa reports BLEU-4 $f_\theta : (I_A, I_B) \mapsto M,$7, CIDEr-D $f_\theta : (I_A, I_B) \mapsto M,$8, and $f_\theta : (I_A, I_B) \mapsto M,$9, outperforming strong Transformer-based captioning baselines (Liu et al., 2024). On WHU-CDC, HiSem reports a $M_{ij}=1$0 BLEU-4 improvement over the previous best method and reaches $M_{ij}=1$1, showing the effect of hierarchical semantic disentangling (Wang et al., 14 May 2026). On SECOND, Semantic-CD reports OA $M_{ij}=1$2, F1 $M_{ij}=1$3, mIoU $M_{ij}=1$4, and SeK $M_{ij}=1$5, improving over strong semantic change detection baselines (Zhu et al., 12 Jan 2025). On interactive tasks, DeltaVLM reports F1 $M_{ij}=1$6 for binary change classification, road-localization F1 $M_{ij}=1$7, building-localization F1 $M_{ij}=1$8, and strong captioning and open-ended QA results on ChangeChat-105k (Deng et al., 30 Jul 2025).

These results also reveal a change in what counts as performance. For binary detection, low false positive rate and accurate object counts may matter as much as headline F1, as shown by weak temporal supervision experiments that emphasize false-alarm control and zero-shot robustness (Bou et al., 5 Jan 2026). For language tasks, decoder structure and visual-text alignment become decisive: RSCaMa’s decoder ablation shows that explicit cross-attention outperforms simple visual-prefix strategies for caption generation (Liu et al., 2024).

6. Challenges, misconceptions, and future directions

A common misconception is that RSICA is equivalent to binary mask generation. The literature here shows a broader picture: RSICA includes binary change detection, semantic change detection, open-vocabulary semantic mapping, change captioning, and interactive question answering over bi-temporal imagery. RSICC and interactive RSICA are not peripheral variations but direct extensions of the same analytic problem (Liu et al., 2024, Deng et al., 30 Jul 2025, Liu et al., 2024).

Another misconception is that raw differencing is sufficient whenever two images are available. Several papers explicitly show why this fails. Seasonal variation can cause vegetation to be detected as changed even when no land-cover change occurred; heterogeneous sensor spaces make direct subtraction invalid; pseudo-changes from illumination, shadows, or registration noise can dominate simple thresholds; and view-angle or side-looking effects complicate rural and urban building datasets (Zhang et al., 2021, Luppino et al., 2018, Dong et al., 19 Jan 2025). This suggests that robust RSICA requires explicit nuisance handling, whether via translation, domain alignment, uncertainty modeling, or architectural bias toward cross-temporal consistency.

The field’s central constraints remain data and generalization. Pixel-level annotation is costly and scarce, motivating self-supervised contrastive learning, weak temporal supervision, and foundation-model transfer. Domain shift is repeatedly identified as a limitation: methods trained on urban Texas, Chinese cities, or specific benchmark distributions may degrade on rural scenes, different climates, new sensors, or very long temporal baselines (Bou et al., 5 Jan 2026, Yu et al., 2024, Wang et al., 14 May 2026, Deng et al., 30 Jul 2025). Open-vocabulary semantic change detection is promising, but current datasets still have limited category inventories, so truly unseen-category evaluation remains constrained (Zhu et al., 12 Jan 2025).

Several forward directions recur across the literature. Foundation models and parameter-efficient fine-tuning are presented as a major trajectory for RSICA, especially for multimodal and large-scale scenarios (Yu et al., 2024). Multi-temporal extensions beyond $M_{ij}=1$9 are explicitly proposed for TT-SSM-style temporal serialization, weak temporal supervision, and hierarchical semantic disentangling (Liu et al., 2024, Bou et al., 5 Jan 2026, Wang et al., 14 May 2026). Multimodal fusion—particularly optical with SAR, and more generally vision-language integration—appears in both methodological proposals and surveys as a key path toward robustness under clouds, sensor gaps, and richer semantic reasoning (Yu et al., 2024, Deng et al., 30 Jul 2025). Interactive systems point toward unified outputs in which change masks, semantic labels, captions, counts, and explanations are produced within a single agentic framework (Liu et al., 2024, Deng et al., 30 Jul 2025).

Taken together, the literature portrays RSICA as a progression from pixel-wise differencing toward structured, multimodal, and semantically explicit change understanding. The technical trend is toward models that jointly encode space, time, semantics, and user intent while remaining robust to nuisance variability, label scarcity, and domain shift.