Unified Change Detection Framework (UniCD)

• Unified Change Detection Framework (UniCD) is a comprehensive paradigm that integrates diverse sensor modalities and supervision levels into a single, scalable system.
• It employs shared backbones, modality-adaptive routing, and multi-branch decoders to effectively align cross-modal data such as optical and SAR imagery.
• The framework improves deployment flexibility and resource utilization while addressing challenges in urban mapping, disaster monitoring, and heterogeneous data fusion.

A Unified Change Detection Framework (UniCD) refers to an architectural and algorithmic paradigm that consolidates heterogeneous change detection modalities, supervision regimes, and data sources under a single, deeply integrated system. UniCD frameworks directly address the intrinsic diversity of real-world change detection experiments—spanning cross-modalities (e.g., optical versus SAR), varied supervision levels (supervised, weakly supervised, unsupervised), and deployment scenarios ranging from homogeneous urban mapping to data-scarce disaster monitoring. Recent advances have crystallized the essential components and design strategies for UniCD, with rigorous experimental and theoretical backing.

1. The Rationale for Unification in Change Detection

Traditional change detection models are tailored to specific data types or annotation regimes, such as optical bi-temporal imagery under pixel-level supervision, leading to limited adaptability in operational and research contexts. Modal distribution discrepancies between sensor types (e.g., passive optical versus active SAR), geometric misalignments, supervision gaps, and conflicting semantic definitions are endemic challenges. A unified framework seeks to overcome these barriers by:

• Sharing feature extraction across modalities and tasks (Liu et al., 25 Mar 2025, Jiang et al., 25 Jan 2026, Shu et al., 21 Jan 2026).
• Enabling supervision-agnostic collaborative optimization (Jiang et al., 25 Jan 2026, Wu et al., 2022).
• Integrating multi-source data and prompts, including open-vocabulary semantic priors (Zhu et al., 15 Dec 2025, Madani et al., 11 Nov 2025, Zhang et al., 4 Nov 2025).
• Building modular architectures that are both scalable and efficient.

This paradigm is motivated by improvements in deployment flexibility, transferability, and resource utilization observed in recent benchmark studies.

2. Core Architectural Patterns in UniCD

The architecture of UniCD frameworks is characterized by several universal and modality-specific design principles:

• Shared Backbone: A common encoder (e.g., CNN, Transformer, or foundation model) processes all input modalities and supervision types, yielding latent multi-scale representations (Liu et al., 25 Mar 2025, Jiang et al., 25 Jan 2026, Zhu et al., 24 Mar 2025).
• Modality-Adaptive Routing: Mixture-of-Experts (MoE) modules enable pixel-wise or block-wise specialization for different modality branches, such as optical and SAR (Liu et al., 25 Mar 2025, Shu et al., 21 Jan 2026). Gating networks dynamically select experts conditioned on either input modality or local feature statistics.
• Multi-Branch Heads: Supervision-specific decoder branches (supervised, weakly supervised, unsupervised) are attached atop the shared encoder, each with custom regularization and task inference logic (Jiang et al., 25 Jan 2026, Wu et al., 2022).
• Cross-Modal Alignment: Privileged training streams (e.g., simulated SAR from optical images via speckle synthesis) and self-distillation mechanisms enforce latent space consistency between modalities (Liu et al., 25 Mar 2025).
• Registration–Detection Integration: Some frameworks (e.g., DiffRegCD) include dense geometric alignment modules for misregistered inputs (Madani et al., 11 Nov 2025).
• Open-Vocabulary and Prompt-Driven Modules: Foundation-model approaches (e.g., UniVCD, UniChange) utilize frozen vision/text encoders and prompt-based inference to generalize to new semantic change descriptors (Zhu et al., 15 Dec 2025, Zhang et al., 4 Nov 2025).

Table: Principal UniCD Components Across Leading Frameworks

Framework	Backbone Type	Cross-Modal Module	Unsupervised Path	Prompt/Token Support
M$^2$CD (Liu et al., 25 Mar 2025)	CNN/Transformer	MoE, O2SP	Self-distillation	No
UniRoute (Shu et al., 21 Jan 2026)	CNN (ResNet)	AR$^2$-MoE, MDR-MoE	CASD	No
UniCD (v2) (Jiang et al., 25 Jan 2026)	CNN	STAM, CRR, SPCI	SPCI	No
UniVCD (Zhu et al., 15 Dec 2025)	Foundation (SAM2, CLIP)	SCFAM	Full	Text prompt
Change3D (Zhu et al., 24 Mar 2025)	Video models	Perception frames	Captioning	No
UniChange (Zhang et al., 4 Nov 2025)	MLLM (LLaVA 7B)	Token-driven vision	Full	Text prompt, token

3. Modality and Fusion Strategy: Mixture-of-Experts, O2SP, Routing

In cross-modal scenarios (e.g., optical/SAR, multispectral) the principal technical challenge is the subspace shift between input domains. UniCD frameworks employ the following strategies:

• Mixture-of-Experts (MoE): At each backbone stage, a set of modality-adaptive experts is activated via sparse gating, implementing specialty functions (e.g., 1×1 conv for dimensional alignment, MLP for multimodal fusion). Top-$k$ softmax selection ensures coverage and avoids mode collapse (Liu et al., 25 Mar 2025, Shu et al., 21 Jan 2026).
• Optical-to-SAR Guided Path (O2SP): A synthetic SAR image ($\hat X_2 = X_1 \odot S$) is generated from the optical pre-event input, infusing SAR-style features and enabling cross-stream alignment through self-distillation (Liu et al., 25 Mar 2025).
• Pixel-wise Routing MoE (UniRoute): The AR$^2$-MoE module disentangles local and global feature representations by binary (hard) routing. The decoder MDR-MoE selects among fusion primitives (subtraction, concatenation, multiplication) per pixel, suppressing incompatible operations under heterogeneous settings (Shu et al., 21 Jan 2026).
• Domain-Specific BatchNorm: Modality tags condition normalization statistics, enhancing stability across sensor domains (Shu et al., 21 Jan 2026).

Self-distillation, entropy minimization, and cosine consistency losses further enforce alignment.

4. Supervision-Agnostic Collaborative Optimization

Unified Change Detection systems generalize across supervision regimes via multi-branch learning and shared representation:

• Supervised Branch: Employs spatial-temporal attention modules (e.g., STAM) and balanced contrastive/dice losses for end-to-end pixel-level label propagation (Jiang et al., 25 Jan 2026).
• Weakly Supervised Branch: Uses image-level classification heads, global CAMs, spatial-coherency and contrastive regularization to extract coarse change indicators (Jiang et al., 25 Jan 2026, Wu et al., 2022).
• Unsupervised Branch: Incorporates semantic prior-driven pseudo-label generation via external foundation models (e.g., FastSAM, CLIP), feeding results to weakly supervised pathways (Jiang et al., 25 Jan 2026, Zhu et al., 15 Dec 2025).
• GAN-based Schemes: Some architectures use an adversarial generator–segmentor–discriminator triad for cross-task coverage (Wu et al., 2022).

Multi-branch training losses are jointly scheduled and balanced to maintain optimization stability and consistency.

5. Open-Vocabulary, Prompt-Driven and MLLM-Based UniCD

Emerging UniCD frameworks employ foundation models and multimodal LLMs (MLLMs):

• Frozen Foundation Models: CLIP and SAM2 encoders provide robust high-level semantics and detailed segmentation priors; lightweight adapters align spatial and contextual features (Zhu et al., 15 Dec 2025).
• Prompt/Token Coordination: Text queries guide inference, with flexible class definitions achieved via prompt engineering; UniChange relies on special tokens ([T1], [T2], [CHANGE]) in autoregressive sequence generation (Zhang et al., 4 Nov 2025).
• Open-Vocabulary Operation: Category-agnostic change inference is performed by contrasting CLIP-derived text embeddings and vision features, associating arbitrary semantic classes without retraining (Zhu et al., 15 Dec 2025).
• Cross-Source Knowledge Integration: MLLM-driven models can merge annotations from BCD and SCD sources, handling label conflicts strictly through token and prompt design (Zhang et al., 4 Nov 2025).

This direction expands UniCD’s scope to unlabeled settings and flexible semantic querying.

6. Experimental Evidence and Performance Evaluation

Extensive benchmarking across remote sensing and video datasets substantiates the superiority and versatility of UniCD frameworks:

• Cross-modal (Optical/SAR) CD: MiT-b1 M$^2$CD achieves OA=96.19%, mF1=91.96%, and mIoU=85.66% on CAU-Flood, exceeding all prior baselines (Liu et al., 25 Mar 2025).
• Modality-adaptive Routing: UniRoute matches or surpasses specialist ensembles on LEVIR-CD, WHU-CD, HTCD, with $<$40% parameters and $<$11% FLOPs (Shu et al., 21 Jan 2026).
• Supervision Domain: UniCD (v2) provides +12.72% F1 gain over best weakly-supervised competitors (Jiang et al., 25 Jan 2026); GAN-based FCD frameworks maintain robust results under unsupervised and regional supervision (Wu et al., 2022).
• Open-Vocabulary and MLLMs: UniChange sets new state-of-the-art on WHU-CD, S2Looking, LEVIR-CD+ and SECOND across both BCD and SCD tasks, handling semantic conflicts arising from diverse annotation schemes (Zhang et al., 4 Nov 2025).
• Captioning and Complex Tasks: Change3D achieves ultra-light performance in change captioning and damage assessment with 6–13% of SOTA parameter cost (Zhu et al., 24 Mar 2025).

Ablation studies reveal the critical impact of MoE specialization, cross-modal alignment, spatial-temporal fusion, and prompt engineering.

7. Limitations, Controversies, and Future Directions

Despite substantive progress, UniCD frameworks face several recognized limitations:

• Increased model complexity and training time due to multi-branch and MoE modules (Liu et al., 25 Mar 2025, Shu et al., 21 Jan 2026).
• Dependence on foundation model priors in unsupervised and open-vocabulary settings; domain adaptation remains challenging (Zhu et al., 15 Dec 2025, Jiang et al., 25 Jan 2026).
• Feature space collapse and routing instability may occur without careful balancing of MoE gating (Liu et al., 25 Mar 2025, Shu et al., 21 Jan 2026).
• Extension to >2 temporal frames and multi-class (versus binary) change detection is an active area of research (Jiang et al., 25 Jan 2026).
• Registration–detection integration needs further development for highly misaligned data and multi-modal observations (Madani et al., 11 Nov 2025).
• Computational efficiency and inference speed trade-offs, especially in foundation model-based pipelines (Zhu et al., 15 Dec 2025, Zhu et al., 24 Mar 2025).

This suggests future UniCD advances will likely focus on adaptive regularization, uncertainty quantification, scalable temporal modeling, and deeper integration of multimodal priors. A plausible implication is the formation of unified monitoring frameworks for real-time and longitudinal Earth observation, robust to annotation scarcity and sensor heterogeneity.