Binary Change Detection (BCD)

• Binary Change Detection (BCD) is a process that generates binary masks to identify pixel-level changes between temporally separated remote-sensing images, supporting diverse real-world applications.
• Recent frameworks such as UniChange and MapFormer leverage multimodal data and conditional fusion strategies, achieving significant improvements in IoU and robustness across heterogeneous datasets.
• Innovative loss functions and state-space models, including Lovász-softmax, enhance model training by addressing challenges like class imbalance, registration errors, and subtle change detection.

Binary change detection (BCD) refers to the task of identifying, at the pixel level, whether a change has occurred between two co-registered remote-sensing images of the same geographical area taken at different dates. In the canonical BCD setting, the output is a binary mask indicating for each pixel if a change has occurred (1) or not (0). BCD is foundational for Earth observation applications such as land cover monitoring, urban expansion analysis, infrastructure update, and disaster assessment. Recent deep learning-based frameworks have established the state-of-the-art performance across a spectrum of benchmarks and modalities, yet significant challenges in generalization, robustness, and data efficiency persist.

1. Formal Definition and Evaluation Protocol

Let $$X^{t_1}, X^{t_2} \in \mathbb{R}^{H\times W\times C}$$ denote two co-registered remote-sensing images at times $t_1$ and $t_2$, respectively. The BCD problem seeks a binary change mask $Y \in \{0,1\}^{H\times W}$, where $Y_{ij} = 1$ if pixel $(i,j)$ changed between $t_1$ and $t_2$, and $0$ otherwise.

Standard quantitative metrics for BCD include: $$\begin{align*} \text{Precision} &= \frac{\text{TP}}{\text{TP} + \text{FP}}, \ \text{Recall} &= \frac{\text{TP}}{\text{TP} + \text{FN}}, \ \text{F1-score} &= 2\frac{\text{Precision} \times \text{Recall}}{\text{Precision}+\text{Recall}}, \ \text{IoU} &= \frac{|\hat Y\,\cap\,Y|}{|\hat Y\,\cup\,Y|} = \frac{\text{TP}}{\text{TP}+\text{FP}+\text{FN}}, \end{align*}$$ where TP, FP, and FN denote true positives, false positives, and false negatives, respectively.

2. Model Architectures and Algorithmic Innovations

Multimodal LLM-based Unification (UniChange)

UniChange (Zhang et al., 4 Nov 2025) demonstrates a paradigm shift by leveraging Multimodal LLMs (MLLMs) for BCD. The model integrates a ViT-based vision encoder (RSBuilding-ViT-L) to extract dual-temporal features and fuses them with textual prompts using special tokens ([T1], [T2], [CHANGE]) constructed for CD tasks. The token-driven decoder employs self- and cross-attention, aligning image and instruction-level semantics. A notable feature is the replacement of classical classification heads with the [CHANGE] special token, allowing unification of BCD and SCD tasks and mitigating conflicts from dataset-specific class definitions.

Ablation experiments confirm:

• Dual-temporal semantic supervision (+0.95 IoU improvement on WHU-CD),
• Optimal LoRA rank (8) for language decoder,
• Robustness to train/test splits via joint training on multiple BCD and SCD datasets.

State Space Models in Vision (ChangeMamba, AtrousMamba)

The MambaBCD (Chen et al., 4 Apr 2024) and AtrousMambaBCD (Wang et al., 22 Jul 2025) frameworks employ Visual State-Space Models (VSSM) for efficient global-spatial modeling.

• MambaBCD uses a patchwise VMamba encoder with four downsampling scales. The change decoder fuses corresponding feature maps from the bi-temporal inputs using three spatio-temporal relationship mechanisms (sequential, cross, and parallel/channel modeling). Each pathway is processed by a SSM block with 2D cross-scan, providing $O(N)$ complexity and full field-of-view context. The final output is generated from the fused features using a $1\times1$ convolution and softmax.
• AtrousMambaBCD introduces Atrous-Window Selective Scan (AWVSS) modules into the decoder, integrating local and global dependencies through dilation and multi-scale windowing. The AWVSS module processes each feature patch with multiple dilation rates/windows, fuses the outputs via SE-style reweighting, and passes the result through further layers.

A consistent theme across both models is the use of pixelwise cross-entropy and (for MambaBCD) Lovász-softmax loss to counter class imbalance.

Conditional Fusion via Side Information (MapFormer)

MapFormer (Bernhard et al., 2023) utilizes pre-change semantic maps ($m^{(1)}$) as side-information to condition bi-temporal feature fusion, thus defining “conditional change detection.” The network employs a shared MixVisionTransformer encoder for $I^{(1)}$ and $I^{(2)}$, a shallow CNN encoder for $m^{(1)}$, and a fusion module at each scale that computes multiple latent “views” followed by semantic-feature–driven soft attention. Training includes a supervised cross-modal contrastive loss to align latent representations across modalities and time.

Ablations reveal:

• The contrastive loss adds ~2.6pp IoU on DynamicEarthNet.
• Robustness to noisy or low-resolution pre-change maps (IoU remains >5–10pp above SOTA bi-temporal approaches).
• Conditional frameworks perform well even in cross-modal (map+1 image) scenarios, supporting potential real-time applications.

Self-Supervised Pixel-Level Learning

The fully self-supervised method of (Chen et al., 2021) employs a dual-branch Siamese ResUnet with a vector-quantized pixel representation, trained via patch-shifted pixelwise InfoNCE contrastive loss and a codebook-usage entropy term. Binary change maps are obtained by cosine distance followed by Rosin thresholding, and an uncertainty-aware teacher-student refinement is introduced for robustness against seasonal effects.

Quantitative results show competitive or superior performance to supervised baselines on both homogeneous (OSCD, MUDS) and heterogeneous (California Flood) datasets.

Specialized Designs for Building Change Detection

CGCCE-Net (Wang, 3 Aug 2025) augments Siamese PVT-based encoders with four dedicated modules:

• Change-Guided Residual Refinement (CGRR): early texture-based priors,
• Global Cross Correlation Module (GCCM): channel/spatial attention and linear-angle cross-temporal correlation,
• Semantic Cognitive Enhancement Module (SCEM): local–global feature reweighting,
• Cross Fusion Decoder (CFD): progressive, attentive reconstruction.

Empirical results indicate consistently higher F1 and IoU on LEVIR-CD (IoU=84.91%), WHU-CD (IoU=90.29%), and GZ-CD (IoU=81.65%) compared to all tested baselines.

3. Datasets and Benchmarking

Major benchmarks for BCD include WHU-CD, LEVIR-CD(+), S2Looking, SYSU-CD, DynamicEarthNet, HRSCD, and, recently, the large-scale, high-challenge ChangeNet dataset (Ji et al., 2023).

ChangeNet Highlights

• 31,000 multi-temporal sets, 0.3 m, 1900×1200 px, 100 cities,
• Paired binary masks derived from six-class semantic annotations,
• Real-world asymmetric distortions, class imbalance up to 20:1,
• Stronger practical difficulty: peak IoU for best models $\leq$ 32.5% (vs. 82.5% on LEVIR-CD).

Method	IoU (ChangeNet)	F1 (ChangeNet)
IFNet	31.2	39.9
SNUNet	30.8	39.3
BIT	31.4	39.8
ChangeSTAR	32.4	41.1
ChangeFormer	32.5	40.8
Siamese-FCN	29.7	37.5

Key recommendations from ChangeNet include explicit misalignment correction and the use of imbalance-aware loss functions.

4. Loss Functions, Training Schemes, and Inference

Common loss formulations comprise:

• Pixelwise binary cross-entropy: $$L_{BCE} = - \sum_{i,j} [y_{ij} \log p_{ij} + (1-y_{ij})\log(1-p_{ij})]$$,
• Dice loss: $$L_{Dice} = 1 - \frac{2\sum p_{ij}y_{ij}}{\sum p_{ij}+\sum y_{ij}}$$,
• Lovász-softmax loss for class-imbalance mitigation,
• Cross-modal or pixelwise contrastive loss when leveraging side-information or unsupervised settings.

Composite objectives may disable or zero-out certain loss terms for “pure” BCD datasets during multi-task learning (as in UniChange), thus avoiding label conflicts.

Data augmentation includes random rotations and flips across most benchmarks. Batch sizes and optimizer choices (AdamW, learning rates $1\times10^{-4}$–$5\times10^{-4}$) are aligned with modern vision standards.

5. Robustness, Generalization, and Practical Recommendations

Several studies report experiments on robustness to input degradation:

• MambaBCD exhibits only an 8.7% F1 drop under severe Gaussian blur ($\sigma=2$) as compared to $>$14% for Transformer backbones (Chen et al., 4 Apr 2024).
• MapFormer is robust to coarse or noisy pre-change side information, still outperforming bi-temporal-only SOTA (Bernhard et al., 2023).

Pretraining on large, heterogeneous datasets such as ChangeNet is beneficial for transfer learning, often yielding 2–3pp IoU improvement on other BCD tasks (Ji et al., 2023).

Recommended architectural and algorithmic strategies include:

• Explicit spatial misalignment correction (deformable convolutions, optical flow),
• Multi-scale, attention-based feature fusion,
• Imbalance-aware and boundary-sensitive loss functions,
• Integration of semantic priors when available,
• Augmenting supervised pipelines with self- or cross-modal supervision for label-scarce or noisy domains.

6. Limitations, Open Challenges, and Extensions

BCD remains challenged by:

• Robustness to registration errors and viewpoint variation, especially in real-world “asymmetric” datasets,
• Sensitivity to class imbalance and subtle changes,
• Scalable annotation for large-scale and temporally extended datasets,
• Generalization across sensor modalities, geographic domains, and change types.

Emerging unified architectures (e.g., UniChange (Zhang et al., 4 Nov 2025)) that synthesize language, modality, and supervision via special token conditioning enable learning from composite BCD and SCD datasets and have demonstrated superior performance.

Conditional change detection frameworks (e.g., MapFormer) extend naturally to settings with incomplete or auxiliary side-information (e.g., GIS, LIDAR), establishing a link between classical map-modification analysis and deep BCD pipelines.

Self-supervised and uncertainty-aware techniques (Chen et al., 2021) indicate strong potential for BCD scenarios lacking exhaustive labels or suffering from frequent non-semantic (seasonal) changes.

Overall, binary change detection continues as a central, challenging domain for multimodal, robust, and scalable learning in remote-sensing analytics.