Seg2Change: OVCD Adapter for Remote Sensing

• Seg2Change is an adapter that couples open-vocabulary semantic segmentation with dense, category-agnostic change localization for bi-temporal remote sensing images.
• It leverages a dedicated change head and the CA-CDD dataset to overcome limitations of traditional proposal-based OVCD methods.
• The framework achieves state-of-the-art metrics on multiple benchmarks while reducing inference latency and model complexity.

Seg2Change is an adapter for open-vocabulary change detection (OVCD) in remote sensing that adapts open-vocabulary semantic segmentation models to bi-temporal change analysis by decoupling dense, category-agnostic change localization from semantic recognition. Its central design is to detect where a change occurs with a category-agnostic change head and to determine what changed with an open-vocabulary semantic segmentation model. The framework is introduced together with CA-CDD, a category-agnostic change detection dataset, and is reported to achieve state-of-the-art OVCD performance, including +9.52 IoU on WHU-CD and +5.50 mIoU on SECOND (Su et al., 13 Apr 2026).

1. OVCD formulation and the problem Seg2Change addresses

Seg2Change is defined in the setting of remote-sensing OVCD, where changes between bi-temporal images are described by arbitrary textual or semantic labels rather than a predefined closed set. In this formulation, conventional change detection pipelines are inadequate because they are typically limited to predefined classes in training datasets, which constrains scalability in real-world scenarios. The framework therefore treats OVCD as a joint vision-language problem in which semantic recognition must remain open-vocabulary while change localization must remain reliable across heterogeneous categories (Su et al., 13 Apr 2026).

The work explicitly positions itself against two instance-based OVCD paradigms. The first is M–C–I—Mask proposal → Comparator → Identifier—and the second is I–M–C—Identifier → Mask proposal → Comparator. Both are described as vulnerable to error propagation from mask generation, brittle fixed-threshold decision rules, lower parallelizability because proposals are processed instance by instance, and task-dependent paradigm switching. Seg2Change replaces these proposal-centric pipelines with direct dense prediction of a category-agnostic change map, thereby removing proposal generation and threshold-based instance classification from the primary localization path.

A recurrent misconception is that open-vocabulary change detection can be solved by applying an open-vocabulary segmenter independently at two times and comparing the outputs post hoc. Seg2Change rejects that assumption. Its architecture treats category-agnostic localization as a distinct subproblem and introduces a dedicated change head for it, rather than asking the semantic segmenter to solve localization implicitly.

2. CA-CDD and category-agnostic supervision

A major component of Seg2Change is CA-CDD (Category-Agnostic Change Detection Dataset), which is constructed by re-annotating coarse change labels from existing datasets so that change regions are no longer limited to a small set of predefined semantic categories. CA-CDD is derived from the training splits of SECOND, JL1-CD, and CNAM-CD. Instead of preserving only category-restricted change definitions, the re-annotation marks regions where the category changes between time $t_1$ and time $t_2$ as change regardless of which specific classes are involved (Su et al., 13 Apr 2026).

This supervision is used to train the category-agnostic change head. The stated rationale is that CA-CDD is closer to real OVCD because it supports arbitrary category transitions rather than only building change, land-cover subsets, or a fixed semantic inventory. The paper also specifies that the SECOND test set is not included in CA-CDD, which is important for avoiding leakage in evaluation.

The dataset has methodological significance beyond its immediate training role. It recasts the supervision target from “detect change among known categories” to “detect category transition itself.” This suggests a different inductive bias: the model is encouraged to learn change evidence that is semantically neutral at the localization stage and only semantically specific at the indexing stage.

3. Architecture: category-agnostic localization and semantic indexing

Seg2Change has two branches. The first is the Category-Agnostic Change Head (CACH), which predicts a category-agnostic change map $M_{\text{ca}}$. The second is an Open-Vocabulary Semantic Segmentation (OVSS) model, which takes the same bi-temporal images together with text prompts and outputs semantic segmentation maps $M^1$ and $M^2$. The two outputs are combined as

$M_{\text{ch}}^t = M_{\text{ca}} \cdot M^t$

for semantic change detection, and

$$M_{\text{ch}} = M_{\text{ca}} \cdot M^1 + M_{\text{ca}} \cdot M^2$$

for binary change detection (Su et al., 13 Apr 2026).

CACH contains four named components: Feature Modulation Module (FMM), Bi-temporal Difference Fusion Module (BDFM), Effective Difference Query Attention (EDQA), and Residual Upsampler (ResUp). Using a DINOv2-Base encoder, the framework selects multi-layer features

$$\{F_2^t, F_5^t, F_8^t, F_{11}^t\} = \Phi_{\text{DINO}}(I^t), \qquad t \in \{1,2\}.$$

FMM aligns these features by a $1 \times 1$ convolution and scale adjustment,

$$\tilde{F}_i^t = \text{Resize}_i(\mathrm{Conv}_{1\times1} * F_i^t).$$

BDFM then computes an attention map from absolute feature differences,

$t_2$0

uses it to enhance each temporal feature,

$t_2$1

and fuses the two streams into

$t_2$2

The paper states that this enhancement can also amplify noise and create pseudo-changes, which motivates EDQA.

EDQA calibrates these discrepancies by using the difference features $t_2$3 to query the modulated features $t_2$4. It first forms

$t_2$5

then applies sliding-window attention,

$t_2$6

The calibrated outputs are fused as

$t_2$7

Within EDQA, the MLP is augmented with a Mixture-of-Experts (MoE) to handle heterogeneous sensors, seasons, and imaging conditions.

ResUp progressively upsamples and aggregates refined differences,

$t_2$8

and produces the final dense category-agnostic prediction

$t_2$9

The OVSS branch is not replaced by Seg2Change; it is wrapped by the adapter. In the main experiments the OVSS model is SegEarth-OV3, and text prompts include class-group formulations such as building: {building, roof, house}, tree: {tree, forest}, and water: {water, river}. The paper’s own shorthand for the interface is effectively “segmentation for semantics + change head for change localization.”

4. Training objectives and optimization regime

CACH is pre-trained on CA-CDD with a weighted combination of three losses. The first is the primary change-map loss,

$M_{\text{ca}}$0

where $M_{\text{ca}}$1 is the category-agnostic change label and $M_{\text{ca}}$2 is the ResUp decoder. The second is a multi-layer upsampling loss,

$M_{\text{ca}}$3

The third is a similarity loss for unchanged regions,

$M_{\text{ca}}$4

where $M_{\text{ca}}$5 denotes unchanged-region labels (Su et al., 13 Apr 2026).

The total training objective is

$M_{\text{ca}}$6

with $M_{\text{ca}}$7, $M_{\text{ca}}$8, and $M_{\text{ca}}$9. The unchanged-region similarity term is specifically intended to suppress pseudo-change predictions, complementing the denoising role of EDQA.

The reported optimization configuration is: DINOv2-Base backbone; selected layers $M^1$0; feature channels $M^1$1; window size $M^1$2; experts $M^1$3; crop size $M^1$4; batch size $M^1$5; learning rate $M^1$6; training for 20 epochs on a single NVIDIA RTX 4090. These details indicate that the adapter is lightweight relative to many full end-to-end change detection systems, while still relying on a strong OVSS backbone for semantics.

5. Evaluation, benchmarks, and reported performance

Seg2Change is evaluated on six datasets spanning three task families. For building change detection, the benchmarks are WHU-CD and LEVIR-CD. For land-cover change detection, they are DSIFN and CLCD. For semantic change detection, they are SC-SCD and SECOND. The reported metrics include $M^1$7, $M^1$8, Precision $M^1$9, Recall $M^2$0, Overall Accuracy, and Kappa; on semantic change detection the reported summaries are $M^2$1 and $M^2$2 (Su et al., 13 Apr 2026).

The baselines listed in the comparison are PCA_KMeans, CVA, DCVA, UCD-SCM, AnyChange, Inst-CEG, and DynamicEarth. Some baselines are also re-run with SegEarth-OV3 as the identifier and marked with an asterisk for fairness. The numerical results reported for Seg2Change are:

Benchmark group	Dataset	Reported Seg2Change result
Building CD	WHU-CD	86.18 F1 / 75.72 IoU / 0.8562 Kappa
Building CD	LEVIR-CD	78.72 F1 / 64.91 IoU / 0.7742 Kappa
Land-cover CD	DSIFN	58.56 F1 / 41.40 IoU
Land-cover CD	CLCD	47.89 F1 / 31.48 IoU
Semantic CD	SC-SCD	35.68 mF1 / 23.22 mIoU / 95.82 OA / 0.3385 Kappa
Semantic CD	SECOND	42.89 mF1 / 29.08 mIoU / 95.17 OA / 0.4045 Kappa

The headline gains reported in the paper are +9.52 IoU on WHU-CD, +3.87 F1 on DSIFN, +9.73 F1 on CLCD, +5.25 mIoU and +6.57 mF1 on SC-SCD, and +5.50 mIoU and +5.38 mF1 on SECOND. The framework is also described as more efficient than DynamicEarth on semantic tasks, with lower GPU memory, lower inference latency, and only about 3.9M learnable parameters.

Ablation studies attribute performance gains to each CACH component. The progression FMM only $M^2$3 +BDFM $M^2$4 +EDQA $M^2$5 +ResUp is reported to improve performance on WHU-CD and DSIFN, and the best backbone feature combination is stated to be $M^2$6. The loss ablation further indicates that $M^2$7 is especially helpful for suppressing pseudo-changes, while CA-CDD pretraining gives the best results on SECOND compared with pretraining on the original coarse datasets.

6. Position within the literature, distinctions, and limitations

Seg2Change belongs to a rapidly expanding line of work that adapts segmentation or foundation-model machinery to change detection, but its target task and decomposition are distinct. It is not identical to SEG2CD, the conversion recipe introduced in “Exchange Is All You Need for Remote Sensing Change Detection,” where ordinary semantic segmentation models become change detectors by inserting a parameter-free exchange mechanism into a Siamese Encoder–Exchange–Decoder framework (Dong et al., 12 Jan 2026). SEG2CD is centered on feature exchange and information-preserving bi-temporal fusion, whereas Seg2Change is centered on open-vocabulary semantic segmentation models, category-agnostic dense change localization, and semantic indexing for OVCD.

It is also distinct from AnyChange, which is a zero-shot, training-free change detection framework built on SAM via bitemporal latent matching and an optional point query mechanism (Zheng et al., 2024). AnyChange compares object proposals across time in SAM’s latent space using bidirectional negative cosine similarity, whereas Seg2Change explicitly rejects proposal-centric and threshold-driven OVCD as its primary design principle. Similarly, it differs from SCM, which combines FastSAM and CLIP for unsupervised building change detection through Recalibrated Feature Fusion and Piecewise Semantic Attention but is specialized to building-oriented UCD rather than open-vocabulary bi-temporal change labeling (Tan et al., 2023).

Several limitations are stated or implied in the Seg2Change paper itself. The framework still depends on the quality of the chosen OVSS model for semantic labeling. CACH is trained on re-annotated datasets derived from existing labels, so annotation coverage still reflects available data sources. The design is for bi-temporal change detection rather than arbitrary multi-temporal sequences. It also requires prompt or class-name mappings, although the paper states that it does not rely on heavy prompt engineering. The broader implication suggested by the work is that future improvements in OVSS backbones and larger category-agnostic change corpora could improve OVCD without redesigning the localization head (Su et al., 13 Apr 2026).