Seg2Change: Adapting Open-Vocabulary Semantic Segmentation Model for Remote Sensing Change Detection

You Su; Yonghong Song; Jingqi Chen; Zehan Wen

cs.CVApr 13, 2026

Seg2Change: Adapting Open-Vocabulary Semantic Segmentation Model for Remote Sensing Change Detection

Authors: You Su, Yonghong Song, Jingqi Chen, Zehan Wen

Organizations: Xi’an Jiaotong University, Xi’an, Shaanxi, China

Abstract

Change detection is a fundamental task in remote sensing, aiming to quantify the impacts of human activities and ecological dynamics on land-cover changes. Existing change detection methods are limited to predefined classes in training datasets, which constrains their scalability in real-world scenarios. In recent years, numerous advanced open-vocabulary semantic segmentation models have emerged for remote sensing imagery. However, there is still a lack of an effective framework for directly applying these models to open-vocabulary change detection (OVCD), a novel task that integrates vision and language to detect changes across arbitrary categories. To address these challenges, we first construct a category-agnostic change detection dataset, termed CA-CDD. Further, we design a category-agnostic change head to detect the transitions of arbitrary categories and index them to specific classes. Based on them, we propose Seg2Change, an adapter designed to adapt open-vocabulary semantic segmentation models to change detection task. Without bells and whistles, this simple yet effective framework achieves state-of-the-art OVCD performance (+9.52 IoU on WHU-CD and +5.50 mIoU on SECOND). Our code is released at https://github.com/yogurts-sy/Seg2Change.

Figures & tables

Figure 1. Previous paradigms vs. our paradigm. Previous paradigms rely on change proposals and are constrained by the segmentation performance of the proposal generator. Meanwhile, both paradigms (a–b) distinguish changed instances using a predefined threshold, which makes it difficult to accurately delineate change boundaries. In contrast, our paradigm, based on category-agnostic change maps, can effectively leverage more powerful open-vocabulary segmentation models.

Figure 2. Visual comparison between predefined change category datasets (a) and our category-agnostic change detection dataset (b). The former focuses only on limited categories.

Figure 3. Annotation process of the category-agnostic change detection dataset. Regions where the category changes between the bi-temporal images are annotated as change maps. The actual recorded categories are not limited to those shown in the figure.

Figure 4. Category-Agnostic Change Head (CACH). Bi-temporal image features are extracted and processed through feature differences fusion and difference query modules to obtain calibrated discrepancies, which are then concatenated across multiple dimensions via our residual upsampler to produce the final category-agnostic change map.

Figure 5. Bi-temporal Difference Fusion Module (BDFM) and Effective Difference Query Attention (EDQA) Module.

Figure 6. Open-vocabulary building and land-cover change detection examples. In each group: images at T1 ( $I^{1}$ ), T2 ( $I^{2}$ ), labels, the results of DynamicEarth and our Seg2Change. Color rendering: "Building/Land-cover" .

Method	Identifier	Comparator	WHU-CD			LEVIR-CD			DSIFN			CLCD
Method	Identifier	Comparator	F1 c	IoU c	Kappa	F1 c	IoU c	Kappa	F1 c	IoU c	Kappa	F1 c	IoU c	Kappa
PCA_KMeans ( Celik, 2009b )	/	KMeans ( Celik, 2009b )	14.33	7.72	0.0846	9.96	5.24	0.0116	35.48	21.57	0.1865	18.42	10.14	0.0780
CVA ( Bruzzone and Prieto, 2002 )	/	CVA Match ( Bruzzone and Prieto, 2002 )	7.17	3.72	0.0023	9.27	4.86	0.0017	27.88	16.20	0.0142	13.39	7.17	0.0017
DCVA ( Bergamasco et al., 2022 )	/	CVA Match ( Bruzzone and Prieto, 2002 )	20.55	11.45	0.1537	13.85	7.44	0.0507	38.52	23.85	0.1758	20.81	11.62	0.0918
UCD-SCM ( Tan et al., 2024 )	SAM	OTSU ( Otsu, 1979 )	32.13	19.14	0.2792	32.36	19.30	0.2734	40.13	25.10	0.2443	23.31	13.19	0.1522
AnyChange ( Zheng et al., 2024 )	SAM	Latent Match ( Zheng et al., 2024 )	28.13	16.37	0.2330	32.68	19.53	0.2672	39.19	24.37	0.2280	31.96	19.02	0.2342

Table 1. Performance comparison on binary OVCD datasets.

\dagger

and

\ddagger

denote the M – C – I and I – M – C configurations of DynamicEarth. * denotes the variants implemented with SegEarth-OV3 (SegE-OV3 for short). The best and the second-best results are, respectively, marked in BOLD and in underline . Except for Kappa, all results are expressed in percentage (%).

Figure 7. Open-vocabulary semantic change detection examples. In each group: images at T1 ( $I^{1}$ ), T2 ( $I^{2}$ ), labels, the results of DynamicEarth and our Seg2Change. Color rendering: "Building" , "Vegetation" , "Tree" , "Playground" , "Water" , "Bareland" .

Method	GPU Memory	Inference	SC-SCD				SECOND
Method	Usage (GB) $\downarrow$	(ms/sample) $\downarrow$	mF1 c	mIoU c	mOA	mKappa	mF1 c	mIoU c	mOA	mKappa
UCD-SCM ( Tan et al., 2024 )	9.46 (100%)	3225 (100%)	12.06	6.51	85.95	0.0619	14.63	8.40	78.61	0.0795
AnyChange ( Zheng et al., 2024 )	6.59 (69%)	3988 (124%)	16.51	9.32	78.27	0.1056	19.54	11.84	76.07	0.1252
AnyChange* ( Zheng et al., 2024 )	10.15 (107%)	2683 (83%)	16.64	9.48	88.06	0.1192	21.30	13.81	77.97	0.1492
Inst-CEG ( Li et al., 2024 )	14.95 (158%)	5672 (176%)	6.90	3.68	94.40	0.0551	17.82	10.46	92.80	0.1604
Inst-CEG* ( Li et al., 2024 )	7.08 (75%)	2928 (91%)	23.57	14.35	94.49	0.2139	29.35	18.40	93.62	0.2670

Table 2. Performance, computational resource usage, and inference time cost comparison on semantic OVCD datasets. GPU memory usage and inference time are evaluated on a pair of bi-temporal remote sensing images (

[2,3,512,512]

Figure 8. Visualization of feature outputs across our modules. Panel (e) presents the output features after all components, illustrating the progressive refinement of difference features.

Components in CACH				WHU-CD			DSIFN
The Effectiveness of Difference Modules in CACH
FMM	BDFM	EDQA	ResUp	F1 c	IoU c	Kappa	F1 c	IoU c	Kappa
✓	✗	✗	✗	69.68	53.47	0.6870	49.65	33.02	0.3581
✓	✓	✗	✗	78.10	64.06	0.7732	53.22	36.26	0.4758
✓	✓	✓	✗	84.18	72.68	0.8354	55.57	38.47	0.4966
✓	✓	✓	✓	86.18	75.72	0.8562	58.56	41.40	0.5075

Table 3. Investigation on the impact of the proposed CACH components on the WHU-CD and DSIFN datasets.

Pre-train	SECOND (Semantic Change Detection Categories)
Pre-train	Building	Tree	Water	Vegetation	Bareland	Playground	$\text{mF1}^{c}$
SECOND ( Yang et al., 2020 )	75.50	21.58	28.32	25.36	49.88	27.15	37.96
JL1-CD ( Liu et al., 2025 )	70.46	16.96	24.90	23.54	48.02	28.78	35.44
CNAM-CD ( Zhou et al., 2023 )	66.46	22.94	28.09	26.85	46.33	24.49	35.86
CA-CDD (ours)	76.36	27.73	37.23	31.49	54.29	30.23	42.89

Table 4. Ablation experimental results (

\text{mF1}^{c}

) of the different CACH pre-train datasets on the SECOND dataset.

OVSS Model	SC-SCD				SECOND
OVSS Model	mF1 c	mIoU c	mOA	mKappa	mF1 c	mIoU c	mOA	mKappa
SegE-OV ( Li et al., 2025d )	20.74	12.28	95.03	0.1866	34.24	22.25	94.49	0.3175
RSKT-Seg ( Li et al., 2025a )	24.90	14.22	91.34	0.2192	34.98	23.27	94.44	0.3219
SAM3 ( Carion et al., 2025 )	26.07	15.49	94.92	0.2379	35.73	23.81	94.37	0.3303
SegE-OV3 ( Li et al., 2025e )	35.68	23.22	95.82	0.3385	42.89	29.08	95.17	0.4045
Method	Replace Comparator with our Proposed CACH

Table 5. Investigation on the impact of different implementation of OVSS model and the effectiveness of our CACH.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Category	Class Name
WHU-CD ( Ji et al., 2019 ) (Single-Category)	Building	{bareland,barren},grass,car,
		{tree,forest},{water,river},
		cropland, {building,roof,house}
LEVIR-CD ( Chen and Shi, 2020 ) (Single-Category)	Building	{bareland,barren},grass,car,
		{tree,forest},{water,river},
		cropland, {building,roof,house}

Appendix

Table 6. The prompt class names for the evaluation datasets. {} indicates multiple prompt vocabularies for a single class. Foreground classes are highlighted in BOLD .

CACH Backbone	WHU-CD			DSIFN
CACH Backbone	F1 c	IoU c	Kappa	F1 c	IoU c	Kappa
ResNet-18 ( He et al., 2015 )	81.28	68.47	0.8062	46.38	30.19	0.4519
ResNet-50 ( He et al., 2015 )	84.16	72.66	0.8355	55.84	38.74	0.5238
DINOv2-Small ( Oquab et al., 2023 )	84.44	73.07	0.8384	54.99	37.93	0.5148
DINOv2-Base ( Oquab et al., 2023 )	86.18	75.72	0.8562	58.56	41.40	0.5075
DINOv2-Large ( Oquab et al., 2023 )	86.58	76.33	0.8603	58.87	41.71	0.5693

Appendix

Table 7. Investigation on the impact of the CACH backbone on the WHU-CD and DSIFN datasets.

OVSS Model	WHU-CD			LEVIR-CD
OVSS Model	F1 c	IoU c	Kappa	F1 c	IoU c	Kappa
SegEarth-OV ( Li et al., 2025d )	74.53	59.40	0.7367	65.97	49.22	0.6438
RSKT-Seg ( Li et al., 2025a )	79.13	65.47	0.7840	72.40	56.75	0.7103
SAM3 ( Carion et al., 2025 )	83.17	71.19	0.8255	77.82	63.69	0.7658
SegEarth-OV3 ( Li et al., 2025e )	86.18	75.72	0.8562	78.72	64.91	0.7742

Appendix

Table 8. Investigation on the impact of different implementation of OVSS model on WHU-CD and LEVIR-CD datasets.

OVSS Model	DSIFN			CLCD
OVSS Model	F1 c	IoU c	Kappa	F1 c	IoU c	Kappa
SegEarth-OV ( Li et al., 2025d )	42.47	26.96	0.4051	33.81	20.35	0.3209
RSKT-Seg ( Li et al., 2025a )	50.97	34.20	0.4900	43.31	27.64	0.4210
SAM3 ( Carion et al., 2025 )	52.00	35.13	0.5004	43.42	27.73	0.4228
SegEarth-OV3 ( Li et al., 2025e )	58.56	41.40	0.5075	47.89	31.48	0.4239

Appendix

Table 9. Investigation on the impact of different implementation of OVSS model on DSIFN and CLCD datasets.

Feature Channel	SC-SCD				SECOND
( $C_{2},C_{5},C_{8},C_{11}$ )	mF1 c	mIoU c	mOA	mKappa	mF1 c	mIoU c	mOA	mKappa
(16, 32, 48, 64)	27.18	15.72	80.85	0.2224	32.29	19.25	90.47	0.2764
(32, 48, 54, 80)	27.64	16.04	80.83	0.2273	42.49	26.98	87.59	0.3670
(48, 64, 80, 96)	35.68	23.22	95.82	0.3385	42.89	29.08	95.17	0.4045
(32, 64, 128, 256)	30.89	18.27	97.44	0.2970	47.30	30.98	87.65	0.4092

Appendix

Table 10. Investigation on the impact of different input channel at BDFM and EDQA on SC-SCD and SECOND datasets.

Feature	SC-SCD				SECOND
Difference	mF1 c	mIoU c	mOA	mKappa	mF1 c	mIoU c	mOA	mKappa
feature abs	28.93	16.91	99.52	0.2875	33.21	19.91	99.52	0.3301
TFIM ( Li et al., 2022c )	28.10	16.35	99.53	0.2793	33.17	19.88	99.52	0.3296
CDEM ( Zhan et al., 2025b )	29.68	17.42	99.52	0.2949	35.44	21.54	54.41	0.1545
MDFM ( Ying et al., 2024b )	33.58	20.18	99.52	0.3338	44.10	28.29	87.70	0.3821
BDFM (ours)	35.68	23.22	95.82	0.3385	42.89	29.08	95.17	0.4045

Appendix

Table 11. Investigation on the impact of different implementations of the feature difference module.

Method	SC-SCD ( $\text{mF1}^{c}$ ) $\uparrow$	GPU Memory Usage (GB) $\downarrow$	Inference Time (ms/sample) $\downarrow$	Learnable Parameters (M)
DynamicEarth $\dagger$ ( Li et al., 2025c )	29.11	7.33	5035	/
DynamicEarth $\ddagger$ ( Li et al., 2025c )	9.19	15.33	6784	/
Seg2Change	35.68	6.08	1521	3.9

Appendix

Table 12. Computational resource usage, and inference time cost comparison on semantic OVCD datasets. GPU memory usage and inference time are evaluated on a pair of bi-temporal remote sensing images (

[2,3,512,512]

Method	WHU-CD						LEVIR-CD
Method	F1 c	IoU c	Pre c	Rec c	OA	Kappa	F1 c	IoU c	Pre c	Rec c	OA	Kappa
PCA_KMeans ( Celik, 2009b )	14.33	7.72	8.24	55.05	75.71	0.0846	9.96	5.24	5.70	39.50	63.60	0.0116
CVA ( Bruzzone and Prieto, 2002 )	7.17	3.72	3.81	62.03	40.74	0.0023	9.27	4.86	5.01	61.27	38.88	0.0017
DCVA ( Bergamasco et al., 2022 )	20.55	11.45	12.40	59.96	82.88	0.1537	13.85	7.44	7.63	75.50	52.17	0.0507
UCD-SCM ( Tan et al., 2024 )	32.13	19.14	20.31	76.81	88.02	0.2792	32.36	19.30	23.87	50.21	89.30	0.2734
AnyChange ( Zheng et al., 2024 )	28.13	16.37	16.48	96.16	81.86	0.2330	32.68	19.53	20.46	81.06	82.98	0.2672

Appendix

Table 13. Performance comparison on building change detection datasets.

\dagger

and

\ddagger

denote the M – C – I and I – M – C configurations of DynamicEarth. * denotes the variants implemented with SegEarth-OV3. The best and the second-best results are, respectively, marked in BOLD and in underline . Except for Kappa, all results are expressed in percentage (%).

Method	DSIFN						CLCD
Method	F1 c	IoU c	Pre c	Rec c	OA	Kappa	F1 c	IoU c	Pre c	Rec c	OA	Kappa
PCA_KMeans ( Celik, 2009b )	35.48	21.57	29.14	45.35	71.98	0.1865	18.42	10.14	11.90	40.76	73.13	0.0780
CVA ( Bruzzone and Prieto, 2002 )	27.88	16.20	17.65	66.29	41.72	0.0142	13.39	7.17	7.53	60.42	41.82	0.0017
DCVA ( Bergamasco et al., 2022 )	38.52	23.85	25.76	76.31	58.60	0.1758	20.81	11.62	12.09	74.77	57.67	0.0918
UCD-SCM ( Tan et al., 2024 )	40.13	25.10	32.82	51.61	73.83	0.2443	23.31	13.19	18.18	32.49	84.09	0.1522
AnyChange ( Zheng et al., 2024 )	39.19	24.37	31.36	52.25	72.45	0.2280	31.96	19.02	21.33	63.63	79.84	0.2342

Appendix

Table 14. Performance comparison on land-cover change detection datasets.

\dagger

and

\ddagger

Method	Bareland		Water		Building		Structure		Farmland		Vegetation		Road		Average
Method	F1 c	IoU c	F1 c	IoU c	F1 c	IoU c	F1 c	IoU c	F1 c	IoU c	F1 c	IoU c	F1 c	IoU c	mF1 c	mIoU c	mOA	mKappa
UCD-SCM ( Tan et al., 2024 )	16.13	8.77	3.32	1.69	15.00	8.11	18.33	10.09	9.18	4.81	16.52	9.01	5.95	3.06	12.06	6.51	85.95	0.0619
AnyChange ( Zheng et al., 2024 )	28.01	16.29	4.22	2.16	19.88	11.03	18.30	10.07	11.19	5.93	30.29	17.85	3.71	1.89	16.51	9.32	78.27	0.1056
AnyChange* ( Zheng et al., 2024 )	24.38	13.88	3.42	1.74	35.91	21.88	17.94	9.85	9.27	4.86	21.92	12.31	3.63	1.85	16.64	9.48	88.06	0.1192
Inst-CEG ( Li et al., 2024 )	0.00	0.00	14.68	7.92	10.34	5.45	0.00	0.00	0.00	0.00	12.71	6.79	10.58	5.58	6.90	3.68	94.40	0.0551
Inst-CEG* ( Li et al., 2024 )	18.41	10.14	38.14	23.57	53.88	36.87	2.48	1.25	16.53	9.01	21.88	12.28	13.70	7.36	23.57	14.35	94.49	0.2139

Appendix

Table 15. Performance comparison on the SC-SCD dataset. The best and the second-best results are, respectively, marked in BOLD and in underline . Except for Kappa, all results are expressed in percentage (%).

Method	Building		Tree		Water		Vegetation		N.v.g Surface		Playground		Average
Method	F1 c	IoU c	F1 c	IoU c	F1 c	IoU c	F1 c	IoU c	F1 c	IoU c	F1 c	IoU c	mF1 c	mIoU c	mOA	mKappa
UCD-SCM ( Tan et al., 2024 )	36.15	22.06	6.50	3.36	2.20	1.11	18.77	10.35	22.83	12.88	1.31	0.66	14.63	8.40	78.61	0.0795
AnyChange ( Zheng et al., 2024 )	42.79	27.22	7.49	3.89	2.19	1.11	24.82	14.17	38.90	24.15	1.03	0.52	19.54	11.84	76.07	0.1252
AnyChange* ( Zheng et al., 2024 )	58.58	41.42	3.93	2.01	0.97	0.49	21.46	12.02	41.69	26.33	1.17	0.59	21.30	13.81	77.97	0.1492
Inst-CEG ( Li et al., 2024 )	38.27	23.66	17.59	9.64	18.24	10.03	1.14	0.57	0.00	0.00	31.70	18.84	17.82	10.46	92.79	0.1604
Inst-CEG* ( Li et al., 2024 )	63.61	46.63	28.01	16.28	22.67	12.78	25.61	14.69	14.12	7.60	22.08	12.41	29.35	18.40	93.62	0.2670

Appendix

Table 16. Performance comparison on the SECOND dataset. The best and the second-best results are, respectively, marked in BOLD and in underline . Except for Kappa, all results are expressed in percentage (%).

Figure 9. Overall of Seg2Change. The bi-temporal remote sensing images are first fed into the category-agnostic change head (CACH) to generate a category-agnostic change map. In parallel, the same images, along with textual descriptions, are fed into an OVSS model to generate semantic segmentation maps. The final change map is derived by combining them.

Dataset	Study Site	Number of Image Pairs	Image Size**	Resolution	Evaluation Task
WHU-CD ( Ji et al., 2019 )	Christchurch, New Zealand	660 (Test)	$512\times 512$	0.3 m	Building CD
LEVIR-CD ( Chen and Shi, 2020 )	Texas, U.S.	512 (Test)	$512\times 512$	0.5 m	Building CD
DSIFN ( Zhang et al., 2020 )	Xi’an, China	48 (Test)	$512\times 512$	0.03–1 m	Land-cover CD
CLCD ( Liu et al., 2022 )	Guangdong, China	120 (Test)	$512\times 512$	0.5–2 m	Land-cover CD
SECOND ( Yang et al., 2020 )	Hangzhou, Chengdu, Shanghai, China	1694 (Test)	$512\times 512$	0.5–3 m	Semantic CD
SC-SCD ( Tan et al., 2025a )	Longwen, Zhangzhou, China	322 (Test)	$512\times 512$	0.5 m	Semantic CD

Appendix

Table 17. Information on the evaluation benchmark datasets and the CA-CDD coarse label source datasets. * denotes that the initially coarse, category-restricted change annotations are refined to produce category-agnostic change maps. ** indicates that all images have been preprocessed and cropped to a resolution of 512

\times

512.

Figure 10. Example images from binary change detection datasets (WHU-CD, LEVIR-CD, DSIFN, and CLCD) and semantic change detection datasets (SC-SCD and SECOND). The binary change detection datasets contain only binary change maps for specific categories. In contrast, the semantic change detection datasets include both binary change maps indicating changed regions between bi-temporal images and multi-class change maps that capture category transitions across time.

Figure 11. Visual comparison between CA-CDD category-agnostic change labels and the original coarse labels. We have improved the limited category range of the original labels. At the same time, we have refined the coarse-grained range annotations in the labels to fine-grained annotations. The gray markings in the figure represent the refinements we made to the original labels.

Figure 12. Qualitative comparison of several representative UCD and OVCD methods on three binary change detection datasets, i.e. , WHU-CD, LEVIR-CD, and DSIFN. White represents a true positive, black is a true negative, green indicates a false positive, and red is a false negative. Fewer green and red pixels represent better performance.

Figure 13. Qualitative comparison of several representative UCD and OVCD methods on CLCD.

Figure 14. Qualitative comparison of the state-of-the-art OVCD method DynamicEarth on a semantic change detection dataset, i.e. , SC-SCD. Black represents non-change. Color rendering: "Bareland" , "Building" , "Farmland" , "Road" , "Structure" , "Vegetation" , "Water" .

Figure 15. Qualitative comparison of the state-of-the-art OVCD method DynamicEarth on a semantic change detection dataset, i.e. , SECOND. Black represents non-change. Color rendering: "Building" , "Vegetation" , "Bareland" , "Playground" , "Tree" , "Water" .

Remote Sensing Change Detection

Remote sensing change detection finds and interprets differences across Earth-observation images acquired at different times.

See all 89 papers

ReA-OVCD: Training-Free Open-Vocabulary Change Detection via Semantic-Spatial Reliability Assessment
Jun 18, 2026Hongming Zhu, Huaji Chen, Bowen Du +2Remote Sensing Change Detection Semantic Change Detection
School of Computer Science and Technology, Tongji University, Shanghai 200092, China · College of Surveying and Geo-Informatics, Tongji University, Shanghai 200092, China
CoRegOVCD: Consistency-Regularized Open-Vocabulary Change Detection
Apr 2, 2026Weidong Tang, Hanbin Sun, Zihan Li +2Remote Sensing Change Detection Semantic Change Detection
China Agricultural University, College of Information and Electrical Engineering, Beijing, China · China Agricultural University, College of Science, Beijing, China
OpenDPR: Open-Vocabulary Change Detection via Vision-Centric Diffusion-Guided Prototype Retrieval for Remote Sensing Imagery
Mar 29, 2026Qi Guo, Jue Wang, Yinhe Liu +1Remote Sensing Change Detection
Wuhan University · Beijing Institute of Technology

Jun 18, 2026J/K move · Enter open · S save

ReA-OVCD: Training-Free Open-Vocabulary Change Detection via Semantic-Spatial Reliability Assessment

Hongming Zhu, Huaji Chen, Bowen Du, Sicong Liu+1

School of Computer Science and Technology, Tongji University, Shanghai 200092, China · College of Surveying and Geo-Informatics, Tongji University, Shanghai 200092, China

Remote Sensing Change Detection Semantic Change Detection

Open paper AskPDF arXiv

Unlike traditional remote sensing change detection that relies on predefined categories, Open-Vocabulary Change Detection (OVCD) identifies land cover changes flexibly using arbitrary text prompts. However, most existing OVCD methods rely on instance-level matching for stable correspondence but may overlook fine-grained variations (e.g., partial building extensions). Dense pixel-level comparison is more flexible, yet direct semantic comparison often produces unreliable candidate changes due to semantic ambiguity and spatial inconsistency. To this end, we propose ReA-OVCD, an efficient training-free framework that revisits pixel-level OVCD from a reliability assessment perspective. It first derives candidate change regions from pixel-wise semantic discrepancies to retain flexible localization. Instead of directly trusting these candidates, ReA-OVCD applies a two-stage semantic-spatial reliability assessment. The semantic stage evaluates whether a label discrepancy is supported by meaningful distributional and response-level changes, while the spatial stage validates whether a candidate region contains stable interior evidence rather than only boundary-induced responses. Extensive experiments across LEVIR-CD, WHU-CD, DSIFN, and SECOND show that the proposed framework improves the reliability of pixel-level OVCD and consistently outperforms state-of-the-art approaches, achieving

\mathrm{F}_{1}^{C}

improvements of 3.54% to 8.45% while maintaining superior computational efficiency. The code is available at https://github.com/Funny0101/ReA-OVCD.

Abstract

Remote Sensing Change Detection

Explore similar work

ReA-OVCD: Training-Free Open-Vocabulary Change Detection via Semantic-Spatial Reliability Assessment

CoRegOVCD: Consistency-Regularized Open-Vocabulary Change Detection

OpenDPR: Open-Vocabulary Change Detection via Vision-Centric Diffusion-Guided Prototype Retrieval for Remote Sensing Imagery