Real-World Knowledge-Guided Change Data Synthesis for Remote Sensing

Yaoyi Qi; Xingxing Weng; Chao Pang; Yongkang Cui; Xiangyu Hao; Xiaokang Zhang; Guibo Zhu; Gui-Song Xia

cs.AIAug 25, 2026

Real-World Knowledge-Guided Change Data Synthesis for Remote Sensing

Authors: Yaoyi Qi, Xingxing Weng, Chao Pang, Yongkang Cui, Xiangyu Hao, Xiaokang Zhang, Guibo Zhu, Gui-Song Xia

Organizations: School of Artificial Intelligence, Wuhan University · Wuhan AI Research · Institute of Automation, University of Chinese Academy of Sciences · Institute for Math & AI, Wuhan

Abstract

Change data synthesis provides a cost-effective solution for expanding training data and improving the performance of change detection models. However, existing synthesis methods typically rely on handcrafted rules to simulate changes, where limited coverage of class transitions restricts the diversity of synthesized data, while predefined transition designs limit their flexibility in accommodating varied change types. In this work, we introduce KnowChange, a knowledge-guided change data synthesis framework that leverages pretrained vision-language models as knowledge sources to reason about plausible change locations and class transitions from pre-change scenes and desired change types. By integrating knowledge-guided change simulation with generalizable synthesis models, KnowChange enables flexible synthesis of diverse change types within a unified framework. Extensive experiments demonstrate that KnowChange-generated data consistently outperforms existing synthetic datasets in both synthetic-to-real transfer and synthetic data augmentation, despite being generated at a compact scale. Further analyses show that the knowledge-guided change simulation can be seamlessly integrated into existing synthesis pipelines and enhance the downstream utility of synthesized data.

Figures & tables

Figure 1: Top : Existing synthesis methods rely on handcrafted rules to simulate changes. Limited coverage of class transitions restricts the diversity of synthesized data, while predefined transition designs limit flexibility in accommodating varied change types. Bottom : KnowChange leverages pretrained vision-language models as knowledge sources to reason about plausible change regions and class transitions given pre-change scene contexts and user-specified change types, enabling flexible synthesis of diverse change data.

Figure 2: Downstream performance comparison of models trained with change data synthesized by different methods.

Figure 3: Overview of KnowChange, a knowledge-guided framework for change data synthesis. Given a pre-change image $I_{\mathrm{pre}}$ , its semantic mask $S_{\mathrm{pre}}$ , and a textual prompt $P$ , Knowledge-Guided Change Simulation reasons about actual and pseudo changes to construct the change layout $\mathcal{G}_{\mathrm{chg}}$ . Generalizable Semantic-Guided Synthesis generates $\mathcal{M}_{\mathrm{sa}}$ , combines it with $\mathcal{M}_{\mathrm{sp}}$ to obtain $S_{\mathrm{post}}$ and $\mathcal{M}$ , and uses them to guide the synthesis of the post-change image $I_{\mathrm{post}}$ .

Figure 4: Examples from Know-BCD, Know-SEC, and Know-HR (top to bottom).

Dataset	Resolution (m)	Bands	# Classes	# Samples
OpenEarthMap [ 33 ]	0.25 $\sim$ 0.5	RGB	8	11,662
FLAIR [ 12 ]	0.2	RGB + NIR	19	61,712
Vaihingen [ 23 ]	0.09	IR-R-G	6	864
Potsdam [ 23 ]	0.05	RGB + IR	6	7,406
GID [ 32 ]	1	RGB + NIR	15	23,100
SkySA [ 45 ]	Variable	RGB	1,763	33,775

Table 1: Summary of remote sensing datasets used for training the layout-to-mask and mask-to-image models.

Dataset	Source Task	# Samples	LEVIR-CD		WHU-CD		DSIFN-CD		SEC-BCD		Average
Dataset	Source Task	# Samples	IoU	F1	IoU	F1	IoU	F1	IoU	F1	IoU	F1
WHU-GCD	SCD	25K	1.26	2.48	1.67	3.30	2.91	5.66	5.36	10.19	2.80	5.40
Changen2-S9	SCD	27K	1.99	3.90	3.80	7.32	4.73	9.04	6.69	12.54	4.30	8.20
FSC-180k	SCD	180K	11.19	20.13	33.50	50.19	5.49	10.40	16.40	28.18	16.64	27.22
Know-SEC	SCD	10K	30.81	47.10	34.95	51.79	32.38	48.92	29.54	45.61	31.92	48.36
SyntheWorld	BCD	40K	28.66	44.55	23.11	37.54	5.08	9.66	14.23	24.91	17.77	29.17

Table 2: Synthetic-to-real transfer results on building change detection benchmarks. Average indicates the mean performance over four benchmarks. The best and second-best results are shown in bold and underline, respectively.

Dataset	#Samples	SECOND			HRSCD			Average			$\Delta$ F1 (vs. ours)
Dataset	#Samples	F1	mIoU	SCS	F1	mIoU	SCS	F1	mIoU	SCS	$\Delta$ F1 (vs. ours)
WHU-GCD	25K	15.32	19.85	18.06	41.71	39.10	36.12	28.52	29.48	27.09	-10.97
FSC-180k	180K	22.95	50.40	23.84	39.36	53.98	31.46	31.16	52.19	27.50	-8.33
Changen2-S9	27K	34.06	57.75	26.70	31.35	41.59	17.81	32.71	49.67	22.26	-6.78
Know-SEC/HR	10K	37.03	55.15	28.31	41.94	50.78	32.16	39.49	52.97	30.24	-

Table 3: Synthetic-to-real transfer results on semantic change detection benchmarks. Average indicates the mean performance over two benchmarks. The best and second-best results are shown in bold and underline, respectively.

Dataset	5% LEVIR-CD		5% WHU-CD		5% DSIFN-CD		5% SEC-BCD		Dataset	5% SECOND
Dataset	IoU	F1	IoU	F1	IoU	F1	IoU	F1	Dataset	F1	mIoU	SeK	SCS
FSC-180k	15.80	26.32	17.81	29.77	21.96	35.84	21.96	33.56	No Syn.	50.84	65.05	10.53	38.81
SyntheWorld	28.97	43.27	29.02	41.75	28.95	44.62	36.09	52.79	FSC-180k	41.34	60.00	4.79	26.95
Changen2-S1	32.61	48.01	36.62	51.54	17.49	29.74	42.56	58.66	Changen2-S9	52.37	66.21	11.85	39.93
Know-BCD	36.28	52.88	37.13	53.59	29.74	45.77	43.20	59.94	Know-SEC	55.27	67.24	13.89	39.87

Table 4: Performance comparison of synthetic data augmentation with 5% real training data from different benchmarks. For BCD tasks, models are evaluated on four BCD benchmarks, and results are reported as average IoU and F1. For SECOND, models are evaluated on the SECOND test set. The best and second-best results are shown in bold and underline, respectively.

Figure 5: FID and KID comparison of synthetic datasets.

Method	LEVIR-CD	WHU-CD	Average
Method	(IoU)	(IoU)	Average
Copy-Paste	4.10	3.82	3.96
SAT	33.95	41.20	37.58
+ SPT	35.33	40.31	37.82
+ PCS	44.07	44.62	44.35
+ VLM	44.81	47.87	46.34

Table 5: Ablation study of knowledge-guided change simulation. Generative models are fixed, and different components are progressively added to evaluate their contributions.

VLM Guidance	F1	mIoU	SCS	$\Delta$ F1 (w/o VLM)
w/o VLM	24.07	49.93	22.97	-
GLM-4.6V	29.84	53.16	25.90	5.77
DouBao-2.0-mini	31.72	53.66	25.98	7.65
Qwen3-VL	33.41	50.39	29.50	9.34

Table 6: Impact of VLMs on knowledge-guided change simulation evaluated on the SECOND benchmark.

Figure 6: Performance comparison of existing synthesis methods with and without our change simulation.

Panel of Figure 7 — Figure 7: Scaling analysis of synthetic training data. Left and middle: Know-BCD scaling results on binary change detection benchmarks. Right: relative improvements on SECOND under varying Know-SEC training data ratios.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Resolution (m)	# Samples	Image Size	Train/Val/Test	Task
LEVIR-CD	0.5	637	$1024\times 1024$	2,548/–/1,392	BCD
WHU-CD	0.075	1	$32507\times 15354$	1,260/–/690	BCD
DSIFN-CD	0.03–1	3,940	$512\times 512$	3,600/340/48	BCD
SECOND	0.5–3	4,662	$512\times 512$	2,968/–/1,694	BCD & SCD
HRSCD	0.5	291	$10000\times 10000$	905/116/123	SCD

Appendix

Table 7: Detailed statistics of the datasets used for binary change detection (BCD) and semantic change detection (SCD). Train/Val/Test denotes the number of samples after preprocessing.

Figure 11: Visualization of generated samples from our three synthetic datasets: Know-BCD, Know-SEC, and Know-HR. In each dataset, the first row uses source images from FLAIR and the second row uses source images from OpenEarthMap.

Synthetic Dataset	Base Dataset	#Samples	Task
WHU-GCD	LoveDA, Evlab-SS, LandCover.ai, Google Earth	25K	SCD
SyntheWorld	Fully Synthetic, OpenEarthMap	40K	BCD
FSC-180k	FLAIR	180K	SCD
Changen2-S1/S9	xView2/OpenEarthMap	15K/27K	BCD/SCD
Know-BCD/SEC/HR	OpenEarthMap, FLAIR	10K/10K/10K	BCD/SCD/SCD

Appendix

Table 8: Overview of synthetic change detection datasets. BCD and SCD denote binary and semantic change detection.

Figure 12: Examples of shape-altering transitions generated by the Layout-to-Mask model. The model fills masked regions with specified categories while maintaining shape plausibility and contextual consistency.

OpenEarthMap Categories	FLAIR Categories	Target (SECOND)	Target ID
Building	Building	Building	5
Agriculture land, Rangeland	Agri. veg., Plowed land, Vineyard, Grass, Brushwood	Low vegetation	1
Road, Developed space, Bareland	Road, Pervious, Bare soil, Cut, Mixed	Non-vegetated ground	2
Tree	Coniferous, Deciduous trees, Lignous	Tree	3
Water	Water, Swimming pool	Water	4
Background	Greenhouse, Snow	Non-change	0

Appendix

Table 9: Label mapping from the source datasets (OpenEarthMap and FLAIR) to the target categories of SECOND.

OpenEarthMap Categories	FLAIR Categories	Target (HRSCD)	Target ID
Building, Road, Developed space	Building, Road, Pervious surface	Artificial surface	1
Agriculture land Rangeland, Bareland	Agri. veg., Plowed land, Vineyard Grass, Brushwood, Bare soil, Cut, Mixed	Agricultural land	2
Tree	Coniferous, Deciduous trees, Lignous	Forest	3
–	Snow	Wetland	4
Water	Water, Swimming pool	Water	5
Background	Greenhouse, Unknown	No info	0

Appendix

Table 10: Label mapping from the source datasets (OpenEarthMap and FLAIR) to the target categories of HRSCD.

Figure 13: Overview of the statistical comparison between the real SECOND dataset and synthetic datasets regarding change proportions and category transition differences.

Figure 14: Impact of 4 $\times$ in-domain augmentation under different labeled data settings.

Figure 15: Visualization of multi-task capabilities using our framework. These samples are generated directly from semantic maps without pre-temporal image inputs ( $I_{\text{pre}}$ ). The visualization demonstrates applicability to downstream tasks including semantic segmentation, building extraction, road extraction, and water detection.

Model	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
M2I	22.3340	0.7042	0.1877
HySCDG	21.3064	0.6900	0.1960
Changen2	22.1237	0.6985	0.1968

Appendix

Table 11: Image reconstruction quality comparison on the SECOND dataset.

Figure 16: Long-term urban evolution synthesized from a single-temporal scene. KnowChange iteratively generates semantic maps (top) and their corresponding remote sensing images (bottom), producing multi-category changes while maintaining semantic and spatial consistency across time.

Synthetic Data Generation

Synthetic data generation creates artificial examples, simulations, and datasets to train, test, or protect machine-learning systems.

See all 625 papers

Grounding Synthetic Data Generation With Vision and Language Models
Mar 10, 2026Ümit Mert Çağlar, Alptekin TemizelSynthetic Data Augmentation Synthetic Data Evaluation
Graduate School of Informatics METU, Turkey
Make Some Noise: Unsupervised Remote Sensing Change Detection Using Latent Space Perturbations
Feb 23, 2026Blaž Rolih, Matic Fučka, Filip Wolf +1Synthetic Data Augmentation Remote Sensing Change Detection
University of Ljubljana, Faculty of Computer and Information Science, Slovenia
ChangeFlow -- Latent Rectified Flow for Change Detection in Remote Sensing
May 14, 2026Blaž Rolih, Matic Fučka, Filip Wolf +1Remote Sensing Change Detection
University of Ljubljana, Faculty of Computer and Information Science, Slovenia

Mar 10, 2026J/K move · Enter open · S save

Grounding Synthetic Data Generation With Vision and Language Models

Ümit Mert Çağlar, Alptekin Temizel

Graduate School of Informatics METU, Turkey

Synthetic Data Augmentation Synthetic Data Evaluation Remote Sensing Benchmarks Semantic Segmentation Image Captioning

Open paper AskPDF arXiv

Deep learning models benefit from increasing data diversity and volume, motivating synthetic data augmentation to improve existing datasets. However, existing evaluation metrics for synthetic data typically calculate latent feature similarity, which is difficult to interpret and does not always correlate with the contribution to downstream tasks. We propose a vision-language grounded framework for interpretable synthetic data augmentation and evaluation in remote sensing. Our approach combines generative models, semantic segmentation and image captioning with vision and language models. Based on this framework, we introduce ARAS400k: A large-scale Remote sensing dataset Augmented with Synthetic data for segmentation and captioning, containing 100k real images and 300k synthetic images, each paired with segmentation maps and descriptions. ARAS400k enables the automated evaluation of synthetic data by analyzing semantic composition, minimizing caption redundancy, and verifying cross-modal consistency between visual structures and language descriptions. Experimental results indicate that while models trained exclusively on synthetic data reach competitive performance levels, those trained with augmented data (a combination of real and synthetic images) consistently outperform real-data baselines. Consequently, this work establishes a scalable benchmark for remote sensing tasks, specifically in semantic segmentation and image captioning. The dataset is available at zenodo.org/records/18890661 and the code base at github.com/caglarmert/ARAS400k.

Abstract

Synthetic Data Generation

Explore similar work

Grounding Synthetic Data Generation With Vision and Language Models

Make Some Noise: Unsupervised Remote Sensing Change Detection Using Latent Space Perturbations

ChangeFlow -- Latent Rectified Flow for Change Detection in Remote Sensing