Research map / Detection and segmentation / Open-set and few-shot detection
The Change You Want To Detect: Semantic Change Detection In Earth Observation With Hybrid Data Generationf
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HySCDG, a Stable Diffusion + ControlNet pipeline that inpaints semantically selected objects in FLAIR VHR images to build the hybrid semantic change dataset FSC-180k, and shows pretraining on it beats SyntheWorld across binary and semantic change detection transfer settings.
Problem
Bi-temporal semantic change detection on very high resolution (VHR) Earth observation imagery lacks large, diverse, annotated training data, and existing synthetic change datasets are either fully synthetic with limited realism or hybrid but narrow in scope.
Approach
Fine-tune Stable Diffusion 2 inpainting (including the VAE) with a ControlNet conditioned on land-cover maps and geographic prompts, then randomly select instances in a single-date land-cover map, relabel them, and inpaint to create realistic bi-temporal semantic change pairs with change maps.
Key results
- Pretraining the Dual UNet on FSC-180k improves SECOND semantic SCS from 0.83 (no pretraining) to 0.89 and HiUCD-mini SCS from 0.73 to 0.78 in sequential training.
- FSC-180k pretraining outperforms SyntheWorld in every sequential binary setting, with gains of 2-3% IoU across five target datasets.
- In zero-shot evaluation, FSC-180k pretraining reaches IoU 0.37 on SECOND and 0.36 on HiUCD-mini, but only 0.33 on LEVIR-CD versus 0.83 for fully supervised.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Pretrain Dual UNet on FSC-180k instead of no pretraining | Dual UNet · SECOND | SCS: 0.83 -> 0.89 | Helped |
| Pretrain Dual UNet on FSC-180k instead of SyntheWorld | Dual UNet · SECOND | SCS: 0.84 -> 0.89 | Helped |
| Pretrain Dual UNet on FSC-180k instead of no pretraining | Dual UNet · HiUCD-mini | SCS: 0.73 -> 0.78 | Helped |
| Pretrain Dual UNet on FSC-180k instead of SyntheWorld | Dual UNet · HiUCD-mini | SCS: 0.73 -> 0.78 | Helped |
| Pretrain SCanNet on FSC-180k instead of no pretraining | SCanNet · SECOND | SCS: 0.84 -> 0.89 | Helped |
| Pretrain SCanNet on FSC-180k instead of no pretraining | SCanNet · HiUCD-mini | SCS: 0.66 -> 0.71 | Helped |
| Pretrain on FSC-180k instead of no pretraining | HiUCD-XL model · HiUCD-XL | Chg. mIoU: 0.17 -> 0.19 | Helped |
| Pretrain on FSC-180k instead of SyntheWorld | HiUCD-XL model · HiUCD-XL | mIoU: 0.58 -> 0.60 | Helped |
| Use Dual UNet rather than SCanNet after FSC-180k pretraining | Change detection model · HiUCD-mini | SCS: 0.71 -> 0.78 | Helped |
| Use SCanNet rather than Dual UNet after FSC-180k pretraining | Change detection model · SECOND | SCS: 0.89 -> 0.89 | No clear effect |
| Pretrain on FSC-180k rather than SyntheWorld | Change detection model · Binary change detection target datasets | IoU: +2-3% of IoU | Helped |
| Increase the target-data share in mixed training from 50% toward 90% | Mixed training with FSC-180k · Levir-CD and HiUCD-mini | performance | Mixed |
Surprising findings
- Optimal mixed training ratio is around 50% target data for several datasets rather than always increasing with target data, contrary to SyntheWorld conclusions.
- The authors could not reproduce SCanNet's reported results from the original paper.
Limitations
- ControlNet may fail to respect the semantic map (about 20% error rate between segmentations of real and generated images).
- Zero-shot transfer is weak on LEVIR-CD (0.33 vs 0.83 IoU) due to limited realism and sparse changes in FSC-180k.
- Mixed-training optimal ratio varies inconsistently across datasets (around 50% vs 90%).
Experimental setup
Datasets: FLAIR FSC-180k LEVIR-CD S2Looking SECOND HiUCD-mini HiUCD-XL SyntheWorld OpenStreetMap
Benchmarks: LEVIR-CD S2Looking SECOND HiUCD-mini HiUCD-XL
Metrics: IoU F1 SCS SeK Chg. mIoU mIoU Inception Score FID
Baselines: SyntheWorld Dual UNet without pretraining SCanNet ChangeFormer MambaCD
Builds on: Stable Diffusion ControlNet Changen2 SyntheWorld
Closest papers by meaning
- UniChange: Unifying Change Detection with Multimodal Large Language Model (CVPR 2026 · similarity 0.88)
- The STVchrono Dataset: Towards Continuous Change Recognition in Time (CVPR 2024 · similarity 0.87)
- RSRCC: A Remote Sensing Regional Change Comprehension Benchmark Constructed via Retrieval-Augmented Best-of-𝑁 Ranking (NeurIPS 2026 · similarity 0.87)
- ChangeDiff: A Multi-Temporal Change Detection Data Generator with Flexible Text Prompts via Diffusion Model (AAAI 2025 · similarity 0.87)
- TEOChat: A Large Vision-Language Assistant for Temporal Earth Observation Data (ICLR 2025 · similarity 0.87)
- Towards Generalizable Scene Change Detection (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.