Research map / Detection and segmentation / CNN features
Interpretation-Oriented Cloud Removal via Observation-Anchored Residual Flow with Geo-Contextual Alignment
Ziyao Wang, Maonan Wang, Yucheng He, Xianping Ma, Ziyi Wang, Hongyang Zhang, Yirong Chen, Man On Pun
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace MRDM diffusion with OAR-Flow on HDiT (ablation) | HDiT backbone · CUHKCR-EXT-CS | PSNR (dB): 23.736 -> 23.986 | Improved |
| replace MRDM diffusion with OAR-Flow (ablation) | HDiT backbone · CUHKCR-EXT-CS | SSIM: 0.712 -> 0.698 | Hurt |
| add GCPA (GCI loss) to OAR-Flow (ablation) | HDiT + OAR-Flow · CUHKCR-EXT-CS | PSNR (dB): 23.986 -> 24.230 | Improved |
| add GCPA (GCI loss) to OAR-Flow (ablation) | HDiT + OAR-Flow · CUHKCR-EXT-CS | CLS accuracy: 0.690 -> 0.781 | Improved |
| use DiT instead of HDiT with MRDM (ablation) | MRDM diffusion · CUHKCR-EXT-CS | GFLOPs: 166.72 -> 697.20 | Hurt |
| reduce patch size from 2 to 1 (ablation) | GACR · Vaihingen-CR-thick | PSNR (dB): 33.018 -> 33.799 PSNR; SEG 0.737 -> 0.720; GFLOPs 56.05 -> 223.60 | Mixed |
| increase patch size from 2 to 4 (ablation) | GACR · Vaihingen-CR-thick | PSNR (dB): 33.018 -> 30.303 | Hurt |
| vary GCI loss weight lambda away from 0.5 (ablation) | GACR · Vaihingen-CR-thick | SEG mIoU: 0.737 (lambda 0.5) -> 0.728 (0.2), 0.727 (5.0) | Hurt |
| observation-anchored flow with GCPA vs EMRDM diffusion | GACR · Vaihingen-CR-thick | PSNR (dB): 28.979 -> 33.018 | Improved |
| apply cloud removal before downstream classification instead of training end-to-end on cloudy images | downstream CLS · CLS-1 (DINOv3-LVD encoder) | accuracy: 0.833 -> 0.833 | No clear effect |
| apply GACR cloud removal before segmentation | downstream SEG · SEG-4 (DINOv3-LVD encoder) | mIoU: 0.684 -> 0.737 | Improved |
| replace DINOv3 LVD downstream weights with SAT weights | downstream evaluation · CLS-1 (EMRDM-processed) | accuracy: 0.776 -> 0.642 | Hurt |
Closest papers by meaning
- GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation (NeurIPS 2026 · similarity 0.85)
- SkySense-O: Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling (CVPR 2025 · similarity 0.84)
- Learnable Earth Parser: Discovering 3D Prototypes in Aerial Scans (CVPR 2024 · similarity 0.84)
- Towards Faithful Reasoning in Remote Sensing: A Perceptually-Grounded GeoSpatial Chain-of-Thought for Vision-Language Models (ICLR 2026 · similarity 0.84)
- SkyMoE: A Vision-Language Foundation Model for Enhancing Geospatial Interpretation with Mixture of Experts (AAAI 2026 · similarity 0.83)
- SemStereo: Semantic-Constrained Stereo Matching Network for Remote Sensing (AAAI 2025 · similarity 0.83)
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.
