Research map / Label-efficient and robust learning / Recognition and long-tail
Feature Spectrum Learning for Remote Sensing Change Detection
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 FeaSpect, which embeds a trainable feature-space spectrum transformation (with learnable masks and strides) into siamese change-detection networks to suppress pseudo-changes caused by imaging conditions.
Problem
Pseudo-changes between bitemporal remote sensing images caused by imaging environment variations (sun elevation, seasons) hurt change detection; prior GAN-based style alignment is hard to train and distorts images, and fixed-box spectrum transformation has limited space and inefficient parameter sea
Approach
Moves Fourier spectrum transformation from image space into the shallow feature space of a siamese network and makes the low-frequency extraction box trainable, via adaptive attention over mask values (EB-AA) and learnable horizontal/vertical strides (EB-LS), so style alignment is learned end-to-end by backpropagation.
Key results
- On SVCD, FeaSpect with ResNet-18 backbone reaches 96.08% F1 and 92.46% IoU, versus 92.83% F1 for ESCNet and 94.75% F1 for ChangeSTAR (ResNet-101).
- FeaSpect with MiT-B2 reaches 97.27% F1 on SVCD and 87.09% F1 on DSIFN, outperforming ChangeFormer (86.47% F1 on DSIFN).
- Plugging FST into existing siamese methods such as STANet, DSAMNet and ChangeSTAR raises SVCD F1 by 3.25-5.53 points without adding parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FeaSpect instead of the strongest listed baseline | SVCD change detection · SVCD | F1(%): 94.75 -> 97.33 | Helped |
| Use FeaSpect instead of the strongest listed baseline | DSIFN change detection · DSIFN | F1(%): 86.47 -> 87.09 | Helped |
| Add FST with fixed content and stride (ablation) | FC-Siam-diff-res · SVCD | F1(%): 86.02 -> 93.02 | Helped |
| Add adaptive attention to FST (ablation) | FeaSpect-fixed · SVCD | F1(%): 93.02 -> 94.74 | Helped |
| Add learnable strides to FST (ablation) | FeaSpect-fixed · SVCD | F1(%): 93.02 -> 95.33 | Helped |
| Combine adaptive attention and learnable strides in FST (ablation) | FC-Siam-diff-res · SVCD | F1(%): 86.02 -> 95.81 | Helped |
| Share horizontal and vertical strides during learning (ablation) | FeaSpect · SVCD | F1(%): 95.67 -> 94.56 | Hurt |
| Apply FST to shallow layers L1 and L2 (ablation) | FC-Siam-diff-res · SVCD | F1(%): 86.0 -> 95.8(+9.8) | Helped |
| Apply FST only to deeper layers L3 and L4 (ablation) | FC-Siam-diff-res · SVCD | F1(%): 86.0 -> 87.1(+1.1) | Helped |
| Add FST to deeper layers after applying it to shallow layers (ablation) | FC-Siam-diff-res with FST at L1 and L2 · SVCD | F1(%): 95.8 -> 95.7 | No clear effect |
| Initialize learnable strides at different values and optimize them (ablation) | FeaSpect · SVCD | F1(%): 95.71 | No clear effect |
| Add FeaSpect to existing change-detection methods | Existing ConvNet- and transformer-based methods · SVCD | F1(%): +13.29; +15.30; +8.85; +5.53; +3.96; +4.42; +3.25; +2.92; +3.38; +3.53 | Helped |
Surprising findings
- Learned strides converge to the same accuracy from different initializations but to very different stride configurations, suggesting many equivalent optima.
- Sharing the learnable stride across horizontal and vertical dimensions reduces F1 (95.74 to 94.00 under one initialization), indicating anisotropic style regions.
Limitations
- Gains on DSIFN are smaller for some backbones and cross-dataset generality is only partially explored.
- Applying FST to deeper layers did not add improvement, so the method is restricted to shallow layers.
Experimental setup
Datasets: SVCD DSIFN
Benchmarks: SVCD DSIFN
Metrics: precision recall F1 IoU
Baselines: FC-EF FC-Siam-conc FC-Siam-diff-res FCN-PP W-Net CDGAN STANet DSAMNet ESCNet SEIFNet ChangeSTAR BIT
Builds on: FDA (Fourier domain adaptation) siamese change detection IBN-Net FC-Siam-diff
Closest papers by meaning
- The Change You Want To Detect: Semantic Change Detection In Earth Observation With Hybrid Data Generationf (CVPR 2025 · similarity 0.86)
- FM-ChangeNet: Learning Change through Pathwise Feature Transport (NeurIPS 2026 · similarity 0.84)
- RSRCC: A Remote Sensing Regional Change Comprehension Benchmark Constructed via Retrieval-Augmented Best-of-𝑁 Ranking (NeurIPS 2026 · similarity 0.83)
- Towards Generalizable Scene Change Detection (CVPR 2025 · similarity 0.83)
- A General Adaptive Dual-level Weighting Mechanism for Remote Sensing Pansharpening (CVPR 2025 · similarity 0.83)
- Learning De-Biased Representations for Remote-Sensing Imagery (NeurIPS 2024 · similarity 0.82)
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.