Research map / Image generation / Fast diffusion sampling
MDFL: Multi-Domain Diffusion-Driven Feature Learning
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 MDFL, a diffusion-driven multi-domain (spectral, spatial, frequency) feature learning network with a feature reuse module, evaluated for multimodal HSI+LiDAR/DSM land-cover classification on Houston2013, Trento and MUUFL.
Problem
Classifying high-dimensional multimodal remote sensing images (HSI with LiDAR/DSM) is limited by single-domain (spatial) feature learning and by fusion approaches that treat multimodal fusion as a black box, which ignores spectral, spatial and frequency information.
Approach
Uses a U-Net diffusion model trained to predict multi-noise-level inputs (t=0,50,100,200,400) with spectral-spatial attention and a frequency-domain learnable filter, plus a feature reuse module that fuses shallow and deep features across the two modalities before an MLP classifier.
Key results
- MDFL reports OA 99.16% on Houston2013, 99.43% on Trento and 96.96% on MUUFL, versus 94.34% best baseline OA (MFT) on MUUFL.
- Ablation: removing frequency-domain analysis drops average OA from 98.51% to 97.75%.
- Abstract claims average OA of 98.25% across three datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MDFL instead of the strongest listed baseline, MFT | Remote-sensing multimodal classification · Houston2013 | OA (%): 89.80 -> 99.16 | Helped |
| Use MDFL instead of the strongest listed baseline, Cross | Remote-sensing multimodal classification · Houston2013 | AA (%): 92.70 -> 96.72 | Helped |
| Use MDFL instead of the strongest listed baseline, CNN-2D | Remote-sensing multimodal classification · Houston2013 | κ (×100): 91.65 -> 94.09 | Helped |
| Use MDFL instead of CNN-2D | Remote-sensing multimodal classification · Trento | OA (%): 98.65 -> 99.43 | Helped |
| Use MDFL instead of Cross | Remote-sensing multimodal classification · Trento | AA (%): 97.35 -> 96.13 | Hurt |
| Use MDFL instead of CNN-2D | Remote-sensing multimodal classification · Trento | κ (×100): 98.19 -> 94.41 | Hurt |
| Use MDFL instead of MFT | Remote-sensing multimodal classification · Muufl | OA (%): 94.34 -> 96.96 | Helped |
| Use MDFL instead of MFT | Remote-sensing multimodal classification · Muufl | AA (%): 81.48 -> 82.63 | Helped |
| Use MDFL instead of MFT | Remote-sensing multimodal classification · Muufl | κ (×100): 92.51 -> 94.63 | Helped |
| Add joint frequency-domain analysis (ablation) | MDFL · Three-dataset average | OA (%): 97.75 -> 98.51 | Helped |
| Add original information to feature reuse (ablation) | MDFL · Three-dataset average | OA (%): 98.10 -> 98.51 | Helped |
| Fuse multiple noise time steps rather than use single-step training (ablation) | MDFL · Houston2013, Trento, and Muufl | classification performance | Helped |
Limitations
- Paper states that traditional methods perform better than MDFL when combining HSI and LiDAR on some settings, and the write-up is unclear on several design details.
- Ablation and comparison numbers are inconsistent across the text (e.g. MDFL OA reported as 98.51% in the ablation vs 98.25% in the abstract).
Experimental setup
Datasets: Houston2013 Trento MUUFL
Benchmarks: Houston2013 Trento MUUFL
Metrics: overall accuracy average accuracy kappa coefficient
Baselines: RNN Cross CNN-2D CALC ViT MFT
Builds on: denoising diffusion probabilistic models U-Net multi-head self-attention
Closest papers by meaning
- Diff-2-in-1: Bridging Generation and Dense Perception with Diffusion Models (ICLR 2025 · similarity 0.86)
- SatSynth: Augmenting Image-Mask Pairs through Diffusion Models for Aerial Semantic Segmentation (CVPR 2024 · similarity 0.85)
- Exploiting Diffusion Prior for Real-World Image Dehazing with Unpaired Training (AAAI 2025 · similarity 0.85)
- Can Diffusion Models Learn Hidden Inter-Feature Rules Behind Images? (ICML 2025 · similarity 0.85)
- U-Know-DiffPAN: An Uncertainty-aware Knowledge Distillation Diffusion Framework with Details Enhancement for PAN-Sharpening (CVPR 2025 · similarity 0.85)
- LGDM: Latent Guidance in Diffusion Models for Perceptual Evaluations (ICML 2025 · similarity 0.85)
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.