Research map / Detection and segmentation / CNN features
Dust-Mamba: An Efficient Dust Storm Detection Network with Multiple Data Sources
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 Dust-Mamba, a UNet++-based network fusing satellite and meteorological data with CBAM-enhanced VSS blocks and structure-sharing transfer learning for dust storm occurrence and intensity detection.
Problem
Dust storm detection is limited by occurrence-only tasks, lack of multi-source data integration, and CNN difficulty capturing long-range dependencies; intensity detection is understudied.
Approach
Combines a VMamba-style VSS block in parallel with CBAM attention inside a UNet++ backbone, uses multi-resolution data fusion for satellite and ERA5/MERRA-2 meteorological inputs, and transfers occurrence-model weights to intensity detection.
Key results
- Dust-Mamba reaches Dice 0.963 for occurrence detection vs 0.913 for the best baseline Unet++.
- Dust-Mamba reaches Dice 0.560 for intensity detection vs 0.458 for Unet++ (7-level intensity).
- Structure-sharing transfer (fine-tuning the entire model from the occurrence model) reaches the from-scratch performance in roughly 1/3 of the training time.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add MRDF to Unet++ (ablation) | Occurrence detection · LSDSSIMR test | Dice: 0.9132 -> 0.9286 | Helped |
| Add MRDF to Unet++ (ablation) | Intensity detection · LSDSSIMR test | Dice: 0.4581 -> 0.4735 | Helped |
| Add VSS BLOCK (ablation) | Occurrence detection with MRDF · LSDSSIMR test | Dice: 0.9286 -> 0.9562 | Helped |
| Add VSS BLOCK (ablation) | Intensity detection with MRDF · LSDSSIMR test | Dice: 0.4735 -> 0.5099 | Helped |
| Add CBAM to VSS BLOCK (ablation) | Occurrence detection with MRDF and VSS BLOCK · LSDSSIMR test | Dice: 0.9562 -> 0.9637 | Helped |
| Add CBAM to VSS BLOCK (ablation) | Intensity detection with MRDF and VSS BLOCK · LSDSSIMR test | Dice: 0.5099 -> 0.5158 | Helped |
| Initialize from the occurrence model and train only the final output layer (ablation) | Dust-Mamba intensity detection · LSDSSIMR test | Dice: 0.5158 -> 0.2743 | Hurt |
| Initialize from the occurrence model and train the entire model (ablation) | Dust-Mamba intensity detection · LSDSSIMR test | Dice: 0.5158 -> 0.5604 | Helped |
| Jointly train occurrence and intensity models with separate outputs (ablation) | Dust-Mamba intensity detection · LSDSSIMR test | Dice: 0.5158 -> 0.5529 | Helped |
| Concatenate satellite and meteorological data at the input instead of using MRDF | Segmentation baselines · LSDSSIMR test, occurrence detection | Dice: 0.9134 -> 0.9132; 0.8945 -> 0.8806; 0.8190 -> 0.6561; 0.8257 -> 0.7847 | Mixed |
| Use Dust-Mamba instead of Unet++ | Occurrence detection · LSDSSIMR test | Dice: 0.9132 -> 0.9637 | Helped |
| Use Dust-Mamba initialized from the occurrence model and train the entire model instead of Unet++ | Intensity detection · LSDSSIMR test | Dice: 0.4581 -> 0.5604 | Helped |
Surprising findings
- Simply concatenating meteorological data with satellite data did not improve and sometimes degraded baseline performance.
- Training only the final output layer during transfer performed poorly (Dice 0.274), worse than training from scratch.
Limitations
- Evaluated on a single dataset (LSDSSIMR) covering one region and spring months.
- Intensity levels 1 and 6 have limited training data and lower performance.
Experimental setup
Datasets: LSDSSIMR FY-4A satellite data ERA5 MERRA-2
Benchmarks: LSDSSIMR
Metrics: Dice IoU Recall Precision Kappa
Baselines: U-net Unet++ Attention U-net TransUNet Swin-UNet VM-UNet
Builds on: Mamba VMamba VM-UNet UNet++ CBAM
Closest papers by meaning
- MambaAD: Exploring State Space Models for Multi-class Unsupervised Anomaly Detection (NeurIPS 2024 · similarity 0.79)
- MambaOut: Do We Really Need Mamba for Vision? (CVPR 2025 · similarity 0.79)
- RoMA: Scaling up Mamba-based Foundation Models for Remote Sensing (NeurIPS 2025 · similarity 0.78)
- Point Cloud Mamba: Point Cloud Learning via State Space Model (AAAI 2025 · similarity 0.77)
- AirRadar: Inferring Nationwide Air Quality in China with Deep Neural Networks (AAAI 2025 · similarity 0.77)
- IDOL: Meeting Diverse Distribution Shifts with Prior Physics for Tropical Cyclone Multi-Task Estimation (NeurIPS 2025 · similarity 0.77)
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.