This repository contains the official implementation of MMFNet, a novel Multi-Modal Foundation Model designed for generalizable flood mapping using Sentinel-1 (Synthetic Aperture Radar - SAR) and Sentinel-2 (Multi-Spectral Imagery - MSI) remote sensing data.
Note: This work is currently under review and the code is not publicly available at this time. We appreciate your understanding.
| FT Label | FT Inputs | Direct Download | Baidu Netdisk |
|---|---|---|---|
| MSI | MSI | Download (1.46 GB) | Download |
| MSI | MSI+SAR | Download (1.79 GB) | - |
| SAR | SAR | Download (1.42 GB) | - |
| SAR | MSI+SAR | Download (1.79 GB) | - |
| MSI+SAR | MSI+SAR | Download (1.79 GB) | - |
MMFNet addresses the critical need for robust and adaptable flood mapping solutions by leveraging the complementary information from both SAR and MSI data. Our model is built upon a multi-phase framework:
- Phase I: Pre-training on a large, unlabeled multi-modal dataset (SSL4EO-S12) using a self-supervised learning approach to learn rich, generalized representations from both SAR and MSI data.
- Phase II: Fine-tuning on a labeled flood dataset (Sen1Floods11) to adapt the pre-trained model for the specific task of flood mapping. MMFNet is capable of handling mono-modal (SAR or MSI) and multi-modal (SAR + MSI) inputs during this phase.
- Phase III: Application to real-world scenarios, demonstrating MMFNet's strong generalization capabilities across diverse flood events.
- Multi-Modal Handling: MMFNet is designed to seamlessly process and fuse information from both Sentinel-1 SAR and Sentinel-2 MSI data, allowing for more comprehensive flood detection.
- Global-Scale Pre-training: The foundation model is pre-trained on a vast, unlabeled dataset, enhancing its generalization ability and label efficiency for downstream tasks.
- Generalizable Flood Mapping: Our experiments demonstrate robust performance in various and practical flood scenarios, highlighting MMFNet's ability to generalize to unseen flood events and geographical regions.
- Foundation Model Approach: By learning generalizable representations through pre-training, MMFNet reduces the need for extensive labeled data for specific flood mapping tasks.
If you find this work relevant to your research, please consider citing our paper (once published). The current draft is under review.
For any inquiries regarding this work, please contact:
We will update this repository with the code and pre-trained models upon acceptance and publication of our paper. Thank you for your patience and interest!