Deep Learning for Remote Sensing to Improve Flood Inundation Mapping

Technical Summary: Deep Learning for Remote Sensing to Improve Flood Inundation Mapping

Problem Statement
Flooding is a pervasive global natural disaster causing significant economic loss and human casualties. While optical satellite missions like the European Space Agency's Copernicus Sentinel-2 provide high-resolution, multispectral data essential for flood detection, their operational utility is severely compromised by cloud cover during extreme precipitation events. In tropical and subtropical regions, over 70% of optical imagery becomes unusable during flood seasons.

Conventional cloud-removal techniques, such as temporal compositing or interpolation, fail in flood monitoring because floodwaters are highly dynamic; images from days prior or post-event do not reflect the current inundation state. While Synthetic Aperture Radar (SAR) can penetrate clouds, it suffers from speckle noise and reduced interpretability in complex urban environments. Furthermore, existing deep learning solutions based on Generative Adversarial Networks (GANs) often suffer from training instability and mode collapse, leading to erroneous reconstructions (e.g., disconnecting river channels or misidentifying dry land) that lack the hydrological continuity required for emergency management.

Methodology
This study introduces a cloud-removal framework based on Denoising Diffusion Probabilistic Models (DDPMs), specifically leveraging the Masked Diffusion Transformer (MDT) architecture. The approach is designed to generate cloud-free multispectral realizations of Sentinel-2B imagery while preserving visual fidelity and hydrological consistency.

• Data: The model was trained on post-2017 Sentinel-2B imagery (10m resolution) covering nine flood events across different locations (2018–2025). The dataset includes 62 images paired with classification maps (Invalid, Land, Water, Cloud, Flooded) and Digital Elevation Models (DEM). Input patches of 256×256 pixels were normalized and augmented via geometric transformations.
• Architecture: The MDT model integrates diffusion-based generative capabilities with Transformer attention mechanisms.
  • Encoder: Processes all input tokens (multispectral bands and DEM features) through 12 transformer blocks. It utilizes Adaptive Layer Normalization (AdaLN-Zero) to incorporate diffusion timestep information.
  • Decoder: An asymmetric network that processes only masked tokens (cloud-obscured regions) while attending to unmasked tokens as context. This reduces computational cost and explicitly learns to reconstruct missing patches.
  • Cloud Detection Head: A separate semantic segmentation branch that classifies pixels into five land cover classes.
• Loss Functions: The training objective combines segmentation and reconstruction losses:
  • Segmentation: A weighted sum of Cross-Entropy, Dice, and Focal losses to handle class imbalance and focus on challenging examples.
  • Reconstruction: A combination of Mean Squared Error (MSE), Perceptual loss, and Structural Similarity Index (SSIM) loss to ensure pixel accuracy, texture similarity, and structural coherence.
• Training & Inference: The model uses an AdamW optimizer with a linear noise schedule over 1000 timesteps. During inference, the reverse diffusion process iteratively denoises masked tokens over 50 steps, keeping unmasked (visible) regions fixed to ensure only cloud-affected areas are regenerated.

Key Contributions

1. Architecture Adaptation: The MDT architecture was adapted for multispectral satellite imagery, explicitly integrating flood-specific spectral characteristics and DEM data to guide the generation of water bodies.
2. Masked Token Modeling: By employing masked token modeling, the framework aligns with the physical problem of cloud obstruction, learning to predict missing image patches based on global spatial context provided by the Transformer's self-attention mechanisms.
3. Hydrological Consistency: The study evaluates the model not just on standard image quality metrics but also on hydrological measures, demonstrating the preservation of water body continuity and spectral signatures critical for water detection indices.
4. Robustness: The approach offers a stable alternative to GANs, avoiding mode collapse and providing physically consistent reconstructions.

Results
The model was trained for over 100 epochs, showing stable convergence with no significant overfitting.

• Performance Metrics: The validation Intersection over Union (IoU) for segmentation improved from 0.3 to a peak of 0.57. Reconstruction loss stabilized, indicating the model effectively restores masked patches while incorporating elevation data.
• Case Study (2021 Tennessee Flood): Applied to the August 2021 Central Tennessee flood, the model successfully reconstructed water bodies in an image where clouds covered 7.91% of the area.
  • The model reduced the residual cloud coverage from 7.91% to 6.09% in the reconstructed output.
  • This performance is noted as a major improvement over traditional threshold-based methods like Sen2Cor (which typically retain 10–15% residual contamination) and simpler CNN-based classifiers (U-Net) which achieved approximately 8.5% residual coverage in the same context.
  • The model detected 0.53% of the image as flooded area, maintaining spatial continuity of water bodies.

Significance and Claims
The paper claims that diffusion-based generative modeling offers a robust and physically consistent alternative for cloud removal in optical flood monitoring. By leveraging global spatial context and iterative refinement, the MDT approach mitigates the temporal gaps in earth observation caused by cloud cover, enabling more reliable, continuous observations for disaster risk management.

The authors acknowledge limitations, noting that the model was trained on a limited number of localized events and relies primarily on imagery and elevation data rather than explicit physics-based hydrodynamic processes. They suggest that future work should expand the training dataset to diverse flood scenarios, incorporate additional physical constraints (e.g., drainage topology, soil moisture), and integrate physics-based models with fused optical and SAR data to further improve predictive robustness. The study positions this work as a step toward enabling high-resolution optical imagery usage under adverse atmospheric conditions, potentially accelerating damage assessment and emergency response timelines.