High-Resolution Flood Mapping With Sentinel-1 and Sentinel-2 via Misalignment-Robust Cross-Sensor Learning and Generative Despeckling

Technical Summary: High-Resolution Flood Mapping With Sentinel-1 and Sentinel-2 via Misalignment-Robust Cross-Sensor Learning and Generative Despeckling

Problem Statement
Reliable high-resolution flood extent mapping is currently constrained by data fidelity issues and sensor-specific artifacts. Multispectral optical imagery (Sentinel-2) is frequently degraded by clouds, shadows, and urban confounders (e.g., asphalt, rooftops) that mimic water signatures. Conversely, Synthetic Aperture Radar (SAR) imagery (Sentinel-1), while independent of weather, suffers from speckle noise and sensor co-registration uncertainty. Existing public flood mapping datasets often rely on semi-automated thresholding or weak labels that lack pixel-level fidelity, particularly in challenging conditions like dense urban areas or under cloud cover. Furthermore, training deep learning models on misaligned SAR and optical data introduces spatial biases that hinder generalization.

Methodology
The authors propose an integrated framework addressing these limitations through curated datasets and novel learning strategies:

1. Dataset Curation:

   • Sentinel-2 (S2): A high-quality dataset of 146 manually annotated 10m water masks across the contiguous United States (CONUS). Annotations were generated using rigorous geospatial protocols, incorporating multi-source data (TCI, NDWI, DEM, roads) to ensure pixel accuracy in challenging scenes (haze, shadows, urban).
   • Sentinel-1 (S1): A dataset of 7,846 SAR tiles. Labels were transferred from temporally coincident S2 imagery (within 24 hours). While 61 tiles were manually verified, the majority utilized "weak labels" generated by a high-performing S2 model, filtered for quality.
   • Multitemporal Composite: A dataset of 38,252 stacks of 7 consecutive non-flood SAR images used exclusively to train a despeckling model.

2. Model Architectures:

   • Segmentation: UNet and UNet++ architectures were employed. UNet++ was selected for its dense skip connections to improve multi-scale feature fusion.
   • Despeckling: A Conditional Variational Autoencoder (CVAE) was trained to map noisy single-temporal SAR patches to clean multitemporal composites. This generative approach models the conditional distribution of despeckled outputs, preserving structural details better than classical filters.

3. Training Innovations:

   • Shift-Invariant Loss: To address residual geolocation uncertainty (approx. 20m at 2σ) between S1 and S2, a shift-invariant loss function was introduced. This allows the model to find the optimal spatial alignment (up to ±2 pixels) between the prediction and the ground truth during backpropagation, preventing the learning of systematic spatial biases.
   • Loss Functions: Tversky and Focal Tversky losses were utilized to handle severe class imbalance (water vs. non-water).
   • Training Strategy: The CVAE was pre-trained on general CONUS SAR data and fine-tuned on flood-adjacent regions. The segmentation models were trained with extensive data augmentation and hyperparameter tuning via Bayesian optimization.

Key Results

• Multispectral Performance: The UNet++ model trained on all available channels (reflectance, water indices, and supplementary geophysical data) achieved a state-of-the-art Area Under the Precision-Recall Curve (AUPRC) of 0.956 on the S2 test set. It demonstrated robustness in challenging conditions, achieving F1 scores of 0.877 in haze and 0.787 in urban areas.
• SAR Performance: On the S1 dataset, the UNet++ model achieved an AUPRC of 0.840. While the shift-invariant loss did not significantly increase the AUPRC score compared to regular loss when evaluated on uncorrected labels, it produced predictions that better respected coherent SAR backscatter structures.
• Despeckling Efficacy: The CVAE despeckler significantly outperformed the Enhanced Lee filter across all metrics (PSNR, SSIM, ENL) on the multitemporal dataset. Qualitatively, it preserved fine structures and edges while smoothing homogeneous regions.
• Downstream Impact: When CVAE-despeckled SAR data was fed into the flood segmentation model, it showed statistically significant improvements over models trained on Enhanced Lee-filtered data. However, the paper notes that CVAE despeckling did not yield statistically significant gains over raw SAR data in terms of AUPRC. The authors attribute this to geometric distortions inherent in SAR (slant range, layover) that remain incompatible with nadir-view optical labels, even after despeckling.

Significance and Claims
The paper claims that the primary bottleneck in high-resolution flood mapping is often data fidelity rather than model architecture. By introducing a dataset with pixel-accurate 10m masks and addressing cross-sensor misalignment through shift-invariant training, the work demonstrates that models can achieve high fidelity in operational scenarios.

The authors emphasize that while generative despeckling (CVAE) significantly improves image quality and structure preservation, its direct translation to improved flood mapping metrics is limited by geometric mismatches between SAR and optical sensors. Consequently, the paper advocates for a holistic approach that prioritizes dataset curation, robust training objectives (like shift-invariant loss), and acknowledges the physical limitations of cross-sensor alignment. The work establishes the viability of generative despeckling for operational use but calls for further research into geometric coregistration techniques to fully leverage despeckled SAR data for flood mapping.