A SAR-based flood mapping approach: application of SAR-SIFT registration and modified DeepLabV3 segmentation in flood hazard assessment

17 Jun 2025

---

Zhang, Z., Xiong, J., Li, X., Li, Y., & Liu, J. (2025). A SAR-based flood mapping approach: application of SAR-SIFT registration and modified DeepLabV3 segmentation in flood hazard assessment. Geocarto International, 40(1), 2512188.

Article Link

1. The highlight of this article
2. Study area and dataset
3. Methods
4. Results

1. The highlight of this article

• Proposes an automated, high-resolution flood mapping framework using SAR satellite imagery, SIFT feature-based registration, and a modified DeepLabV3+ deep learning segmentation model.

• Demonstrates that the model is effective in detecting flood extents even in complex urban and vegetation-covered environments, outperforming traditional unsupervised methods.

• The approach provides timely and accurate flood hazard assessment that can assist in emergency response and planning.

2. Study area and dataset

Study Area: Hainan Province, China, which experienced severe flooding from August 29 to September 1, 2022.

Focused on areas affected by Typhoon Ma-on, including both urban and agricultural zones.

Data Sources:

Sentinel-1 SAR imagery: Dual-temporal VV polarization (10 m resolution), pre-flood (August 23) and post-flood (August 31) scenes, sourced from ESA Copernicus Open Access Hub.

Sentinel-2 optical imagery: Used for reference and labeling.

Land cover data from the Chinese Academy of Sciences (CLCD).

Manual labeling of flooded areas based on Sentinel-2 images and field knowledge, used to create ground truth masks.

3. Methods

The proposed SAR-based flood mapping framework includes:

1. SAR-SIFT Image Registration:

SAR-SIFT algorithm used for co-registration of multi-temporal SAR images.

Improves alignment accuracy and maintains local details better than conventional correlation-based approaches.

1. Modified DeepLabV3+ Model:

DeepLabV3+ with a ResNet-101 backbone is modified to handle binary segmentation of flooded vs. non-flooded areas.

Input: dual-temporal SAR image composites (pre-flood + post-flood).

Ground truth masks are generated from manual interpretation of Sentinel-2 and FEMA data.

Model trained with data augmentation and evaluated using precision, recall, F1-score, and IOU metrics.

1. Comparison with Unsupervised Methods:

Benchmarked against change detection techniques (e.g., logarithmic ratio, histogram thresholding).

Also compared to raw DeepLabV3+ without SAR-SIFT registration.

4. Results

SAR-SIFT Performance:

Achieved better co-registration accuracy than OpenCV and ENVI methods, with lower RMSE and misalignment.

Flood Segmentation Accuracy:

The modified DeepLabV3+ model with SAR-SIFT registered inputs achieved:

F1-score: 0.853, IOU: 0.743

Outperformed unsupervised methods (e.g., F1-score of histogram thresholding ~0.4).

Visual Accuracy:

The deep learning-based method produced flood maps that aligned better with FEMA boundaries, including in urban areas and complex terrains.

Generalizability:

The workflow is modular and scalable to other regions and flood events using SAR imagery.