UrbanSARFloods: Sentinel-1 SLC-Based Benchmark Dataset for Urban and Open-Area Flood Mapping (2406.04111v1)

Abstract: Due to its cloud-penetrating capability and independence from solar illumination, satellite Synthetic Aperture Radar (SAR) is the preferred data source for large-scale flood mapping, providing global coverage and including various land cover classes. However, most studies on large-scale SAR-derived flood mapping using deep learning algorithms have primarily focused on flooded open areas, utilizing available open-access datasets (e.g., Sen1Floods11) and with limited attention to urban floods. To address this gap, we introduce \textbf{UrbanSARFloods}, a floodwater dataset featuring pre-processed Sentinel-1 intensity data and interferometric coherence imagery acquired before and during flood events. It contains 8,879 $512\times 512$ chips covering 807,500 $km^2$ across 20 land cover classes and 5 continents, spanning 18 flood events. We used UrbanSARFloods to benchmark existing state-of-the-art convolutional neural networks (CNNs) for segmenting open and urban flood areas. Our findings indicate that prevalent approaches, including the Weighted Cross-Entropy (WCE) loss and the application of transfer learning with pretrained models, fall short in overcoming the obstacles posed by imbalanced data and the constraints of a small training dataset. Urban flood detection remains challenging. Future research should explore strategies for addressing imbalanced data challenges and investigate transfer learning's potential for SAR-based large-scale flood mapping. Besides, expanding this dataset to include additional flood events holds promise for enhancing its utility and contributing to advancements in flood mapping techniques.

• The paper demonstrates the added value of combining SAR intensity with interferometric coherence data for enhanced flood mapping in urban contexts.
• It details a large-scale dataset spanning 807,500 sq km and 18 flood events, comprising 8,879 image chips prepared for deep learning.
• Benchmark evaluations reveal that standard segmentation models struggle with urban flood detection, highlighting the need for advanced techniques to handle data imbalance and complex backscatter.

UrbanSARFloods: Sentinel-1 SLC-Based Benchmark Dataset for Urban and Open-Area Flood Mapping

The paper "UrbanSARFloods: Sentinel-1 SLC-Based Benchmark Dataset for Urban and Open-Area Flood Mapping" addresses a significant gap in the domain of large-scale flood mapping using Synthetic Aperture Radar (SAR) data. Existing datasets and methodologies have largely concentrated on mapping floods in open areas, with limited exploration of urban contexts where the complexity and variability of land cover pose substantial challenges.

Dataset Construction and Composition

UrbanSARFloods is a newly developed, comprehensive dataset comprising pre-processed Sentinel-1 SAR imagery. This dataset includes intensity data and interferometric coherence images collected before and during flood events. It spans a significant geographic scope, covering 807,500 square kilometers across 5 continents and 18 distinct flood events. The dataset consists of 8,879 chips, each measuring 512×512 pixels, and represents an extensive variety of land cover classes.

The dataset's construction involved a meticulous preprocessing protocol: Sentinel-1 Level-1 Interferometric Wide Swath SLC data were utilized to extract interferograms, calibrated, and transformed into intensity data. This preprocessing step ensures the dataset's readiness for deep learning applications. The importance of including both intensity and InSAR coherence data was underscored, given their complementary utility in detecting floodwater, especially in urban settings where backscatter characteristics can be complex and varied.

Challenges Highlighted

Two primary challenges were identified from the analysis and application of previous flood mapping techniques:

1. Data Imbalance: Flood mapping datasets often exhibit extreme imbalances, with flood pixels constituting a minor fraction of the overall data. This characteristic was prominently addressed by employing the Weighted Cross-Entropy (WCE) loss in the evaluation phase to mitigate class imbalance effects.
2. Complex Backgrounds: Urban areas present highly heterogeneous environments where SAR backscatter can be influenced by various structures and surfaces, making flood detection more complex than in open areas.

Evaluation of State-of-the-Art Models

The paper benchmarked several semantic segmentation models using UrbanSARFloods, including Unet, Unet++, MANet, Linknet, FPN, PSPNet, PAN, DeepLabV3, and DeepLabV3+. The evaluation revealed that these state-of-the-art models struggled with urban flood detection due to the aforementioned challenges. For instance, the F1 scores for flooded open areas (FO) varied significantly, highlighting a tendency towards overestimation in certain scenarios. This variability underscored the need for more sophisticated handling of imbalanced and complex data.

Transfer Learning and Model Generalization

Additionally, transfer learning techniques were applied using models pre-trained on ImageNet. However, minimal performance gains were observed compared to models trained from scratch. This outcome was attributed to the substantial feature space differences between RGB images typical of ImageNet and the SAR-specific characteristics of UrbanSARFloods' data. The findings suggest that more domain-specific pre-training or advanced transfer learning approaches, such as domain adaptation, could be explored in future research to improve model efficacy.

Implications and Future Research Directions

The implications of this research are both practical and theoretical. Practically, the dataset serves as a critical resource for developing and validating flood detection methodologies that can operate at a global scale, incorporating both urban and rural contexts. Theoretically, the paper advances understanding of SAR data's utility in complex environments and highlights areas for methodological improvement, such as advanced handling of data imbalance and feature space discrepancies in transfer learning.

Future research should focus on:

• Advanced Data Augmentation: Leveraging synthetic data generation and domain adaptation techniques to ameliorate feature space discrepancies.
• Model Robustness: Developing models specifically addressing the heterogeneous nature of urban landscapes to improve flood detection accuracy.
• Scalable Solutions: Exploring scalable methodologies that can produce reliable flood maps in real-time, assisting in disaster response and mitigation efforts.

In conclusion, UrbanSARFloods presents a major step forward in flood mapping research by providing a rich, diversified dataset that enables the development of more robust and generalized flood detection models. The insights gained from benchmarking existing deep learning models underscore the need for continued innovation in handling data imbalance and complex SAR features in urban flood mapping applications.