DAM-Net: Global Flood Detection from SAR Imagery Using Differential Attention Metric-Based Vision Transformers (2306.00704v1)

Abstract: The detection of flooded areas using high-resolution synthetic aperture radar (SAR) imagery is a critical task with applications in crisis and disaster management, as well as environmental resource planning. However, the complex nature of SAR images presents a challenge that often leads to an overestimation of the flood extent. To address this issue, we propose a novel differential attention metric-based network (DAM-Net) in this study. The DAM-Net comprises two key components: a weight-sharing Siamese backbone to obtain multi-scale change features of multi-temporal images and tokens containing high-level semantic information of water-body changes, and a temporal differential fusion (TDF) module that integrates semantic tokens and change features to generate flood maps with reduced speckle noise. Specifically, the backbone is split into multiple stages. In each stage, we design three modules, namely, temporal-wise feature extraction (TWFE), cross-temporal change attention (CTCA), and temporal-aware change enhancement (TACE), to effectively extract the change features. In TACE of the last stage, we introduce a class token to record high-level semantic information of water-body changes via the attention mechanism. Another challenge faced by data-driven deep learning algorithms is the limited availability of flood detection datasets. To overcome this, we have created the S1GFloods open-source dataset, a global-scale high-resolution Sentinel-1 SAR image pairs dataset covering 46 global flood events between 2015 and 2022. The experiments on the S1GFloods dataset using the proposed DAM-Net showed top results compared to state-of-the-art methods in terms of overall accuracy, F1-score, and IoU, which reached 97.8%, 96.5%, and 93.2%, respectively. Our dataset and code will be available online at https://github.com/Tamer-Saleh/S1GFlood-Detection.

• The paper presents DAM-Net, a novel approach using a weight-sharing Siamese backbone with differential attention for precise flood detection.
• It employs a Temporal Differential Fusion module to integrate multi-temporal SAR features, effectively mitigating noise and detecting changes.
• Experiments on the S1GFloods dataset demonstrate that DAM-Net outperforms state-of-the-art methods with accuracies up to 97.8% and an IoU of 93.2%.

Overview of "DAM-Net: Global Flood Detection from SAR Imagery Using Differential Attention Metric-Based Vision Transformers"

The paper entitled "DAM-Net: Global Flood Detection from SAR Imagery Using Differential Attention Metric-Based Vision Transformers," presents a novel approach to detecting floods using synthetic aperture radar (SAR) imagery. This work highlights the importance of accurate flood detection, an essential task for disaster management and environmental planning. The proposed method, named DAM-Net, utilizes a vision transformer framework tailored for the specific challenges presented by SAR images, such as speckle noise and similar backscatter signals between different land covers.

Methodological Innovations

DAM-Net offers a structured approach to processing SAR imagery for flood detection. The network consists of two main components: a weight-sharing Siamese backbone and a temporal differential fusion (TDF) module. The former is responsible for extracting multi-scale change features of multi-temporal images, whereas the latter integrates these features to generate precise flood maps.

The Siamese backbone of DAM-Net leverages differential attention metrics, allowing the model to capture high-level semantic information of water-body changes effectively. It employs a series of modules:

1. Temporal-Wise Feature Extraction (TWFE): This module aims to extract features from each temporal image separately, leveraging both local and global contexts.
2. Cross-Temporal Change Attention (CTCA): CTCA plays a crucial role in identifying changes across temporal images by modeling long-range dependencies between them.
3. Temporal-Aware Change Enhancement (TACE): This module enhances the extracted change features by retaining high-level semantic information captured through an attention mechanism.

The TDF module further processes these enhanced features, emphasizing the relationship between the semantic tokens and change features, culminating in the generation of high-quality flood maps. The utilization of a class token within the TDF module aids in capturing the semantic intricacies of water-body changes, thereby facilitating accurate flood detection.

Significance of the S1GFloods Dataset

A prominent feature of this work is the introduction of the S1GFloods dataset, which addresses a gap in existing flood detection databases. The dataset comprises high-resolution Sentinel-1 SAR image pairs from 46 global flood events, enriching the research landscape with diverse flood causes and scenes. This dataset, covering a broad spectrum of flooded environments and events, is crucial for developing and testing robust flood detection models.

Experimental Results

The authors substantiate the efficacy of DAM-Net through comprehensive experiments conducted on the S1GFloods dataset. The model outperforms several state-of-the-art techniques, including both CNN-based and ViT-based methods, achieving an overall accuracy of 97.8%, an F1-score of 96.5%, and an IoU of 93.2%. These metrics demonstrate DAM-Net's capability to accurately identify flooded areas while mitigating common pitfalls such as false positives and missed detections.

Implications and Future Directions

DAM-Net represents a significant advance in SAR-based flood detection, demonstrating the potential of vision transformers in this domain. The architecture effectively addresses the challenges posed by SAR data, suggesting broader applications of transformer models in remote sensing tasks. The creation of the S1GFloods dataset also sets a precedent for future works, encouraging the development of more diverse and representative datasets.

Potential future directions could involve the integration of additional data sources, such as optical or LiDAR imagery, to further enhance model performance and resilience. Moreover, adapting DAM-Net for near-real-time applications could significantly impact disaster response strategies, enabling timely and efficient management of flood events.

In summary, the DAM-Net framework, combined with the S1GFloods dataset, provides a sophisticated and effective methodology for global flood detection, contributing both practically and theoretically to advancements in remote sensing and environmental monitoring technologies.