Enhancing Flood Detection: A New Benchmark and Model Paradigm for Urban Waterlogging

Urban waterlogging is a pressing concern, posing major risks to public safety and infrastructure. Traditional methods for detecting waterlogged areas typically involve installing water-level sensors, but these methods are often impractical due to high maintenance requirements and their inability to achieve comprehensive coverage. In recent years, there has been a shift towards utilizing surveillance camera imagery coupled with deep learning techniques for detecting urban waterlogging. However, these advanced methods also face significant challenges, particularly due to the scarcity of relevant data and adverse environmental conditions that affect detection accuracy.

Innovative Approach to Urban Waterlogging Detection

In response to these challenges, a team of researchers from Chongqing University and Huawei Technologies has developed an innovative approach to enhance urban waterlogging detection. Their work introduces a challenging Urban Waterlogging Benchmark (UW-Bench) and proposes a Large-Small Model co-adapter (LSM-adapter) paradigm. The UW-Bench is designed to address the limitations of current datasets by including a diverse range of conditions that waterlogged areas might present. This benchmark comprises 7,677 images that capture waterlogging under various challenging scenarios such as strong-light reflections, low-light environments, and clear water, which makes the task of waterlogging detection even more complex.

Harnessing the Power of Large and Small Models

The LSM-adapter paradigm is a novel approach that combines the strengths of both large and small models. Large models are known for their substantial generic segmentation capabilities, while small models offer task-specific guidance. By integrating these two models, the researchers aim to leverage the broad segmentation power of the large model and the specific, detailed guidance of the small model. The LSM-adapter includes several key components to facilitate this integration: a Triple-S Prompt Adapter module, a Dynamic Prompt Combiner, and a Histogram Equalization Adapter. The Triple-S Prompt Adapter module is designed to generate and merge multiple prompts for mask decoder adaptation, while the Dynamic Prompt Combiner dynamically weighs and blends these prompts to create a more accurate and robust model. Additionally, the Histogram Equalization Adapter infuses image-specific information into the encoder to enhance the overall performance of the system.

Challenges in Waterlogging Detection

Urban waterlogging detection is inherently challenging due to several factors. Waterlogged areas can vary significantly in shape, size, and depth, making it difficult to develop a uniform set of features for detection. The reflective nature of water surfaces and the presence of shallow or clear standing water can obscure water texture information, further complicating detection efforts. Moreover, low-light conditions can make waterlogging features less prominent, increasing the difficulty of accurate detection. Existing methods often struggle to provide accurate segmentation in real-world urban scenarios, particularly due to the limited scale and insufficient diversity of labeled data, which hampers the generalizability of these models.

Advancements Through the UW-Bench Dataset

To address these challenges, the UW-Bench provides a comprehensive dataset that captures a wide range of waterlogging scenarios. This dataset is a significant advancement in the field, offering a robust platform for developing and testing new detection algorithms. The LSM-adapter paradigm enhances the capabilities of existing models by integrating large and small models in a complementary manner. The large model, guided by the small model's specific task-directed prompts, can achieve more accurate and reliable waterlogging detection.

Proven Results and Future Prospects

The researchers conducted extensive experiments to evaluate the performance of their proposed method. The results demonstrate that the LSM-adapter significantly outperforms existing methods in terms of precision, recall, F1-score, and Intersection over Union (IoU). Specifically, the LSM-adapter shows an increase in F1-score by 6.8% to 18.28% and an increase in IoU by 8.22% to 19.89% compared to the respective small models. When compared to the large model SAM-Adapter, the LSM-adapter shows improvements of 7.45% to 9.02% in F1-score and 6.57% to 8.53% in IoU. These results highlight the effectiveness of the LSM-adapter in enhancing detection performance under diverse and challenging conditions.

The development of the UW-Bench and the LSM-adapter represents a significant advancement in the field of urban waterlogging detection. By addressing the limitations of traditional methods and existing datasets, this research provides a more robust and reliable framework for detecting waterlogged areas in urban environments. The integration of large and small models offers a promising solution to the challenges of waterlogging detection, paving the way for more effective flood management systems in the future. This work not only contributes to the academic field but also has practical implications for improving urban infrastructure resilience and public safety in the face of flooding events.