Introduction
Accurate flood extent mapping is a cornerstone of effective disaster response and insurance risk assessment. While deep learning models offer the potential to automate this process, their performance is strictly bound by the availability of high-quality "ground truth" labels. Currently, generating these labels is a significant bottleneck; domain experts must manually delineate flooded areas in high-resolution aerial imagery—a process that is both labor-intensive and expensive.
Crowdsourcing to non-experts presents a scalable alternative, yet untrained users often struggle to distinguish between wet and dry land, particularly when features are obscured by tree canopies or shadows. This paper introduces FloodTrace, a web-based tool designed to bridge this gap by enabling non-experts to produce expert-quality flood annotations with unprecedented speed.
Methodology
FloodTrace departs from traditional "brush and polygon" annotation tools by incorporating elevation-aware features that leverage the physical properties of water flow:
- Topological Segmentation: Utilizing persistent homology and contour trees, the tool automatically groups pixels into regions based on their elevation profile. This allows users to label entire depressions or river basins with a single click rather than painting pixel-by-pixel.
- Elevation-Guided Breadth-First Search (BFS): This tool allows users to select a "seed" pixel, after which the algorithm automatically labels connected pixels at lower (downstream) or higher (upstream) elevations, respecting the physical constraints of terrain.
- 3D Interactive Visualization: To assist in terrain comprehension, aerial imagery is rendered over a 3D mesh derived from Digital Elevation Models (DEM), allowing users to rotate and zoom to identify natural water collection points.
- Uncertainty-Aware Aggregation: The system includes a review mode for researchers that highlights areas of disagreement among crowdworkers, allowing experts to focus their limited time on correcting only the most uncertain regions.
Data and Evaluation
The study utilized high-resolution aerial imagery from NOAA’s National Geodetic Survey and DEM data from the University of North Carolina Libraries, focusing on the aftermath of Hurricane Matthew (2016). The authors have open-sourced a new benchmark dataset consisting of 1,321 unique annotations across eight regions, created by 266 participants and covering over 3.5 billion pixels.
Results and Key Insights
A large-scale user study involving 266 graduate students demonstrated the efficacy of the elevation-aware approach:
- Significant Speedup: The topological segmentation tool was 2.29x faster than manual brushing and significantly outperformed previous state-of-the-art elevation tools.
- Efficiency Gains: Median annotation time was reduced by more than half compared to existing methods.
- Expert Parity: Machine learning models trained on aggregated and corrected crowdsourced data achieved performance metrics (accuracy and F-score) equal to those trained on expert-only labels. Crucially, this was achieved with only 24 minutes of expert "correction" time compared to 116 minutes for full manual labeling.
Beyond raw metrics, the study revealed that topological segments align remarkably well with flooded regions because water naturally follows the mathematical "breaks" in terrain. This approach effectively solves the "canopy problem," where flooded forests appear green in standard imagery; because the tool incorporates elevation data, it can accurately label the ground beneath the trees.
Discussion and Future Work
While FloodTrace represents a significant leap forward, several areas for improvement remain. The current iteration performs binary classification (flooded vs. dry) and does not yet distinguish between permanent water bodies and temporary floodwater. Additionally, the reliance on DEMs (bare earth models) means the tool does not account for man-made structures like buildings that might obstruct water flow. Future work will explore the integration of Digital Surface Models (DSM) or LiDAR data to improve accuracy in urban environments and address potential temporal misalignments between elevation scans and flood imagery.
Summary updated