Abstract
Accurate flood mapping is critical for disaster response, yet the scarcity of labeled high-resolution satellite imagery creates a significant bottleneck. Manual annotation is labor-intensive, often taking hours per image due to complex terrain and obstructions. This paper introduces ALFA (Active Learning for Flood Annotation), an interactive framework that combines a physics-aware neural network with a novel consistency-enhanced active learning strategy. By integrating elevation data and identifying informative regions through prediction inconsistencies, ALFA reduces annotation time from hours to approximately 20 minutes per image while achieving high accuracy.
Introduction
During a disaster, flood mapping is a race against time. While satellite imagery is increasingly available, the labels required to train AI models are scarce. Existing datasets are often too small for robust generalization, and manual labeling remains the primary bottleneck. Experts must painstakingly outline jagged flood boundaries, often obscured by buildings or tree canopies. ALFA addresses this "annotation bottleneck" by facilitating a collaborative workflow between humans and AI, maximizing the impact of every user interaction.
Methodology
The ALFA framework is built upon four primary pillars designed to streamline the labeling process:
- Elevation-Guided BFS: Utilizing the physical properties of water, the system employs a Breadth-First Search (BFS) algorithm. When a user marks a pixel as "flooded," the system automatically labels all connected lower-elevation pixels (pit-filling). Conversely, marking a "dry" pixel propagates the label to higher elevations (hill-climbing).
- Consistency-Enhanced Active Learning: Rather than relying solely on model uncertainty, ALFA identifies informative regions by looking for View Consistency (robustness across image rotations) and Temporal Consistency (stability of predictions across training cycles).
- Superpixel Recommendations: The system recommends clusters of similar pixels (superpixels) that the model finds most confusing, allowing users to label large areas with single clicks.
- EvaNet: A lightweight U-Net variant that integrates Digital Elevation Model (DEM) data directly into its architecture, ensuring the model respects physical flow constraints.
Datasets and Evaluation
The researchers evaluated ALFA using high-resolution aerial imagery and DEM data from two major events: Hurricane Matthew (2016) in North Carolina and Hurricane Harvey (2017) in Texas. To handle visual obstructions, a "tree score" derived from a separate model (ForestNet) was used to filter out ambiguous areas under dense canopy, keeping the user focused on high-impact regions.
Key Results
Experimental results demonstrate that ALFA significantly outperforms traditional manual and automated baselines:
- Massive Speedup: In user studies with domain experts, the median time to annotate a region dropped from several hours to approximately 20 minutes.
- Superior Accuracy: The ALFA-PO variant (using probability offset) consistently achieved higher F1 scores and mIoU than random sampling or standard active learning methods like PixelPick.
- Rapid Convergence: The model typically reaches high performance (F1 > 90%) very early in the annotation process, requiring only a small fraction of the total image to be labeled.
Discussion and Future Work
ALFA demonstrates that incorporating physical laws and prediction robustness can dramatically accelerate geospatial data labeling. While the current system uses a lightweight model for interactivity, future work could involve using ALFA to generate high-quality labels for fine-tuning massive geo-foundation models. Additionally, while the elevation-guided BFS is highly effective in rural and suburban areas, more sophisticated modeling may be required for dense urban environments with complex man-made drainage systems.
Summary updated