🤖 machine learning

Geometric Flood Depth Estimation: Fusing Transformer-Based Segmentation with Digital Elevation Models

This paper proposes a geometric method that fuses Mask2Former-based flood segmentation with Digital Elevation Models to estimate 3D flood depth and water surface elevation from monocular aerial imagery, enabling rapid post-disaster situational awareness without the latency of hydrodynamic simulations.

Original authors: Nhut Le, Ehsan Karimi, Maryam Rahnemoonfar

Published 2026-05-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Nhut Le, Ehsan Karimi, Maryam Rahnemoonfar

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a disaster responder looking at a satellite photo of a flooded city. You can clearly see where the water is (the dark blue patches), but the photo is flat. You don't know if the water is just a puddle you can step over, or a deep river that will swallow a car.

This paper presents a clever, fast way to figure out how deep the water is without needing complex physics simulations or waiting for surveyors to walk through the mud.

Here is how their method works, broken down into simple steps:

1. The "Bathtub" Idea

The core concept is what the authors call the "Bathtub Principle."
Think of a flooded city like a giant bathtub. When you fill a bathtub with water, the water doesn't stay deep in one corner and shallow in another; it naturally settles into a flat, level surface.

The Paper's Logic: If we know exactly where the water touches the dry land (the "waterline"), we can measure the height of the land at that edge. Because the water is flat, that edge height tells us the height of the water everywhere inside that flooded area.
The Math: Once we know the water's surface height, we just subtract the height of the ground (from a digital map) to find the depth.
- Water Height – Ground Height = Water Depth.

2. The Two-Step Pipeline

To make this work, the authors built a system that combines two types of data:

Step A: The "Sharp-Eyed" AI (The Segmentation)
First, they need to know exactly where the water ends and the land begins.

The Problem: Older computer vision models are a bit blurry. They might think a flooded road is just part of a big lake, or they might get confused by shadows.
The Solution: They used a super-smart AI model called Mask2Former. Think of this model as a high-definition artist that can draw a razor-sharp line around the water. It uses a "Transformer" (a type of AI good at understanding the whole picture at once) to distinguish between a flooded street, a building, and a tree.
Why it matters: If the line is blurry, the depth calculation is wrong. This model draws the line perfectly.

Step B: The "Digital Terrain" (The DEM)
Next, they take that sharp water line and overlay it on a Digital Elevation Model (DEM).

What is a DEM? Imagine a 3D map of the city where every single pixel knows its exact height above sea level, even if it's currently underwater.
The Magic: The system finds the "waterline" on this 3D map, reads the height of the land there, and assumes the water is flat at that height all the way across the flood. Then, it calculates the depth for every single pixel.

3. Why This is Better Than Old Methods

The paper compares their method to two other ways of doing things:

The "Physics Engine" Way: Traditional methods try to simulate how water flows using complex math (like a video game physics engine). The authors say this is too slow. Simulating one flood event could take days of computer time. Their method takes seconds.
The "Guessing" Way: Some AI models try to guess the depth directly from the photo. The authors say these models often "hallucinate" (make things up) or create smooth, fake-looking depths that don't match reality.
The "Geometric" Way (Their Method): By using the "Bathtub" logic and a sharp AI outline, they get a result that is fast (no waiting days) and physically grounded (it follows the laws of how water sits on land).

4. The Results

They tested this on real flood data from Hurricane Harvey in Texas.

They compared their AI-generated depth maps against official government survey data (FEMA).
The Outcome: Their method was surprisingly accurate, with an average error of about 2 meters (roughly 6.5 feet). Given that they didn't use any complex physics simulations or real-time weather data, this is considered a very strong result.
Visual Proof: They showed images where the "depth map" looks realistic: shallow near the edges where the water touches the land, and getting deeper as you move toward the center of the flooded area, just like a real bathtub.

Summary

In short, this paper teaches a computer to:

See the flood with laser-sharp precision.
Look up the height of the land at the edge of the flood.
Assume the water is flat like a bathtub.
Calculate the depth instantly.

This gives emergency responders a quick, 3D view of the flood depth from a simple 2D photo, helping them decide which roads are passable and which buildings are in danger, all without waiting for slow simulations.

Technical Summary: Geometric Flood Depth Estimation

Problem Statement
Post-disaster flood response requires not only the identification of flood extent but also the estimation of water depth to assess navigability and structural risk. While deep learning-based semantic segmentation effectively identifies 2D flood masks, it lacks the vertical dimension necessary for volumetric analysis. Existing alternatives face significant limitations:

Hydrodynamic solvers (e.g., those solving Shallow Water Equations) offer physical rigor but suffer from high computational latency (days for a single event) and a reliance on real-time boundary conditions (inflow rates, rainfall) that are rarely available immediately after a disaster.
Machine learning surrogate models reduce inference time but often introduce smoothing artifacts or hallucinate floods in physically impossible locations.
Reference products (e.g., FEMA depth rasters) provide accurate ground truth but are released weeks to months post-event, rendering them useless for immediate response.

Methodology
The paper proposes a Geometric Depth Estimation pipeline that bypasses hydrodynamic simulation and data-driven surrogates by fusing high-precision 2D segmentation masks with pre-existing Digital Elevation Models (DEMs). The approach operates on the "bathtub principle," assuming standing floodwater creates a locally flat surface. The pipeline consists of three stages:

Data Acquisition and Preprocessing:
- Datasets: The system utilizes FloodNet for training segmentation models and CRASAR-U-DROIDS (containing GeoTIFF imagery with geospatial metadata) for deployment and validation.
- Alignment: DEMs are fetched from OpenTopography and reprojected/upsampled via bilinear interpolation to match the Coordinate Reference System (CRS) and resolution of the UAV imagery.
- Tiling: Large orthomosaics are cropped into 1024×1024 tiles centered on building centroids, filtered for major hurricane events, and quality-controlled to remove void pixels.
2D Flood Segmentation (Stage 1):
- Model Selection: The authors fine-tuned and benchmarked three architectures (Attention U-Net, SegFormer, and Mask2Former) on FloodNet. Mask2Former was selected as the optimal backbone due to its transformer-based attention mechanism, which provides sharper waterline delineation critical for geometric accuracy.
- Class Aggregation: For urban scenarios, classes are simplified into "Flood" (combining water and flooded roads), "Buildings," and "Background" to ensure connected water bodies are correctly identified.
Geometric Depth Calculation (Stage 2):
- Connected Component Analysis (CCA): The binary flood mask is processed to identify disjoint water bodies as independent physical systems, preventing erroneous water level calculations across different elevations.
- Waterline Identification: A "Strict Dryland Contact" logic is applied. Morphological erosion identifies the inner boundary pixels where water meets dry land, rejecting image-edge artifacts.
- Depth Calculation: For each independent water body, the global water surface elevation ( $Z_{water}$ ) is estimated by computing the 95th percentile of the DEM values at the identified boundary pixels. This statistical approach filters out high-frequency outliers (e.g., misclassified vegetation or building walls) better than using maximum elevation. The depth for any pixel $(x, y)$ is calculated as:
  $Depth(x, y) = \max(0, Z_{water}^{(k)} - DEM(x, y))$

Key Contributions

Transformer Integration: The integration of Mask2Former into geometric depth estimation, leveraging its attention mechanism to achieve sharper waterline delineation compared to traditional CNNs or coarser SAR-derived masks.
Strict Dryland Contact Logic: The introduction of a boundary logic with edge-artifact rejection, improving the precision of the water-land interface identification, which is the critical bottleneck in geometric depth accuracy.
Efficient Volumetric Inference: A pipeline that extracts 3D volumetric data from 2D monocular imagery without the latency of hydrodynamic simulations or the instability of data-driven surrogates.

Results
The method was evaluated on 136 Hurricane Harvey tiles from the CRASAR-U-DROIDS dataset, covering over 52 million openly-flooded pixels, using FEMA flood depth rasters as quantitative ground truth.

Segmentation Performance: Mask2Former achieved a mean Intersection over Union (mIoU) of 73.6% on the FloodNet test set, outperforming SegFormer (70.4%) and Attention U-Net (69.7%), with significant gains in "Road Flooded" and "Water" classes.
Depth Accuracy: The pipeline achieved a Root Mean Square Error (RMSE) of 1.999 meters and a global $\sqrt{MSE}$ of 2.169 meters.
Qualitative Validation: Visual results demonstrated physically plausible hydrostatic gradients (shallow at waterlines, deepening toward basin centers) and correct masking of buildings to prevent erroneous depth inference on vertical structures.

Significance and Claims
The authors claim that this geometric framework offers a computationally efficient and physically grounded approach for rapid post-disaster situational awareness. By relying on direct observation (segmentation + DEM) rather than complex simulation, the method delivers actionable volumetric insights suitable for immediate response. The paper asserts that the method is robust enough to recover realistic depth gradients in complex urban environments without site-specific calibration. Future work is identified as integrating these derived depth estimates into damage assessment networks to quantify structural damage levels for individual buildings.