How SAR Satellites Detect Flooded Areas

In short
- SAR satellites can observe the ground with radar waves, even through clouds and at night.
- Water surfaces reflect radar waves like a mirror, so backscatter is low and they look dark.
- Change detection, which compares images from before and after a flood, separates the flooded extent.
Imagine that heavy rain has been falling for 48 hours. The sky is full of thick cloud, and the disaster response agency cannot tell which areas are flooded or how deeply, so it cannot even position rescue vehicles. An ordinary optical satellite cannot see the ground at all because of the cloud.
SAR (Synthetic Aperture Radar) satellites come into their own in this situation. A SAR satellite transmits its own radar signal in the microwave band and records the signal that is reflected back, so it can obtain ground information even under thick cloud or at night. This post explains, step by step, the principle of finding flooded extent with SAR images in flood response practice.
How does a SAR satellite observe the ground?
SAR sends microwaves from its antenna toward the ground at an angle and measures the strength of the signal that returns after hitting objects on the ground (backscatter). The stronger the backscatter, the brighter it appears in the image, and the weaker, the darker. Buildings and rough forests reflect the signal in all directions, so a lot of it returns to the antenna and they appear bright. Conversely, the smoother and more even the surface, the more the signal bounces away like off a mirror and hardly returns to the antenna. The observation frequency (C-band, L-band and so on) and the resolution differ between satellite operators, so you must check the metadata of the images you plan to use before analysis. In flood response, this property makes repeated observation possible even in cloudy disaster conditions, and it is also used to monitor the spread of flooding at night.
Why does water look dark in a SAR image?
A calm water surface behaves much like a mirror. When the SAR signal arrives at an angle, the water surface reflects it away in the opposite direction from the incoming angle (specular reflection). Very little energy returns to the antenna, so the backscatter value is low, and it is drawn in a dark color in the image. This principle applies equally to calm lakes, flooded rice paddies and flooded urban areas. However, if the wind is strong or waves form on the water, the roughened surface scatters the signal, so backscatter rises and the surface may look less dark. In a typhoon with strong winds at a disaster site, you need to allow in advance for the possibility that the flooded extent is detected as narrower than it really is.
What is the procedure for finding flooded extent by change detection?
Change detection is a method of comparing images of the same area from two different times to find what has changed. In flood detection, you proceed in the following order. First, prepare a SAR image from before the flood (normal conditions) and one from after the flood. Second, register the two images to the same coordinate system so the pixel positions line up. Third, calculate the backscatter difference (backscatter before subtracted from backscatter after). Places where water pooled during the flood were bright before and dark after, so the difference appears as a large negative number. Fourth, set a threshold and classify pixels below it as flood candidates. This threshold can vary with the region and the season, so it is something you have to check yourself. In practice, flood candidate areas extracted in this way are overlaid with administrative boundaries and population data to estimate the number of affected households.
Below is a simple example that uses Python and numpy to obtain a backscatter difference image.
import numpy as np
import rasterio
with rasterio.open("sar_before.tif") as src:
before = src.read(1).astype(float)
with rasterio.open("sar_after.tif") as src:
after = src.read(1).astype(float)
diff = after - before
threshold = -3.0 # threshold in dB; adjust it yourself for each region
flood_mask = diff < threshold
What are the main causes of false detections?
There are two main kinds of false positives in SAR change detection (cases where an area is flagged as flooded when it is not). The first is shadow. In radar shadow areas behind mountains or tall buildings, the signal hardly arrives, so backscatter comes out low, and these areas can be mistaken for water. The second is smooth ground. Even surfaces such as unpaved roads, closely mown lawns and dry salt flats can show low backscatter similar to a calm water surface. To filter these out, you need to mask steep shadow areas using a DEM (digital elevation model), or overlay land cover data and cross-check dense building areas. When you report detection results to disaster response decision makers, stating the possibility of these false detections is a basic requirement of a trustworthy report.
How does this connect to real disaster response work?
Once the flooded extent is fixed as a vector polygon, the data feeds into decisions at several stages. In the disaster situation rooms of the Ministry of the Interior and Safety and local governments, flood polygons are overlaid with road network data to find areas that cannot be reached and to re-plan rescue vehicle routes. If you cross them with residence data based on resident registration, you can quickly estimate the population that may be isolated. In this process, the time delay between when the image was taken and when the analysis result is delivered largely determines how useful it is at the site. For this reason, it is becoming increasingly important in practice to build an automated analysis pipeline in advance so that results come out as soon as a satellite image is received. A later post in this series looks at how to add a deep learning model to this kind of automated flow.
Step by step on screen
The figure below is a diagram that summarizes the order of the steps.

- Search for images
Go to the ESA Copernicus Data Space (https://dataspace.copernicus.eu), enter your area of interest in the search box and select the Sentinel-1 GRD filter - Set the period
In the date range fields, enter the 30 days before the flood and the date after it, and download the two images - Open SNAP
Start ESA SNAP > File > Open Product and load the two images (.zip) one after the other (menu positions can differ by version, so see the official SNAP documentation) - Run preprocessing
Select Radar > Geometric > Terrain Correction, set the output coordinate system to EPSG:4326 and click Run - Calculate the difference
Click Raster > Band Maths, enter 'after_VV - before_VV' in the expression box and click OK - Classify by threshold
Run Band Maths again, enter '(after_VV - before_VV) < -3.0' to create the flood mask, and export the result layer
Menu names may differ slightly between program versions.
Prompts you can copy
Use this to turn SAR flood detection results into an explanation for a disaster situation room report
You are a GIS analyst in the disaster safety field. Based on the SAR change detection results below, write a short report paragraph that a non-specialist officer in a disaster situation room can understand. Analysis overview: [flood-affected area], [before and after dates], Sentinel-1 GRD images used Estimated flooded area: [number] km² Main flooded administrative districts: [list of districts] Areas with possible false detections: [the areas, or "none"] Conditions: - Write one paragraph of about 200 characters in English - Add a one-line explanation in parentheses for technical terms such as SAR and backscatter - Always include the uncertainty of the analysis (possible false detections, wind effects and so on) in the last sentence - Never add numbers that have not been confirmed Items a person must check: review whether the area figure matches the actual input value and whether the uncertainty sentence was left out.
Replace the bracketed parts for your situation. Check the facts in any answer before you use it.
Try it yourself
- On the Copernicus Data Space, search for Sentinel-1 GRD images from two dates over an area near a river in Korea, and look for places where you can see a brightness difference by eye
- With the Band Maths function of ESA SNAP, make a difference image of the VV band from the two dates, check the histogram, adjust the threshold yourself and watch how the flood candidate area changes
- Load the flood mask you extracted into QGIS as a layer, compare it with a background map, and note down at least three places where you suspect false detections, such as shadows and roads
Common questions
Q. How is the NDWI from optical satellites different from SAR flood detection?
A. NDWI uses the difference in visible and near-infrared reflectance, so it can be used only when there is no cloud. SAR transmits microwaves itself, so it can observe through cloud and at night. Cross-checking the results of the two methods can raise the reliability of detection.
Q. How do I decide the threshold?
A. There is no fixed value. It depends on the land cover, the season and the satellite viewing angle, so first pick several places in the study area that are clearly flooded and clearly not flooded, and check the distribution of backscatter difference values yourself. The process of setting reference points like this and adjusting the threshold is called ground truthing.
Q. Where can I get SAR images for free?
A. You can download Sentinel-1 images after registration from the Copernicus Data Space of ESA (European Space Agency). The terms and the way to access may change with platform policy, so check the latest terms of use on the site before you use it.
What comes next
The reason water looks dark in a SAR image is physically simple: a calm water surface bounces the radar signal far away. When you add a comparison procedure called change detection on top of this simple principle, it becomes a practical tool for estimating flooded extent even in a cloudy disaster situation. However, false detections caused by shadows and smooth ground are hard to filter out completely by automated analysis alone, and a step in which a person reviews the results together with terrain and land cover data is essential. Once you have found the flooded extent pixel by pixel, the next question follows naturally: if you add a deep learning model to this flooded area, can it detect it faster and more accurately?
This is an English adaptation of a post in the Korean GeoAI basics series. The original is here, and more learning materials are on GeoAI for Disaster Safety.
This post was drafted with the help of AI tools and reviewed by the author. Tool features and data portals change over time, so check the official documentation when you practice.
Comments
Post a Comment