Lab 27: Mapping Floods Through the Clouds

Part 1: Setting the scene

Cyclone Idai made landfall near Beira, Mozambique on March 14, 2019. Let's define our study area over the low-lying floodplain where the Buzi and Pungwe rivers overflowed.

// Study area: Beira and surrounding floodplain, Mozambique
var beira = ee.Geometry.Point([34.87, -19.84]);
var studyArea = beira.buffer(60000); // 60 km radius

Map.centerObject(studyArea, 9);
Map.addLayer(studyArea, {color: 'yellow'}, 'Study area');

Part 2: Building our "before" baseline

Let's create a clean baseline image from before the cyclone hit. We will take the temporal median of Sentinel-1 images from February 2019. Using a median reduces the "salt and pepper" speckle noise that is common in radar images.

// Helper function to apply standard Sentinel-1 filters
function filterS1(collection) {
  return collection
    .filter(ee.Filter.eq('instrumentMode', 'IW'))
    .filter(ee.Filter.listContains('transmitterReceiverPolarisation', 'VV'))
    .filter(ee.Filter.listContains('transmitterReceiverPolarisation', 'VH'))
    .filter(ee.Filter.eq('orbitProperties_pass', 'DESCENDING'));
}

// Before image: February 2019 (pre-cyclone baseline)
var beforeCollection = filterS1(
  ee.ImageCollection('COPERNICUS/S1_GRD')
    .filterBounds(studyArea)
    .filterDate('2019-02-01', '2019-02-28')
);

var before = beforeCollection.select('VV').median();
print('Before images:', beforeCollection.size());
Map.addLayer(before, {min: -25, max: 0}, 'Before (Feb 2019) VV');

Part 3: Capturing the flood

Now let's load the Sentinel-1 image acquired during peak flooding (March 19 to 25, 2019). We use a single mosaic rather than a median here, because we want to capture the actual flood state. A median would smooth it away.

// During image: March 19-25, 2019 (peak flooding)
var duringCollection = filterS1(
  ee.ImageCollection('COPERNICUS/S1_GRD')
    .filterBounds(studyArea)
    .filterDate('2019-03-19', '2019-03-25')
);

var during = duringCollection.select('VV').mosaic();
print('During-flood images:', duringCollection.size());
Map.addLayer(during, {min: -25, max: 0}, 'During flood (Mar 2019) VV');

Part 4: Spotting the flood with subtraction

Here is the core idea: we subtract the during-flood image from the before image. Wherever the backscatter dropped significantly (land became smooth water), the difference will be large and positive. We set a threshold to create our flood mask. Think of it like finding everywhere the "volume" of the radar signal went quiet.

// Change detection: before minus during
var difference = before.subtract(during);

// Visualize the difference (positive = backscatter decreased = possible flood)
Map.addLayer(difference, {min: -5, max: 10, palette: ['blue', 'white', 'red']}, 'Backscatter difference');

// Apply threshold: backscatter dropped by more than 3 dB = flooded
var floodMask = difference.gt(3);
Map.addLayer(floodMask.selfMask(), {palette: ['cyan']}, 'Flood extent (threshold = 3 dB)');

Part 5: How much land was flooded?

Let's put a number on the destruction. We will use ee.Image.pixelArea() to calculate the total flooded area in square kilometers.

// Calculate flooded area in square kilometers
var floodArea = floodMask.multiply(ee.Image.pixelArea());

var stats = floodArea.reduceRegion({
  reducer: ee.Reducer.sum(),
  geometry: studyArea,
  scale: 10,
  maxPixels: 1e9
});

var areaKm2 = ee.Number(stats.get('VV')).divide(1e6);
print('Flooded area (km²):', areaKm2);

Part 6: Cleaning up our flood map

Our raw flood mask may include areas that are always water (rivers, lakes). Let's remove those using the JRC dataset so we only see the new flooding caused by the cyclone. We will also add a Sentinel-2 true-color image for context.

// Remove permanent water using JRC Global Surface Water
var jrc = ee.Image('JRC/GSW1_4/GlobalSurfaceWater');
var permanentWater = jrc.select('occurrence').gt(50);

// Refined flood: only areas that are NOT permanent water
var floodRefined = floodMask.updateMask(permanentWater.not());
Map.addLayer(floodRefined.selfMask(), {palette: ['ff6600']}, 'Flood (excluding permanent water)');

// Add true-color context (Sentinel-2 before the event)
var s2Before = ee.ImageCollection('COPERNICUS/S2_SR')
  .filterBounds(studyArea)
  .filterDate('2019-02-01', '2019-02-28')
  .sort('CLOUDY_PIXEL_PERCENTAGE')
  .first()
  .divide(10000);
Map.addLayer(s2Before, {bands: ['B4', 'B3', 'B2'], min: 0, max: 0.3}, 'Context (S2 true color)');

Key Takeaways

• SAR can map floods through clouds, which is critical for disaster response.
• The standard workflow is: baseline (temporal median), event (single image), difference, threshold.
• VV polarization is most effective for detecting open water.
• A 3 dB backscatter drop is a reasonable initial flood detection threshold.
• Removing permanent water bodies with JRC GSW isolates event-specific flooding.
• SAR and optical data are complementary: use SAR for flood detection, optical for context.