HOW TO MAP FLOODWATER FROM RADAR IMAGERY USING SEMANTIC SEGMENTATION - BENCHMARK¶

Welcome to the benchmark solution tutorial for our newest competition run in partnership with Microsoft AI for Earth and Cloud to Street.

In this challenge, you are tasked with detecting the presence of floodwater in Sentinel-1 global synthetic aperture radar (SAR) imagery. While traditional optical sensors measure reflected solar light, SAR operates in the microwave band of the electromagnetic spectrum. It is therefore uniquely positioned to provide a day-and-night supply of images of the Earth's surface through the coverage of clouds, smoke, vegetation, and precipitation. With the increasing prevalence of extreme weather events due to climate change, your machine learning models and insights have the potential to strengthen flood mapping algorithms and improve disaster risk management efforts around the world.

The training data for this competition consist of 542 "chips" from flood events around the world, which each contain two polarization bands. Each chip has a corresponding label mask that indicates which pixels in a scene contain water. You can optionally supplement this training data with information on permanent water sources and elevation, available through the Microsoft AI for Earth STAC API (see the competition STAC resources page for more information on bringing in this data). To enter this challenge, you will package and submit your code to generate predictions as single-band tifs with binary pixel values.

In this post, we'll cover:

• Exploring the data
• Preparing the data
• Explaining the benchmark model
• Fitting the benchmark model
• Generating a submission in the correct format

Let's get started!

Explore the Data¶

Metadata¶

First we'll download the training data to a new subdirectory called training_data/ in our current working directory. If you are working on the Planetary Computer Hub, consider using Azure Blob Storage to store your data using the West Europe region, which is where your server is running. See Writing data to Azure Blob Storage for an example.

$ tree training_data/
training_data/
├── flood-training-metadata.csv
├── train_features
│   ├── awc00_vh.tif
│   ├── awc00_vv.tif
        ...
│   ├── wvy31_vh.tif
│   └── wvy31_vv.tif
└── train_labels
    ├── awc00.tif
    ├── awc01.tif
        ...
    ├── wvy30.tif
    └── wvy31.tif

2 directories, 1627 files

To better understand what's in our data, we can begin by exploring trends across different locations and flood events. Let's load the metadata and look at what we have.

The training data for this competition consist of two polarization bands that bring out distinct physical properties in a scene: VV (vertical transmit, vertical receive) and VH (vertical transmit, horizontal receive). Each band is identified by a unique image_id, which is equivalent to {chip_id}_{polarization}. A chip_id consists of a three letter flood_id and a two digit chip number. Each unique chip has both a _vv and _vh band. For example, awc00_vh represents the vv band of chip number 00 for flood awc.

Let's take a look at how many flood events are in the training data and how many chips we have per event.

The training data include chips from 13 flood events. We have anywhere from 15 to 69 chips (30 to 138 images) per unique event, with half of events containing fewer than 32 chips (64 images).

We can also take a look at the distribution of chips by location.

The data cover a wide geographic range. We have more than 60 chips (120 images) for floods in the United States, India, Paraguay, and Slovakia.

Let's also take a look at when the images were originally captured.

The training data cover flood events that occurred between 2016 and 2020. Images for each event were captured on the same day.

Images¶

Next, we can begin exploring the image data!

A GeoTIFF is a raster image file that contains geographic metadata describing the location of the image. This metadata can include bounding coordinates, an affine transform, and a coordinate reference system (CRS) projection. The package rasterio makes it easy to interact with our geospatial raster data.

Recall that each set of two polarizations (vh and vh) correspond with a single water label.

We will need to be able to identify pixels with missing data, since we will only be evaluated on predictions made for valid input pixels. The metadata tells us that a value of 0.0 represents missing data for an input image. In rasterio, you can access two different kinds of missing data masks. The first mask is a GDAL-style mask, in which non-zero elements (typically 255) indicate that the corresponding data elements are valid.

The second mask is a numpy masked array, which has the inverse sense: True values indicate that the corresponding data elements are invalid. To load the data as a numpy masked array and access this type of missing data mask, simply pass a masked flag to read.

Pixel values represent energy that was reflected back to the satellite measured in decibels. To better visualize the bands or channels of Sentinel-1 images, we will create a false color composite by treating the two bands and their ratio as red, grean, and blue channels, respectively. We will prepare a few helper functions to visualize the data.

Let's inspect a few chips and their water labels.