Skip to content
My final project for the Metis Data Science Bootcamp, satellite image segmentation, in partnership with Digital Globe.
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
images
FCN-VGG19-master.ipynb
FCN-resnet-master.ipynb
MetisNet-master.ipynb
Prepare-Train-Val-Sets.ipynb
README.md
U-Net-master.ipynb
download-ground-truth-from-osm-shanghai.ipynb
exploration_and_build_ground_truth.ipynb

README.md

satellite-image-segmentation

This was my final project at the Metis Data Science Bootcamp. It was done as part of a partnership with Digital Globe, utilizing images from their WorldView-3 satellite to perform image segmentation.

Project Goal

In my project, I performed image segmentation on satellite imagery from Shanghai in an attempt to map out the location of farmland. In the course of the project, I was able to extend to segmenting multiple classes, but for the purposes of the presentation, my focus was on farmland.

I decided to map out farmland in Shanghai because I feel it highlights a powerful aspect of satellite image segmentation. It gives us the ability to track changes in critical changes in environment and landscape over time. As a city that used to be surrounded by farmland, which has exploded in population in recent years, Shanghai was a natural city to begin my exploration.

What is Image Segmentation?

Many people are familiar with computer vision applications in image classification, in which the program determines what object is represented in an image. This requires predicting a single value for each image.

With segmentation, however, we need to predict a representation for each individual pixel.

As with classification, convolutional neural networks are commonly used for segmentation tasks. It's also standard practice to utilize classification architectures (minus the final output layers) to develop a sense of what is in an image.

By converting any dense layers to convolutional, we're also able to get a better sense of spatial representation for which classes are most probable in each individual region. Unfortunately, the pooling layers of the classification phase leave us with only a few large regions on which to predict. To get more precise layers, we need to add several upsampling (or transverse convolutional) layers to our model.

You can see some of the results of this process below.

Workflow

To accomplish my objective, I had to:

  1. Locate images containing farmland using Open Street Maps data.
  2. Create ground truth masks using QGIS.
  3. Train several convolutional neural networks to predict the segmentation.
  4. (Optionally) ensemble model predictions.

The models used are available in this github repository.

Some Segmentation Results

| | |

You can’t perform that action at this time.