A keras segnet implementation for building detection in the spacenet dataset

The spacenet challenge

A few weeks ago topcoder, a website hosting programming competitions, launched the spacenet challenge. Sponsored by cosmiq works, digitalglobe and nvidia, its goal was to detect buildings in satellite images and build the polygons describing their contours.
We will discuss here of one possible solution to this problem using deep learning techniques.

Automatic detection of points of interests in satellite images is a great challenge because it is both technically difficult and can have a significant impact. Indeed, until a reliable process can be set up, a major part of those analyses is performed manually, which is time and money consuming. Also, this is something that we want to use at Qucit in order to automatically enrich our contextual data.

The process of classifying each part of an image in different categories is called “image segmentation”. A satellite image can be segmented into parts representing either man made structures such as buildings and roads, natural structures such as forests, fields and beaches or dynamic events such as roadworks and gatherings. Here we will just segment images in two categories: building and nonbuilding.

The segnet neural network

Currently the very best method used for image segmentation is the segnet neural network architecture, a deep learning techniques based on a set of several artificial neural networks architectures.