U-Net Based Classification for Urban Areas in Algeria

Abstract

Nowadays, researchers in the field of remote sensing and image classification have to face the challenge of the massive amount of information contained in satellite images, especially in urban areas. These types of areas contain numerous classes, where each class is made of several groups of pixels that are not adjacent, and that are rich in texture. Convolutional Neural Networks possess the ability to handle these problems. However, CNNs require quite a very large number of annotated training samples. U-Net came as a revolutionary solution for this major drawback. This paper aims to study the ability of a pre-trained U-Net to classify a satellite image and is then compared to the performance of a Support Vector Machine classifier