Using Deep Networks for Semantic Segmentation of Satellite Images

摘要

In this paper we aim to investigate different deep learning techniques for automatic extraction of valuable information from large sized satellite image data. We focus on the problem of semantic segmentation which attaches a class label to each pixel from the image. We investigate two semantic segmentation architectures based an convolutional neural networks: segnet and u-net. We analyse different tiling strides with reverse aggregation methods. We compare two classical methods (averaging and maximum) and propose a new method based on entropy. We test the models with distinct types of images, emphasizing the need to predict the results using information from all of them. We discuss various fusion strategies and introduce a fusion strategy based on the observations obtained from separately analysing the distinct image types.