Going Deep into Remote Sensing Spatial Feature Learning

Abstract

A lot of information may be extracted from the Earth’s surface through aerial images. This information may assist in myriad applications, such as urban planning, crop and forest management, disaster relief, etc. However, the process of distilling this information is strongly based on efficiently encoding the spatial features, a challenging task. Facing this, Deep Learning is able to learn specific data-driven features. This PhD thesis1 introduces deep learning into the remote sensing domain. Specifically, we tackled two main tasks, scene and pixel classification, using Deep Learning to encode spatial features over high-resolution remote sensing images. First, we proposed an architecture and analyze different strategies to exploit Convolutional Networks for image classification. Second, we introduced a network and proposed a new strategy to better exploit multi-context information in order to improve pixelwise classification. Finally, we proposed a new network based on morphological operations towards better learning of some relevant visual features.