Unsupervised Change Detection in Heterogeneous Remote Sensing Imagery

Abstract

<p><i>Change detection</i> is a thriving and challenging topic in remote sensing for Earth observation. The goal is to identify changes that happen on the Earth by comparing two or more satellite or aerial images acquired at different times. Traditional methods rely on homogeneous data, that is, images acquired by the same sensor, under the same geometry, seasonal conditions,and recording configurations. <p>However, the assumption of homogeneity does not hold true for many practical examples and applications, and in particular when different sensors are involved. This represents a significant limitation, both in terms of response time to sudden events and in terms of temporal resolution when monitoring long-term trends. <p>The alternative is to combine heterogeneous data, which on one hand allows to fully exploit the capabilities of all the available sensors, but on the other hand raises additional technical challenges. Indeed, heterogeneous sources imply different data domains, diverse statistical distributions and inconsistent surface signatures across the various image acquisitions. <p>This thesis tries to explore the kinds of techniques meant to cope with these issues, which are referred to as heterogeneous change detection methods. Specifically, the effort is dedicated to unsupervised learning, the branch of machine learning which does not rely on any prior knowledge about the data. This problem setting is as challenging as important, in order to tackle the task in the most automatic way without relying on any user interaction. <p>The main novelty driving this study is that the comparison of affinity matrices can be used to define crossdomain similarities based on pixel relations rather than the direct comparison of radiometry values. Starting from this fundamental idea, the research endeavours presented in this thesis result in the formulation of three methodologies that prove themselves reliable and perform favourably when compared to the state-of-the-art. These methods leverage this affinity matrix comparison and incorporate both conventional machine learning techniques and more contemporary deep learning architectures to tackle the problem of unsupervised heterogeneous change detection.