All You Need to Know about Learning with Limited Labels for Remote Sensing

OVERVIEW

In the last years, machine learning has experienced an astonishing development in all domains. The advent of Deep Learning (DL) – since the pioneering work of Hinton in 2012 – changed the perspective of the community, adopting in this way DL as the go-to technique for different remote sensing tasks such as classification, segmentation and detection. However, a major drawback of these techniques is the high dependence on a large and well-representative corpus of labelled data. In several real world problems, this might be a strong assumption for a solution, as annotated data contains strong human bias, and for many domains, including remote sensing, it is expensive and time consuming to obtain labels. Motivated by these drawbacks, different paradigms that rely on less labels have experience a fast development including Semi-Supervised Learning and Weakly Supervised Learning, in which one aims to exploit the inherent relationship between a very small set of label data and a huge amount of unlabelled data. In this tutorial, we aim to draw attention to current developments in learning with fewer labels for remote sensing data analysis. We will start by introducing the topic and giving an overview of the body of the literature in the area. We will then present current challenges when dealing with remote sensing data including video street level, hyperspectral and time series data. We will close our tutorial by summarising the current challenges and opportunities in this domain. Some open questions related to the topic will also be discussed in the end.

Registration