Transfer learning for image information mining applications

Abstract

The recent explosion of data from various Earth observation (EO) systems requires new ways to rapidly harness the information and synthesise it for decision-making. Currently, several image information mining (IIM) systems have some form of supervised statistical learning models that relate the image content to the various semantic classes. However, this kind of approach is constrained by the paucity of training information in several EO domains due to limited ground truth. Although semi-supervised learning methods alleviate this problem to a certain extent by using unlabelled data from various spatial databases, these methods require that the training data and future unseen data should conform to the same statistical distribution and feature space. To overcome this problem, a more recent approach, known as transfer learning, is focused on using small amounts of labelled information from closely related or similar learning task and somehow adapts that information in developing new semantic models. Transfer learning can be applied in several processes of supervised and unsupervised learning. In this article, we propose transfer learning methods for IIM and discuss various techniques and their implications for content-based retrieval in the EO domain. Specifically, we explore three domains; the first two illustrate the heterogeneity in land cover and coastal zone thematic representations, and the third is on rapid disaster response during coastal events. This study demonstrates the usefulness of knowledge transfer between closely related tasks in the EO data synthesis. The adopted methodology for knowledge transfer is based on harnessing prior knowledge from similar concepts to learn new ones and uses a modified weighted least squares support vector machine.