Hyperspectral remote sensing image classification using active learning

Abstract

Hyperspectral remote sensing images capture a large number of narrow spectral bands ranging between visible and infrared spectrum. The abundant spectral data provides huge land cover information that helps in accurate classification of land use land cover of earth’s surface. However, obtaining labelled training data in hyperspectral images (HSIs) is labour expensive and time consuming. Therefore, designing a classifier that uses fewer labelled samples as possible for classification is highly desirable. Active learning (AL) is a branch of machine learning that finds most uncertain samples in an iterative way from unlabelled dataset resulting relatively smaller training set to achieve adequate classification accuracy. Support vector machine (SVM) has been extensively used as a classifier in AL approach. However, it has high computational complexity. Recently, a non-iterative learning algorithm based on least square solution known as extreme learning machine (ELM), has been integrated in AL framework for HSI classification. It provides a comparable classification accuracy while reducing computation time drastically.