Hyperclass: A Framework for Graph-Based Semi-Supervised Classification of Hyperspectral Imagery

Hyperclass is an interactive workbench supporting visual data analysis of sensor data. It provides an extendable interactive interface and toolsuite which can be used to jumpstart the development of novel methods for addressing data analysis challenges in the earth and space sciences. This talk presents a number of hyperclass applications, including innovative semi-supervised machine learning methods for landscape classification using hyperspectral imagery (HSI).

Hyperspectral sensors provide a vast amount of spectral and spatial information, which is exploited in HSI classification applications for agriculture, environmental management, urban planning, mineral detection and urban mapping. HSI classification deals with the problem of pixel-wise labeling of the hyperspectral spectrum. While huge volumes of HSI data are increasingly becoming available, ground truth labels remain scarce, due to the immense manual efforts required to collect them. Semi-supervised machine learning (SSL) is an effective method for addressing the shortage-of-labels problem.

The Hyperclass SSL algorithm uses a relatively small number of interactively labeled points to infer the labels of other similar points, thus greatly increasing the number of training samples for the classifier. It represents an image as a collection of spatially-referenced spectra forming a (relatively low dimensional) manifold within the (high dimensional) spectral metric space. The SSL algorithm expands, with interactive user feedback, a small set of labels by adding nearby points on the data manifold, which is represented by a nearest-neighbor graph. A support vector machine is then used to construct a classification mapping using the expanded set of labels.

The Hyperclass framework enables this analysis and training process by providing a linked set of interactive visualization panels, including a 3D visualization of the UMAP embedding of the data manifold, spatial plots of the HSI bands, spectral plots for selected points, and satellite views of selected regions. Applications include studies that detect and quantify changes in land cover and land use; examine their impact on the environment, climate, and society; or model future scenarios of LCLUC and its various impacts and feedbacks.

*