Interactive visualisation techniques for data mining of satellite imagery

Abstract

Supervised classification of satellite imagery largely removes the user from the information extraction process. Visualisation is an often ignored means by which users may interactively explore the complex patterns and relationships in satellite imagery. Classification can be considered a 'hypothesis testing' form of analysis. Visual Data Mining allows for dynamic hypothesis generation, testing and revision based on a human user's perception. In this study Visual Data Mining was applied to the classification of satellite imagery. After reviewing appropriate techniques and literature a tool was developed for the visual exploration and mining of satellite image data. This tool augments existing semi-automatic data mining techniques with visualisation capabilities. The tool was developed in IDL as an extension to ENVI, a popular remote sensing package. The tool developed was used to conduct a visual data mining analysis of high-resolution imagery of Heard Island. This process demonstrated the positive impacts of visualisation and visual data mining when used in the analysis of satellite imagery. These impacts consist of: increased opportunity for understanding and hence confidence in classification results, increased opportunity for the discovery of subtle patterns in satellite imagery and the ability to create, test and revise hypotheses based on visual assessment.