CS student awarded 1st place in GSA Annual Research Symposium

Publish Date: 04/07/2009

Rhonda Phillips, a Computer Science PhD candidate under Dr. Layne Watson, was awarded 1st place in the Engineering category of the recent GSA Annual Research Symposium. Ms. Phillips presented a poster entitled “A Probabilistic Classification Algorithm with Soft Classification Output.” The poster presented her research on using the continuous iterative guided spectral class rejection (CIGSCR) classification method to classify remotely sensed images.

The classification of remotely sensed images is essential to many applications such as land use monitoring, natural resource management, and global climate change analysis and prediction. However, classification is often difficult because training data are difficult and expensive to acquire, and identifying good training classes within those data is problematic.

The CIGSCR classification method addresses these issues by using semisupervised techniques in machine learning to supplement limited training data and automatically locate training classes in the form of clusters. While this technique (clustering to produce classification) is not new, the method of determining which clusters should be used and refining “rejected” clusters is unique to CIGSCR and its predecessor, IGSCR. CIGSCR features a statistical hypothesis test used to select clusters that are representative of classes in the classification scheme. Once a cluster has been targeted for refinement, CIGSCR uses this continuous data to seed additional clusters. Upon termination of the algorithm, multiple clusters associated with classes in the classification scheme are located, enabling classification.

The classification results indicate that CIGSCR has advantages over IGSCR, improves classification results based on clustering alone, and produces high quality, informative soft classifications of remotely sensed data.