Interactive Training of Advanced Classifiers for Remote Sensing Image Analysis

Incorporating supplemental GIS information and human expert knowledgeinto digital image processing has been acknowledged as a necessityfor improving remote sensing image analysis. However, no commercialimage analysis system can currently translate expert knowledgeautomatically to a computer-usable format. The objective of this projectis to develop a system for seamless and transparent integration ofmachine learning algorithms for interactive classification of remotelysensed imagery. We are proposing to develop a hierarchical scenerepresentation to create pixel level, region level and scene levelland cover models. This hierarchical scene representation aims to bridgethe gap between low-level image data and high-level expert interpretation.Each level in the representation will include classification modelsto create rules that will be used in subsequent levels. Spectral values,ancillary GIS layers and other image measurements like textural featureswill be used in learning. Easy-to-use interfaces will support user relevancefeedback and enable rule generation with a small amount of training data.The system will be evaluated using quantitative and qualitative criteriaon independent training and test data. The final product will be a machinelearning library with a powerful graphical user interface that has fullinput/output compatibility with existing COTS software packages.The proposed research on the hierarchical scene representation willbridge the gap between low-level image data and high-level expertinterpretation, and will significantly benefit the remote sensingcommunity in discovering information that is not evident in pixellevel data. We will also extend the application domain of ourclassification software to biomedical image analysis systems byproviding tools for automatically transfering expert knowledge tomedical knowledge bases.