Knowledge Discovery for Classification with Mixed Spatial Data Types - A Rough Set Approach

Principal Investigator:
Leung Yee

Co-investigator(s):
Manfred M. Fischer, Zhang Wenxiu

Summary:
Classification has long been a corner stone in spatial analysis. Its significance becomes even more prominent with the availability of large volume of geo-referenced data captured in geographic information systems (GIS) and remotely sensed images. Being able to discover non-trivial, previously unknown and potentially useful knowledge from a data set for a specific classification task, such as land covers in hyperspectral images, is thus of great importance for real-life applications. Methods such as statistics and fuzzy sets have to rely on external parameters and prior model assumptions, e.g. probability distributions in statistics and membership functions in fuzzy sets. Rough set, on the other hand, only uses internal knowledge embedded in a raw information system to discover classification rules. Out of all features (attributes) employed for a classification, rough set models can automatically select the minimal set of features necessary and sufficient for a classification task. It is especially instrumental in hyperspectral analysis where a very large number of spectral bands is employed for image analysis. Through the process of knowledge reduction, the rough set approach can also discover the optimal set of rules. This can sharpen our knowledge and reduce the dimension and complexity of a classification task.

In this research, we will develop novel rough set models capable of discovering knowledge in (1) purely qualitative, (2) purely quantitative, and (3) mixed spatial databases. The approach generalizes existing rough set models and will advance the research frontier of rough set in general and spatial data mining and knowledge discovery in particular. To validate and evaluate, we will develop efficient algorithms for the implementation of the proposed rough set models. A real-life application in hyperspectral classification with mixed data types will be made for substantiation and assessment.