Developing a Seamless Integration Between Machine Learning Techniques and Rule-Based Classification of Remotely Sensed Imagery

The proposed system will give an approach for increasing the speed, accuracy, and validation of feature extraction models while simultaneously creating an environment that allows both the flexibility for power users and the simplicity for the majority ofusers (novice users that can be trained and effective in a day) - and a mechanism that allows the two levels of users to seamlessly work together and communicate. In Phase I, we have demonstrated the value of converting black box learners into open(human-understandable) learners in the form of decision trees and the rule sets that are derived from them. While classification accuracy has always been of utmost importance, speed has become an increasingly important issue with the proliferation ofterabyte- and even petabyte-sized digital image databases. Theory refinement has the advantages of both rule-based expert systems and inductive learning, and takes advantage of uniquely human perception in extracting features from imagery. Mostimportantly, it makes feature extraction models more portable and reduces expert user time and effort in porting models to new and different problem domains. Geospatial information technology is the cornerstone that supports the United States defense andintelligence mission to maintain information superiority over enemy forces. The connectionist theory refinement system proposed here advances VLS' award winning Feature Analystr software by increasing the speed, accuracy, and portability of AFE models.The convergence between GIS software technology and the high-resolution commercial satellite imagery market provides substantial opportunities for the Feature Analyst technology in the following market spaces:1. Defense and Intelligence solutions to support the war on global terrorism. The events of September 11, 2001 have created a heightened sense of awareness on the value of timely and accurate GIS data. The DoD budget for 2004 and beyond will includesubstantial funds going towards improvements in intelligence gathering including advancements in automated feature extraction and target recognition.2. Homeland Security for creation and maintenance of GIS data layers. The newly created Office of Homeland Security has a budget of over $30 billion. Identification and mapping of high-value assets (pipelines, power plants, etc), monitoring of borders,and preparation of disaster and emergency services all require GIS mapping.3. Forestry to support timber management applications, wildfire modeling, and land-use analysis.4. Civil Government applications such as pervious-impervious surface mapping, creation and maintenance of GIS data layers for roads and structures, identification of urban