Effective Exploitation of Related Domain Knowledge for Large-Scale Sensor Data
In this project, we seek to develop new automatic methodologies for effective exploitation of knowledge in large-scale sensor data with emphasis on spatial information. This work addresses a common problem that occurs when systems create knowledge from complex sensor data: there are not enough labeled examples of the phenomena of interest for an automated system to learn to accurately classify and label it. We will research knowledge adaptation that leverages unlabeled data through exploitation of knowledge in multiple related domains. This is done through a proposed structural adaptive regression algorithm that exploits shared classification information by explicitly imposing a regularization term that enforces classification structure sharing between the two domains. The fundamental results will be applied and validated in specific applications emphasizing reducing the labor required to build models for analyzing imagery of geographical scenes and a concomitant increase in classification accuracy. Our proposed work will develop cross-sensor processing techniques to extract descriptive metadata information using all components of multiple sensor data.