Active learning for detecting a spectrally variable subject in color infrared imagery

To classify Egeria densa, Brazilian waterweed, in scan-digitized color infrared aerial photographs, we are developing an interactive computer system based on datamining techniques with Active Learning capabilities. Key components of the system are: feature extraction, automatic classification, Active Learning, and experimental evaluation. Key words—active learning, data mining, feature extraction, CIR imagery. I. BACKGROUND AND PROBLEM The need for airborne data collection, including aerial photography, still exists for many applications. Improvements in orthorectification techniques have allowed aerial photography, recent and historic, to be digitized, geometrically corrected, and integrated into databases more rapidly and cost-effectively. However, human visual/manual image interpretation and analysis procedures are often time-consuming and costly, not repeatable, and dependent on the varying abilities of the interpreters. Individual expertise is hard to transfer from one interpreter to another, which contributes to high training costs. There is a significant gap between fast routine data collection and the slow interpretation and analysis of complex and detailed images in multidimensional (spatial and temporal) space. Monitoring Egeria densa, an invasive submergent weed, by remote sensing represents such a case. Egeria densa, commonly called Brazilian waterweed, has grown uncontrolled in the Sacramento-San Joaquin Delta of Northern California for over 35 years displacing native flora and now covers about 2400 hectares of waterways. The presence of this exotic weed is