|
Abstract:
|
Classification of geographical regions has an important impact on various planning and conservation efforts. For example, governments allocate and monitor forestry regulations based on the perceived amount of resources available. Often, these decisions are made using remote sensing data, a method of surveying land via satellite or remote sensors. To classify land pixels, i.e., differentiate subsets of forest cover, a reference data set is created. This is often done at the ground level, by experts surveying different areas. Classification rules, such as discriminant analysis, are then trained using the reference set. Usually, this is done under the assumption that label assignment is free from misclassifications. Forested areas also change over time through disturbances such as fire, insects, disease, and harvesting. This presentation will explore methods of classification when group labels are assigned stochastically. Solutions for a geographical classifier where imperfect class labels exist will be shown. Probabilistic models for training the classifier will be presented under frequentist and Bayesian perspectives.
|
