Detection of Smoke in Satellite Images Using Autologistic Regression
Datetime：2014-04-14 Clicks：423
Title: Detection of Smoke in Satellite Images Using Autologistic Regression
Speaker: Dr. Mark Wolters (University of Western Ontario)
Time: 4:15pm, 18 April 2014
Speaker: Dr. Mark Wolters (University of Western Ontario)
Time: 4:15pm, 18 April 2014
Place: Room 1018, School of Management
Hosted by Department of Statistics and Finance
Abstract: Satellite imagery provides a rich data source for studying the spread of smoke produced by forest fires. The present work considers hyperspectral images, where each of 36 image planes holds a measurement in a different spectral band. The objective is to develop an automatic image segmentation routine: a classifier that assigns a new image's pixels to either the "smoke" or "nonsmoke" category. To incorporate association between nearby pixels, the true scene is modeled as a binary random field. An autologistic regression approach is used to model the joint probability mass function of the scene, with the hyperspectral data as predictors. This model poses computational challenges for feature selection, estimation, and prediction; methods for overcoming these challenges will be discussed.
Abstract: Satellite imagery provides a rich data source for studying the spread of smoke produced by forest fires. The present work considers hyperspectral images, where each of 36 image planes holds a measurement in a different spectral band. The objective is to develop an automatic image segmentation routine: a classifier that assigns a new image's pixels to either the "smoke" or "nonsmoke" category. To incorporate association between nearby pixels, the true scene is modeled as a binary random field. An autologistic regression approach is used to model the joint probability mass function of the scene, with the hyperspectral data as predictors. This model poses computational challenges for feature selection, estimation, and prediction; methods for overcoming these challenges will be discussed.
News
