Presentation / Speaker Listing
View All Presentations | Report a Problem With This Abstract
Tuesday 18 October 2005: Session 1, 10:30:00 AM
Per-segment Decision Tree Classification in the Arkansas 2004 Land Cover Project
Presentation Abstract
Multi-resolution image segmentation can enhance the remote sensor data classification process by taking advantage of spatial autocorrelation and by facilitating derivative landscape ecology metrics. An important research area is implementation design of inductive classification tools, especially in the automated classification of image segments. While machine learning decision trees have proven useful for predicting classes in non-normally distributed data, the optimization of these predictive models remains a challenge. The greater the number of optional input variables (e.g. attribute fields in a GIS data table) and image segmentation structures, the greater the need for this optimization. The Arkansas 2004 Land Cover Project, based on three seasons of Landsat 5 imagery and ancillary datasets demonstrated a flexible implementation of the C5.0 machine learning decision tree with the Arcview 3.x Avenue customization environment. A significant advantage over other similar decision tree applications was an auxiliary relational database (separate from the input segment and associated attribute datasets) used to store global model parameters and accuracy results. The expanding number of parameters and scenarios that might be tested for a given parameter set implies the need for this higher organization level in classification implementation. Optimization cannot reasonably be attempted without proper tracking of parameters and storage of results. Additionally, a relational database should permit secondary data mining opportunities (“the most accurate models favor variable x”). Successes (e.g. whole fields of corn accurately classified) and challenges (e.g. dilution of some spectral signatures in the spatial aggregation process) relating to the Arkansas 2004 per-segment classification are presented.
Speaker Biographical Information
Jackson Cothren Assistant Professor, Geosciences: RGIS Mid-South Center for Advanced Spatial Technologies (CAST) - University of Arkansas
Dr. Jackson Cothren is an Assistant Professor in the Department of Geosciences at the University of Arkansas and a research scientist at the Center for Advanced Spatial Technologies (www.cast.uark.edu), also at the University of Arkansas. He has worked extensively with Quickbird imagery as a source of imagery to update aging orthophotos and quickly generate impervious surface maps. Dr. Cothren has worked with high school students across seven states to introduce fundamental and advanced GIS and remote sensing technical and planning skills as a member of the EAST Geospatial Support team at CAST (www.cast.uark.edu/cast/east and www.eastproject.org), develop curriculum for more than seven different two to five day short courses and visited more than 50 EAST labs in schools across California and Arkansas. Before joining CAST and the University of Arkansas, Dr. Cothren spend almost 12 years as an Air Force officer and as a civilian photogrammetric engineer working for the Air Force responsible for the direction and management of all photogrammetric and geodetic research at the National Air Intelligence Center. He has worked as professional consultant to private industry to develop digital photogrammetry software.
