HR: 1340h
AN: GC13A-0944    [Abstracts]
TI: A Megamodel-Classifier in Support of a Sampling Strategy Using Landsat Data to Estimate Deforestation in the Brazilian Amazon
AU: * Broich, M
EM: Mark.Broich@sdstate.edu
AF: Geographic Information Science Center of Excellence (GIScCE), South Dakota State University, 1021 Medary Ave, Wecota Hall, Box 506B, Brookings, SD 57007, United States
AU: Hansen, M C
EM: Matthew.Hansen@sdstate.edu
AF: Geographic Information Science Center of Excellence (GIScCE), South Dakota State University, 1021 Medary Ave, Wecota Hall, Box 506B, Brookings, SD 57007, United States
AB: The rainforest of the Brazilian Amazon is subject to high rates of deforestation. Estimates of deforestation are needed for management decisions and policy support. A strategy that provides timely estimates of basin-wide forest cover change would be beneficial to policy makers and land managers alike and provide useful information more quickly than current, large-scale efforts. This study employed a megamodel based on a decision tree algorithm using Landsat data from 2000 and 2005 in an automated, sampling block pair approach to estimate deforestation for the Brazilian Amazon. The megamodel, unlike locally derived models, is generic and can systematically characterize a large number of sampling block pairs automatically over large geographical regions. We trained the megamodel on a subset of pixels taken from sampling block pairs that were interpreter-classified into deforested area and unchanged forest. We assessed the megamodel in a three stage procedure: (1) We tested the model on the sampling block pairs that we derived the training from, (2) we used a bootstrapping approach to evaluate the classification accuracy, and (3) we applied the megamodel to new sampling block pairs and evaluated the classification accuracy against PRODES digital deforestation estimates. The first test of the megamodel resulted in 96% overall classification accuracy suggesting that the megamodel is a robust classifier capable of producing accurate estimations of forest cover change for the Brazilian Amazon. Misclassification occurred (A) because of spectral classes that were underrepresented in the training and (B) because of spectral classes that the interpreter labeled arbitrarily from one to another sampling block pair in the training. The Brazilian Amazon provides a test bed for the megamodel classifier because deforestation and unchanged forest thematic classes are spectrally homogenous across the forested area of the basin.
DE: 1632 Land cover change
DE: 1640 Remote sensing (1855)
SC: Global Environmental Change [GC]
MN: 2007 Fall Meeting