Date of Award

Spring 5-1-2005

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Geography, Geology, and Anthropology

First Advisor

Ryan Jensen

Second Advisor

Susan Berta

Third Advisor

James Speer

Abstract

This research compared a traditional statistical technique with artificial neural networks to determine their effectiveness for modeling LAI in the Santarem region of the Brazilian Amazon. In addition, different satellite sensors in the form of ETM+, ASTER, MO DIS, and IKONOS were used to assess if spatial and spectral resolutions play a role in the ability to estimate LAI. This study focused on the following: (1) Determining if neural networks perform more accurately than multiple regression for LAI modeling (2) the effectiveness of different resolution satellite data including ETM+, ASTER, MODIS, and IKONOS for LAI estimation; (3) determining if spatial or spectral resolutions are important for predicting LAI. This research involved the collection of field data in and around the Santarem region of Brazil. LAI values were collected at 76 different locations characteristic of many different cover types including mature forest, secondary succession, pasture, cropped land, barren land, and urban area. Each LAI measure was collected in a defined 20 x 20 m quadrat on the ground. The field data was used in conjunction with the satellite data to develop multiple regression and artificial neural network regression models for LAI prediction. Assessments of model accuracy were determined by calculating RMSE. This analysis demonstrated that neural networks perform better than the traditional multiple regression techniques for LAI modeling. The ASTER sensor provided the most accurate models because it was most characteristic of the 20 x 20 m sampling scheme. With its 15 and 30 m spatial resolutions, ASTER produced the lowest overall residual error of 1.94. In contrast, MODIS was found to be the overall poorest modeler ofLAI with no residuals less than 3.14.

Download

Share

COinS