Semester

Summer

Date of Graduation

1999

Document Type

Thesis

Degree Type

MA

College

Eberly College of Arts and Sciences

Department

Geology and Geography

Committee Chair

Timothy Warner.

Committee Co-Chair

Joseph Gardner

Committee Member

Robert Hanham

Abstract

This thesis is a theoretical and empirical examination of the relationship between texture and scale, and its effects on image classification. This study involved creating a model that automatically selected windows of optimal size according to the location of a pixel within a land cover region and the texture of the surrounding pixels. Large windows were used to get a representative sample of within-class variability in the interior of these regions. Smaller windows were used near the boundaries of land cover regions in order to reduce "edge effect" errors due to between-class variability. This program was tested using a Maximum Likelihood classification scheme against spectral data and texture from fixed-size windows to determine if there were any improvements in classification accuracy. Three different types and scales of data, including SPOT, SIR-C, and ADAR, were used to test the robustness of this program.;The results from this research indicate that the addition of texture can improve classification accuracy, especially in land cover regions with high local variability among the pixels. The 21 x 21 texture image achieved a Kappa Index of Agreement (KIA) of 0.97 for the highly textured Sunlit Forest (leaf off) class in the ADAR data, compared to 0.92 using the spectral data alone. However, texture windows of fixed-size created some errors due to "between-class" texture. This was most evident in the SPOT interior test data where the 21 x 21 texture window achieved a KIA of 0.70, compared to 0.92 for the spectral data. In many cases, images that incorporated the Optimal Size Window Program were superior in accuracy to all of the other images. In the radar data, the image created from the Optimal Size Window Program improved the overall KIA from 0.51 in the spectral data, to 0.71.

Share

COinS